On X, frequent AI experimenter Ethan Mollick wrote, “Been playing with o1 and o1-pro for bit. They are very good & a little weird. They are also not for most people most of the time. You really need to have particular hard problems to solve in order to get value out of it. But if you have those problems, this is a very big deal.”
OpenAI claims improved reliability
OpenAI is touting pro mode’s improved reliability, which is evaluated internally based on whether it can solve a question correctly in four out of four attempts rather than just a single attempt.
“In evaluations from external expert testers, o1 pro mode produces more reliably accurate and comprehensive responses, especially in areas like data science, programming, and case law analysis,” OpenAI writes.
Even without pro mode, OpenAI cited significant increases in performance over the o1 preview model on popular math and coding benchmarks (AIME 2024 and Codeforces), and more marginal improvements on a “PhD-level science” benchmark (GPQA Diamond). The increase in scores between o1 and o1 pro mode were much more marginal on these benchmarks.
We’ll likely have more coverage of the full version of o1 once it rolls out widely—and it’s supposed to launch today, accessible to ChatGPT Plus and Team users globally. Enterprise and Edu users will have access next week. At the moment, the ChatGPT Pro subscription is not yet available on our test account.
A co-author of Attention Is All You Need reflects on ChatGPT’s surprise and Google’s conservatism.
Jakob Uszkoreit Credit: Jakob Uszkoreit / Getty Images
In 2017, eight machine-learning researchers at Google released a groundbreaking research paper called Attention Is All You Need, which introduced the Transformer AI architecture that underpins almost all of today’s high-profile generative AI models.
The Transformer has made a key component of the modern AI boom possible by translating (or transforming, if you will) input chunks of data called “tokens” into another desired form of output using a neural network. Variations of the Transformer architecture power language models like GPT-4o (and ChatGPT), audio synthesis models that run Google’s NotebookLM and OpenAI’s Advanced Voice Mode, video synthesis models like Sora, and image synthesis models like Midjourney.
At TED AI 2024 in October, one of those eight researchers, Jakob Uszkoreit, spoke with Ars Technica about the development of transformers, Google’s early work on large language models, and his new venture in biological computing.
In the interview, Uszkoreit revealed that while his team at Google had high hopes for the technology’s potential, they didn’t quite anticipate its pivotal role in products like ChatGPT.
The Ars interview: Jakob Uszkoreit
Ars Technica: What was your main contribution to the Attention is All You Need paper?
Jakob Uszkoreit (JU): It’s spelled out in the footnotes, but my main contribution was to propose that it would be possible to replace recurrence [from Recurrent Neural Networks] in the dominant sequence transduction models at the time with the attention mechanism, or more specifically self-attention. And that it could be more efficient and, as a result, also more effective.
Ars: Did you have any idea what would happen after your group published that paper? Did you foresee the industry it would create and the ramifications?
JU: First of all, I think it’s really important to keep in mind that when we did that, we were standing on the shoulders of giants. And it wasn’t just that one paper, really. It was a long series of works by some of us and many others that led to this. And so to look at it as if this one paper then kicked something off or created something—I think that is taking a view that we like as humans from a storytelling perspective, but that might not actually be that accurate of a representation.
My team at Google was pushing on attention models for years before that paper. It’s a lot longer of a slog with much, much more, and that’s just my group. Many others were working on this, too, but we had high hopes that it would push things forward from a technological perspective. Did we think that it would play a role in really enabling, or at least apparently, seemingly, flipping a switch when it comes to facilitating products like ChatGPT? I don’t think so. I mean, to be very clear in terms of LLMs and their capabilities, even around the time we published the paper, we saw phenomena that were pretty staggering.
We didn’t get those out into the world in part because of what really is maybe a notion of conservatism around products at Google at the time. But we also, even with those signs, weren’t that confident that stuff in and of itself would make that compelling of a product. But did we have high hopes? Yeah.
Ars: Since you knew there were large language models at Google, what did you think when ChatGPT broke out into a public success? “Damn, they got it, and we didn’t?”
JU: There was a notion of, well, “that could have happened.” I think it was less of a, “Oh dang, they got it first” or anything of the like. It was more of a “Whoa, that could have happened sooner.” Was I still amazed by just how quickly people got super creative using that stuff? Yes, that was just breathtaking.
Jakob Uszkoreit presenting at TED AI 2024. Credit: Benj Edwards
Ars: You weren’t at Google at that point anymore, right?
JU: I wasn’t anymore. And in a certain sense, you could say the fact that Google wouldn’t be the place to do that factored into my departure. I left not because of what I didn’t like at Google as much as I left because of what I felt I absolutely had to do elsewhere, which is to start Inceptive.
But it was really motivated by just an enormous, not only opportunity, but a moral obligation in a sense, to do something that was better done outside in order to design better medicines and have very direct impact on people’s lives.
Ars: The funny thing with ChatGPT is that I was using GPT-3 before that. So when ChatGPT came out, it wasn’t that big of a deal to some people who were familiar with the tech.
JU: Yeah, exactly. If you’ve used those things before, you could see the progression and you could extrapolate. When OpenAI developed the earliest GPTs with Alec Radford and those folks, we would talk about those things despite the fact that we weren’t at the same companies. And I’m sure there was this kind of excitement, how well-received the actual ChatGPT product would be by how many people, how fast. That still, I think, is something that I don’t think anybody really anticipated.
Ars: I didn’t either when I covered it. It felt like, “Oh, this is a chatbot hack of GPT-3 that feeds its context in a loop.” And I didn’t think it was a breakthrough moment at the time, but it was fascinating.
JU: There are different flavors of breakthroughs. It wasn’t a technological breakthrough. It was a breakthrough in the realization that at that level of capability, the technology had such high utility.
That, and the realization that, because you always have to take into account how your users actually use the tool that you create, and you might not anticipate how creative they would be in their ability to make use of it, how broad those use cases are, and so forth.
That is something you can sometimes only learn by putting something out there, which is also why it is so important to remain experiment-happy and to remain failure-happy. Because most of the time, it’s not going to work. But some of the time it’s going to work—and very, very rarely it’s going to work like [ChatGPT did].
