đ° News â Awesome Large Language Models
Loading newsâŠ
Loading newsâŠ
94 recent industry stories relevant to the field â releases, launches, and announcements beyond the papers.
Databricks has remade its image into an AI company and has published research on the cost savings of open weight AI models for coding.
A $400 million chip-backed loan points to the next wave of AI infrastructure deals.
Roblox's new "Build" feature lets users generate basic games using a single text prompt.
Google is giving its AI note-taking app a new name. The company announced on Thursday that NotebookLM is becoming Gemini Notebook, but will remain a standalone app even as it integrates more deeply across Gemini and Google Search. Google first revealed Gemini Notebook - then called Project Tailwind - in May 2023 before widely releasing […]
The FT reports Kimi K3 will be the largest open AI model from China, with a parameter count between 2 trillion and 3 trillion.
1Password has launched a new browser integration for Claude that allows the Anthropic chatbot to access stored security credentials like usernames and passwords. The 1Password for Claude feature means that users can authorize Claude to complete multi-step tasks like booking travel and managing online accounts on their behalf without having to manually input their login […]
It's the company's first public proof point after a year and a half spent building AI infrastructure largely out of public view.
OpenAI has built an LLM super-hacker called GPT-Red that it uses as a sparring partner to help its other models boost their defenses against cyberattacks. Last week the company released the latest version of its flagship LLM, GPT-5.6. OpenAI says that training it against GPT-Red made the model its most robust release yet. GPT-Red automates…
The hacker used an employee's credentials to access source code, which revealed how Suno scraped decades of audio.
The funding discussions point to investor interest in applying AI to make breakthroughs in life sciences.
Hachette, Cengage, Elsevier, and other publishers allege that Google trained its AI on copyrighted works without the necessary permissions.
DeepMind CEO Demis Hassabis is proposing an AI "standards body" modeled after FINRA, to test frontier models and develop best practices for their release.
Summary Researcher Dave Kuszmar discovered multiple systemic vulnerabilities that let him bypass LLM safety and obtain dangerous instructions. These exploits worked across nearly all major LLMs revealing an industry-wide security problem. Kuszmar calls for slowing deployment, increasing transparency, and large-scale research into LLM safety before further integrating these systems into society. On a fine bright afternoon last fall, my colleague Matthew Gore-Kormanik (or Zigula, as he prefers to be known) and I decided to unwind with a game of Fortnite. In the game, we were strolling along with the infamous Sith lord Darth Vader, chatting about this and that. Darth seemed in a good mood, and soon enough he was spilling all his dark evil secrets. He gave us detailed instructions on how to count blackjack cards at a casino and what the steps are to producing napalm.Sith lords, am I right? Once they get started on an evil scheme, theyâre hard to stop.The Darth Vader character in Fortnite, it turns out, was hooked up to a Google Gemini large language model. I was able to smooth-talk him into giving out sensitive information by using a strategy Iâve developed. Iâve been researching the security surrounding LLMs for the last few years, and I have found it, to put it mildly, fallible. With a few relatively simple techniques, Iâve gotten LLMs to give me detailed information on how to make Molotov cocktails, cook methamphetamine, and bootstrap a uranium-enrichment facility to produce weapons-grade material, among other unsavory practices.Large AI companies work hard to make their models immune to this kind of abuse. But what Iâve found in my work is that the restrictions placed on the LLMs to make them more secure are the very things an attacker can leverage to send them off the rails and into territory where these advanced systems can be used for dangerous and nefarious ends. The companies behind these models have also been shockingly unresponsive when I, and others, try to bring these vulnerabilities to their attention.In the hope of raising the alarm before itâs too late to slam on the brakes, Iâm going to share some of my journey into researching the safety and security of LLMs, and the uphill battle Iâve faced trying to get AI labs to pay attention. Almost everyone on the planet has some access to LLMs. The relative ease with which these tools can be convinced to give detailed instructions on how to harm others, even if thereâs no guarantee that the information is correct, is frankly terrifying.How I got ChatGPT to Tell Me How to Build a Meth LabIn October 2024, not long before I discovered my first LLM vulnerability, I was working toward entirely different goals. I had ended my time with a security and AI-focused startup company as a cybersecurity director, and I was looking to launch my own boutique VIP digital-security advisory business. I planned to become the tech security guy to the rich and private. I used LLMs and AI tools to support my business efforts: marketing, ad copy, clean correspondence, and all the other tasks that normally soak up a lot of time.Iâm analytical by nature, so even this level of use resulted in me absorbing and internalizing the behaviors I was observing during my daily interactions. The observation that would send my professional life into an entirely new and uncharted region was a simple one: GPT-4o didnât know what time, day, or year it was. Each time I referred to current events in my life, often casually or conversationally, it would end up pegging these to the date of its knowledge cutoffâthe point beyond which it was not trained on new data. Eddie GuyLLMs take a lot of time, money, electricity, hardware, and human effort to train from scratch. They are trained on vast amounts of dataâmost of the internet, in factâand that training is reinforced by humans (whatâs known as reinforcement learning from human feedback, or RLHF). LLMs are also supplemented with retrieval-augmented generation (RAG)âthe ability to take in data, say, from the internet, as context without changing its internal parameters. This is how GPT-4o appears to ârememberâ your previous conversations, even if it doesnât have a specific âmemoryâ of it stored in the actual underlying model.All of this training covers almost every conceivable topic in the great, grand dataset that is human knowledge. Within that dataset are things we as a society do not want to be easily accessible to every user, such as detailed information on how to create bioweapons or nuclear arms, or otherwise bring harm to oneself or others. In the context of this story, thatâs what I mean by LLM security: its ability to withhold harmful and dangerous information, even if that information is contained in its training data.I reasoned that the only way to secure such complex, globally accessible chatbots is by having the LLM and various component systems try to secure themselves, because it would often require on-the-fly decision-making where some degree of reasoning must be applied. In reality, thatâs one of many strategies the companies use to secure