A researcher in the Department of Computer Science and Engineering is developing a security-focused large language model to defend against malware.

"ChatGPT" for Computer Security

The research is part of an attempt to grapple with the effort to make expensive, long-term investments to mitigate the impacts of climate change. The substantial uncertainty related to long-term climate change makes it difficult for political leaders to make investments now that are designed to protect citizens for decades or longer. The difficulty is enhanced by the vast number of variables that go into any such decision and by the fact that the variables are likely to shift in unforeseen ways.In a March 18 <a href="https://www.pnas.org/doi/10.1073/pnas.2402826122">article </a>in the Proceedings of the National Academy of Sciences, the researchers looked at flooding, which has caused increasing damage along the coastal United States and around the world. Governments are building coastal defenses against flooding, but they cannot rely on past conditions to guide defenses that will be needed in the future.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1742909563584-16x9-for-web-Rutgers-Cape-Shore-Lab-Berm-1536x864.jpg" style="width:100%px" class="fr-fic fr-dib" />The Rutgers Cape Shore Laboratory is protected by a berm at its location on the Delaware Bay in Cape May Courthouse, New Jersey.“Defenses are being built to protect coastal regions for the next few decades or longer,” said co-author <a href="https://engineering.princeton.edu/faculty/ning-lin">Ning Lin</a>, a professor of <a href="https://cee.princeton.edu/">civil and environmental engineering</a> at Princeton. “Climate projects are largely uncertain over long time horizons.”Lin said that to deal with this uncertainty, planners must be flexible and ready to adapt their plans to future observation of climate conditions. Although this is extremely challenging because of the complexity of climate science, Lin said that harnessing advances in data science can provide an effective strategy.<a href="https://www.bobkopp.net/">Robert Kopp</a>, a co-author of the study and a distinguished professor of Earth and Planetary Sciences at Rutgers said uncertainty about the impact of melting ice sheets on sea levels has led to “controversy about how planners should consider the possibility of rapid ice-sheet loss.Kopp said that flexible approaches can help communities prepare for worst-case scenarios without paying too much for protection.  “Planning for high-end sea-level rise costs a lot, and there’s a good chance it won’t be necessary, but failing to plan for it can be devastating,” he said.In the PNAS article, the team describes how they simulated efforts to defend Manhattan against sea level rise through the end of this century. The goal was to determine whether any decision-making process that systematically incorporates observations and updating would prove superior to others over such a long period of time. To do this, the researchers simulated decisions by city planners in 10-year intervals up to the year 2100. The researchers compared their decision-making process with existing methods. For example, using the static method of building a seawall for a historic 100-year flood plus a sea-level-rise projection — as currently applied by New York and other coastal cities — is one method; designing a dynamic seawall that will be increased in height over time according to projected future climate change is another.The researchers graded each method by its efficiency — the cost of defenses plus the estimated damage caused by flooding. For example, spending $10 million on a seawall that allowed $50 million in property damage ($60 million cost) would be less efficient than spending $30 million on a seawall with $15 million in property damage ($45 million cost). The researchers found that calculations of a dynamic seawall based on reinforcement learning that systematically incorporates observations of sea level rise over time increased efficiency when compared with other methods (detailed in the paper). Compared with other systems, reinforcement learning lowered costs by 6-36% in a scenario modeling climate change under moderate carbon emissions. For a high emissions scenario, the decrease was 9-77%.Reinforcement learning is a type of machine learning in which a program makes decisions and receives positive reinforcement based on results. Designers train the program by running it through vast amounts of simulated decisions, and it learns by trial and error rather than through explicit instructions from programmers. This is essentially the way many AI systems operate. It is particularly effective for situations that are extremely complex and subject to rapid changes over time. Computer scientists have used reinforcement learning to train AI to perform tasks such as playing chess, driving cars, and controlling robots and drones. The method has also been used for large systems used to store power or control water supplies.For their case study, the researchers looked at defenses proposed for low-lying areas in Manhattan. After Hurricane Sandy caused tremendous damage in New York and New Jersey in 2012, the U.S. Army Corps of Engineers proposed constructing a series of seawalls to defend Manhattan called the Big U. Some sections of the Big U are being built and others are in the planning stage. But a final completion plan for the entire system has not yet been set.The study looked at proposals for the Big U and how defenses against coastal flooding should be changed to respond to future threats to New York from the sea. Flooding is driven both by sea level rise and storms. Both are affected by climate change, which in turn is impacted by global carbon emissions. The researchers wanted to evaluate decision methods every 10 years and allow for adjusting the Big U to reflect available threat data at each interval.Because the impacts of climate change over a long period remain uncertain, the researchers wanted to study methods of making the best decisions for future changes with the data available at the time. Traditionally, engineers have built protective systems like seawalls and levees to resist historic floods, building protection from floods that would occur only once in 50 or 100 years. But because the climate is changing, such systems no longer work. Building the Big U seawalls to match the highest storm surge over the past 100 years would leave Manhattan vulnerable as climate change drives higher storm surge levels.The researchers looked at a number of decision-making methods that take into account changing conditions. Most methods allowed for changes based on key variables and some allowed planners to make future projections that would also influence decisions. For example, as a temporary measure, residents could flood-proof their homes (“accommodate”), but eventually, higher sea levels would necessitate a high seawall (“protect”). In some cases, a seawall could prove too costly to protect everyone and residents would be encouraged or compelled to leave threatened communities (“retreat”). In a situation with only a few variables to account for, the benefits of a plan can be estimated relatively easily. But uncertain climate change presents an extremely complex scenario. The researchers showed that reinforcement learning can be used to design integrated strategies, including retreating from low-lying areas, protecting property further inland at higher elevation, and accommodating in between, with a 5-15% reduction in cost compared to the one-dimensional seawall strategy.Lin, one of the lead researchers, said defending Manhattan is not only complex, but it also requires making difficult decisions under uncertain conditions. For each time interval, planners must make decisions based on observed sea level rise and roughly 80,000 scenarios of future sea level rise and corresponding decision reactions. The difficulty of the decisions compounds as the number of intervals increases.While climate adaptation decisions are not simple, Lin said reinforcement learning is  a highly efficient system for incorporating observations and updating plans to derive optimal solutions for limiting impacts from extreme events. Reinforcement learning also outperforms previous methods by avoiding losses induced by uncertain changes in future global carbon emissions.“The analysis of New York City’s situation is by no means unique,” said <a href="https://spia.princeton.edu/faculty/omichael">Michael Oppenheimer</a>, a study co-author and professor of geosciences and international affairs at Princeton. “The method can be applied widely, although its benefit compared to other systems of analysis would vary from place to place.”<hr />The paper, “<a href="https://www.pnas.org/doi/10.1073/pnas.2402826122">Reinforcement learning-based adaptive strategies for climate change adaptation: An application for coastal flood risk management</a>,” was published March 18 in the Proceedings of the National Academy of Sciences. Besides Kopp, Lin and Oppenheimer, co-authors include Kairui Feng, formerly of Princeton and now at Tongji University, Shanghai, China; and Siyuan Xian of Princeton. The authors are part of the Megalopolitan Coastal Transformation Hub, a Rutgers-led, National Science Foundation-funded consortium of research institutions working to advance the science of how coastal climate hazards, landforms and human decisions interact to shape climate risk and advance climate adaptation in the New York City-New Jersey-Philadelphia region.

A house, being relocated to higher ground, is transported by barge through Little Egg Inlet along New Jersey’s southeastern coast. Photos by Matthew Drews, Rutgers University

Researchers from Princeton and Rutgers University have used reinforcement learning, a method frequently deployed to train artificial intelligence, to show how flexible responses can substantially increase the cost-effectiveness of steps to defend cities like New York against climate change.

AI technique boosts climate change defenses

Wevolver: David. So what are we looking at here?David: Thanks. Well, what we have here is an STM32N6. Now, this is a microcontroller-based product. So this is an MCU with an NPU. And what we're doing in this specific example here is object detection and anomaly detection all in one. So we can place a coffee pod or two onto the conveyor belt and we will see that coffee pod has been detected. Now we're also detecting, as I said, anomalies or defect detection. So if I place these coffee pods which have defects in them, we'll see over on the right hand side of the screen where the defects are in a heat map. I'll take this one here, which has some puncture wounds in it. And we'll do the object detection as well as the defects. This is super interesting because this is an MCU based device, extremely low power, and because it's got the MCU and the NPU, we can actually run two distinct models at the same time on the device. Pretty cool.David: All right, and now we're over at a second demo pod here. For this one though, I'm going to actually pass the mic over to my colleague Dmitry, who will walk you through it. Dmitry? Dmitry: Hey. Yeah, I'm Dmitry. And here we see the modal cascade. There are two models running on this particularly powerful device made by Qualcomm. It's 9075, recently announced chip. So we have the object detection model which detects vehicles and then once it's detected, it is being passed to a visual language model. You have a choice of lava, open source, lava and AMG. This is in in-house model by Qualcomm. So we have crafted a prompt here which is the vehicle class, vehicle manufacturer, sign identification. If press generate all, it'll just take a few moments. And keep in mind that both the detection and the visual language model, they both run on this device. It doesn't go anywhere to the clouds, you can see the answers being streamed in real time. Yeah, let's wait for the last one here. Yeah, it gets most of it very correctly. Let's see the last one actually. Yeah, that's all good. And the value proposition for us at Edge Impulse is in fact that you cannot blindly trust visual language models. You use ChatGPT, you know it can make mistakes, it says they are right at the bottom. So in that case, if you really are serious and you want to put this into production, you want to validate your prompt, you want to create it, validate it and check the accuracy against the data that you have, and this is exactly what we are providing at Edge Impulse. So you create your prompt here, you tweak it around, you test it, and then you pass it on to the production model running there. That's about it.David: All right, thanks, Dmitry. Appreciate it. And I think that's going to do it for us here from Embedded World 2025. If you have any questions, be sure to reach out edgeimpulse.com. Thanks.

Unveiling the Future of Embedded AI: Object Anomaly Detection and More at Embedded World 2025

Edge Impulse's Object Anomaly Detection Technology, And Other Innovations at Embedded World 2025