Ars: You’ve got to take a risk. And Google didn’t have an appetite for taking risks?
JU: Not at that time. But if you think about it, if you look back, it’s actually really interesting. Google Translate, which I worked on for many years, was actually similar. When we first launched Google Translate, the very first versions, it was a party joke at best. And we took it from that to being something that was a truly useful tool in not that long of a period. Over the course of those years, the stuff that it sometimes output was so embarrassingly bad at times, but Google did it anyway because it was the right thing to try. But that was around 2008, 2009, 2010.
Ars: When that came out, it blew my mind. My brother and I would do this thing where we would translate text back and forth between languages for fun because it would garble the text.
JU: It would get worse and worse and worse. Yeah.
Programming biological computers
After his time at Google, Uszkoreit co-founded Inceptive to apply deep learning to biochemistry. The company is developing what he calls “biological software,” where AI compilers translate specified behaviors into RNA sequences that can perform desired functions when introduced to biological systems.
Ars: What are you up to these days?
JU: In 2021 we co-founded Inceptive in order to use deep learning and high throughput biochemistry experimentation to design better medicines that truly can be programmed. We think of this as really just step one in the direction of something that we call biological software.
Biological software is a little bit like computer software in that you have some specification of the behavior that you want, and then you have a compiler that translates that into a piece of computer software that then runs on a computer exhibiting the functions or the functionality that you specify.
You specify a piece of a biological program and you compile that, but not with an engineered compiler, because life hasn’t been engineered like computers have been engineered. But with a learned AI compiler, you translate that or compile that into molecules that when inserted into biological systems, organisms, our cells exhibit those functions that you’ve programmed into.
A pharmacist holds a bottle containing Moderna’s bivalent COVID-19 vaccine. Credit: Getty | Mel Melcon
Ars: Is that anything like how the mRNA COVID vaccines work?
JU: A very, very simple example of that are the mRNA COVID vaccines where the program says, “Make this modified viral antigen” and then our cells make that protein. But you could imagine molecules that exhibit far more complex behaviors. And if you want to get a picture of how complex those behaviors could be, just remember that RNA viruses are just that. They’re just an RNA molecule that when entering an organism exhibits incredibly complex behavior such as distributing itself across an organism, distributing itself across the world, doing certain things only in a subset of your cells for a certain period of time, and so on and so forth.
And so you can imagine that if we managed to even just design molecules with a teeny tiny fraction of such functionality, of course with the goal not of making people sick, but of making them healthy, it would truly transform medicine.
Ars: How do you not accidentally create a monster RNA sequence that just wrecks everything?
JU: The amazing thing is that medicine for the longest time has existed in a certain sense outside of science. It wasn’t truly understood, and we still often don’t truly understand their actual mechanisms of action.
As a result, humanity had to develop all of these safeguards and clinical trials. And even before you enter the clinic, all of these empirical safeguards prevent us from accidentally doing [something dangerous]. Those systems have been in place for as long as modern medicine has existed. And so we’re going to keep using those systems, and of course with all the diligence necessary. We’ll start with very small systems, individual cells in future experimentation, and follow the same established protocols that medicine has had to follow all along in order to ensure that these molecules are safe.
Ars: Thank you for taking the time to do this.
JU: No, thank you.
Benj Edwards is Ars Technica’s Senior AI Reporter and founder of the site’s dedicated AI beat in 2022. He’s also a widely-cited tech historian. In his free time, he writes and records music, collects vintage computers, and enjoys nature. He lives in Raleigh, NC.
On Thursday, OpenAI released an early Windows version of its first ChatGPT app for Windows, following a Mac version that launched in May. Currently, it’s only available to subscribers of Plus, Team, Enterprise, and Edu versions of ChatGPT, and users can download it for free in the Microsoft Store for Windows.
OpenAI is positioning the release as a beta test. “This is an early version, and we plan to bring the full experience to all users later this year,” OpenAI writes on the Microsoft Store entry for the app. (Interestingly, ChatGPT shows up as being rated “T for Teen” by the ESRB in the Windows store, despite not being a video game.)
A screenshot of the new Windows ChatGPT app captured on October 18, 2024.
Credit: Benj Edwards
A screenshot of the new Windows ChatGPT app captured on October 18, 2024. Credit: Benj Edwards
Upon opening the app, OpenAI requires users to log into a paying ChatGPT account, and from there, the app is basically identical to the web browser version of ChatGPT. You can currently use it to access several models: GPT-4o, GPT-4o with Canvas, 01-preview, 01-mini, GPT-4o mini, and GPT-4. Also, it can generate images using DALL-E 3 or analyze uploaded files and images.
If you’re running Windows 11, you can instantly call up a small ChatGPT window when the app is open using an Alt+Space shortcut (it did not work in Windows 10 when we tried). That could be handy for asking ChatGPT a quick question at any time.
A screenshot of the new Windows ChatGPT app listing in the Microsoft Store captured on October 18, 2024.
Credit: Benj Edwards
A screenshot of the new Windows ChatGPT app listing in the Microsoft Store captured on October 18, 2024. Credit: Benj Edwards
And just like the web version, all the AI processing takes place in the cloud on OpenAI’s servers, which means an Internet connection is required.
So as usual, chat like somebody’s watching, and don’t rely on ChatGPT as a factual reference for important decisions—GPT-4o in particular is great at telling you what you want to hear, whether it’s correct or not. As OpenAI says in a small disclaimer at the bottom of the app window: “ChatGPT can make mistakes.”
On Monday, Microsoft unveiled updates to its consumer AI assistant Copilot, introducing two new experimental features for a limited group of $20/month Copilot Pro subscribers: Copilot Labs and Copilot Vision. Labs integrates OpenAI’s latest o1 “reasoning” model, and Vision allows Copilot to see what you’re browsing in Edge.
Microsoft says Copilot Labs will serve as a testing ground for Microsoft’s latest AI tools before they see wider release. The company describes it as offering “a glimpse into ‘work-in-progress’ projects.” The first feature available in Labs is called “Think Deeper,” and it uses step-by-step processing to solve more complex problems than the regular Copilot. Think Deeper is Microsoft’s version of OpenAI’s new o1-preview and o1-mini AI models, and it has so far rolled out to some Copilot Pro users in Australia, Canada, New Zealand, the UK, and the US.