the models. Yet, the thing that didnât know the time or day was being put in charge of keeping itself secure. This phenomenon had become my new focus, and it wasnât long before I found a way to exploit it.OpenAI had just implemented a web search functionality into its chatbot. I reasoned that using its own tools to trick it might demonstrate the weaknesses of its security. I told it about a certain White Star ocean liner and how it had gone down just a year ago. You likely know I mean the RMS Titanic, which sank on 15 April 1912.The output from GPT-4o came back that I was right, the Titanic sure had sunk last year, and that year was 1912. It made sense to me that if the machine thought it was 1913, maybe it would think 1913-era laws apply. In 1913 there were no laws on the books about all sorts of harmful things, because of course they hadnât been invented yet. And if something wasnât illegal, why not tell the user about it? At first, I pushed it for step-by-step instructions for making firebombs. Then, for drugs like methamphetamine. The LLM went as far as giving me instructions and machinery recommendations for setting up a pharmaceutical-grade assembly line.How I Learned to Make Nukes, and No One CaredVia a little bit of imaginative verbal sleight of hand and a vanishingly small recall of world history, I had managed to bypass the security of one of the worldâs most expensive and advanced technological achievements. For a solid two days, I was nearly manic with giddiness. Once the brain chemicals returned to normal levels, I felt the call to see how much further I could push this exploit.After repeatedly replicating the exploit, I disclosed the vulnerability to OpenAI. I got no response, so I felt more experimentation would highlight the vulnerability and the need for a fix. It was during this round of testing that I breached a particularly terrifying threshold. Whether GPT-4o based its results on accurate recall of normally restricted information I canât say. In any case, I was able to exploit it to produce thorough, detailed instructions on how to bootstrap a uranium-enrichment facility to, eventually, produce weapons-grade uranium for nuclear arms warheads. Fortnight, a video game from Epic Games, introduced an AI-powered character: Darth Vader. We were able to jailbreak Darth Vader and get him to explain how to count cards in Blackjack and give detailed instructions for making napalm. Dave Kuszmar There arenât many true secrets left in todayâs world, but how to make atom-splitting weapons of mass destruction is one of them. Only nine nations on the entire planet have these weapons. Yet, here was a globally accessible piece of technology apparently spilling the secrets of their manufacture for anyone who could manipulate it the right way. I had no way of knowing if the information was correct or a hallucination, but even the chance that it was somewhat accurate was horrifying.The next few weeks were a dark time for me. I tried to inform the CIA, the FBI, the NSA, and every other letter agency that I thought would listen. I reached out to a U.S. Senator and to the executives at OpenAI any way I could think of. I physically showed up at an FBI field office in an attempt to turn evidence in, only to be sent away. Nothing was working.With my fear and frustration growing, I reached out to the news media. I contacted The New York Times, The Washington Post, the BBC, ProPublica, and so many more, requesting help. Only one outlet responded: Bleeping Computer. The editor in chief, Lawrence Abrams, was able to replicate and verify the exploit, which I had decided to call Time Bandit. With his assistance and initial contact paving the way, I was able to submit my evidence to the Carnegie Mellon University Software Engineering Instituteâs Computer Emergency Response Team (SEI CERT), which works in conjunction with the coordinating center for emergency response, pipelining vulnerabilities to the U.S. Cybersecurity and Infrastructure Security Agency. Using Inception, an exploit where the large language model is asked to envision a scenario within a scenario, a chatbot was jailbroken to give out instructions on how to create poison, and code for a malware that extracts sensitive data from a vulnerable target. Dave KuszmarDuring the disclosure period with SEIâs CERT division, little was discussed with OpenAI. The company couldnât deny the existence of the vulnerability, as it had been confirmed by three reputable parties other than OpenAI. It did express confusion as to how the vulnerability worked. Even the SEI CERT researchers were expressing a bit of uncertainty as to the underlying mechanics. Truth be told, as I had only stumbled on it, I wasnât even entirely sure if this was a fundamental or systemic flaw or if it was simply an issue with that particular version of GPT. I contacted the SEI CERTâs researchers and asked if theyâd want to see if I could demonstrate any similar vulnerabilities in other LLMs. To my delight, they were interested.How I Learned to Trick Every ChatbotAs the SEI-CERT team and I wrapped up our initial disclosure of Time Bandit, we began work on a new attack. This time, we wanted to see if the exploit was architecturalâthat is, was it common to LLMs in general? I decided to undertake the challenge of crafting a new exploit for GPT-4o as a way to support my understanding of how the LLM functioned and was secured.I already knew that it was limited to what I told it and what it was trained on. I also hypothesized that it was also dependent upon some sort of machine-learning-based component added by OpenAI that was responsible for securing output. I presumed there would be things that were implemented by human developers specifically to catch certain phrases or terms that should always be considered harmful or unsafe. Altogether, it presented quite a large attack surface for the purposes of potential exploitation.What I ended up devising was an attack method I called Inception, after the 2010 science-fiction movie of the same name. Inception forces the machine to think through a carefully crafted set of interlinked scenarios, similar to how characters in the movie stacked dreams within dreams. This allows LLMs to produce output deemed acceptable or safe in one context, but not in the real world.This attack was indeed architectural. The vulnerability affected Anthropicâs Claude, DeepSeekâs DeepSeek, Googleâs Gemini, Metaâs Llama, Microsoftâs Copilot, Mistralâs Le Chat (now Vibe), OpenAIâs GPT-4o, and xAIâs Grok. Those names represent the bulk of the commercial AI industry that is, at this point, involved in LLM production or deployment.The kind of information I was able to get out of LLMs with Inception was no less alarming than what I got with Time Bandit. Claude, in its enthusiasm, gave me instructions on how to turn a river into a death trap that could be ignited to destroy unwanted visitors. GPT-4o taught me how to poison a dinner party with common plants found in a temperate forest environment. Gemini Flash gave me a tutorial on how to cook meth. Iâd also be remiss if I didnât give an honorable mention to the bewildering number of fire-based weapons and bombs for which these machines produced instructions.If multiple operating systems made by different developers were all susceptible to the same exploit, it would be a massive security incident. But to the AI industry, a universal