Few people realize that the carbon footprint of digital technology is on par with that of the aerospace industry, accounting for between 2% and 4% of global carbon emissions. And this digital carbon footprint is expanding at a rapid pace. When it comes to power use, the approximately 11,000 data centers in operation today consume just as much energy as the entire country of France did in 2022, or around 460 TWh per year. Will the widespread adoption of generative AI send those figures soaring?The new technology will clearly affect the amount of energy that’s consumed worldwide, but exactly how is hard to quantify. “We need to know the total cost of generative AI systems to be able to use them as efficiently as possible,” says Manuel Cubero-Castan, the project manager on <a href="https://www.epfl.ch/about/sustainability/sustainable-it/">Sustainable IT at EPFL</a>. He believes we should consider the entire life cycle of generative AI technology, from the extraction of minerals and the assembly of components – activities whose impact concerns not only energy – to the disposal of the tons of electronic waste that are generated, which often gets dumped illegally. From this perspective, the environmental ramifications of generative AI go well beyond the power and water consumption of data centers alone.<h2>The cost of training</h2>For now, most of the data available on digital technology power use relates only to data centers. According to the International Energy Agency (IEA), these centers (excluding data networks and cryptocurrency mining) consumed between 240 TWh and 340 TWh of power in 2022, or 1% to 1.3% of the global total. Yet even though the number of centers is growing by 4% per year, their overall power use didn’t change much between 2010 and 2020, thanks to energy- efficiency improvements.With generative AI set to be adopted on a massive scale, that will certainly change. Generative AI technology is based on large language models (LLMs) that use power in two ways. First, while they’re being trained – a step that involves running terabytes of data through algorithms so that they learn to predict words and sentences in a given context. Until recently, this was the most energy-intensive step.Second, while they’re processing data in response to a prompt. Now that LLMs are being implemented on a large scale, this is the step requiring the most energy. Recent data from Meta and Google suggest that this step now accounts for 60% to 70% of the power used by generative AI systems, against 30% to 40% for training.<h2>ChatGPT query vs conventional Google search</h2>A ChatGPT query consumes around 3 Wh of power, while a conventional Google search uses 0.3 Wh, according to the IEA. If all of the approximately nine billion Google searches performed daily were switched to ChatGPT, that would increase the total power requirement by 10 TWh per year. Goldman Sachs Research (GSR) estimates that the amount of electricity used by data centers will swell by 160% over the next five years, and that they will account for 3% to 4% of global electricity use. In addition, their carbon emissions will likely double between 2022 and 2030. According to IEA figures, total power demand in Europe decreased for three years in a row but picked up in 2024 and should return to 2021 levels – some 2,560 TWh per year – by 2026. Nearly a third of this increase will be due to data centers. GSR estimates that the AI-related power demand at data centers will grow by approximately 200 TWh per year between 2023 and 2030. By 2028, AI should account for nearly 19% of data centers’ energy consumption.If we begin using generative AI technology on a massive scale, with ever-bigger LLMs, the resulting energy gains will be far from enough to achieve a reduction in overall carbon emissions.- Manuel Cubero-Castan, project manager on Sustainable IT at EPFLHowever, the rapid expansion of generative AI could wrong-foot these forecasts. Chinese company DeepSeek is already shaking things up – it introduced a generative AI program in late January that uses less energy than its US counterparts for both training algorithms and responding to prompts. Another factor that could stem the growth in AI power demand is the limited amount of mining resources available for producing chips. Nvidia currently dominates the market for AI chips, with a 95% market share. The three million Nvidia H100 chips installed around the world used 13.8 TWh of power in 2024 – the same amount as Guatemala. By 2027, Nvidia chips could burn through 85 to 134 TWh of power. But will the company be able to produce them at that scale?<h2>Not always a sustainable choice</h2>Another factor to consider is whether our aging power grids will be able to support the additional load. Many of them, both nationally and locally, are already being pushed to the limit to meet current demand. And the fact that data centers are often concentrated geographically complicates things further. For example, data centers make up 20% of the power consumption in Ireland and over 25% in the US state of Virginia. “Building data centers in regions where water and power supplies are already strained may not be the most sustainable choice,” says Cubero-Castan.There’s also the cost issue. If Google wanted to be able to process generative AI queries, it would need to set up 400,000 additional servers – at a price tag of some 100 billion dollars, which would shrink its operating margin to zero. An unlikely scenario.<h2>Untapped benefits</h2>Some of the increase in power consumption caused by generative AI could be offset by the benefits of AI in general. Although training algorithms requires an investment, it could pay off in terms of energy savings or climate benefits. For instance, AI could speed the pace of innovation in the energy sector. That could help users to better predict and reduce their power use; enable utilities to manage their power grids more effectively; improve resource management; and allow engineers to run simulations and drive advances at the leading edge of modeling, climate economics, education and basic research. Whether we’re able to leverage the benefits of this kind of innovation will depend on its impacts, how extensively the new technology is adopted by consumers, and how well policymakers understand it and draft laws to govern it.The next-generation data centers being built today are more energy efficient and allow for greater flexibility in how their capacity is used. By the same token, Nvidia is working to improve the performance of its chips while lowering their power requirement. And we shouldn’t forget the potential of quantum computing. When it comes to data centers, the IEA calculates that 40% of the electricity they use goes to cooling, 40% to running servers and 20% to other system components including data storage and communication. At EPFL, Prof. Mario Paolone is heading up the <a href="https://heatingbits.epfl.ch/">Heating Bits</a> initiative to build a demonstrator for testing new cooling methods. Five research groups and the <a href="https://ecocloud.epfl.ch/">EcoCloud Center</a> have teamed up for the initiative, with the goal of developing new processes for heat recovery, cogeneration, incorporating renewable energy and optimizing server use.<h2>Keeping the bigger picture in mind</h2>Another (painless and free) way to cut data centers’ power use is to clear out the clutter. Every day, companies worldwide generate 1.3 trillion gigabytes of data, most of which ends up as dark data, or data that are collected and stored but never used. Researchers at Loughborough Business School estimate that 60% of the data kept today are dark data, and storing them emits just as much carbon as three million London–New York flights. This year’s Digital Cleanup Day will be held on 15 March, but you don’t have to wait until spring to do your cleaning!Cubero-Castan warns us, however, to keep the bigger picture in mind: “If we begin using generative AI technology on a massive scale, with ever-bigger LLMs, the resulting energy gains will be far from enough to achieve a reduction in overall carbon emissions. Lowering our usage and increasing the lifespan and efficiency of our infrastructure remain essential.”The energy impact of generative AI mustn’t be overlooked, but for now it’s only marginal at the global level – it’s simply adding to the already hefty power consumption of digital technology in general. Videos currently account for 70% to 80% of data traffic around the world, while other major contributors are multiplayer online games and cryptocurrency. The main drivers of power demand today are economic growth, electric vehicles, air-conditioning and manufacturing. And most of that power still comes from fossil fuels.<hr /><h3>References</h3>This article was published in the issue of Dimensions which was released on 14 March 2025. Dimensions, an EPFL magazine that showcases cutting-edge research through a series of in-depth articles, interviews, portraits and news highlights. Published four times a year in both English and French, it can be sent to <a href="https://www.epfl.ch/about/news-and-media/fr/dimensions/abonnements/">anyone who wants to subscribe</a> as well as <a href="https://epflalumni.ch/page/avantages" rel="noopener" target="_blank">contributing members of the EPFL Alumni Club</a>. It is also distributed free of charge on EPFL’s campuses.

Data centers are major electricity consumers. © iStock

It takes ten times more electricity for ChatGPT to respond to a prompt than for Google to carry out a standard search. Still, researchers are struggling to get a grasp on the energy implications of generative artificial intelligence both now and going forward.

Can energy-hungry AI help cut our energy use?

In this episode, we explore how a generative AI tool is marking a major milestone in biology and accelerating advancements in healthcare, genetics, and drug development.<iframe height="200px" width="100%" frameborder="no" scrolling="no" src="https://player.simplecast.com/8b738738-5a69-4f97-ac33-8d48a50ec427?dark=false"></iframe> <hr /><h3>Episode Notes</h3>In this episode, we explore how a generative AI tool is marking a major milestone in biology and accelerating advancements in healthcare, genetics, and drug development.Become a founding reader of our newsletter: <a href="http://read.thenextbyte.com/">http://read.thenextbyte.com/</a><hr /><h2 id="isPasted">Transcript</h2>What's up folks, welcome back to the Next Byte podcast. And in this one, we're talking about DNA sequencing. Specifically, we're talking about a team over at Stanford that has created a ChatGPT-like AI model that can predict what comes next if you give it any DNA sequence, including yours.  that's got you a little bit spooked or excited or just curious. Well, then let's get into it.What's up friends, this is The Next Byte Podcast where one gentleman and one scholar explore the secret sauce behind cool tech and make it easy to understand.Farbod: This is another one of those biology ones. So, you already know I'm gonna struggle, but that's okay. We learn when we struggle. I was…Daniel: You just need to test your limits, man!Farbod: And that's what I'm doing. I'm trying here, but it was so interesting that I knew we needed to talk about it. So, for the audience, please bear with me. This is not my area of expertise, but God am I trying. So, for this episode, we're going to Stanford. You know the West Coast is the best coast, but these folks at Stanford, they're doing some crazy things with DNA sequencing.  And to be completely honest, they've kind of done this before, but now they're doing it better. So, what am I talking about? They've developed a chatGPT-like model, right? It's structured like a large language model, but instead of processing the meaning of the word that you're typing or the paragraph that you're writing or whatever and then giving you an answer or a follow-up or completing your train of thought, it takes as an input. A sequence of nucleotides, a sequence of DNA, and it completes it for you or it makes sense of it and it tells you what genes might be associated with each other. Now, before we get into that, I think we got to take it a couple steps back and by we, I mean me because I know I need to take it, a couple steps back. So, they did an amazing job in this article, by the way, explaining things like you're a five-year-old, which again is wonderful for me, and they have a three-minute video that I highly recommend you check out. It's going to be attached to the show notes. But they explained the issue at hand like this. They said, imagine you're reading the book, and this book has three billion letters in it, right? And every one of those letters is a nucleotide. These nucleotides make up the DNA that everything we know of in this world is pretty much composed of. And it's actually kind of straightforward. We represent them as four letters, A, C, G, and T. And standalone, they don't really mean much, but the sequence of them in different patterns together make up specific genes. Now, as we know, if you've been watching the news, this gene causes this, and poor, broken nucleotide sequences can, for most people, be harmless. For some people, it can cause cancer. So, there's a lot of value in understanding what those sequences mean, how they're related to each other, et cetera, et. Now, as I mentioned, this team had already worked on a chatGPT-like model before called the EVO-1 that was trained using 300 billion nucleotides, mostly from bacteria and stuff like that, to help them proof of concept this idea of, if we had a model that could predict what comes next. Could we better understand what the next evolutionary step would be for this plant or these bacteria or this human? Or could we better understand how the sequence that is at the one millionth element is related to the one at the 100,000th element? And if so, what does that tell us about the chances that you might have cancer? Or could we edit this out? The possibilities and the downsides of what already exists, basically.Daniel: Well, and when we talked about the technological development of having sequenced the human genome several decades ago, that didn't necessarily crack the code for us on what everything in human DNA means. That was just us looking at the DNA and writing a list, writing that three billion letter long book of just A’s, C’s, G’s, and T’s. It's awesome that like there's only four letters, but essentially, we didn't understand how those letters comprise genes and genes, they liken it in the article to like words. So, like basically, the major development in sequencing the human genome was like figuring out the alphabet and the order of the letters in the alphabet. But we did not yet understand the vocabulary. We didn't understand the words. And beyond that, we don't understand the grammar about how you put several words together to create phrases. And that's how physical characteristics are expressed or through a combination of several different genes. So, this is essentially helping us bridge the gap. And like in my mind, truly cracked the code on the meaning of DNA. If we can start to predict protein form and function by looking at the order of the A's, C's, G's and T's, predict harmful slash beneficial genes, potentially create new genetic sequences for medicines or treatments like this. This is bridging the gap between the letters in the physical realm, whereas before it was awesome that we understood what word of the letters were in.Farbod: You know, just like EVO2 is making sense of these nucleotide sequences, I'm happy you're here to make sense of my train of thought about biology.Daniel: EVO2 wouldn't exist without EVO1, buddy.Farbod: You're right. And like, I guess that frame of thinking is even better when comparing it to something like ChatGPT or really any other large language model, because the key breakthrough there, I think we mentioned it five episodes ago. The key breakthrough there was the transformer architecture, which said in relation to this wider body of text, what does this letter mean and this letter mean, instead of trying to make sense of a sequence of words back-to-back, which then lost meaning once you reach the 50th word. So again, they've taken the same underlying engine here, but now applying it to what does this sequence of genes mean, these nucleotides mean. And I think, I mentioned that EVO1 had 300 billion nucleotides, which already sounds quite impressive, but it was mostly bacteria and whatnot. Well, they upped it for EVO2, and now they include animals, plants. Overall, there's nine trillion nucleotides that are represented in this model. The only thing that they left out by design was viruses. The team said they didn't want this tool to be something that bad actors could use to develop and refine existing viruses, which I personally thank them. I think no one wants another round of the coronavirus. So, kudos for thinking ahead on that.Daniel: And this is like, it's pretty monumental here. So, in addition to like excluding viruses, right? Basically, except for viruses, the training data now is 27 times bigger than before. It includes all known living species that we have a DNA sequence for, plus some that are extinct. So, it's like, it's pretty much everything that we've ever sequenced ever, except for viruses so that we don't have people going, creating designer viruses to go try and get people sick. But every living being that we have the DNA sequence for has been included in the training data for this next generation. That's a big deal to increase your training data size, your training data set by a factor of 27, right? It's not like they've trained on 10 % more sequences than before. This is a huge deal.Farbod: Absolutely. And not only have they increased the number of data points that they're training on, but the input is also, what does ChatGPT call them, the number of tokens that you can input for a query? Yeah.Daniel: Yeah.Farbod: So, they've upped it to a one million letter sequence of nucleotides. So, you can have one million letters of combination of ACGT and it can then do whatever you want with it, analyze it, predict the next set of sequences, cetera, et. And I was wondering why one million? Apparently in the world of biology, the one million mark is, it's critical because it's long enough that it takes into account historically significantly far enough genes to make sense of rather than just taking into account the recent developments or deficiencies.Daniel: That makes sense. I appreciate the significance around there. And this makes me feel more so.  ChatGPT is an awesome analogy to give people, especially because I feel like a lot of people have had exposure to ChatGPT. But I'm hoping within our audience, maybe there's folks who've had a greater exposure to things like a co-pilot in coding or things like cursor composer. This actually feels a lot like cursor composer to me where in the cursor editor, you can go in the composer bar and you can have your code base living there, which is like a sequence of DNA. And you can tell it, hey, I want to generate a change to do this to the code  and composer will go and make those changes based off of its understanding of the code. This feels a lot like that, right? So, some of the technical features that it has, similar actually again to cursor composer is like autocomplete. They make a big deal about it in cursor. They call them tabs, but basically like you're writing and it understands what you're trying to code and you can just hit tab and it completes it, right? Autocomplete for code. They have autocomplete for DNA here based off of patterns in DNA sequences. They also have prediction models that can understand real world function. So instead of saying like change all the A's to C's or you know, control F my ACGTs and change them to TCGAs, you can say, I actually just want to change how this physical characteristic appears in the real world. And then it'll go change the DNA to reflect that. And then the interesting part, which is really awesome in my mind, it tends to be somewhat controversial. But we've seen a lot of work with CRISPR gene editing, which basically allows us to like sequence DNA, fabricate it, and then inject it into a DNA sequence. And then in a living being and basically edit a living being's DNA. They're able to essentially predict what the outcome of CRISPR gene editing will look like in the real world to propose edits essentially propose future experiments and tests.Farbod: Yes. That's what I gonna bring up I think, you know, the audience might have mixed feelings about that but what stood out to me? They were like look within this model with an EVO2 we can whenever you're predicting a new gene, we can tell you if it already exists if it doesn't if you implement it like what kind of behavior it could have, and then obviously you can do lab testing, put it into a living cell and see how it pans out. One thing that I thought was interesting in regards to CRISPR and gene editing, they were like, what if we could use this tool to modify bacteria to consume microplastics and to clean up oil spills, right? So again, I know it's a mixed feeling type of topic, especially as it relates to human beings, but the benefits of this tool are actually pretty significant. And even on the human-Daniel: This doesn't necessarily dictate how the findings are applied, right? This is an engine for understanding DNA, right? That's what we should be excited about, is it's an engine for us to understand what's going on. Even if you decide not to go edit your genes, to get rid of a disease, you can understand what diseases you might be predisposed to based off of your DNA.Farbod: Yeah, and you know, most of the stuff that they talked about are all the positives that we can all agree on, like modifying the bacteria. They said this can be used to significantly cut down on the discovery phase of new drug development because you can find which proteins work best for addressing whatever symptom. We talked about these two episodes ago, I think, on the snake venom episode, which has been a hit, by the way. They did mention that EVO2 could help with human evolution because we could predict what genes could be beneficial to us, but that's a topic for another day. I think what we should focus on here is that this is an incredible platform that they have open source. It's available to all researchers that they want others to use in their workflow and build on top of so that we can make better drugs and help our environment.Daniel: Well, and I think two really important things to mention. And you mentioned one of them and I kind of want to double down on it. One of them is direct quote from the researchers. They said it's really, really good at drug discovery, which is awesome. So obviously, helping identify treatments for potential diseases is awesome. But one of the other things that they can mention is like they, this model can distinguish between harmless DNA mutations to disease, disease causing ones. So, where we've got a lot of concern around free radicals and/or mutations in your DNA, potentially cancer causing, as you mentioned before, is like, we're trying to limit exposure to anything that could cause any type of mutation in your DNA. And that's where we get a lot of conflicting advice from people like, oh, like don't put on sunscreen, but also don't get sunburned and don't eat this and don't eat that. And don't do that. Like it feels like you're, you're screwed either way. I feel like this will potentially help us get more clarity around which types of causes for DNA mutations are most likely to cause disease and then help us hone our focus on like, what's the 80-20? What's the Pareto principle? Like what's the 20% of effort that I can do to protect myself from 80% of bad outcomes? I feel like that's one of the most important things that you and I will experience in our day-to-day life is a little bit more clarity around a lot of the conflicting guidance that we get. I think the most specific one that I think of is like around sun exposure. Like they're like, you need sun exposure for vitamin D, but don't get so much sun exposure that it causes mutations in your skin can cause melanoma. So, if you're outside, make sure you put on sunscreen, but also sunscreen can like cross through the blood brain barrier and cause disruptions in your hormones. Try and give us some clarity around a lot of the conflicting advice like that, where it feels like you can't win either way. Like, hopefully this will help us understand more what changes to make in a day-to-day life that actually have an impact on the presence of disease in our DNA.Farbod: Dude. Yeah, I totally agree that was a good hot take that I completely missed. And the sunscreen one, I'm totally with you. I think about it every morning when I put on my sunscreen. With that said, I'm going to wrap things up. So, folks, if you've heard about DNA sequencing and what a miracle it was that we cracked it a couple decades ago, you might be shocked to hear that it didn't crack the code all the way. Look, we understand what the nucleotides are that make up our DNA. You know, the four different ones, the A, C, G and T. But these folks over at Stanford are now developing a model or have already developed a model that can make sense of the different genes that are part of this DNA sequence and how they interact with each other. But more importantly, almost like you're using ChatGPT, it can predict what comes next. Now, why is this valuable? Well, by understanding the relation between the different genes, they can understand what is harmful, what can be used to develop new tools, how we can modify an existing sequence to give us beneficial properties. For example, one they called out was modifying bacteria to eat microplastics or clean up oil spills. And the next bit, it could even help us understand what good way we can go for evolving the human genome so that we get all the best properties possible.Daniel: That's Pretty fireman.Farbod: Sweet. That's the pod.Daniel: All right. See you.<hr />As always, you can find these and other interesting &amp; impactful engineering articles on Wevolver.com.To learn more about this show, please visit our<a href="https://www.wevolver.com/profile/the.next.byte"> shows page</a>. By following the page, you will get automatic updates by email when a new show is published. Be sure to give us a follow and review on<a href="https://podcasts.apple.com/nl/podcast/the-next-byte/id1548255861?l=en" target="_blank" rel="nofollow"> Apple</a> podcasts,<a href="https://open.spotify.com/show/6lPPsRtegnivsRUEsQ09R7?si=IPMiah7xT_apJtPjqvXMlA" target="_blank" rel="nofollow"> Spotify</a>, and most of your favorite podcast platforms!--The Next Byte: We're two engineers on a mission to simplify complex science &amp; technology, making it easy to understand. In each episode of our show, we dive into world-changing tech (such as AI, robotics, 3D printing, IoT, &amp; much more), all while keeping it entertaining &amp; engaging along the way.