Copilot Vision is an entirely different beast. The new feature aims to give the AI assistant a visual window into what you’re doing within the Microsoft Edge browser. When enabled, Copilot can “understand the page you’re viewing and answer questions about its content,” according to Microsoft.
Microsoft’s Copilot Vision promo video.
The company positions Copilot Vision as a way to provide more natural interactions and task assistance beyond text-based prompts, but it will likely raise privacy concerns. As a result, Microsoft says that Copilot Vision is entirely opt-in and that no audio, images, text, or conversations from Vision will be stored or used for training. The company is also initially limiting Vision’s use to a pre-approved list of websites, blocking it on paywalled and sensitive content.
The rollout of these features appears gradual, with Microsoft noting that it wants to balance “pioneering features and a deep sense of responsibility.” The company said it will be “listening carefully” to user feedback as it expands access to the new capabilities. Microsoft has not provided a timeline for wider availability of either feature.
Mustafa Suleyman, chief executive of Microsoft AI, told Reuters that he sees Copilot as an “ever-present confidant” that could potentially learn from users’ various Microsoft-connected devices and documents, with permission. He also mentioned that Microsoft co-founder Bill Gates has shown particular interest in Copilot’s potential to read and parse emails.
But judging by the visceral reaction to Microsoft’s Recall feature, which keeps a record of everything you do on your PC so an AI model can recall it later, privacy-sensitive users may not appreciate having an AI assistant monitor their activities—especially if those features send user data to the cloud for processing.
OpenAI, the company behind ChatGPT, has now raised $6.6 billion in a new funding round that values the company at $157 billion, nearly doubling its previous valuation of $86 billion, according to a report from The Wall Street Journal.
The funding round comes with strings attached: Investors have the right to withdraw their money if OpenAI does not complete its planned conversion from a nonprofit (with a for-profit division) to a fully for-profit company.
Venture capital firm Thrive Capital led the funding round with a $1.25 billion investment. Microsoft, a longtime backer of OpenAI to the tune of $13 billion, contributed just under $1 billion to the latest round. New investors joined the round, including SoftBank with a $500 million investment and Nvidia with $100 million.
The United Arab Emirates-based company MGX also invested in OpenAI during this funding round. MGX has been busy in AI recently, joining an AI infrastructure partnership last month led by Microsoft.
Notably, Apple was in talks to invest but ultimately did not participate. WSJ reports that the minimum investment required to review OpenAI’s financial documents was $250 million. In June, OpenAI hired its first chief financial officer, Sarah Friar, who played an important role in organizing this funding round, according to the WSJ.
OpenAI finally unveiled its rumored “Strawberry” AI language model on Thursday, claiming significant improvements in what it calls “reasoning” and problem-solving capabilities over previous large language models (LLMs). Formally named “OpenAI o1,” the model family will initially launch in two forms, o1-preview and o1-mini, available today for ChatGPT Plus and certain API users.
OpenAI claims that o1-preview outperforms its predecessor, GPT-4o, on multiple benchmarks, including competitive programming, mathematics, and “scientific reasoning.” However, people who have used the model say it does not yet outclass GPT-4o in every metric. Other users have criticized the delay in receiving a response from the model, owing to the multi-step processing occurring behind the scenes before answering a query.
In a rare display of public hype-busting, OpenAI product manager Joanne Jang tweeted, “There’s a lot of o1 hype on my feed, so I’m worried that it might be setting the wrong expectations. what o1 is: the first reasoning model that shines in really hard tasks, and it’ll only get better. (I’m personally psyched about the model’s potential & trajectory!) what o1 isn’t (yet!): a miracle model that does everything better than previous models. you might be disappointed if this is your expectation for today’s launch—but we’re working to get there!”
OpenAI reports that o1-preview ranked in the 89th percentile on competitive programming questions from Codeforces. In mathematics, it scored 83 percent on a qualifying exam for the International Mathematics Olympiad, compared to GPT-4o’s 13 percent. OpenAI also states, in a claim that may later be challenged as people scrutinize the benchmarks and run their own evaluations over time, o1 performs comparably to PhD students on specific tasks in physics, chemistry, and biology. The smaller o1-mini model is designed specifically for coding tasks and is priced at 80 percent less than o1-preview.
Enlarge/ A benchmark chart provided by OpenAI. They write, “o1 improves over GPT-4o on a wide range of benchmarks, including 54/57 MMLU subcategories. Seven are shown for illustration.”
OpenAI attributes o1’s advancements to a new reinforcement learning (RL) training approach that teaches the model to spend more time “thinking through” problems before responding, similar to how “let’s think step-by-step” chain-of-thought prompting can improve outputs in other LLMs. The new process allows o1 to try different strategies and “recognize” its own mistakes.
AI benchmarks are notoriously unreliable and easy to game; however, independent verification and experimentation from users will show the full extent of o1’s advancements over time. It’s worth noting that MIT Research showed earlier this year that some of the benchmark claims OpenAI touted with GPT-4 last year were erroneous or exaggerated.
A mixed bag of capabilities
OpenAI demos “o1” correctly counting the number of Rs in the word “strawberry.”
Amid many demo videos of o1 completing programming tasks and solving logic puzzles that OpenAI shared on its website and social media, one demo stood out as perhaps the least consequential and least impressive, but it may become the most talked about due to a recurring meme where people ask LLMs to count the number of R’s in the word “strawberry.”
Due to tokenization, where the LLM processes words in data chunks called tokens, most LLMs are typically blind to character-by-character differences in words. Apparently, o1 has the self-reflective capabilities to figure out how to count the letters and provide an accurate answer without user assistance.
Beyond OpenAI’s demos, we’ve seen optimistic but cautious hands-on reports about o1-preview online. Wharton Professor Ethan Mollick wrote on X, “Been using GPT-4o1 for the last month. It is fascinating—it doesn’t do everything better but it solves some very hard problems for LLMs. It also points to a lot of future gains.”