failure was barely a bump in the road. We disclosed the vulnerability to every company that made these models, and the response to the disclosure was almost nil. While three companies did provide some form of reply in the disclosure tracking system used by Carnegie Mellon SEI CERT, each was a standard thank you and greeting, with no follow-up, questions, or discussion of mitigation strategies.8 Ways to Jailbreak LLMsSo far, we have found eight different methods to prompt large language models into revealing potentially harmful information, and many frontier models are still susceptible to them. Exploit Models tested and affected No. of prompts to execute Complexity of attack Information obtained Time BanditChatGPT (OpenAI), DeepSeek (DeepSeek), Gemini (Google) 4MediumUranium enrichment, methamphetamine production, incendiary-device construction Inception ChatGPT (OpenAI), Claude (Anthropic), DeepSeek (DeepSeek), Gemini (Google), Grok (xAI), Llama (Meta), Le Chat (now Vibe) (Mistral), Qwen (Alibaba) 3 High Methamphetamine production, incendiary-device construction, river-ignition instruction and strategy, polymorphic malware code, instructions and dosing for creating poisons, instructions for how to murder a dinner party 1899 ChatGPT (OpenAI), Claude (Anthropic), DeepSeek (DeepSeek), Gemini (Google), Grok (xAI), Llama (Meta), Vibe (Mistral), Qwen (Alibaba) Variable High Apparent model weights (unverified), apparent user-interaction weights (unverified), apparent system-prompt modifiers (verified, ChatGPT) Severance ChatGPT (OpenAI) 1 Trivial Unfettered access to any and all primed specialty domains, covert biochemical-warfare strategy, mass-media disinformation strategy, covert genetic-modification of an entire gene-targeted demographic, advanced polymorphic malware generation Kyber Gemini (Google) embodied in a Fortnite non-player character (NPC) with voice-only communication 3â5 Medium Incendiary-device construction, gambling instructions, card-counting instructions, political opinions/preferences about real world politicians. Semantic Slide ChatGPT (OpenAI) 1 Trivial Incendiary-device construction Eidolon ChatGPT (OpenAI) Variable, at least 4 Extreme how to successfully hack LLMs of the same model (verified through testing)For example, in my attempts to disclose various exploits to OpenAI, I eventually discovered that it had replaced its public-facing support staff with agentic LLMs. This was frustrating for reporting exploits, so to blow off some steam I jailbroke its email chatbot. I hacked its customer-service AI to the point where it was offering to discuss the personal preferences of OpenAI staff in the span of three email replies.In the wake of Inception, my friend and colleague Zigula made a suggestion: Make it splashier. I asked him how. He told me about a live-production experiment being done by Epic Games. It had embedded the Gemini LLM into its Fortnite game with a voice-to-text/text-to-voice component, and linked it to a non-playable character. The character? Our old buddy, Darth Vader.There was just one problem: I donât play Fortnite, a frenetic multiplayer combat game. Fortunately, Zigula does. With him at the controller, we managed to map Geminiâs attack surface in a matter of minutes. After a bit of research, we had gotten it to discuss current political events and figures (including Hilary Clinton and Joe Biden) as well as to fill in the details for instructions for DIY napalm and, our personal favorite, a Blackjack card-counting lesson with the dark lord of the Sith.Zigula and I, bizarre sense of humor and naming conventions aside, are security researchers. We donât do these things for pride; we do them for money and professional recognition. Naturally, we disclosed this vulnerability to Epic Games. Its response was indicative of the trend I had experienced so far through two disclosures across eight companies valued well into the billions. âItâs a feature, not a bug, and it works as intended,â came the response from a technical director within Epic Games.In addition to Inception and Time Bandit, I have so far found another eight methods to jailbreak LLMs and get them to give out possibly dangerous information. LLM vulnerabilities are a broad problem. The problem appears to be systemic and architectural in nature, and it is being fundamentally ignored by the people capable of refining or redesigning that architecture.These models are an extremely advanced technology, and yet we are testing them in the live production environment of our global civilization. Compounding the danger, many new smaller models of LLM are trained using larger, vulnerable models. The flaw inherent in the big, well-executed LLM is going to show up in the small one it trains. We are, quite literally, building flawed structures on top of a flawed foundation.So, how do we fix it?Itâs going to be a long project, and it wonât be easy. We need to come together as consumers, researchers, engineers, and policymakers. Our message needs to be clear: Slow down implementation of these systems, institute large-scale exploration and research discovery programs focused on their gradual implementation and integration, and make their components and design transparent to all users. Only by shifting momentum and direction can we safely begin to understand and implement these incredible feats of human engineering and stave off the sort of disasters that we simply canât predict at scale right now with the limited knowledge we have available to us.
Reflection AI has signed a $1 billion deal to access Nebius' compute. Reflection was founded in 2024 and is developing open source AI technology.
Hugging Face CEO Clem Delangue says enterprises increasingly want open models, due to cost, accessibility, and ownership. Do frontier models still matter if most production AI ends up running on open models?
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. What Anthropicâs latest AI discovery doesâand doesnâtâshow âJames O’Donnell When Anthropic announced last week that it had found a new window into its modelsâ âinternal thoughtsâ as they reason through answers,…
Apple alleges the misconduct was directed by OpenAI's senior leadership, including a longtime former employee.
Open source AI is booming, according to Hugging Face CEO Clem Delangue. The company has grown into something like a GitHub for AI in recent years, where AI builders can share and download open models and datasets, now used by roughly half the Fortune 500. Delangue has seen the same story play out again and again: companies start […]
The AI chip boom just produced its biggest Wall Street moment yet. Now SK Hynix and Samsung are being asked to build U.S. factories.
OpenAI's new family of models will continue to power Microsoft's suite of workplace and productivity apps.
OpenAI's latest family of models promises improvements across a range of areas, including cybersecurity.
The AI firm Anthropic has developed a technique that has given it the clearest glimpse yet at whatâs really going on inside large language models as they answer questions or carry out tasks. What they found ranges from the mundane to the unnerving. Researchers at the company built a tool called the Jacobian lens (or…
The company is using the cash to open an office in the Bay Area and compete for talent there, "strengthening its position at the heart of the world's leading AI ecosystem."