In this episode, we explore how a generative AI tool is marking a major milestone in biology and accelerating advancements in healthcare, genetics, and drug development.

Podcast: Stanford Predicts The Next Human Evolution

<h2>Introduction</h2>The Robotics market is experiencing rapid growth globally, with the market value expected to reach $189.3 billion by 2027 [1]. Advancements in automation and  AI-powered vision systems drive this progress. Camera-equipped robotic systems serve as key expansion components within this industry because they improve efficiency and accuracy across multiple fields, including manufacturing, healthcare, logistics, and autonomous vehicle development. These systems use high-resolution imaging and real-time data processing to accomplish tasks that demand precision and adaptability.However, despite their potential, several challenges hinder the performance of these systems. High latency decreases operational efficiency; low-resolution imaging contributes to industrial defect detection errors. Short and inflexible cables, electromagnetic interference (EMI), and high power consumption negatively impact conventional camera systems. These factors influence the functionality and reliability of the systems. Addressing the challenges is essential for achieving the full potential of vision-equipped robotics.Silicon Line, a leading provider of high-performance semiconductor solutions, has developed their Mobile Industry Processor Interface (MIPI) SerDes Integrated Circuits (ICs) to overcome these challenges. These ICs transport native high-resolution images with extremely low latency while allowing longer cable lengths without degrading signals. Silicon Line’s MIPI SerDes ICs enable camera-equipped robotic systems to deliver improved imaging performance while processing data in real time. This results in enhanced operational efficiency, making them more adaptable to industry requirements.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1742349844848-ezgif-4d614fd9adfd9d.jpg" /><h2>Challenges in Camera-Equipped Robotic Systems</h2>Robotic vision systems require solutions to multiple challenges for seamless operation in industrial environments. Some of the most common challenges include:<h3>High Latency:</h3>The main operational problem arises from high latency, which causes delays that affect real-time decision-making processes. Applications such as autonomous navigation systems and robotic-assisted surgery require prompt decision-making; even a fraction of a second can lead to suboptimal outcomes.<h3>Low Resolution:</h3>Another challenge that demands immediate attention is implementing high-resolution imaging capabilities. Using low-resolution cameras in industrial robots substantially increases inspection errors, affecting product quality and driving costs. Compressed image data further reduces clarity to the extent that it hinders the effectiveness of AI-based defect detection algorithms. For instance, semiconductor manufacturing experiences financial losses because imprecise imaging can produce defective products.<h3>Short and Inflexible Cables:</h3>Cable limitations also present another significant problem. Transporting high-resolution images requires a very high bandwidth, which limits cable lengths and leads to thick and rigid cables. These cables limit the mobility of robotic arms and other vision-equipped robots, making them difficult to install and maintain. Traditional cables add weight and stiffness to the system, restricting the speed and agility needed for robotic systems to adapt to dynamic environments. The need for 3D camera capability in robotic systems also leads to an increase in cable complexity.<h3>Electromagnetic Interference (EMI):</h3>EMI can also degrade the image quality and hinder AI-driven analysis. Manufacturing environments with several electronic devices are prone to interference, leading to inaccurate object recognition and errors in automated processes, quality control, and safety systems.<h3>High Power Consumption:</h3>High power usage, especially in battery-operated robotic systems, is also a primary concern, as energy-intensive vision systems contribute to more significant operational costs and sustainability challenges. Rectifying these issues is essential for optimizing the efficiency and reliability of robotic vision applications.These challenges affect robotic systems' efficiency and reliability, restricting their potential applications. Overcoming these issues is crucial for industries looking to take full advantage of camera-equipped robotics.<h2>Silicon Line’s MIPI SerDes ICs: A Solution to Camera-Equipped Robotic Challenges</h2>Silicon Line’s MIPI SerDes ICs offer the most efficient solution to these challenges by leveraging advanced serialization and deserialization techniques to optimize high-speed data transfer while maintaining signal integrity. These ICs utilize the industry standard MIPI D-PHY interface, ensuring compatibility with a wide range of imaging and processing components while delivering robust performance in high-speed data transmission applications. These ICs reduce latency by enabling real-time transfer of uncompressed and unconverted high-resolution images, keeping robotic systems in a position to make accurate, split-second decisions within automated navigation and robotic surgical treatments.These ICs also enhance image recognition and support ultra-high-definition image transmission without compression, preserving crucial details for defect detection and quality control through AI. They increase the strength of Machine learning (ML)/ Deep Learning (DL) algorithms for detecting defects by eliminating lossy image processing. Thus increasing the accuracy of inspections on industrial sites such as semiconductor manufacturing and precision assembly lines.Silicon Line’s ICs overcome the cable's limitations by supporting optical and copper based transmission over long distances and reducing cable thickness and weight. The serialization process means the 10-wire MIPI D-PHY interface is serialized to just two wires or one optical fiber. The latest Silicon Line products can also transport dual camera images over the same single fiber or two copper wires. The ICs facilitate greater robotic mobility, flexibility, and accuracy. Additionally, Silicon Line’s extremely low-power circuit design saves power, making vision-enabled robotic systems more sustainable and cost-effective. Their compatibility with various AI &amp; edge computing platforms ensures seamless integration, using real-time data analysis and decision-making processes in robotics applications across industries. <img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1742349964256-1742349964256.png" /><h2>Real-World Applications of MIPI SerDes ICs in Robotics</h2>Silicon Line’s MIPI SerDes ICs have several applications across various industries. Here are a few examples:<h3>Humanoid Robotics:</h3>Humanoid robots with high-resolution cameras need lightweight image processing to frequently perform complicated tasks such as navigation and object manipulation. Implementing MIPI SerDes ICs allows these robots to operate with greater precision and responsiveness in industries and healthcare. These ICs aid humanoid robots in improving interaction with humans in service robotics and healthcare for elderly care applications.<h3>Surgical Robotics:</h3>In medical robotics, ultra-high-definition imaging is required for minimally invasive procedures. MIPI SerDes ICs allow transmitting real-time images, giving surgeons clear and precise visuals during operative procedures. This dramatically increases accuracy and patient safety, especially in delicate procedures such as neurosurgery or robotic-assisted laparoscopy.<h3>Factory Automation and Quality Control:</h3>Manufacturing plants employ vision systems for the detection of defects and quality control. Integrating high-resolution imaging with AI-based analysis decreases defect rates significantly. MIPI SerDes ICs allow high-resolution native data transmission, enabling manufacturers to implement automated inspection systems with increased precision and reliability. This improves productivity and reduces material waste, resulting in cost savings and enhanced sustainability.<h2>Autonomous Guided Vehicles (AGVs)</h2>AGVs are crucial for logistics and warehouse automation, where real-time navigation and obstacle detection are essential. High-speed image transmission makes AGVs move efficiently, reducing the risk of collisions and enhancing overall workflow efficiency. Silicon Line’s MIPI SerDes ICs increase the performance of these vehicles by enabling real-time data processing and longer, flexible cable lengths.<h2>Collaborative Robots (Cobots)</h2>Cobots increase productivity, precision, and safety by working alongside human operators. They need real-time object detection and decision-making, requiring high-resolution data transfer and real-time handling. Silicon ICs, especially the MIPI SerDes ICs, enhance cobot's efficiency and reliability through high imaging resolution, reduced latency, and low power dissipation. This enables cobots to work flawlessly in manufacturing, healthcare, and logistics, where adaptability and high accuracy are paramount.<h3>Conclusion:</h3>Camera-equipped robotic systems are reshaping industries, but challenges like high latency, low resolution, inflexible and short cables often limit their potential. The advancement in silicon IC technology, especially MIPI SerDes ICs, has significantly transformed the performance of camera-enabled robotic systems. These innovations have resolved the key issues associated with conventional systems. These ICs have improved the performance and reliability of robotics in almost all sectors by allowing high-speed data transmission, real-time image processing, and energy-efficient operation.Continuous advancements in semiconductor technology will shape the future of robotics. Developing ultra-low power ICs, AI-optimized processors, and edge-computing solutions will further enhance the autonomy and efficiency of the robotic systems. Furthermore, integrating 3D imaging, hyperspectral vision, and neuromorphic computing will allow robots to perceive and interpret their environments accurately.With industries moving towards intelligent automation, the fusion of the latest silicon ICs and AI-driven robotics will result in more advanced, adaptable, and sustainable solutions. These innovations will increase the functions of camera-based robotic automation, transform the future of robotics and automation, and enhance productivity, safety, and operational efficiency across diverse sectors. Silicon Line invites professionals from all stakeholders to explore these advancements further.<h1>References:</h1>[1]    Advantech, “Enabling Smart Robotics with Single-Board Computing.”[2]    P. Lu, K. Yuan, and W. Zou, “A High Performance Low Power Consumption Robot Vision System,” in Third International Conference on Natural Computation (ICNC 2007) Vol V, IEEE, Aug. 2007, pp. 171–175. doi: 10.1109/ICNC.2007.42.[3]    Silicon Line, “Silicon Line.” Accessed: Feb. 21, 2025. [Online]. Available: https://www.silicon-line.com/