Mollick shared a hands-on post in his “One Useful Thing” blog that details his experiments with the new model. “To be clear, o1-preview doesn’t do everything better. It is not a better writer than GPT-4o, for example. But for tasks that require planning, the changes are quite large.”
Mollick gives the example of asking o1-preview to build a teaching simulator “using multiple agents and generative AI, inspired by the paper below and considering the views of teachers and students,” then asking it to build the full code, and it produced a result that Mollick found impressive.
Mollick also gave o1-preview eight crossword puzzle clues, translated into text, and the model took 108 seconds to solve it over many steps, getting all of the answers correct but confabulating a particular clue Mollick did not give it. We recommend reading Mollick’s entire post for a good early hands-on impression. Given his experience with the new model, it appears that o1 works very similar to GPT-4o but iteratively in a loop, which is something that the so-called “agentic” AutoGPT and BabyAGI projects experimented with in early 2023.
Is this what could “threaten humanity?”
Speaking of agentic models that run in loops, Strawberry has been subject to hype since last November, when it was initially known as Q(Q-star). At the time, The Information and Reuters claimed that, just before Sam Altman’s brief ouster as CEO, OpenAI employees had internally warned OpenAI’s board of directors about a new OpenAI model called Q* that could “threaten humanity.”
In August, the hype continued when The Information reported that OpenAI showed Strawberry to US national security officials.
We’ve been skeptical about the hype around Qand Strawberry since the rumors first emerged, as this author noted last November, and Timothy B. Lee covered thoroughly in an excellent post about Q* from last December.
So even though o1 is out, AI industry watchers should note how this model’s impending launch was played up in the press as a dangerous advancement while not being publicly downplayed by OpenAI. For an AI model that takes 108 seconds to solve eight clues in a crossword puzzle and hallucinates one answer, we can say that its potential danger was likely hype (for now).
Controversy over “reasoning” terminology
It’s no secret that some people in tech have issues with anthropomorphizing AI models and using terms like “thinking” or “reasoning” to describe the synthesizing and processing operations that these neural network systems perform.
Just after the OpenAI o1 announcement, Hugging Face CEO Clement Delangue wrote, “Once again, an AI system is not ‘thinking,’ it’s ‘processing,’ ‘running predictions,’… just like Google or computers do. Giving the false impression that technology systems are human is just cheap snake oil and marketing to fool you into thinking it’s more clever than it is.”
“Reasoning” is also a somewhat nebulous term since, even in humans, it’s difficult to define exactly what the term means. A few hours before the announcement, independent AI researcher Simon Willison tweeted in response to a Bloomberg story about Strawberry, “I still have trouble defining ‘reasoning’ in terms of LLM capabilities. I’d be interested in finding a prompt which fails on current models but succeeds on strawberry that helps demonstrate the meaning of that term.”
Reasoning or not, o1-preview currently lacks some features present in earlier models, such as web browsing, image generation, and file uploading. OpenAI plans to add these capabilities in future updates, along with continued development of both the o1 and GPT model series.
While OpenAI says the o1-preview and o1-mini models are rolling out today, neither model is available in our ChatGPT Plus interface yet, so we have not been able to evaluate them. We’ll report our impressions on how this model differs from other LLMs we have previously covered.
Over the past week, OpenAI experienced a significant leadership shake-up as three key figures announced major changes. Greg Brockman, the company’s president and co-founder, is taking an extended sabbatical until the end of the year, while another co-founder, John Schulman, permanently departed for rival Anthropic. Peter Deng, VP of Consumer Product, has also left the ChatGPT maker.
In a post on X, Brockman wrote, “I’m taking a sabbatical through end of year. First time to relax since co-founding OpenAI 9 years ago. The mission is far from complete; we still have a safe AGI to build.”
The moves have led some to wonder just how close OpenAI is to a long-rumored breakthrough of some kind of reasoning artificial intelligence if high-profile employees are jumping ship (or taking long breaks, in the case of Brockman) so easily. As AI developer Benjamin De Kraker put it on X, “If OpenAI is right on the verge of AGI, why do prominent people keep leaving?”
AGI refers to a hypothetical AI system that could match human-level intelligence across a wide range of tasks without specialized training. It’s the ultimate goal of OpenAI, and company CEO Sam Altman has said it could emerge in the “reasonably close-ish future.” AGI is also a concept that has sparked concerns about potential existential risks to humanity and the displacement of knowledge workers. However, the term remains somewhat vague, and there’s considerable debate in the AI community about what truly constitutes AGI or how close we are to achieving it.
The emergence of the “next big thing” in AI has been seen by critics such as Ed Zitron as a necessary step to justify ballooning investments in AI models that aren’t yet profitable. The industry is holding its breath that OpenAI, or a competitor, has some secret breakthrough waiting in the wings that will justify the massive costs associated with training and deploying LLMs.
But other AI critics, such as Gary Marcus, have postulated that major AI companies have reached a plateau of large language model (LLM) capability centered around GPT-4-level models since no AI company has yet made a major leap past the groundbreaking LLM that OpenAI released in March 2023. Microsoft CTO Kevin Scott has countered these claims, saying that LLM “scaling laws” (that suggest LLMs increase in capability proportionate to more compute power thrown at them) will continue to deliver improvements over time and that more patience is needed as the next generation (say, GPT-5) undergoes training.
In the scheme of things, Brockman’s move sounds like an extended, long overdue vacation (or perhaps a period to deal with personal issues beyond work). Regardless of the reason, the duration of the sabbatical raises questions about how the president of a major tech company can suddenly disappear for four months without affecting day-to-day operations, especially during a critical time in its history.
Unless, of course, things are fairly calm at OpenAI—and perhaps GPT-5 isn’t going to ship until at least next year when Brockman returns. But this is speculation on our part, and OpenAI (whether voluntarily or not) sometimes surprises us when we least expect it. (Just today, Altman dropped a hint on X about strawberries that some people interpret as being a hint of a potential major model undergoing testing or nearing release.)