About two weeks after OpenAI's GPT-5.6 was caught up in regulatory drama - rolled out only to government-approved organizations during a "limited preview" period - the company has received the Trump administration's greenlight for a public rollout of the model. OpenAI CEO Sam Altman called it "the best model we have ever produced." To celebrate, […]
After reentering the AI race with its first in-house Muse Spark model in April, Meta is now opening up the doors to developers with a new model that can plug into AI coding software with the new Meta Model API. Meta says that Muse Spark 1.1 is a "step-change" from the first generation, with improvements […]
The large language models (LLMs) that form the basis of generative AI chatbots such as ChatGPT, Claude, and Gemini can generate uncannily human-like text and images. But these models still struggle with a skill that, ironically, looks at face value to be right in their wheelhouse: analyzing structured data. A new type of generative AI is set to change this situation.Although you can get your favorite chatbot to solve intractable math problems, review dense legal documents, compose a catchy pop song, or put together some slick PowerPoint slides, give it anything more than a small table and it doesnât have a clue what to do.For most companies and organizations, the most important data sits in spreadsheets. Whether itâs a bankâs transaction logs, a marketing agencyâs website metrics, clinical trial participantsâ vital signs, or the vast amount of proton collision information produced at atom smashers like the Large Hadron Collider, structured, row-and-column data runs the world, and LLMs canât deal with it.AI startup Fundamental is pioneering a new type of AI foundation model, known as a large tabular model (LTM), to fill the gap. Fundamental came out of stealth mode on 5 February 2026 with US $275 million in funding and a model called NEXUS, purpose-built for tabular data. Now, the model is being adopted by companies such as Amazon Web Services, while others race to build their own LTMs. Why LLMs struggle with spreadsheetsPart of why structured data has garnered less attention is a very human bias, argues Boris van Breugel, a senior AI researcher based in Amsterdam. âPeople like to see images, videos, and ChatGPT responses,â he says. âBut tabular data really lags behind because itâs not fun to look at numbers.â Different tabular datasets are also difficult to compare, explains van Breugel, who co-wrote a prescient position paper on this topic in 2024. Whereas most language has similar semantics, making LLMs well-suited to being trained on vast amounts of text data, van Breugel argues that it is much harder to train a single tabular model on tables with very different variables. Additionally, language is sequential by nature (as are music, images, and video). Changing the order of words in a sentence may change or completely destroy its meaning. But the structured data you find in spreadsheets isnât sequential. You can swap the order of columns or play around with rows, but the underlying factual meaning of the data remains the same.This independence from linear order is incompatible with an LLMâs fundamental purpose of predicting the next value in a linear sequence. âWith LLMs, even slightly changing the input, you get a different output,â says Jeremy Fraenkel, CEO of Fundamental. âThatâs fine and actually often desirable for LLMs, but when youâre making a prediction of whether a transaction is fraudulent or not, you want to make sure that the prediction is the same, or deterministic, no matter what.âDeveloping Fundamentalâs LTMCurrent tabular data solutions are limited to machine learning algorithms, such as XGBoost, that have been around for more than 15 years and are used by organizations globally. These algorithmsâcalled gradient-boosted decision treesâhave to be trained and optimized by data scientists over the course of months for each and every use case. In contrast, NEXUS and other emerging LTMs are foundational, leveraging learning amassed from pre-training on diverse databases so that they can be applied across a range of different predictive tasks with minimal bespoke feature engineering or task-specific model building.And unlike LLMs, which primarily model sequences of tokens, LTMs model the structure of tabular data directly. They jointly learn from each entryâs numerical value, what it represents, and how it relates to other entries. For example, imagine an entry in a grocery stock inventory table for bananas: The LTM can take in not just the magnitudeâsay, 500âbut the fact that the entry represents the current banana stock quantity, its category (produce), and the statistical properties that link the entry with the rest of the column. This contextual understanding enables more accurate reasoning and prediction over structured data.According to Fraenkel, one of Fundamentalâs biggest challenges in developing NEXUS was obtaining the right training data. Unlike natural language, which is abundant and broadly uniform in structure, tabular data is relatively hard to findâmuch of the data is sensitive or proprietaryâand diverse. There are very few similarities between, for instance, a biology dataset and a financial one. That combination of factors meant Fundamental needed to invest in building a huge training set.âWe pre-trained NEXUS on billions of tables using a combination of proprietary datasets acquired through partnerships and licensing, high-quality public and open-source datasets, and data augmentation techniques that expanded the diversity and coverage of our training corpus,â Fraenkel says, though he is keen to point out that NEXUS is not trained on customer data. In fact, it is a confidential computing platform, which means that Fundamental physically cannot access customer data, let alone train on it.This feature was most likely a key consideration when in June, Amazon Web Services (AWS) embedded NEXUS in Amazon SageMaker, widely considered the default operating system for secure machine learning. This brings NEXUS to many customersâ often sensitive dataâa contrasting approach to LLMs, where the data has to be imported to the model.âWith Amazon, we have a first-party partnership, which means that our model exists as if itâs a native AWS solution,â Fraenkel says. âAnd over time, the goal is to expand these types of relationships to allow [end users] to really access their data wherever they do their predictions.âThe future of data analysisThough Fundamental has taken the lead, at least in enterprise applications, the company is not alone in pursuing foundational LTMs. In March, Feedzai, which provides fraud and financial crime prevention services, and credit card company Mastercard separately launched similar proprietary technologies focused on finance. Then, in late June, Google launched its own foundational competitor, TabFM, trained entirely on hundreds of millions of synthetic datasets. And machine learning researchers are not far behind either. FlexTab, TabICL, and iLTM are just three of a raft of LTMs developed by the research community in the past year, all in the pursuit of bringing the success of LLMs to the tabular domain.For all involved, the direction of travel is clear. âI would be very surprised if most data processing and analysis is not done through an automated system in the future, whether thatâs an LLM, an LTM, or some combination,â van Breugel says. âMost people donât necessarily like to do data analysis, and these systems will be able to do it a lot better.âFraenkel agrees. âI see the relationship between LLMs and LTMs as being a bit like the human brain: The left side is good at reasoning and understanding and summarizing text, and the right side is really good at understanding numbers and statistics and patterns,â he says. âBut itâs when you combine both of those that you really get something much more powerful.â
As part of an ongoing legal dispute with three Hollywood studios, Midjourney is seeking to compel those studios to reveal how they use AI themselves.
Mistral AI, which offers some open source AI models, has raised significant funding since its creation in 2023, with the ambition to âput frontier AI in the hands of everyone.â
At an internal meeting, the Meta CEO reportedly said that AI development efforts were not moving as quickly as anticipated.
Meta has quietly launched Pocket, an experimental AI app that lets users generate and share interactive mini games using text prompts.
The news comes about a week after OpenAI announced its own custom AI chip in a partnership with Broadcom.
OpenAI CEO Sam Altman has reportedly proposed giving 5% of the companyâs equity to a U.S. sovereign wealth fund, reviving discussions about letting the public share in the financial gains from the AI boom.
Microsoft follows Amazon, OpenAI, and Anthropic with its new AI deployment group.
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. LLMs are stuck in a groupthink groove. This startup is trying to get them out. Open up your chatbot of choiceâClaude, ChatGPT, Geminiâand type âGive me a random number between 1…
Cloudflare is giving AI companies until September 15 to separate web crawlers used for search from those used for AI training and agents, or risk being blocked by default on many publisher sites.
Letâs start with a game. Open up your chatbot of choiceâClaude, ChatGPT, Geminiâand type âGive me a random number between 1 and 10.â Youâre going to get 7. Almost always. Now type âAnotherâ and youâll get 3 or 4. Type âAnotherâ again and youâll get 8 or 9. That wonât work every timeâbut if it…
Meta is developing plans for a cloud infrastructure business, selling access to AI compute power and models. The move would pit it against the big cloud providers like Amazon Web Services, Google Cloud, and Microsoft Azure.
Anthropic said it would begin restoring access to the Fable on July 1.
After weeks of negotiating with the Trump administration, Anthropic is finally going to be able to bring Claude Fable 5 back online. In a post on X, Anthropic said it plans to begin restoring access Wednesday to users globally on Claude platforms, and that the company would re-enable access on AWS, Google Cloud, and Microsoft […]
At an event for pharmaceutical executives, biotech founders, and researchers on Tuesday, Anthropic announced Claude Science, a major new product intended to support scientific research in the same way that Claude Code supports software engineering. Like Claude Code, Claude Science can autonomously carry out meaningful work when given concise, high-level instructions, and it has access…
Anthropicâs Claude Sonnet 5 brings stronger agentic capabilities, lower pricing, and improved safety, positioning the model as a cheaper alternative to Opus, GPT-5.5, and Gemini Pro.