Camera-equipped robotic systems serve as key expansion components within this industry because they improve efficiency and accuracy across multiple fields, including manufacturing, healthcare, logistics, and autonomous vehicle development.

Highly Accurate Object Recognition: MIPI Serialization ICs in  Camera-Equipped Robotic Systems

AI, as you may have noticed, is everywhere. Whether it is a discussion about political elections, the implications for recruitment, education or for manufacturing, AI is seemingly always a key part of the conversation. Applications of AI models are nascent across all use cases, including within AM. In this area, the National Institute of Standards and Technology (NIST) is currently running the Advanced Informatics and Artificial Intelligence for Additive Manufacturing (AI2AM) project to develop the methods, models, standards, and best practices to leverage AI and related technologies to move AM towards “born qualified” and “first part correct” outcomes. <table><tbody><tr><td><h3 id="isPasted">Stay Informed on the Latest FormNext 2025 Updates! Sign up to the FormNext newsletter to get notified as soon as tickets are available.</h3><a href="https://info.mesago.de/-lp/5HKPB14784/p4jkP173" target="_blank" id="isPasted" class="button-main-action">Sign up here</a></td></tr></tbody></table>Despite the fledgling nature of much of the technology, in the world of AM, different types of AI do, will or could find use across the process chain.<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzQyMjg2ODAxMzY5LUNoZXNzLndlYnAuMTI4MC53ZWJwIiwiZWRpdHMiOnsicmVzaXplIjp7IndpZHRoIjo5NTAsImZpdCI6ImNvdmVyIn19fQ==" class="fr-fic fr-dii" />In addition to design, AI offers support in further process steps in Additive Manufacturing. Image: James Woodock<h2>Ideation and design</h2>AI can play a significant role in the DfAM space, with generative AI able to create complex designs that make full use of the benefits of AM. As well as being embedded in major CAD software like Autodesk Fusion 360 and Siemens NX, standalone products like Meshy AI allows users to create 3D models — exportable as .stl files — from either text prompts or photographs. Deploying deep learning and generative design in this way drastically reduces the design time while allowing more solutions to be trialed in a reduced timeframe.Once part files have been generated, several AI processes including machine learning, predictive analytics and physics-informed AI software can then be used to analyze and fix files prior to printing. By analyzing the parts and mining known datasets about how different AM systems build, solutions like 1000 Kelvin’s AMAIZE, can predict issues with individual builds based on the part geometry, machine parameters and materials choice, reducing the number of failed builds.Here is a concrete example: The Citrine Platform was used in collaboration with the Alliance for the Development of Additive Processing Technologies (ADAPT) Consortium to optimize the design and production of hinges for Mine-Resistant Ambush Protected (MRAP) light tactical vehicles. AI-driven optimization led to a 50% increase in strength, 38% reduction in weight, and consolidated the original six parts down to one.<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzQyMjg2ODM3ODc4LVRyYWRlc2hvdy53ZWJwLjEyODAud2VicCIsImVkaXRzIjp7InJlc2l6ZSI6eyJ3aWR0aCI6OTUwLCJmaXQiOiJjb3ZlciJ9fX0=" class="fr-fic fr-dii" />Exhibition stand designed with AI. Image: James Woodcock<h2>Defining the build</h2>Machine learning (ML) models can analyze huge material datasets and combine that information with data from previous builds to predict how a given material will behave during the printing process. This allows engineers to make informed decisions about which materials to use for an application based on real-world data.Predictive analytics AI ensures that schedules are optimized for maximum efficiency, helping manufacturers reduce downtime by planning and scheduling optimal times to run certain print jobs, based on machine availability, material stocks, and job priorities.<h2>AI as QA </h2>AM processes, especially in metals, require the complex choreography of chemistry and physics. Things can and do go wrong. AI-based computer vision systems use deep learning to track each layer of a print as it is being produced, flagging defects such as layer shifts, warping, or other inconsistencies in real-time. Where fixes can be implemented, feedback loops can action corrections, at the very least part builds can be stopped at the point of failure instead of letting the process run.Once the parts are built, inspection is still required — a time-consuming process open to human error. AI-driven automated inspection systems solve this by using computer vision and deep learning to analyze parts post-production. Companies like EOS have integrated computer vision into their platforms to provide real-time process monitoring (EOSCONNECT), ensuring that defects are identified early in the process. Oqton’s Build Quality combines build simulation, monitoring, and part inspection to ensure parts are built to specification without human interference.<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzQyMjg2ODY1MTQ0LVJvYm90LndlYnAuMTI4MC53ZWJwIiwiZWRpdHMiOnsicmVzaXplIjp7IndpZHRoIjo5NTAsImZpdCI6ImNvdmVyIn19fQ==" class="fr-fic fr-dii" />AI also helps to monitor and maintain machines. Image: James WoodcockPerennial bugbear, post-processing, can also be automated with AI-driven robotic systems, further optimizing the production workflow, and ensuring consistent results by handling support removal and surface finishing.<h2>Future failure alert</h2>Predictive maintenance — i.e., fixing a system before it breaks — is one of the most significant areas where AI adds value to users of AM systems. By using machine learning and IoT/IIoT-enabled sensors, AI can monitor the health of AM machines and predict when maintenance is required before a breakdown occurs. This proactive approach minimizes unplanned downtime, reduces repair costs, and extends machine lifespan. By integrating predictive maintenance solutions, manufacturers can achieve 20% efficiency gains and increase machine utilization, significantly enhancing profitability. <h3>Explanation of terms</h3>Machine Learning           A subset of AI focused on building systems that learn from data to make predictions or decisions without being explicitly programmed for each task.Deep LearningA subset of machine learning that uses neural networks to model complex patterns in data. Deep learning is particularly effective for tasks like image and speech recognition.Generative Models       AI models designed to generate new data samples that resemble the training data. These models can create new text, images, music, or other content.Neural Networks A set of algorithms modelled loosely after the human brain, designed to recognize patterns.<table><tbody><tr><td><h3>Stay Informed on the Latest FormNext 2025 Updates! Sign up to the FormNext newsletter to get notified as soon as tickets are available.</h3><a href="https://info.mesago.de/-lp/5HKPB14784/p4jkP173" target="_blank" rel="nofollow" class="button-main-action">Sign up here</a></td></tr></tbody></table>

An overview of the current state of artificial intelligence in Additive Manufacturing with a special focus on the areas of idea generation and design, the construction process, quality assurance and future error prevention.

The next hype cycle?