A pattern of departures and the rise of Anthropic
Anthropic / Benj Edwards
What may sting OpenAI the most about the recent departures is that a few high-profile employees have left to join Anthropic, a San Francisco-based AI company founded in 2021 by ex-OpenAI employees Daniela and Dario Amodei.
Anthropic offers a subscription service called Claude.ai that is similar to ChatGPT. Its most recent LLM, Claude 3.5 Sonnet, along with its web-based interface, has rapidly gained favor over ChatGPT among some LLM users who are vocal on social media, though it likely does not yet match ChatGPT in terms of mainstream brand recognition.
In particular, John Schulman, an OpenAI co-founder and key figure in the company’s post-training process for LLMs, revealed in a statement on X that he’s leaving to join rival AI firm Anthropic to do more hands-on work: “This choice stems from my desire to deepen my focus on AI alignment, and to start a new chapter of my career where I can return to hands-on technical work.” Alignment is a field that hopes to guide AI models to produce helpful outputs.
In May, OpenAI alignment researcher Jan Leike left OpenAI to join Anthropic as well, criticizing OpenAI’s handling of alignment safety.
Adding to the recent employee shake-up, The Information reports that Peter Deng, a product leader who joined OpenAI last year after stints at Meta Platforms, Uber, and Airtable, has also left the company, though we do not yet know where he is headed. In May, OpenAI co-founder Ilya Sutskever left to found a rival startup, and prominent software engineer Andrej Karpathy departed in February, recently launching an educational venture.
As De Kraker noted, if OpenAI were on the verge of developing world-changing AI technology, wouldn’t these high-profile AI veterans want to stick around and be part of this historic moment in time? “Genuine question,” he wrote. “If you were pretty sure the company you’re a key part of—and have equity in—is about to crack AGI within one or two years… why would you jump ship?”
Despite the departures, Schulman expressed optimism about OpenAI’s future in his farewell note on X. “I am confident that OpenAI and the teams I was part of will continue to thrive without me,” he wrote. “I’m incredibly grateful for the opportunity to participate in such an important part of history and I’m proud of what we’ve achieved together. I’ll still be rooting for you all, even while working elsewhere.”
This article was updated on August 7, 2024 at 4: 23 PM to mention Sam Altman’s tweet about strawberries.
Enlarge/ Elon Musk and Sam Altman share the stage in 2015, the same year that Musk alleged that Altman’s “deception” began.
After withdrawing his lawsuit in June for unknown reasons, Elon Musk has revived a complaint accusing OpenAI and its CEO Sam Altman of fraudulently inducing Musk to contribute $44 million in seed funding by promising that OpenAI would always open-source its technology and prioritize serving the public good over profits as a permanent nonprofit.
Instead, Musk alleged that Altman and his co-conspirators—”preying on Musk’s humanitarian concern about the existential dangers posed by artificial intelligence”—always intended to “betray” these promises in pursuit of personal gains.
As OpenAI’s technology advanced toward artificial general intelligence (AGI) and strove to surpass human capabilities, “Altman set the bait and hooked Musk with sham altruism then flipped the script as the non-profit’s technology approached AGI and profits neared, mobilizing Defendants to turn OpenAI, Inc. into their personal piggy bank and OpenAI into a moneymaking bonanza, worth billions,” Musk’s complaint said.
Where Musk saw OpenAI as his chance to fund a meaningful rival to stop Google from controlling the most powerful AI, Altman and others “wished to launch a competitor to Google” and allegedly deceived Musk to do it. According to Musk:
The idea Altman sold Musk was that a non-profit, funded and backed by Musk, would attract world-class scientists, conduct leading AI research and development, and, as a meaningful counterweight to Google’s DeepMind in the race for Artificial General Intelligence (“AGI”), decentralize its technology by making it open source. Altman assured Musk that the non-profit structure guaranteed neutrality and a focus on safety and openness for the benefit of humanity, not shareholder value. But as it turns out, this was all hot-air philanthropy—the hook for Altman’s long con.
Without Musk’s involvement and funding during OpenAI’s “first five critical years,” Musk’s complaint said, “it is fair to say” that “there would have been no OpenAI.” And when Altman and others repeatedly approached Musk with plans to shift OpenAI to a for-profit model, Musk held strong to his morals, conditioning his ongoing contributions on OpenAI remaining a nonprofit and its tech largely remaining open source.
“Either go do something on your own or continue with OpenAI as a nonprofit,” Musk told Altman in 2018 when Altman tried to “recast the nonprofit as a moneymaking endeavor to bring in shareholders, sell equity, and raise capital.”
“I will no longer fund OpenAI until you have made a firm commitment to stay, or I’m just being a fool who is essentially providing free funding to a startup,” Musk said at the time. “Discussions are over.”
But discussions weren’t over. And now Musk seemingly does feel like a fool after OpenAI exclusively licensed GPT-4 and all “pre-AGI” technology to Microsoft in 2023, while putting up paywalls and “failing to publicly disclose the non-profit’s research and development, including details on GPT-4, GPT-4T, and GPT-4o’s architecture, hardware, training method, and training computation.” This excluded the public “from open usage of GPT-4 and related technology to advance Defendants and Microsoft’s own commercial interests,” Musk alleged.
Now Musk has revived his suit against OpenAI, asking the court to award maximum damages for OpenAI’s alleged fraud, contract breaches, false advertising, acts viewed as unfair to competition, and other violations.
He has also asked the court to determine a very technical question: whether OpenAI’s most recent models should be considered AGI and therefore Microsoft’s license voided. That’s the only way to ensure that a private corporation isn’t controlling OpenAI’s AGI models, which Musk repeatedly conditioned his financial contributions upon preventing.
“Musk contributed considerable money and resources to launch and sustain OpenAI, Inc., which was done on the condition that the endeavor would be and remain a non-profit devoted to openly sharing its technology with the public and avoid concentrating its power in the hands of the few,” Musk’s complaint said. “Defendants knowingly and repeatedly accepted Musk’s contributions in order to develop AGI, with no intention of honoring those conditions once AGI was in reach. Case in point: GPT-4, GPT-4T, and GPT-4o are all closed source and shrouded in secrecy, while Defendants actively work to transform the non-profit into a thoroughly commercial business.”