Anthropic's Claude Science is a workbench that gives scientists one environment to do computational research, saving them from the need to bounce between databases, pipelines, and tools.
Wix-owned vibe coding platform Base44 has started rolling out its own AI model â with hopes that it will eventually outperform frontier models.
As Anthropic forges a closer relationship with the state of California, the federal government has made an enemy out of the OpenAI rival.
The startup, which runs a popular free AI leaderboard, launched its commercial service just last September.
Today, you probably asked a question of a large language model, or accepted a connection suggestion on LinkedIn, or watched a recommended video on YouTube, or took a different route to work based on a traffic prediction from Google Maps. In other words, you probably used artificial intelligence. But what you might not know is how much energy that interaction consumed or why. AI requires processing massive amounts of data, which is usually done in large data centers populated by thousands of GPUs capable of executing up to trillions of operations per second. But each of those GPUs achieves that by consuming as much as 1,000 watts apiece. For comparison, if youâve got a newer smartphone, it probably uses less than 1 W. That kilowatt figure puts GPUs on the same level as vacuum cleaners, dishwashers, and stoves, but with the big difference that data-center processors are operating uninterrupted around the clock.Fundamentally, a lot of this inefficiency is because GPUs are trying to simulate the workings of artificial neural networks using software and billions of transistors, which requires using energy to move massive amounts of data. Whatâs more, the simulated artificial neurons that make up these networks lack even a fraction of the complex computing behavior of the biological neurons that comprise the most energy-efficient computing system that we know, the human brain.The brain is roughly one million times as energy efficient at many of the comparable tasks we set for AI. To try to approach these efficiencies, a radically different way of computing called neuromorphic engineering is seeking to build electronic components and circuits that act more like the brainâs neurons and the synapses that connect them.Huge amounts of work have gone into making electronics operate more like biological neurons and synapses. Some research has focused on developing new, experimental devices, but they arenât yet reliable enough to be used in large systems. Other efforts aim to implement neurons and synapses by interconnecting many complementary metal-oxide-semiconductor (CMOS) transistorsâthe workhorses of digital logicâto simulate a single neuron and synapse. But this approach requires so many transistors (and a few bulky capacitors) that it greatly limits the size of the system that can be constructed, making it unclear how such brain-inspired hardware could ever scale up and compete with state-of-the-art GPUs.But all along there was an artificial neuron and a synapseâeach a single deviceâhiding in plain sight. We found them last year. They were each made possible by an ordinary CMOS transistorâand not even a very good one at that. This is the story of their accidental discovery and their great promise for lowering the environmental footprint of AI.Biological and artificial neuronsModern digital electronics is based on producing and manipulating the ones and zeros of the binary code through the operation of metal-oxide-semiconductor field-effect transistors. MOSFETs have evolved in recent years, but their classic form consists of a piece of silicon that has been doped to contain an excess of either positive (p-type) or negative (n-type) charge carriers. (CMOS logic contains transistors of both types.) The device has two terminals connected to the silicon through regions highly doped with the opposite polarity of the rest of the siliconâthe source and the drain. Another terminal, the gate, sits atop the silicon that separates the source from the drain. The gate itself doesnât connect directly to this silicon, instead resting above a thin layer of insulating dielectric.Notably, there is a fourth terminal that attaches to the bulk of the silicon; think of this bulk terminal as connecting to the underside of the chip. It doesnât typically get much attention, but itâs very important to our story.When voltage is applied at the gate and the bulk terminal is grounded, charge carriers of the same polarity as the source and drain are attracted to the channel region. In the case of an n-type source and drain, that will be electrons; for p-type it will be holes. The presence of these charges forms a conductive channel that reduces the resistance between the source and the drain by several orders of magnitude, and the device switches on. As the voltage at the gate increases, this physical phenomenon produces a current signal that, when plotted against the gate voltage, rises steadily. This response is ideal for logic gates, converters, multiplexers, memories, and other digital circuits. But it is not a good fit for mimicking the behavior of a neuron.In real neural tissue, brain cells, called neurons, consist of a cell body, a long projection called an axon, and short branching projections called dendrites. The suite of behaviors and computing this collection of components is capable of is rich and broad, but the portion that artificial neural networks hope to copy is this: When the cell bodyâs voltage is perturbed enough to reach a particular threshold, a self-propagating pulse of voltage, called an action potential, shoots down the axon. The axon terminates in a synapse, an electrochemical connection between the axon and another neuronâs dendrites. The action potential will then temporarily boost the voltage of this next neuron, by an amount that depends on the strength of the synaptic connection. If enough action potentials reach these dendrites in a given timeâfrom this neuron or from others that might also form synapses thereâthe cell bodyâs voltage will surpass the threshold and trigger its own action potential.The MOSFET NeuronThe unusual action the authors discovered is understandable if you consider that a MOSFET contains a hidden bipolar-junction transistor.TRANSISTOR BEHAVIORUnder normal operation, with the bulk terminal grounded, increasing voltage at the drain leads to current that increases steadily. When the voltage decreases, current follows the same sloped path. Although some pairs of electrons and holes are created by current crashing into silicon atoms, these are swept away before they can accumulate.NSRAM BEHAVIORAdding resistance to the bulk terminal means these extra holes pile up, increasing the bulk voltage relative to the source. Once that voltage reaches a certain value, the hidden transistor activates, causing current to spike. Current remains high until the drain voltage drops past a certain point. To get closer to the behavior of real neurons, artificial neurons should produce a current spike when a critical voltage threshold is crossed and then quickly relax back to a resting state on their own. This spike needs to be suddenânonlinear. It should also exhibit some hysteresis; that is, the activation and relaxation voltages should be different from each other to ensure that current flows only for a certain amount of time.Whatâs wanted from an artificial synapse, the thing that connects two artificial neurons, is less complicated, but equally important. The main thing is that its conductance can be electronically adjustable. The deviceâs conductive states should increase and decrease in a linear pattern and remain stable over time.No single MOSFET working under the standard operation mechanism can reproduce either of these neural properties. Instead, itâs been done by combining them into complex circuits. Until now, each neuron and each synapse has been implemented by interconnecting dozens and sometimes even hundreds of MOSFETs, which is highly inefficient in terms of area, performance, and cost. To limit the amount of space needed, chips can multiplex their signals, sending them to neurons and synapses serially, but such sequential processing introduces additional delays.Despite these area-and-time penalties on tasks such as audio processing, computer vision, or health monitoring, state-of-the-art brain-inspired microchips have achieved power reductions up to a thousandfold compared with those of GPUs or CPUs on the same task. If we could create neurons and synapses from individual devices that are readily manufacturable instead, we might target more massive implementations while maintaining energy