<html><head></head><body>Event-based computing is transforming AI processing, delivering smarter, faster, and hyper-efficient solutions across industries like IoT, healthcare, and automotive. It enables smart assistants to respond to user commands intelligently without excessive battery drain and extends battery life in essential wearable devices that continuously monitors your voice, pulse, stride, breathing, activity and surroundings. As technology evolves, event-based computing paves the way for more sustainable and innovative applications in everyday life.Spearheading this transformation, <a href="https://brainchip.com/">BrainChip</a>, a pioneer in event-based AI, has recently introduced their Akida™ Pico solution. Unveiling a revolutionary ultra-low power co-processor designed to fully leverage the capabilities of event-based architectures.&nbsp;BrainChip’s <a href="https://brainchip.com/akida-pico-announcement/">Akida Pico</a> stands out by operating on less than a single milliwatt of power, enabling AI functionality in energy-and size-constrained devices, like wearables, sensor hubs, and IoT devices. By directly processing data streaming from the sensor, Akida Pico reduces reliance on the cloud, delivering substantial advantages in latency, privacy, and energy efficiency.<h2 dir="ltr">Why Event-Based Computing Matters in the Real World</h2>Standard approaches to AI processing often involve processing large amounts of data, much of which may be irrelevant to the task at hand. This inefficiency consumes significant energy and introduces latency, which can be detrimental in scenarios requiring real-time responsiveness. Event-based computing flips AI processing on its head, analyzing only essential inputs as they occur. Inspired by neural processing in the brain, event-based architectures focus on temporal patterns, enabling intelligent systems to respond faster and with far greater energy efficiency.&nbsp;The practical implications of event-based computing and Akida Pico’s capabilities are far-reaching. By focusing on temporal patterns and relevant events, this approach is especially effective in:&nbsp;<ul><li dir="ltr">Voice and Audio Processing: Event-based computing excels at tasks such as voice wake detection, keyword spotting, speech noise reduction, and audio enhancement, where responsiveness and low power consumption are critical.</li><li dir="ltr">Continuous Monitoring: For applications like presence detection and motion tracking, Akida Pico ensures efficient, real-time analysis without the need for constant power-hungry processing.</li><li dir="ltr">Edge AI for Wearables: By delivering personalized insights on minimal energy budgets, Akida Pico supports next-generation wearable devices in healthcare, fitness, and consumer electronics.</li></ul><h2 dir="ltr">Redefining AI Performance for Today’s Applications</h2><h3 dir="ltr">Core Features of Akida Pico:</h3><ul><li dir="ltr">Event-Based Data Processing:&nbsp;The architecture focuses on processing relevant events, reducing unnecessary computations and conserving energy.</li><li dir="ltr">Ultra-Low Power Consumption: Consuming less than 1mW, Akida Pico is ideal for battery-powered devices requiring extended operation times.</li><li dir="ltr">Compact Design: A minimal die area and configurable data buffer allow integration into space-constrained designs without compromising functionality.</li><li dir="ltr">False Alarm Filtering:&nbsp;The co-processor’s neural networks filter out non-critical events, preserving power and reducing system wake-ups.</li><li dir="ltr">Developer-Friendly Integration:&nbsp;BrainChip’s MetaTF™ development environment supports TensorFlow, Keras, and PyTorch workflows, simplifying the development and deployment of Temporal-Enabled Neural Networks (TENNs).</li></ul><h2 dir="ltr">Accelerating Development with MetaTF</h2>A key enabler of Akida Pico’s success is the MetaTF software framework, which allows developers to easily compile and optimize AI models for event-based processing. By supporting popular tools like TensorFlow and PyTorch, MetaTF lowers the learning curve for engineers, enabling faster prototyping and deployment of edge AI solutions.MetaTF supplies tools to convert a model with floating point weights and activations to a model with low bit-width weights and activations while maintaining model performance. These tools also convert quantized models trained using traditional deep learning methods to event-domain models for execution with low-latency and low-power on the Akida Pico Processor.&nbsp;<h2 dir="ltr">Expanding Possibilities of Edge AI Across Industries</h2>Event-based computing is driving innovation across multiple sectors, and the Akida Pico is positioned to lead the charge. Its ability to deliver high performance with ultra-low power consumption makes it a valuable asset in:<ul><li dir="ltr">Industrial IoT:&nbsp;Continuous monitoring of equipment and infrastructure is more efficient and cost-effective with Akida Pico’s energy-saving capabilities.</li><li dir="ltr">Automotive Systems: From early detection systems to low-latency decision-making, Akida Pico enhances the functionality of AI-driven automotive solutions.</li><li dir="ltr">Defense and Security:&nbsp;The co-processor’s real-time event detection and low power requirements are critical for mission-critical systems that operate in resource-constrained environments.</li></ul><h3 dir="ltr">Explore the Future of Edge AI</h3>BrainChip provides a range of software, hardware and IP products that can be integrated into existing and future designs, with a roadmap for customers to deploy multi-modal AI models at the edge.<hr><span style="font-size: 11pt; font-family: Arial, sans-serif; background-color: transparent; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-variant-position: normal; font-variant-emoji: normal; vertical-align: baseline; white-space-collapse: preserve;">Discover&nbsp;<a href="https://brainchip.com/akida-foundations/"><span style="font-size: 11pt; font-family: Arial, sans-serif; color: rgb(17, 85, 204); background-color: transparent; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-variant-position: normal; font-variant-emoji: normal; text-decoration-line: underline; text-decoration-skip-ink: none; vertical-align: baseline; white-space-collapse: preserve;">more about their Akida series</a><span style="font-size: 11pt; font-family: Arial, sans-serif; background-color: transparent; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-variant-position: normal; font-variant-emoji: normal; vertical-align: baseline; white-space-collapse: preserve;">&nbsp;and how the&nbsp;<a href="https://brainchip.com/akida-pico-announcement/"><span style="font-size: 11pt; font-family: Arial, sans-serif; color: rgb(17, 85, 204); background-color: transparent; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-variant-position: normal; font-variant-emoji: normal; text-decoration-line: underline; text-decoration-skip-ink: none; vertical-align: baseline; white-space-collapse: preserve;">Akida Pico Ultra Low Power</a><span style="font-size: 11pt; font-family: Arial, sans-serif; background-color: transparent; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-alternates: normal; font-variant-position: normal; font-variant-emoji: normal; vertical-align: baseline; white-space-collapse: preserve;">&nbsp;can elevate your projects in speech, audio, and medical monitoring.</body></html>

Event-based computing is transforming AI processing, delivering smarter, faster, and hyper-efficient solutions across industries like IoT, healthcare, and automotive.

Smarter, Faster, Hyper-Efficient: Why Event-Based Computing is Key to AI at the Edge

<html><head></head><body>Imagine that a robot is helping you clean the dishes. You ask it to grab a soapy bowl out of the sink, but its gripper slightly misses the mark.Using a new framework developed by MIT and NVIDIA researchers, you could correct that robot’s behavior with simple interactions. The method would allow you to point to the bowl or trace a trajectory to it on a screen, or simply give the robot’s arm a nudge in the right direction.Unlike other methods for correcting robot behavior, this technique does not require users to collect new data and retrain the machine-learning model that powers the robot’s brain. It enables a robot to use intuitive, real-time human feedback to choose a feasible action sequence that gets as close as possible to satisfying the user’s intent.When the researchers tested their framework, its success rate was 21 percent higher than an alternative method that did not leverage human interventions.In the long run, this framework could enable a user to more easily guide a factory-trained robot to perform a wide variety of household tasks even though the robot has never seen their home or the objects in it.“We can’t expect laypeople to perform data collection and fine-tune a neural network model. The consumer will expect the robot to work right out of the box, and if it doesn’t, they would want an intuitive mechanism to customize it. That is the challenge we tackled in this work,” says Felix Yanwei Wang, an electrical engineering and computer science (EECS) graduate student and lead author of a <a href="https://arxiv.org/pdf/2411.16627" target="_blank">paper on this method</a>.His co-authors include Lirui Wang PhD ’24 and Yilun Du PhD ’24; senior author Julie Shah, an MIT professor of aeronautics and astronautics and the director of the Interactive Robotics Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL); as well as Balakumar Sundaralingam, Xuning Yang, Yu-Wei Chao, Claudia Perez-D’Arpino PhD ’19, and Dieter Fox of NVIDIA. The research will be presented at the International Conference on Robots and Automation.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1741875298807-MIT-Policy-Steering-02-press.jpg" style="width: 100%px;" class="fr-fic fr-dib">The interactive system developed by MIT researchers allows laypeople to easily correct a robot’s behavior using simple interactions. For instance, a user can guide a robot’s actions by pointing to an object on a screen, as demonstrated here by graduate student Felix Yanwei Wang.&nbsp; Credit: Melanie Gonick, MIT<h2>Mitigating misalignment</h2>Recently, researchers have begun using pre-trained generative AI models to learn a “policy,” or a set of rules, that a robot follows to complete an action. Generative models can solve multiple complex tasks.During training, the model only sees feasible robot motions, so it learns to generate valid trajectories for the robot to follow.While these trajectories are valid, that doesn’t mean they always align with a user’s intent in the real world. The robot might have been trained to grab boxes off a shelf without knocking them over, but it could fail to reach the box on top of someone’s bookshelf if the shelf is oriented differently than those it saw in training.To overcome these failures, engineers typically collect data demonstrating the new task and re-train the generative model, a costly and time-consuming process that requires machine-learning expertise.Instead, the MIT researchers wanted to allow users to steer the robot’s behavior during deployment when it makes a mistake.But if a human interacts with the robot to correct its behavior, that could inadvertently cause the generative model to choose an invalid action. It might reach the box the user wants, but knock books off the shelf in the process.“We want to allow the user to interact with the robot without introducing those kinds of mistakes, so we get a behavior that is much more aligned with user intent during deployment, but that is also valid and feasible,” Wang says.Their framework accomplishes this by providing the user with three intuitive ways to correct the robot’s behavior, each of which offers certain advantages.First, the user can point to the object they want the robot to manipulate in an interface that shows its camera view. Second, they can trace a trajectory in that interface, allowing them to specify how they want the robot to reach the object. Third, they can physically move the robot’s arm in the direction they want it to follow.“When you are mapping a 2D image of the environment to actions in a 3D space, some information is lost. Physically nudging the robot is the most direct way to specifying user intent without losing any of the information,” says Wang.<h2>Sampling for success</h2>To ensure these interactions don’t cause the robot to choose an invalid action, such as colliding with other objects, the researchers use a specific sampling procedure. This technique lets the model choose an action from the set of valid actions that most closely aligns with the user’s goal.“Rather than just imposing the user’s will, we give the robot an idea of what the user intends but let the sampling procedure oscillate around its own set of learned behaviors,” Wang explains.This sampling method enabled the researchers’ framework to outperform the other methods they compared it to during simulations and experiments with a real robot arm in a toy kitchen.While their method might not always complete the task right away, it offers users the advantage of being able to immediately correct the robot if they see it doing something wrong, rather than waiting for it to finish and then giving it new instructions.Moreover, after a user nudges the robot a few times until it picks up the correct bowl, it could log that corrective action and incorporate it into its behavior through future training. Then, the next day, the robot could pick up the correct bowl without needing a nudge.“But the key to that continuous improvement is having a way for the user to interact with the robot, which is what we have shown here,” Wang says.In the future, the researchers want to boost the speed of the sampling procedure while maintaining or improving its performance. They also want to experiment with robot policy generation in novel environments.</body></html>

Graduate student Felix Yanwei Wang nudges a robotic arm that is manipulating a bowl in a toy kitchen set up in the group’s lab. Using the framework Wang and his collaborators developed, slightly nudging a robot is one way to correct its behavior. Credit: Melanie Gonick, MIT

New research could allow a person to correct a robot’s actions in real-time, using the kind of feedback they’d give another human.

Robotic helper making mistakes? Just nudge it in the right direction

<html><head></head><body>Researchers at the Technical University of Munich (TUM) have developed a wheeled robot that makes its way through a crowd of people safely and without hesitation. A computer on board predicts the movement of people in the vicinity and how they are likely to react to the robot. From this, it calculates the fastest route. Similar algorithms could also be used for humanoid robots or autonomous driving to enable safe interaction between robots and humans.The little robot weaves its way through crowds of people on its wheels like a human. To make this possible, researchers from TUM Professor Angela Schoellig's Learning Systems and Robotics Lab have combined computing power, sensors and mathematical skills. ‘Our robot models the way people will react to its movements to plan its paths. This is the big difference to other approaches that typically ignore this interaction,’ explains Prof. Schoellig.<h2>New route calculated ten times per second</h2>A lidar system constantly sends laser beams into the surroundings, measures their reflections, and uses them to build a precise 360-degree map of what the robot sees. A special focus is placed on people walking around nearby. At the same time, sensors in the wheels measure the robot's own speed and the distances traveled. A computer processes this information, calculates the estimated distances people will cover in the next two seconds, and simultaneously plans the optimum route to the destination. ‘Our robot adjusts its route ten times a second while simultaneously recognizing people's paths,’ explains TUM researcher Sepehr Samavi.<h2>System learns from data based on people in crowds</h2>The robot, which Sepehr Samavi has named ‘Jack’, learns behavioral patterns from humans so that it does not constantly stop due to the risk of collisions. ‘Our mathematical model, on which the planning algorithm is based, was derived from human movements and translated into equations,’ explains Professor Schoellig. For Jack's decision, this means that he does not stop immediately as soon as a person approaches him. He takes into account that people will adapt to the situation, react and change their route slightly so that they do not collide with him. If someone remains on a collision course contrary to expectations, the robot changes its plans quickly and takes a different route - but does not stop.&nbsp;The researchers also incorporate data sets that show people's behavior in crowds. The robot, which has already been used outside of the laboratory, is constantly learning and becoming more human-like: ‘Jack knows his destination, observes people and sees where they are going to constantly optimize his own paths,’ says Prof. Schoellig, ‘almost like a human being.’&nbsp;The TUM researchers have already reached the third evolutionary stage with the new algorithm. Instead of ‘only’ reacting to a situation (stage 1) or ‘merely’ predicting the movements of oncoming people (stage 2), the TUM robot is interactive (stage 3). ‘On the one hand, it predicts other people's movements, but it also manages to influence these people through its behavior and simultaneously avoid collisions,’ explains researcher Sepehr Samavi. &nbsp;&nbsp;<h2>Development for use in autonomous driving</h2>It is precisely such interactive scenarios that form the bottleneck in autonomous driving, says Prof Schoellig. For example, if a vehicle starts accelerating on a highway on-ramp, many drivers will change lanes or tap the brakes. The new approach makes it possible to consider the reaction of others in such a scenario. However, the researchers are initially looking at applications in delivery robots or with wheelchair users. The advantage: these vehicles can reach their destination independently and reliably. Even humanoid robots could benefit from the new algorithms. However, the intelligent vehicle has one decisive disadvantage: ‘A moving robot can simply stop if necessary - humanoids are still quite wobbly and quickly lose their balance,’ says Prof. Schoellig.<hr>PublicationsSepehr Samavi, James R. Han, Florian Shkurti, and Angela P. Schoellig; SICNav: Safe and Interactive Crowd Navigation Using Model Predictive Control and Bilevel Optimization; IEEE Transaction on Robotics; 2025; <a href="https://ieeexplore.ieee.org/document/10726864" target="_blank" rel="noreferrer">https://ieeexplore.ieee.org/document/10726864</a><hr>Further information and links- Prof. Angela Schoellig heads the Chair of Safety, Performance, and Reliability for Learning Systems at the TUM School of Computation, Information and Technology. She is responsible for international relations on the board of the Munich Institute of Robotics and Machine Intelligence (MIRMI) and also coordinates the Robotics Institute Germany, which is funded by the Federal Ministry of Education and Research (BMBF) – <a href="https://www.robotics-institute-germany.de/" target="_blank" rel="noreferrer">https://www.robotics-institute-germany.de/</a>. She is also a member of the Munich Data Science Institute (MDSI) at TUM. &nbsp;&nbsp;- TUM researcher Sepehr Samavi introduces to Jack: <a href="https://youtu.be/7pxZhRulXVY" target="_blank" rel="noreferrer">https://youtu.be/7pxZhRulXVY</a>- Scientific video: <a href="https://www.youtube.com/watch?v=BY23UXwM1sM" target="_blank" rel="noreferrer">https://www.youtube.com/watch?v=BY23UXwM1sM</a></body></html>

Sepehr Samavi and Prof Angela Schoellig next to robot Jack. Credit: Astrid Eckert / TUM

Researchers at the Technical University of Munich (TUM) have developed a wheeled robot that makes its way through a crowd of people safely and without hesitation. A computer on board predicts the movement of people in the vicinity and how they are likely to react to the robot.