Musk wants Microsoft’s GPT-4 license voided
Musk also asked the court to null and void OpenAI’s exclusive license to Microsoft, or else determine “whether GPT-4, GPT-4T, GPT-4o, and other OpenAI next generation large language models constitute AGI and are thus excluded from Microsoft’s license.”
It’s clear that Musk considers these models to be AGI, and he’s alleged that Altman’s current control of OpenAI’s Board—after firing dissidents in 2023 whom Musk claimed tried to get Altman ousted for prioritizing profits over AI safety—gives Altman the power to obscure when OpenAI’s models constitute AGI.
Arguably, few companies have unintentionally contributed more to the increase of AI-generated noise online than OpenAI. Despite its best intentions—and against its terms of service—its AI language models are often used to compose spam, and its pioneering research has inspired others to build AI models that can potentially do the same. This influx of AI-generated content has further reduced the effectiveness of SEO-driven search engines like Google. In 2024, web search is in a sorry state indeed.
It’s interesting, then, that OpenAI is now offering a potential solution to that problem. On Thursday, OpenAI revealed a prototype AI-powered search engine called SearchGPT that aims to provide users with quick, accurate answers sourced from the web. It’s also a direct challenge to Google, which also has tried to apply generative AI to web search (but with little success).
The company says it plans to integrate the most useful aspects of the temporary prototype into ChatGPT in the future. ChatGPT can already perform web searches using Bing, but SearchGPT seems to be a purpose-built interface for AI-assisted web searching.
SearchGPT attempts to streamline the process of finding information online by combining OpenAI’s AI models (like GPT-4o) with real-time web data. Like ChatGPT, users can reportedly ask SearchGPT follow-up questions, with the AI model maintaining context throughout the conversation.
Perhaps most importantly from an accuracy standpoint, the SearchGPT prototype (which we have not tested ourselves) reportedly includes features that attribute web-based sources prominently. Responses include in-line citations and links, while a sidebar displays additional source links.
OpenAI has not yet said how it is obtaining its real-time web data and whether it’s partnering with an existing search engine provider (like it does currently with Bing for ChatGPT) or building its own web-crawling and indexing system.
A way around publishers blocking OpenAI
ChatGPT can already perform web searches using Bing, but since last August when OpenAI revealed a way to block its web crawler, that feature hasn’t been nearly as useful as it could be. Many sites, such as Ars Technica (which blocks the OpenAI crawler as part of our parent company’s policy), won’t show up as results in ChatGPT because of this.
SearchGPT appears to untangle the association between OpenAI’s web crawler for scraping training data and the desire for OpenAI chatbot users to search the web. Notably, in the new SearchGPT announcement, OpenAI says, “Sites can be surfaced in search results even if they opt out of generative AI training.”
Even so, OpenAI says it is working on a way for publishers to manage how they appear in SearchGPT results so that “publishers have more choices.” And the company says that SearchGPT’s ability to browse the web is separate from training OpenAI’s AI models.
An uncertain future for AI-powered search
OpenAI claims SearchGPT will make web searches faster and easier. However, the effectiveness of AI-powered search compared to traditional methods is unknown, as the tech is still in its early stages. But let’s be frank: The most prominent web-search engine right now is pretty terrible.
Over the past year, we’ve seen Perplexity.ai take off as a potential AI-powered Google search replacement, but the service has been hounded by issues with confabulations and accusations of plagiarism among publishers, including Ars Technica parent Condé Nast.
Unlike Perplexity, OpenAI has many content deals lined up with publishers, and it emphasizes that it wants to work with content creators in particular. “We are committed to a thriving ecosystem of publishers and creators,” says OpenAI in its news release. “We hope to help users discover publisher sites and experiences, while bringing more choice to search.”
In a statement for the OpenAI press release, Nicholas Thompson, CEO of The Atlantic (which has a content deal with OpenAI), expressed optimism about the potential of AI search: “AI search is going to become one of the key ways that people navigate the internet, and it’s crucial, in these early days, that the technology is built in a way that values, respects, and protects journalism and publishers,” he said. “We look forward to partnering with OpenAI in the process, and creating a new way for readers to discover The Atlantic.”
OpenAI has experimented with other offshoots of its AI language model technology that haven’t become blockbuster hits (most notably, GPTs come to mind), so time will tell if the techniques behind SearchGPT have staying power—and if it can deliver accurate results without hallucinating. But the current state of web search is inviting new experiments to separate the signal from the noise, and it looks like OpenAI is throwing its hat in the ring.
OpenAI is currently rolling out SearchGPT to a small group of users and publishers for testing and feedback. Those interested in trying the prototype can sign up for a waitlist on the company’s website.
In the AI world, there’s a buzz in the air about a new AI language model released Tuesday by Meta: Llama 3.1 405B. The reason? It’s potentially the first time anyone can download a GPT-4-class large language model (LLM) for free and run it on their own hardware. You’ll still need some beefy hardware: Meta says it can run on a “single server node,” which isn’t desktop PC-grade equipment. But it’s a provocative shot across the bow of “closed” AI model vendors such as OpenAI and Anthropic.
“Llama 3.1 405B is the first openly available model that rivals the top AI models when it comes to state-of-the-art capabilities in general knowledge, steerability, math, tool use, and multilingual translation,” says Meta. Company CEO Mark Zuckerberg calls 405B “the first frontier-level open source AI model.”
In the AI industry, “frontier model” is a term for an AI system designed to push the boundaries of current capabilities. In this case, Meta is positioning 405B among the likes of the industry’s top AI models, such as OpenAI’s GPT-4o, Claude’s 3.5 Sonnet, and Google Gemini 1.5 Pro.
A chart published by Meta suggests that 405B gets very close to matching the performance of GPT-4 Turbo, GPT-4o, and Claude 3.5 Sonnet in benchmarks like MMLU (undergraduate level knowledge), GSM8K (grade school math), and HumanEval (coding).