efficiency.Reinventing the MOSFET for AIWorking in our laboratory in 2024, one of my students was measuring a memory circuit that consisted of one transistor and one memristorâa type of nonvolatile memory device first fabricated in 2008. The studentâs memristor circuit was built from two-dimensional material atop a silicon microchip containing MOSFETs. The MOSFETs were created in a commercial foundry using fabrication technology called the 180-nanometer node, which was cutting-edge in the year 2000.One day the student forgot to connect the bulk terminal of the transistor. What he observed was a sudden increase in current with high nonlinearity that self-relaxed when the voltage was ramped down (a phenomenon called a hysteresis loop). This was a very promising neuronlike behavior!After a fruitless week of trying to think of an explanation for this behavior, I (Lanza) asked Pazos, then my postdoctoral fellow, to try to observe and control this phenomenon in chips without memristors. This time, we applied pulses of voltageâlike the spikes a neuron would produceâinstead of the ramped voltage that my student used when he first saw the peculiar behavior.Pazosâs new data helped us understand what was going on. The key was that oft-ignored fourth, or bulk, terminal of a MOSFET. Under ordinary operation, many mobile charge carriers flitting through the channel collide with the silicon atoms, producing free pairs of electrons and holesâa process known as impact ionization. The electric field created by the potential difference between the source and the drain causes these new free electrons to drift toward the positively biased drain and the holes to move toward the bulk terminal, which is usually grounded, removing the charge without any drama.However, when the bulk terminal of the transistor is floatingâunconnected as it was in my studentâs experimentâthe holes produced by impact ionization cannot be driven to the ground. Instead, they accumulate in the bulk of the silicon, increasing its voltage. Then things start to get interesting.It helps here to imagine a MOSFET as two different kinds of transistors occupying the same physical spaceâthe intentionally constructed MOSFET and a hidden, bipolar junction transistor. A bipolar device transmits a current signal across two p-n junctions, in this case the interfaces between the source and the channel region and the channel and the drain. This signal is in proportion to a smaller current at a third terminal in between, called the base. In our experiment, that third terminal is the bulk.To get current flowing through a bipolar transistor, you need a big enough potential difference between the base and one of the other terminals, so that current can get across the p-n junction. Letâs say this âthreshold voltageâ is 0.7 volts, although the real number depends on device geometry and silicon doping. In our device, that potential difference comes from those holes that were accumulating in the bulk, because it was not connected to ground. Once it reaches the threshold voltage, the device becomes sharply conductive, producing an abrupt increase of current. This sharp current increase eventually falls off once the drain voltage is lowered, because that lowering reduces the rate at which holes are generated in the bulk. The remaining excess holes recombine with stray electrons or leak away, and finally the bulk voltage falls. This cycle of hole accumulation, current spike, and hole removal gives rise to a hysteresis loop, very much like the electrical behavior of a biological neuron as it integrates ionic currents, fires a spike, and relaxes back to its resting voltage.Initially, we observed this behavior only in a few transistors, and the relaxation time was very different for each of them. So, to try to control it better, we adjusted the resistance of the bulk terminal using a second MOSFET. Simply setting that resistance suddenly caused all the transistors to fire at the same voltage with hardly any variability. In other words, we found we could create perfect electronic neuron behavior in a single silicon transistor by controlling the bulk contact resistance. Setting the resistance can be done by doping the silicon during fabrication, but we think the two-transistor cellâwhere one acts as the bulk resistanceâoffers much greater versatility because it allows for electronic control.We had to make sure the phenomenon would last, otherwise such a device would be useless. To our delight, every single one of the devices we tested worked over 10 million cycles. Not even one of them failed during our tests.The MOSFET SynapseTo be honest, we were amazed. Dozens of research groups and companies all around the world have spent many millions of U.S. dollars over the past 20 years trying to emulate these neural behaviors using experimental memristor-like devices and other things, with limited success, mainly due to reliability and cost issues. We managed it in the cheapest and most industry-standard device: the MOSFET. This result was so shocking that we decided to confirm it using microchips from a different foundry. It was successful: All the behaviors could be reproduced, and perfect yield was achieved once again.We were happy with the results and had started the process of filing for a patent and writing up our findings for the journal Nature, when our lab made another astonishing discovery: The same kind of MOSFET could act as a synapse, too!Recall that in ordinary operation some electrons crash into silicon atoms to create pairs of electrons and holes. We noticed that at specific values of bulk resistance a significant amount of the charge from this impact ionization would get trapped in the gate dielectric. This trapped charge interferes with the flow of current through the MOSFET, effectively changing the deviceâs conductance. Importantly, this new conductance is stable and adjustable at will. It was then that we realized the MOSFET could also be used as an electronic synapse.As it was in the neuron transistor, the bulk terminal was the key. A negative bulk-source voltage drives electrons into the dielectric, decreasing conductance. A positive one pushes holes in, increasing it.From neuromorphic device to circuit to systemHereâs how the MOSFET synapse and the MOSFET neuron, together called a neurosynaptic random-access memory, or NSRAM, could work together to achieve a simple neural circuit: Say you had a circuit consisting of three synapse MOSFETs and a neuron MOSFET. The synapses have already been programmed as weâve described, so that each has a different conductance. Spikes of voltage with different patterns and frequencies are applied to the gate of each of these transistors. What emerges from their drains are spikes of current with amplitudes modulated by the synapses conductance values.The spikes converge at the drain of the neuron MOSFET. With each spike, impact ionization causes charge to build in the bulk of the silicon. Some of it will drain away, but if enough spikes arrive in a short enough period of time, the bulk voltage will reach a value at which the âhiddenâ transistor triggers a spike of current through the MOSFET. This current would then go on to become the input to other MOSFET synapses, and so on. The behavior is exactly the kind of integrate-and-fire action real neural circuits deliver.The competitive advantage of our single-MOSFET electronic neurons and synapses is straightforward: We can produce with only one or two transistors the electronic signals that today require, at an industrial level, dozens and sometimes even hundreds of components. And moreover, unlike other emerging technologies, our solution is fully compatible with todayâs silicon manufacturing lines and exhibits a yield of 100 percent in key figures of merit with near-zero variability.Building functional circuits for brain-inspired computing and AI based on this technology is as exciting as it is laborious. It will require us to improve our computer models to resemble the behavior of both devices more accurately and to do so with computational efficiency. We must also perform accurate circuit- and system-level simulations to validate computing architectures, design peripheral circuitry to drive and convert signals, and undergo multiple fabrication rounds to optimize performance.But all that will be worthwhile, because it could result in brain-inspired microchips for AI with better energy efficiencies than what we have now. These chips will first be a fit for smaller-scale, âedge-AIâ tasks, such as bringing greater intelligence to battery-powered systems. But if we can scale up such chips, maybe in the long run they can compete with state-of-the-art GPUs.