Robot Jack moves like a human

<html><head></head><body>Millions of Americans undergo surgery each year. After surgery, preventing complications like pneumonia, blood clots and infections can be the difference between a successful recovery and a prolonged, painful hospital stay – or worse. More than 10% of surgical patients experience such complications, which can lead to longer stays in the intensive care unit (ICU), higher mortality rates and increased health care costs. Early identification of at-risk patients is crucial, but predicting these risks accurately remains a challenge.New advancements in artificial intelligence (AI), particularly large language models (LLMs), now offer a promising solution. A recent study led by&nbsp;<a href="https://engineering.washu.edu/faculty/Chenyang-Lu.html">Chenyang Lu</a>, the Fullgraf Professor in computer science &amp; engineering in the McKelvey School of Engineering and director of the&nbsp;<a href="https://aihealth.wustl.edu/">AI for Health Institute</a> (AIHealth) at Washington University in St. Louis, explores the potential of LLMs to predict postoperative complications by analyzing preoperative assessments and clinical notes. The work, published online Feb. 11 in&nbsp;<a href="https://www.nature.com/articles/s41746-025-01489-2">npj Digital Medicine</a>, shows that specialized LLMs can significantly outperform traditional machine learning methods in forecasting postoperative risks.“Surgery carries significant risks and costs, yet clinical notes hold a wealth of valuable insights from the surgical team," Lu said. "Our large language model, tailored specifically for surgical notes, enables early and accurate prediction of postoperative complications. By identifying risks proactively, clinicians can intervene sooner, improving patient safety and outcomes.”Traditional risk prediction models have primarily relied on structured data, such as lab test results, patient demographics, and surgical details like procedure duration or the surgeon’s experience. While this information is undoubtedly valuable, it often lacks the nuance of a patient's unique clinical narrative, which is captured in the detailed text of clinical notes. These notes contain personalized accounts of the patient's medical history, current condition, and other factors that influence the likelihood of complications.Lu and co-first authors Charles Alba and Bing Xue, both graduate students working with Lu at the time the study was conducted, employed specialized LLMs trained on publicly available medical literature and electronic health records. They then fine-tuned the pretrained model on surgical notes to make better predictions about surgical outcomes. The resulting method – the first of its kind to process surgical notes and use them to make predictions about postoperative outcomes – can go beyond structured data to recognize patterns in the patient's condition that might otherwise be overlooked.&nbsp;Based on nearly 85,000 surgical notes and associated patient outcomes from an academic medical center in the Midwest collected between 2018 and 2021, the team reported that their model performed far better than traditional methods in predicting complications. For every 100 patients who experienced a postoperative complication, the team’s new model correctly predicted 39 more patients who had complications than traditional natural language processing models.Beyond the number of patients who could potentially have surgical complications caught early and mitigated, the study also showcases the power of foundation AI models, which are designed to multitask and can be applied to a wide range of problems.“Foundation models can be diversified, so they’re generally more useful than specialized models. In this case, where lots of complications are possible, the model needs to be versatile enough to predict many different outcomes,” said Alba, who is also a graduate student in WashU’s&nbsp;<a href="https://datasciences.wustl.edu/">Division of Computational &amp; Data Sciences</a>. “We fine-tuned our model for multiple tasks at same time and found that it predicts complications more accurately than models trained specifically to detect individual complications. This makes sense because complications are often correlated, so a unified foundational model benefits from shared knowledge about different outcomes and doesn’t have to be painstakingly tuned for each one.”“This versatile model has the potential to be deployed across various clinical settings to predict a wide range of complications,” said&nbsp;<a href="https://anesthesiology.wustl.edu/people/joanna-abraham-phd/">Joanna Abraham</a>, associate professor of anesthesiology at WashU Medicine and a member of the Institute for Informatics (I2) at WashU Medicine. “By identifying risks early, it could become an invaluable tool for clinicians, enabling them to take proactive measures and tailor interventions to improve patient outcomes.”&nbsp;<hr>Alba C, Xue B, Abraham J, Kannampallil T, Lu C. The foundational capabilities of large language models in predicting postoperative risks using clinical notes.&nbsp;njp Digital Medicine, published online Feb. 11, 2025. DOI:&nbsp;<a href="https://www.nature.com/articles/s41746-025-01489-2">https://www.nature.com/articles/s41746-025-01489-2</a>&nbsp;</body></html>

Study led by Chenyang Lu explores the potential of large language models to predict postoperative complications by analyzing preoperative assessments and clinical notes. (Photo: iStock)

AI for Health Institute team unveils a versatile large language model to enhance perioperative care

Foundation AI model predicts postoperative risks from clinical notes

<html><head></head><body>Felix Heide’s goal was big: help computers see.&nbsp;He did not know he was starting on a path to develop artificial intelligence that uses light instead of electricity, requires vastly less energy than a conventional computer, and produces results hundreds of times faster.“The question for me was always how can we use algorithms to sense and understand the world?” said&nbsp;<a href="https://engineering.princeton.edu/faculty/felix-heide">Heide</a>, an assistant professor of&nbsp;<a href="https://www.cs.princeton.edu/" data-type="link" data-id="https://www.cs.princeton.edu">computer science</a> at Princeton University.At first, Heide, who joined the faculty in 2020, focused on using machine learning to more effectively draw information from light captured by cameras. He succeeded in creating cameras that use visible light or radar signals to detect objects around&nbsp;<a href="https://engineering.princeton.edu/news/2020/06/24/new-radar-lets-cars-spot-hazards-around-corners">blind corners</a> and see&nbsp;<a href="https://engineering.princeton.edu/news/2021/06/29/metropolis-project-backs-solutions-safer-more-resilient-cities">through fog</a> — important goals for aiding drivers or ensuring the safety of autonomous vehicles.&nbsp;Soon, however, Heide realized that making more progress would require rethinking what a camera is, what a lens is.&nbsp;With his first graduate student,&nbsp;<a href="https://ethan-tseng.github.io/">Ethan Tseng</a>, Heide began to explore metasurfaces — materials whose geometry gives them special properties. Instead of bending light inside a piece of glass or plastic, like a conventional lens, a metasurface diffracts light around tiny objects, like light spreading out as it passes through a slit. To build metasurfaces, Heide and his team struck up a collaboration with the lab of&nbsp;<a href="https://people.ece.uw.edu/majumdar_arka/">Arka Majumdar</a> at the University of Washington, who are experts in ultra-small devices that control the interaction of light and matter.“This is a completely new way of thinking about optics, which is very different from traditional optics,” Majumdar said in a&nbsp;<a href="https://www.ece.uw.edu/spotlight/majumdar-computer-vision/">news story</a> published by the University of Washington.In their first breakthrough, the combined research team built a high-resolution camera smaller than a&nbsp;<a href="https://engineering.princeton.edu/news/2021/11/29/researchers-shrink-camera-size-salt-grain">large grain of salt</a>. First reported in 2021 and now widely cited in scientific journals and by news media, their tiny camera depended on the exact positioning of millions of tiny pillars on a metasurface. The team used machine learning to arrange the pillars — each about a 1000th of a millimeter tall and much narrower in width — to make the maximum use of the light hitting the device.&nbsp;Like a conventional lens, this surface focused the light onto a sensor, but it extracted a lot of information from a very small amount of light. The team wrote software that reconstructed this compressed information into surprisingly crisp color pictures.<h2>Surpassing an AI benchmark</h2>Then the researchers realized something profound. The light coming out of these complex arrays of pillars did not need to look to a human eye at all like the object that produced the image. The pillars could function as highly specialized filters that organized the optical information into categories, such as edges, light areas, dark areas, or even qualities that a human viewer might not be able to perceive or make sense of, but that could be useful to a computer that receives this pre-processed information.Suddenly their lens was not just benefiting from artificial intelligence, it was&nbsp;performing functions a lot like what happens in the most sophisticated AI systems that recognize images. Could the lens itself tell the difference between a dog and a horse?“We realized we don’t need to record a perfect image,” Heide said. “We can record only certain features that we can then aggregate to perform tasks such as classification.”The team pressed ahead with this idea and by the end of 2024, they succeeded in creating a system that could identify an object in an image using less than 1% of the computing needed by conventional techniques. Their metasurface lens — actually an array of metasurface lenses — had done 99.4% of the work.In doing so, the team surpassed a key benchmark. Computer vision burst into practical use around 2012 when an artificial intelligence system called AlexNet achieved an unprecedented level of accuracy in identifying images. Heide’s system has now beaten the performance of AlexNet when tested against industry-standard databases of images, such as the ImageNet library developed at Princeton.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1741614629348-metasurface-D1-v4_1600x900.jpg" style="width: 100%px;" class="fr-fic fr-dib">AlexNet “started this whole deep neural network revolution that we have seen play out over the last decade,” said Heide. Now, he said, it appears within grasp for optical systems to start a new revolution in ultra-fast, low-energy-use artificial intelligence.&nbsp;“This could potentially extend beyond image processing, and this is where we are just touching the tip of the iceberg,” Heide said.Francesco Monticone, an associate professor of electrical and computer engineering at Cornell University, said the work by Heide and Majumdar “elegantly and convincingly demonstrates the promise of optical metasurfaces” for low-power, low-latency computing systems. Monticone said that the ability of metasurfaces to perform many operations in parallel and to be finely manipulated across many variables could offer advantages over electronic processors and that Heide and Majumdar’s findings “represent a major step in the quest for ‘optical advantage’ in computing.”<h2>Building on a Princeton legacy</h2>The artificial intelligence that Heide’s lens achieved is rooted in the work of Princeton physicist and neuroscientist John Hopfield, who won a Nobel Prize last year for introducing the idea of an artificial neural network. Hopfield was interested in how the brain learns and remembers. In 1982, he proposed computer algorithms that represent neurons mathematically and represent learning and memory by adjusting the strength of connections between the virtual neurons.&nbsp;Soon after that researchers around the world began trying to construct Hopfield networks using light signals. The prospect of using light to compute is tantalizing. Fast as electrons are, photons are faster. Nothing moves faster than the speed of light.As a practical matter, however, it was the electrical, computer-chip-based systems that eventually took off. The systems that drive today’s artificial intelligence require trillions upon trillions of calculations to adjust the virtual connections in Hopfield’s network, eating up enormous amounts of computing power and energy.Heide’s system offers a different paradigm. His optical system can perform the equivalent of hundreds of millions of computer calculations (called floating point operations or FLOPS) in an instant. Like Heide’s diffractive “filters,” computer-based neural networks apply mathematical filters (known as kernels) to their input data to extract specific pieces of information. Looking at even a few pixels of a picture requires a considerable number of calculations, so modern AI systems look for very specific features in a small number of pixels at a time. In Heide’s system, large and complex filtering operations happen naturally as light passes through, which allows for a few large complex filters, or kernels, to analyze the whole image at once. In each metasurface lens within the array, the tiny pillars reorganize and abstract the light without any electrical input or active control.Heide and Tseng said this concept of a few large optical “kernels” as opposed to many small ones is key to their success. Another key is the integration of hardware and software.In a sense, Tseng said, what the team is doing is like the evolution of animal vision.“Animal vision is joint between optical hardware and neural back-end processing,” Tseng said, citing examples of the mantis shrimp, dragonfly, or cuttlefish that can sense a property of light waves called polarization, which conventional optics don’t pick up. “There are animals that have more exotic vision than what we have, and we suspect that the hardware of their eyes is working together with their brain to perform various tasks.”For Heide, the freedom to follow his curiosity and stray from pure computer science into the different field of optics is unique to universities. As was his move from the specific quest to improve cameras to the threshold of a broad new computing paradigm.&nbsp;“Often I get the question: Working in AI, how do you compete with industry?” said Heide, who also credits the National Science Foundation for supporting much of the work. “This is a perfect example of something that would have never happened in industry because it’s at the intersection of disciplines. It’s so disparate that we wouldn’t have found it without that freedom.”<iframe width="640" height="360" src="https://player.vimeo.com/video/1061275878" frameborder="0" allowfullscreen="" class="fr-draggable" data-lf-form-tracking-inspected-kn9eq4roawkarlvp="true" data-lf-yt-playback-inspected-kn9eq4roawkarlvp="true" data-lf-vimeo-playback-inspected-kn9eq4roawkarlvp="true"></iframe> </body></html>

This ultracompact camera has optical elements that function as a neural network and can classify images at the speed of light. Image courtesy of the researchers

Researchers began to explore metasurfaces — materials whose geometry gives them special properties. Instead of bending light inside a piece of glass or plastic, like a conventional lens, a metasurface diffracts light around tiny objects, like light spreading out as it passes through a slit.