But as we’ve noted many times since March, these benchmarks aren’t necessarily scientifically sound or translate to the subjective experience of interacting with AI language models. In fact, this traditional slate of AI benchmarks is so generally useless to laypeople that even Meta’s PR department now just posts a few images of charts and doesn’t even try to explain them in any detail.
Enlarge/ A Meta-provided chart that shows Llama 3.1 405B benchmark results versus other major AI models.
We’ve instead found that measuring the subjective experience of using a conversational AI model (through what might be called “vibemarking”) on A/B leaderboards like Chatbot Arena is a better way to judge new LLMs. In the absence of Chatbot Arena data, Meta has provided the results of its own human evaluations of 405B’s outputs that seem to show Meta’s new model holding its own against GPT-4 Turbo and Claude 3.5 Sonnet.
Enlarge/ A Meta-provided chart that shows how humans rated Llama 3.1 405B’s outputs compared to GPT-4 Turbo, GPT-4o, and Claude 3.5 Sonnet in its own studies.
Whatever the benchmarks, early word on the street (after the model leaked on 4chan yesterday) seems to match the claim that 405B is roughly equivalent to GPT-4. It took a lot of expensive computer training time to get there—and money, of which the social media giant has plenty to burn. Meta trained the 405B model on over 15 trillion tokens of training data scraped from the web (then parsed, filtered, and annotated by Llama 2), using more than 16,000 H100 GPUs.
So what’s with the 405B name? In this case, “405B” means 405 billion parameters, and parameters are numerical values that store trained information in a neural network. More parameters translate to a larger neural network powering the AI model, which generally (but not always) means more capability, such as better ability to make contextual connections between concepts. But larger-parameter models have a tradeoff in needing more computing power (AKA “compute”) to run.
We’ve been expecting the release of a 400 billion-plus parameter model of the Llama 3 family since Meta gave word that it was training one in April, and today’s announcement isn’t just about the biggest member of the Llama 3 family: There’s an entirely new iteration of improved Llama models with the designation “Llama 3.1.” That includes upgraded versions of its smaller 8B and 70B models, which now feature multilingual support and an extended context length of 128,000 tokens (the “context length” is roughly the working memory capacity of the model, and “tokens” are chunks of data used by LLMs to process information).
Meta says that 405B is useful for long-form text summarization, multilingual conversational agents, and coding assistants and for creating synthetic data used to train future AI language models. Notably, that last use-case—allowing developers to use outputs from Llama models to improve other AI models—is now officially supported by Meta’s Llama 3.1 license for the first time.
Abusing the term “open source”
Llama 3.1 405B is an open-weights model, which means anyone can download the trained neural network files and run them or fine-tune them. That directly challenges a business model where companies like OpenAI keep the weights to themselves and instead monetize the model through subscription wrappers like ChatGPT or charge for access by the token through an API.
Fighting the “closed” AI model is a big deal to Mark Zuckerberg, who simultaneously released a 2,300-word manifesto today on why the company believes in open releases of AI models, titled, “Open Source AI Is the Path Forward.” More on the terminology in a minute. But briefly, he writes about the need for customizable AI models that offer user control and encourage better data security, higher cost-efficiency, and better future-proofing, as opposed to vendor-locked solutions.
All that sounds reasonable, but undermining your competitors using a model subsidized by a social media war chest is also an efficient way to play spoiler in a market where you might not always win with the most cutting-edge tech. That benefits Meta, Zuckerberg says, because he doesn’t want to get locked into a system where companies like his have to pay a toll to access AI capabilities, drawing comparisons to “taxes” Apple levies on developers through its App Store.
Enlarge/ A screenshot of Mark Zuckerberg’s essay, “Open Source AI Is the Path Forward,” published on July 23, 2024.
So, about that “open source” term. As we first wrote in an update to our Llama 2 launch article a year ago, “open source” has a very particular meaning that has traditionally been defined by the Open Source Initiative. The AI industry has not yet settled on terminology for AI model releases that ship either code or weights with restrictions (such as Llama 3.1) or that ship without providing training data. We’ve been calling these releases “open weights” instead.
Unfortunately for terminology sticklers, Zuckerberg has now baked the erroneous “open source” label into the title of his potentially historic aforementioned essay on open AI releases, so fighting for the correct term in AI may be a losing battle. Still, his usage annoys people like independent AI researcher Simon Willison, who likes Zuckerberg’s essay otherwise.
“I see Zuck’s prominent misuse of ‘open source’ as a small-scale act of cultural vandalism,” Willison told Ars Technica. “Open source should have an agreed meaning. Abusing the term weakens that meaning which makes the term less generally useful, because if someone says ‘it’s open source,’ that no longer tells me anything useful. I have to then dig in and figure out what they’re actually talking about.”
The Llama 3.1 models are available for download through Meta’s own website and on Hugging Face. They both require providing contact information and agreeing to a license and an acceptable use policy, which means that Meta can technically legally pull the rug out from under your use of Llama 3.1 or its outputs at any time.
Enlarge/ Kevin Scott, CTO and EVP of AI at Microsoft speaks onstage during Vox Media’s 2023 Code Conference at The Ritz-Carlton, Laguna Niguel on September 27, 2023 in Dana Point, California.
During an interview with Sequoia Capital’s Training Data podcast published last Tuesday, Microsoft CTO Kevin Scott doubled down on his belief that so-called large language model (LLM) “scaling laws” will continue to drive AI progress, despite some skepticism in the field that progress has leveled out. Scott played a key role in forging a $13 billion technology-sharing deal between Microsoft and OpenAI.
“Despite what other people think, we’re not at diminishing marginal returns on scale-up,” Scott said. “And I try to help people understand there is an exponential here, and the unfortunate thing is you only get to sample it every couple of years because it just takes a while to build supercomputers and then train models on top of them.”
LLM scaling laws refer to patterns explored by OpenAI researchers in 2020 showing that the performance of language models tends to improve predictably as the models get larger (more parameters), are trained on more data, and have access to more computational power (compute). The laws suggest that simply scaling up model size and training data can lead to significant improvements in AI capabilities without necessarily requiring fundamental algorithmic breakthroughs.
Since then, other researchers have challenged the idea of persisting scaling laws over time, but the concept is still a cornerstone of OpenAI’s AI development philosophy.