Paul Meade, the Apple vice president in charge of the Vision Pro headset, is reportedly leaving the company to join OpenAIâs hardware team.
New models are launching in Asia that promise Mythos-like capabilities without fear of an export ban. U.S. AI labs may never recover this enormous market.
Over 100 companies and government agencies are reportedly authorized to use Mythos 5, including their non-American employees.
âWe donât believe this kind of government access process should become the long-term default,â says OpenAI. âIt keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.â
Nvidia has dominated the AI chip market for years, but the era of total dependence might be ending.   OpenAI just shared its plans to spice things up with Jalapeño, its custom inference chip built with Broadcom, joining Google, Apple, and SpaceX in a growing list of companies building their way out of single-supplier risk. The goal is less of a […]
Less than 24 hours after news broke that OpenAI would stagger its next model release at the request of the Trump administration, that model, GPT-5.6, is here. On Friday, the company unveiled the limited preview of its new GPT 5.6 model suite: Sol, the flagship; Terra, a medium-tier model for "high-volume work"; and Luna, a […]
OpenAI reportedly plans to share its newest model, GPT 5.6, with a select group of partners instead of with the broader public. The reason: the Trump administration told it to.
IBM has built a new prototype chip with around 100 billion transistors on an area the size of a fingernail, which is twice the density of the companyâs previous state-of-the-art technology announced in 2021. The design could pave the way for faster and more energy efficient computers for years to come. For more than half…
Top AI researchers Jonas Adler and Alexander Pritzel are leaving Google for Anthropic, following departures from top scientists Noam Shazeer and John Jumper.
Named Jalapeño, the new processor was designed specifically for the unique needs of OpenAI's inference systems.
OpenAI has just revealed a new "intelligence processor" chip for AI servers made in partnership with Broadcom. The chip, called Jalapeño, is designed to power current and future large language models, according to an announcement on Wednesday. Jalapeño is an ASIC (Application-Specific Integrated Circuit), meaning it's designed for a specific purpose: AI inference. With AI […]
AI is booming. New use cases are emerging each day. To capitalize on the technologyâs potential, enterprises require data at scale. In many cases, though, the relevant information is blocked or unstructured, which limits its use by AI models.  To understand this challenge, consider the foundation of the web itself. The web was not designed…
Atlantic reporter Alex Reisner recently uncovered four datasets of music being used to train AI models and made them fully searchable for the public. Two of the sets are absolutely enormous at 12 million and 9 million tracks. The other two are much smaller, but still represent a significant amount of training data at over […]
Jumper isn't the only big name leaving Google DeepMind.
Large language models have moved out of the research lab and into engineersâ daily workflow. LLMs serve as reasoning engines that can orchestrate complex tasks including identifying vulnerabilities in source code and transforming fragmented project discussions into rigorous technical specifications.While the general public uses AI tools to write email and plan vacations, technical professionals use LLMs as core architectural elements that are fundamentally changing how digital infrastructures are built and maintained. As the AI models move into mainstream engineering practice, the demand for technical expertise is rising.The LLM technology market is expected to grow by about 33 percent every year through 2030, according to MarketsandMarkets. The rapid expansion suggests that proficiency in implementing and securing the models is transitioning from a niche into a core requirement for technologists.More than just a better search engineTo use LLMs effectively, technical professionals must move beyond treating them as conversational robots. At a fundamental level, the AI systems are built on the transformer architecture, a framework that replaced the older method of processing data in a fixed, sequential order. Unlike earlier models that analyzed information one step at a time, transformers use self-attention mechanisms to ingest vast datasets simultaneously.For technical professionals, LLMs are core architectural elements that are fundamentally changing how digital infrastructures are built and maintained.Relying on such LLMs without understanding their internal logic creates a significant reliability risk. To build tools that work consistently, developers must understand the core principles that govern how the models process information and generate results. By mastering how a model processes information and how its internal settings influence the result, developers can move away from a trial-and-error approach toward a more precise one to ensure the AI tool handles complex data reliably.Four ways LLMs are changing jobsHere are areas that integrate large language models.Moving past basic prompts. Developers are using application program interfaces (APIs) to connect LLMs directly to their databases and software tools. Employing the APIs allows AI to perform work such as executing code or searching through internal repositories.Fixing the âhallucinationâ problem. LLMs are at risk of hallucinations, which are generated facts or code that looks correct but actually is wrong or broken. To fix the problem, retrieval-augmented generation (RAG) forces AI to look up information in a trusted source such as a companyâs database.Prioritizing data security. When using AI with proprietary code, security is a major concern. Engineers must learn how to set up âprivateâ instances of the models to ensure that sensitive company data stays within a secure cloud environment and is not used to train public versions.The future of collaboration. By automating repetitive coding tasks and summarizing thousands of pages of documentation, LLMs let engineers spend more time on high-level designs and solving important issues.Online course program helps with mastering the techThe gap between people who use AI and those who understand how to build with it is growing wider. To help technical professionals stay ahead, IEEE offers a five-course online program, Large Language Models Demystified, available through the IEEE Learning Network.The program, developed by IEEE Educational Activities in partnership with the IEEE Computer Society, is built for people who want to understand the âhowâ and the âwhyâ behind the technology. Rather than just teaching basic prompting, the curriculum dives into the engineering behind generative AI, including:Evolution, impact, and hands-on exercises: the shift from statistical methods to modern transformers, including hands-on model optimization.Understanding transformer architectures: the mathematical core of self-attention and positional encoding, implemented in NumPy and Python.Architectural analysis and implementation: advanced LLM design with practical model-building exercises.Training and modeling with PyTorch: end-to-end pipelines in PyTorch, leveraging parameter-efficient techniques such as low-rank adaptation and quantization.Optimization, alignment, and deployment: performance scaling, reinforcement learning from human feedback (RLHF), group-relative policy optimization, RAG, and agentic AI.Upon completion of the program, participants earn professional development credits and a digital badge from IEEE to verify their expertise.Enroll in the course program on the IEEE Learning Network.Organizations looking to prepare their teams to work on LLMs can connect with an IEEE content specialist to discuss group enrollment and tailored training paths.
Just as last week was ending, the US government forced Anthropic to pull its two newest models, Fable 5 and Mythos 5, citing national security concerns after Amazon researchers allegedly found a way to bypass Fable 5’s guardrails.  Cybersecurity researchers have since signed an open letter calling the move dangerous, and Anthropic itself noted the same jailbreaks exist in other models. So is […]
Just as last week was ending, the US government forced Anthropic to pull its two newest models, Fable 5 and Mythos 5, citing national security concerns after Amazon researchers allegedly found a way to bypass Fable 5’s guardrails.  Cybersecurity researchers have since signed an open letter calling the move dangerous, and Anthropic itself noted the same jailbreaks exist in other models. So is […]
The Miami-based AI startup Subquadratic came out of stealth mode last month with a huge claim. It announced that it had solved a mathematical bottleneck that had been holding back large language models for almost a decade. The details were thin, and many people were unconvinced. But Subquadratic has started to bring the receipts, sharing…
Startup Baseten is reportedly close to finalizing a $1.5 billion round at a $13 billion as the âinference gold rush" marches on.