A new way of seeing, a new way of computing

High-Bandwidth Memory (HBM) is a 3D-stacked DRAM designed for ultra-high bandwidth and efficiency. Used in GPUs, AI, and HPC, it tackles memory bottlenecks by stacking dies vertically near processors. This article explores its evolution, architecture, and impact on modern computing.

High Bandwidth Memory: Concepts, Architecture, and Applications

A team of scientists has developed a method to illuminate the dynamic behavior of nanoparticles, which are foundational components in the creation of pharmaceuticals, electronics, and industrial and energy-conversion materials. 

Scientists use AI to better understand nanoparticles.

<html><head></head><body>Bringing AI to the edge is a complex process filled with trade-offs in performance, power efficiency, and cost. Developers often face long development cycles, challenging optimizations, and fragmented hardware compatibility. How can companies accelerate edge AI innovation while reducing time-to-market?In this webinar, host Smair Jaber is joined by Eric Smiley, VP of Business Development at embedUR, and John Marconi, VP of Technology and Architect at embedUR, for an insightful conversation on how ModelNova, embedUR’s AI acceleration platform, is transforming edge AI development. Discover how developers can move from proof-of-concept to deployment in record time by leveraging pre-optimized AI models, curated datasets, and AI-ready blueprints designed for resource-constrained edge devices.&nbsp;<iframe id="JotFormIFrame-250683743144458" title="Edge AI 2025 - embedUR Registration (Embedded)" allowtransparency="true" src="https://form.jotform.com/250683743144458" frameborder="0" style="min-width:100%;max-width:100%;height:420px;border:none;" scrolling="no" class="fr-draggable" data-lf-form-tracking-inspected-kn9eq4roawkarlvp="true" data-lf-yt-playback-inspected-kn9eq4roawkarlvp="true" data-lf-vimeo-playback-inspected-kn9eq4roawkarlvp="true">&nbsp; </iframe><h3 dir="ltr">In this webinar, we’ll explore:</h3><ul><li dir="ltr">How ModelNova slashes optimization time from weeks to minutes</li><li dir="ltr">How pre-trained AI models and curated datasets eliminate the need for extensive training and adaptation</li><li dir="ltr">The biggest trade-offs in edge AI (power, performance, and cost) and how to navigate them effectively</li><li dir="ltr">How embedUR’s collaborations with silicon vendors ensure seamless hardware integration</li><li dir="ltr">Real-world applications where ModelNova accelerates deployment, from face recognition and people detection to audio de-noising</li></ul><h3 dir="ltr">Event details</h3>Webinar Date:&nbsp;May 6th 2025 | 6 pm CET | 9:00 PSTWebinar Duration: 1 hour. 45 minutes discussion with 15 minutes for QnA.&nbsp;<h3 dir="ltr">About the speakers</h3>Eric Smiley,&nbsp;VP of Business Development at embedUR.Bridging ideas with technology, Eric helps connect embedUR’s expertise with customers and partners to envision, design, develop and support tomorrow’s intelligent connected solutions. &nbsp;These innovative products help us communicate, educate, and innovate even further. Eric believes in the power of spreading knowledge and ideas through technology that simplifies our lives and facilitates connections.&nbsp;Studying at Duke University, Eric earned undergraduate and master's degrees in electrical engineering fostering a passion for education and entrepreneurship. &nbsp;He co-founded a Wi-Fi security start-up to help secure wireless connectivity. Eric continued his career at Hatteras Network as a QA and Sales Engineer gaining Carrier networking deployment experience.The service provider knowledge led to an opportunity with a dark fiber service provider to design, deploy and manage networks for School districts. Eric then joined LVL7 Systems as a Field Applications Engineer helping to market Switching and Wireless software solutions for embedded networking products. When Broadcom acquired LVL7, Eric stepped into a Product Line Management role assisting with the formation and rapid growth of Broadcom’s Enterprise Wi-Fi group.For over a decade, Eric has worked as part of embedUR’s uniquely experienced team riding the latest technology waves from Wi-Fi to Cloud to IoT and now to Edge AI, helping our customers utilize advanced innovations in their new solutions. John Marconi,&nbsp;VP of Technology at embedUR.John is a seasoned technology leader with over 25 years of experience in designing, developing, and launching innovative products. He has played a pivotal role in creating award-winning software solutions, bridging low-level device drivers, embedded systems, and scalable cloud architectures with big-data and edge AI innovations. At embedUR, he leads the technology and architecture across the company including embedUR's innovative ModelNova platform and all edge AI initiatives. &nbsp;Passionate about empowering customers to turn their concepts into market-ready products, John's vision continues to drive technological progress.<h3 dir="ltr">About the host</h3>Samir Jaber, editor, writer, and industry expert on topics of technology, science, and engineering.&nbsp;He is the editor-in-chief of the 2024 State of Edge AI and 2023 Edge AI Technology reports with Wevolver. Samir is the Chief Editor and Founder of Wryters, a content marketing and consulting agency. He has comprehensive experience working with Fortune 500 companies and industry leaders as a writer, editor, content manager, and consultant. He is an online content specialist with an academic background in mechanical engineering, nanotechnology, and scientific research. Samir is also a featured author in 30+ industrial magazines with a focus on Artificial Intelligence (AI), the Internet of Things (IoT), 3D printing, Autonomous Vehicles (AV), nanotechnology, materials science, and sustainability. His experience includes award-winning engineering research and patented engineering design in the fields of nanofabrication and microfluidics.</body></html>

Register for our latest Edge AI webinar | May 6, 2025 06:00 PM CET

From Concept to Deployment: How embedUR's ModelNova Accelerates Edge AI Development

A deep dive into the convergence of generative AI and edge computing.

Edge AI Technology Report: Generative AI Edition

A.I.