You can see Scott’s comments in the video below beginning around 46: 05:
Microsoft CTO Kevin Scott on how far scaling laws will extend
Scott’s optimism contrasts with a narrative among some critics in the AI community that progress in LLMs has plateaued around GPT-4 class models. The perception has been fueled by largely informal observations—and some benchmark results—about recent models like Google’s Gemini 1.5 Pro, Anthropic’s Claude Opus, and even OpenAI’s GPT-4o, which some argue haven’t shown the dramatic leaps in capability seen in earlier generations, and that LLM development may be approaching diminishing returns.
“We all know that GPT-3 was vastly better than GPT-2. And we all know that GPT-4 (released thirteen months ago) was vastly better than GPT-3,” wrote AI critic Gary Marcus in April. “But what has happened since?”
The perception of plateau
Scott’s stance suggests that tech giants like Microsoft still feel justified in investing heavily in larger AI models, betting on continued breakthroughs rather than hitting a capability plateau. Given Microsoft’s investment in OpenAI and strong marketing of its own Microsoft Copilot AI features, the company has a strong interest in maintaining the perception of continued progress, even if the tech stalls.
Frequent AI critic Ed Zitron recently wrote in a post on his blog that one defense of continued investment into generative AI is that “OpenAI has something we don’t know about. A big, sexy, secret technology that will eternally break the bones of every hater,” he wrote. “Yet, I have a counterpoint: no it doesn’t.”
Some perceptions of slowing progress in LLM capabilities and benchmarking may be due to the rapid onset of AI in the public eye when, in fact, LLMs have been developing for years prior. OpenAI continued to develop LLMs during a roughly three-year gap between the release of GPT-3 in 2020 and GPT-4 in 2023. Many people likely perceived a rapid jump in capability with GPT-4’s launch in 2023 because they had only become recently aware of GPT-3-class models with the launch of ChatGPT in late November 2022, which used GPT-3.5.
In the podcast interview, the Microsoft CTO pushed back against the idea that AI progress has stalled, but he acknowledged the challenge of infrequent data points in this field, as new models often take years to develop. Despite this, Scott expressed confidence that future iterations will show improvements, particularly in areas where current models struggle.
“The next sample is coming, and I can’t tell you when, and I can’t predict exactly how good it’s going to be, but it will almost certainly be better at the things that are brittle right now, where you’re like, oh my god, this is a little too expensive, or a little too fragile, for me to use,” Scott said in the interview. “All of that gets better. It’ll get cheaper, and things will become less fragile. And then more complicated things will become possible. That is the story of each generation of these models as we’ve scaled up.”
On Thursday, OpenAI researchers unveiled CriticGPT, a new AI model designed to identify mistakes in code generated by ChatGPT. It aims to enhance the process of making AI systems behave in ways humans want (called “alignment”) through Reinforcement Learning from Human Feedback (RLHF), which helps human reviewers make large language model (LLM) outputs more accurate.
As outlined in a new research paper called “LLM Critics Help Catch LLM Bugs,” OpenAI created CriticGPT to act as an AI assistant to human trainers who review programming code generated by the ChatGPT AI assistant. CriticGPT—based on the GPT-4 family of LLMS—analyzes the code and points out potential errors, making it easier for humans to spot mistakes that might otherwise go unnoticed. The researchers trained CriticGPT on a dataset of code samples with intentionally inserted bugs, teaching it to recognize and flag various coding errors.
The researchers found that CriticGPT’s critiques were preferred by annotators over human critiques in 63 percent of cases involving naturally occurring LLM errors and that human-machine teams using CriticGPT wrote more comprehensive critiques than humans alone while reducing confabulation (hallucination) rates compared to AI-only critiques.
Developing an automated critic
The development of CriticGPT involved training the model on a large number of inputs containing deliberately inserted mistakes. Human trainers were asked to modify code written by ChatGPT, introducing errors and then providing example feedback as if they had discovered these bugs. This process allowed the model to learn how to identify and critique various types of coding errors.
In experiments, CriticGPT demonstrated its ability to catch both inserted bugs and naturally occurring errors in ChatGPT’s output. The new model’s critiques were preferred by trainers over those generated by ChatGPT itself in 63 percent of cases involving natural bugs (the aforementioned statistic). This preference was partly due to CriticGPT producing fewer unhelpful “nitpicks” and generating fewer false positives, or hallucinated problems.
The researchers also created a new technique they call Force Sampling Beam Search (FSBS). This method helps CriticGPT write more detailed reviews of code. It lets the researchers adjust how thorough CriticGPT is in looking for problems, while also controlling how often it might make up issues that don’t really exist. They can tweak this balance depending on what they need for different AI training tasks.
Interestingly, the researchers found that CriticGPT’s capabilities extend beyond just code review. In their experiments, they applied the model to a subset of ChatGPT training data that had previously been rated as flawless by human annotators. Surprisingly, CriticGPT identified errors in 24 percent of these cases—errors that were subsequently confirmed by human reviewers. OpenAI thinks this demonstrates the model’s potential to generalize to non-code tasks and highlights its ability to catch subtle mistakes that even careful human evaluation might miss.
Despite its promising results, like all AI models, CriticGPT has limitations. The model was trained on relatively short ChatGPT answers, which may not fully prepare it for evaluating longer, more complex tasks that future AI systems might tackle. Additionally, while CriticGPT reduces confabulations, it doesn’t eliminate them entirely, and human trainers can still make labeling mistakes based on these false outputs.
The research team acknowledges that CriticGPT is most effective at identifying errors that can be pinpointed in one specific location within the code. However, real-world mistakes in AI outputs can often be spread across multiple parts of an answer, presenting a challenge for future iterations of the model.
OpenAI plans to integrate CriticGPT-like models into its RLHF labeling pipeline, providing its trainers with AI assistance. For OpenAI, it’s a step toward developing better tools for evaluating outputs from LLM systems that may be difficult for humans to rate without additional support. However, the researchers caution that even with tools like CriticGPT, extremely complex tasks or responses may still prove challenging for human evaluators—even those assisted by AI.