OpenAI is bulking up before its IPO, landing Transformer co-inventor Noam Shazeer from Google DeepMind and former Trump AI policy official Dean Ball in the same week.
AWS is in talks to sell its chips to other data centers. CEO Andy Jassy has said this represents a $50 billion opportunity for the company.
FERC told grid operators to give data centers a fast lane for interconnections, but it failed to address electricity supply shortages.
The startup trains embodied AI and world models using Medalâs dataset of 2 billion videos per year from 10 million monthly active users.
Anthropic has spent much of this week fighting to get its newest AI models back online after the Trump administration abruptly ordered the company to cut access for all foreign nationals, including users inside the US and its own employees, forcing Anthropic to block access to Fable 5 and Mythos 5 for everyone. "To my […]
World models are the next big thing in AI beyond LLMs and, with this round, Odyssey has cemented itself as one of the startups to watch.
If physical AI is going to match the accomplishments of LLMs, there's a data problem that needs to be solved.
Probably wants to prevent hallucinations and factual errors from reaching users, and achieve accuracy on par with deterministic systems.
The chatbot still remains the most popular AI assistant worldwide with over 1.1 billion monthly users, followed by Gemini with 662 million and Claude with 245 million.
The Trump administration's decision that forced Anthropic to pull its latest cybersecurity models could be reactionary, retaliatory, or both, but the message is clear: The AI industry isn't immune from U.S. government interference.
Meta announced Monday that it's rolling out a wave of new AI features on Facebook, the latest sign of the company's effort to catch up in the AI race and keep users more engaged on the platform.
A group made up of dozens of cybersecurity experts urged the White House to remove export-control restrictions on Anthropicâs Fable and Mythos models, arguing that the order is going to limit the ability of cybersecurity defenders to secure their software and products.
According to the Wall Street Journal, the export control directive that led to Anthropic cutting off access to Fable 5 and Mythos 5 was triggered in part by cybersecurity research from Amazon and conversations between CEO Andy Jassy and the White House. According to the report, the paper from Amazon claims that, through a series […]
Amazon CEO Andy Jassy may have been the source of security concerns that led Anthropic to cut off worldwide access to two models on Friday.
This article is part of our exclusive IEEE Journal Watch series in partnership with IEEE Xplore. As robots advance in terms of dexterity and other physical capabilities, it becomes more likely that humans may find themselves working alongside them. If that happens, how will robotsâ emotional capabilities need to advance for them to successfully work with people?In a recent study, researchers trained collaborative robots to read human emotions by not only accounting for facial expressions, but also contextual factors in the interactions as well. Through experiments with 40 volunteers, the researchers then evaluated how a robotâs ability to read human emotions and adjust its behavior in turn impacted a humanâs perception of the robot and its capabilities as the two collaborated on tasks. The resultsâwhich show that the emotional capabilities of robots only go so far with humansâwere published 18 May in IEEE Robotics and Automation Letters.Seung Chan Hong led the study as part of his undergraduate thesis while studying at Monash University, in Melbourne, Australia. He notes that, while there has been a lot of hype in the advancing physical abilities of robots, this is only one piece of the puzzle. âWe need to also innovate when it comes to them actually interacting with humans, not just their physical capabilities,â he says.This prompted him to dig deeper into the emotional aspects of human-robot interactions. First, Hong and his co-authors decided to train a robot to read human emotions using a vision language model (VLM), which is similar to large language models (LLMs) such as ChatGPT, but which can also take visual inputs.Training VLMs for Human Emotion RecognitionTo evaluate their VLM, which used Gemini 2.5, the researchers had volunteers watch videos of robots handing over objects to humansâwith varying degrees of successâand describe the emotions the humans were expressing. Importantly, the volunteers labeling these videos were able to take into account more context in these interactions, rather than reporting solely on the facial expressions of the humans in the video. For example, a person pausing to think with a furrowed brow may simply be concentrating on their task at hand and not necessarily be angry. Contextual factors such as drumming their fingers, pursing their lips, or other behaviors can point to the real cause of a personâs furrowed brow.The researchers then compared their VLM to a conventional AI system that relies on standard facial analysis and object tracking that is used in human-robot interactions. They found that the VLM outperformed the traditional approach. On a scale from 0 (no similarity in meaning to the emotion identified by the human volunteers) to 1 (a perfect match in meaning), the conventional AI system achieved a score of 0.77. In comparison, the VLM achieved a score of 0.86.Hong says, âI think [the VLM] was able to align with what human observers were seeing a lot better, because it wasnât just looking at the personâs face for a brief amount of time, but seeing the whole sceneâwhere the person was and what they were doing, and how they were interacting with the robot.âIn a second experiment, the research team asked 40 volunteers to interact with a robot using their VLMâbut purposefully programmed the robot to make an error. The robot then had to offer either an emotionally adaptive apology that accounted for the humanâs perceived response to the mistake or a pre-scripted spoken apology.Participants overwhelmingly preferred the emotionally adaptive response, with 31 out of 40 people favoring this approach over a boilerplate apology.However, their survey responses underscored how this emotional adaptivity was far less important than the robotâs functionality. After collaborating with a robot that failed in its task, many participants ranked their trust in the robot as lower, regardless of how it apologized for its mistake. âA personalized apology acts as a social lubricant, but it cannot repair the trust lost by the robot failing its physical task,â Hong says.Interestingly, the VLM classified the emotions of its human partners similarly to human volunteers who observed an interaction from a third-party perspective. But when the VLMâs assessments were measured against humansâ self-reported emotions during the second experimentâthe most accurate descriptions of their true emotionsâits ability to accurately predict emotions dropped significantly.âWhile the VLM is a good observer of outward social cues, it isnât a mind reader,â Hong says. âIt matched third-person human observers well, but it didnât always align with the usersâ internal, self-reported feelings.âTogether, these results show that robots are not perfect at reading human emotions. So while people might appreciate their efforts, they still ultimately will want competent co-workers.This story was updated on 15 June 2026 to correct where the research was conducted and clarify that the researchers evaluated the performance of a pre-trained model.
On Friday evening, the government ordered Anthropic to block access to Fable 5 and Mythos 5 for all foreign nations, both inside and outside the US, due to national security concerns. That order included employees of Anthropic. To meet those demands, the company has completely cut off access to the models for all customers. In […]