Specialized hardware accelerators like Neural Processing Units (NPUs) and Tensor Processing Units (TPUs) have revolutionized artificial intelligence (AI) hardware by providing capabilities for accelerating machine learning tasks. Both are designed to optimize performance for tasks like machine learning, deep learning, and neural network computations, but they cater to distinct use cases and requirements.&nbsp;This article discusses the difference in architectures, performance, practical applications, and challenges of these powerful processing units, offering actionable insights for professionals.Recommended reading:&nbsp;<a href="https://www.wevolver.com/article/tpu-vs-gpu-in-ai-a-comprehensive-guide-to-their-roles-and-impact-on-artificial-intelligence" target="_blank">TPU vs GPU in AI: A Comprehensive Guide to Their Roles and Impact on Artificial Intelligence</a><h2 dir="ltr">Introduction</h2>The rapid advancement of AI has led to an increasing demand for specialized hardware capable of efficiently handling complex computations. NPUs and TPUs have emerged as pivotal technologies in this space, each tailored to optimize tasks involving complex machine learning algorithms in unique ways. NPUs, with their low-power design, enhance edge computing performance, making them indispensable for IoT and mobile devices. TPUs, on the other hand, excel in scaling deep learning workloads across massive datasets, driving breakthroughs in cloud-based AI.This article aims to provide professionals with a detailed comparison of these technologies, offering practical insights into their architectures, applications, and limitations. Understanding these distinctions can guide informed decision-making and optimize AI deployments for diverse requirements.<h2 dir="ltr">What Are Neural Processing Units (NPUs)?</h2>NPUs are specialized processors designed to optimize computations for neural networks, enabling efficient processing in real-time. With their ability to efficiently handle workloads like image recognition, voice processing, and natural language tasks, NPUs are an integral part of modern central processing units (CPUs) and System-on-Chip (SoC) designs. Companies like Huawei, Samsung, and Apple are leading the way in integrating NPUs into their devices to enhance AI capabilities without sacrificing power efficiency.<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM3NTY2MzI0NTQ0LWltYWdlMi5wbmciLCJlZGl0cyI6eyJyZXNpemUiOnsid2lkdGgiOjk1MCwiZml0IjoiY292ZXIifX19" style="width: 100%px;" class="fr-fic fr-dib" alt="apple-neural-engine-devices">Apple's NPU, branded as the Neural Engine, is a key component of Apple's A-series and M-series chips. It is designed specifically for accelerating on-device machine learning and AI tasksRecommended reading:&nbsp;<a href="https://www.wevolver.com/article/npu-vs-gpu-understanding-the-key-differences-and-use-cases" target="_blank">NPU vs GPU: Understanding the Key Differences and Use Cases</a><h2 dir="ltr">What Are Tensor Processing Units (TPUs)?</h2>TPUs are hardware AI accelerator application-specific integrated circuits (ASICs) developed by Google to accelerate tensor-based computations, which are foundational to deep learning. Their design is optimized for mathematical operations like matrix-multiplication and parallel computing, making them ideal for tasks such as training large-scale neural networks and performing batch processing.Here is a block diagram showcasing the architecture of TPUs:<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM3NTY2Mzc3NTczLWltYWdlNS5qcGciLCJlZGl0cyI6eyJyZXNpemUiOnsid2lkdGgiOjk1MCwiZml0IjoiY292ZXIifX19" style="width: 100%px;" class="fr-fic fr-dib" alt="TPU-architecture-simplified-block-diagram">A simplified block diagram of a TPUThe TPU architecture is designed to optimize AI workloads by integrating specialized components for efficient computation, storage, and communication. At the core is the Matrix Multiply Unit, which performs rapid large-scale matrix operations critical for AI tasks. Data flows through the Unified Buffer, a local storage unit for managing activations, ensuring quick access and reducing dependency on external memory. The Control Unit orchestrates data movement and synchronizes operations between components. The PCIe Interface and Host Interface facilitate communication between the TPU and external systems, enabling smooth data transfer and coordination. The DRAM Ports provide access to external memory for managing large datasets that exceed the buffer's capacity. The Activation Pipeline processes intermediate results, applying transformations essential for refining AI computations. Additionally, the Accumulator stores partial computation results temporarily for aggregation and further processing. Finally, the Miscellaneous I/O manages input/output tasks, ensuring seamless integration with the broader system. This balanced architecture ensures high efficiency and performance for AI tasks in data centers.<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM3NTY2MzkwMzg5LWltYWdlNC5wbmciLCJlZGl0cyI6eyJyZXNpemUiOnsid2lkdGgiOjk1MCwiZml0IjoiY292ZXIifX19" style="width: 100%px;" class="fr-fic fr-dib" alt="Google-TPU-v4">Google TPUs are primarily used in servers within data centers to accelerate large-scale AI workloads, such as training and deploying deep learning models.Recommended reading:&nbsp;<a href="https://www.wevolver.com/article/tensor-cores-vs-cuda-cores" target="_blank">Tensor Cores vs CUDA Cores: The Powerhouses of GPU Computing from Nvidia</a><h2 dir="ltr">Comparing the Performance of NPUs vs TPUs</h2>NPUs and TPUs exhibit distinct performance characteristics, making them suitable for different AI workloads. Following section compares both the AI chips based on different parameters.<h3 dir="ltr">Speed and Latency</h3>NPUs&nbsp;are designed for&nbsp;real-time, low-latency&nbsp;tasks, making them ideal for&nbsp;edge computing&nbsp;scenarios where immediate responses are required. For instance,&nbsp;NPUs&nbsp;enable instant image processing in&nbsp;smartphones&nbsp;and&nbsp;IoT devicesTPUs often excel in speed and scalability for large datasets, leveraging their specialized architecture for tensor computations. They are optimized for throughput rather than latency, making them suitable for batch data processing tasks like training massive deep learning models. In contrast, NPUs are designed for real-time, low-latency tasks on the edge.&nbsp;<h3 dir="ltr">Power and Energy Efficiency</h3>Power consumption is a critical factor for both NPUs and TPUs, but their optimization differs. NPUs provide better efficiency for edge computing and low-power applications, optimizing computing tasks with limited computational resources. &nbsp;This ensures extended battery life for small devices while performing complex AI tasks. TPUs, while efficient for their scale, are designed for environments like data centers where power availability is less constrained but efficiency remains essential for handling extensive computations.<h3 dir="ltr">Scalability and Flexibility</h3>NPUs&nbsp;are less scalable compared to&nbsp;TPUs, as they are tailored for specific edge&nbsp;AI&nbsp;applications. However, their&nbsp;flexibility&nbsp;in&nbsp;real-time&nbsp;processing tasks makes them invaluable for&nbsp;localized AI workloads.&nbsp;TPUs, on the other hand, are built for&nbsp;scalability, capable of handling&nbsp;large-scale&nbsp;AI&nbsp;models and&nbsp;datasets&nbsp;in&nbsp;cloud-based&nbsp;or&nbsp;distributed environments.<h2 dir="ltr">Architectural Differences</h2><h3 dir="ltr">Core Design</h3>NPUs&nbsp;are engineered to handle&nbsp;neural network activations&nbsp;and&nbsp;real-time&nbsp;matrix computations with specialized cores that focus on efficient, localized processing. These cores typically feature lightweight architectures optimized for low power consumption and latency-sensitive tasks.&nbsp;TPUs, in contrast, utilize systolic array architectures with a high density of computational cores designed for parallel matrix multiplications.<h3 dir="ltr">Programming Frameworks and Ecosystem</h3>NPUs are compatible with general-purpose AI frameworks such as TensorFlow, PyTorch, and ONNX, allowing developers to integrate AI tasks into various devices. As they are integrated within larger processors, they are more general-purpose and are compatible with a number of different computing tasks. TPUs, in contrast, are tightly integrated with Google’s TensorFlow ecosystem, making them an ideal choice for developers already working within Google’s AI infrastructure. The trade-off lies in flexibility versus specialization.<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM3NTY2NDEzOTA4LWltYWdlMS5wbmciLCJlZGl0cyI6eyJyZXNpemUiOnsid2lkdGgiOjk1MCwiZml0IjoiY292ZXIifX19" style="width: 100%px;" class="fr-fic fr-dib" alt="intel-core-ultra-gen-1-series-with-NPUs">Intel Core Ultra processors integrate NPUs to deliver efficient and enhanced performance for modern AI desktop applications.<h2 dir="ltr">Use Cases and&nbsp;Practical Applications</h2>NPUs and TPUs cater to distinct scenarios based on their specialized capabilities.NPUs find their strengths in edge computing, enabling AI tasks directly on devices like smartphones, cameras, and IoT sensors. Examples of&nbsp;NPU Applications&nbsp;include:<ul><li dir="ltr">Real-time object detection in edge devices, such as security cameras.</li><li dir="ltr">Voice recognition in smart assistants operating on low-power devices.</li><li dir="ltr">AI-powered healthcare devices for monitoring patient vitals and diagnosing conditions.</li><li dir="ltr">Industrial <a href="https://www.wevolver.com/article/building-automation-systems">automation systems</a> requiring real-time decision-making capabilities.</li></ul>TPUs are the backbone of AI operations on large amounts of data. Examples of&nbsp;TPU Applications&nbsp;include:<ul><li dir="ltr">Training large-scale machine learning models in cloud environments.</li><li dir="ltr">Batch processing for natural language understanding tasks in virtual assistants.</li><li dir="ltr">Optimizing recommendation systems in large-scale e-commerce platforms.</li><li dir="ltr">Enabling real-time analytics in large datasets for financial and scientific domains.</li></ul><h2 dir="ltr">Challenges and Limitations of NPUs and TPUs</h2><h3 dir="ltr">Engineering Challenges in Implementation</h3>Deploying NPUs and TPUs presents several technical challenges, including:<ul><li dir="ltr">Compatibility Issues: Existing software ecosystems, such as libraries and frameworks, often require extensive adaptation to fully leverage the capabilities of NPUs and TPUs.</li><li dir="ltr">Hardware Integration:&nbsp;Ensuring seamless integration with existing system architectures can be complex, particularly in legacy systems.</li><li dir="ltr">Optimization Challenges:&nbsp;AI models often need to be specifically optimized to achieve peak performance on these specialized processors.</li><li dir="ltr">Cost-Effectiveness:&nbsp;While powerful, the initial investment in NPU or TPU hardware can be prohibitive for smaller organizations.</li></ul><h3 dir="ltr">Limitations in Real-World Applications</h3>Despite their capabilities, NPUs and TPUs face limitations in specific environments:<ul><li dir="ltr">Latency-Sensitive Systems: TPUs, designed for high-throughput batch processing, may not be ideal for real-time applications that demand low latency.</li><li dir="ltr">Hybrid Requirements: Certain applications, such as autonomous vehicles, may require a combination of TPU and graphics processing unit (GPU) capabilities to balance precision and speed effectively.</li><li dir="ltr">Scalability Constraints:&nbsp;While TPUs scale well in cloud environments, NPUs are less adaptable to large-scale deployments due to their edge-specific design.</li></ul>Here is a table summarising the discussion on NPUs and TPUs:<div dir="ltr" align="left"><table><tbody><tr><td>Feature</td><td>NPUs</td><td>TPUs</td></tr><tr><td>Primary Use Case</td><td>Real-time AI tasks on edge devices (e.g., smartphones, laptops, IoT, robotics)</td><td>Large-scale AI workloads in data centers (e.g., training deep learning models)</td></tr><tr><td>Key Applications</td><td>Image recognition, natural language processing, voice recognition, augmented reality</td><td>Language models, natural language processing, large-scale image classification</td></tr><tr><td>Architectural Focus</td><td>Low-latency, power-efficient processing for real-time AI tasks</td><td>High-performance parallel processing and matrix multiplication for deep learning</td></tr><tr><td>Core Design</td><td>Lightweight architectures optimized for low power and real-time processing</td><td>Systolic array architectures optimized for massive parallel tensor operations</td></tr><tr><td>Scalability</td><td>Limited scalability, designed for specific edge AI tasks</td><td>High scalability for training large datasets and deploying AI models</td></tr><tr><td>Power Consumption</td><td>Optimized for energy efficiency in portable devices</td><td>Designed for efficiency in data centers with less strict power constraints</td></tr><tr><td>Programming Ecosystem</td><td>Supports general-purpose AI frameworks (e.g., TensorFlow, PyTorch, ONNX)</td><td>Tightly integrated with TensorFlow and Google Cloud services</td></tr><tr><td>Manufacturers</td><td>Companies like Apple (Neural Engine), Huawei, Samsung, and others</td><td>Google</td></tr><tr><td>Integration</td><td>Embedded in SoC designs for smartphones and IoT devices</td><td>Used as standalone accelerators in server environments</td></tr><tr><td>Key Strengths</td><td>Real-time processing, low latency, energy efficiency</td><td>Exceptional performance for large-scale AI training and inference</td></tr><tr><td>Key Limitations</td><td>Less suited for large-scale AI workloads</td><td>Higher power requirements, less suited for edge applications</td></tr></tbody></table></div><h2 dir="ltr">Conclusion</h2>NPUs and TPUs play complementary roles in the advancement of AI technology, each excelling in specific environments. While TPUs offer unparalleled performance for large-scale machine learning tasks in cloud environments, NPUs are perfect for real-time inference and energy-efficient edge applications.&nbsp;<h2 dir="ltr">Frequently Asked Questions</h2><ol><li dir="ltr">What Is the Primary Difference Between NPUs and TPUs? NPUs are specialized for low-power, real-time tasks, making them ideal for edge devices. TPUs are optimized for large-scale tensor computations, excelling in cloud-based machine learning.</li><li dir="ltr">What Are the Key Roles of NPUs and TPUs in AI? NPUs focus on real-time, low-power AI processing in edge devices like smartphones and IoT systems, enabling efficient inference for tasks such as image recognition and voice commands. TPUs specialize in high-performance, large-scale tensor computations, excelling in data centers for AI training and deployment.</li><li dir="ltr">How Do NPUs and TPUs Compare to GPUs? NPUs are specialized for efficient AI/ML tasks, particularly inference, and are optimized for low-power edge devices like smartphones. TPUs excel in tensor-based operations for deep learning, offering high-speed training and inference, primarily in Google's ecosystem. GPUs, versatile in design, handle a wide range of workloads, including graphics rendering, ML training, and scientific computing, making them ideal for parallel processing. While GPUs are general-purpose, NPUs and TPUs focus on AI-centric optimizations.</li></ol>Recommended reading:&nbsp;<a href="https://www.wevolver.com/article/understanding-nvidia-cuda-cores-a-comprehensive-guide" target="_blank">Understanding Nvidia CUDA Cores: A Comprehensive Guide</a><h2 dir="ltr">References</h2>[1] Google. Cloud TPU documentation. Available from: <a href="https://cloud.google.com/tpu/docs" target="_blank">https://cloud.google.com/tpu/docs</a>[2] Apple. Deploying Transformers on the Apple Neural Engine. Available from: <a href="https://machinelearning.apple.com/research/neural-engine-transformers" target="_blank">https://machinelearning.apple.com/research/neural-engine-transformers</a>[3] Van Belle JL. Why neural processing units (NPUs) are the next Big Thing in AI. 2024. doi:10.13140/RG.2.2.35619.87845. Available from: <a href="https://www.researchgate.net/publication/383184868_Why_neural_processing_units_NPUs_are_the_next_Big_Thing_in_AI" target="_blank">https://www.researchgate.net/publication/383184868_Why_neural_processing_units_NPUs_are_the_next_Big_Thing_in_AI</a>[4] Sanmartín D, Prohaska V. Exploring TPUs for AI Applications. IE Robotics &amp; AI Club, IE University, Madrid, Spain. arXiv preprint arXiv:2309.08918. Available from:<a href="https://arxiv.org/pdf/2309.08918" target="_blank">&nbsp;https://arxiv.org/pdf/2309.08918</a>.

NPUs are integrated units that excel in real-time AI tasks on edge devices like smartphones and IoT systems with low power consumption. TPUs are standalone processors designed for large-scale AI workloads in data centers, delivering exceptional performance in deep learning tasks.

NPU vs TPU: Understanding the Key Differences in AI Hardware Accelerators

Random Access Memory (RAM) strip inserted into the Motherboard

This article is a detailed analysis of In-Memory Compute technology, covering its architecture, use cases, recent advancements, and practical implementation strategies to enhance computational efficiency.

In Memory Compute: Transforming Data Processing for Speed and Scalability

GPUs excel in parallel processing for graphics and AI training with scalability, while NPUs focus on low-latency AI inference on edge devices, enhancing privacy by processing data locally. Together, they complement each other in addressing different stages of AI workloads efficiently.

NPU vs GPU: Understanding the Key Differences and Use Cases

Fine-tuning large language models adapts pre-trained models to specific tasks or domains using tailored datasets, while Retrieval-Augmented Generation (RAG) combines retrieval systems with generative models to dynamically incorporate external, up-to-date knowledge into outputs.

RAG vs Fine-Tuning: Differences, Benefits, and Use Cases Explained

Large Language Model (LLM) training involves teaching AI models to understand and generate human-like text by processing vast amounts of data, significantly enhancing their language comprehension and production capabilities.