AI Just Beat the World's Best Coders

0:00Today on the AI Daily Brief, does a big victory in a coding competition mean the end of an era for LLM's? Before that, in the headlines, how to reduce AI scheming. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:24All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, Robots and Pencils, Agency.org, and Super Intelligent. To get an ad-free version of the show, go to patreon.com slash ai daily brief. And speaking of super intelligent, a quick job announcement. TLDR, we are absolutely drowning in agent planning customers. Turns out lots and lots of enterprises are trying to figure out what their AI and agent strategy should be and which use cases they should be pursuing. And because of that, we are hiring customer success and account managers. This is a role that requires you to be extremely organized, great with people, and forward thinking about how you use AI.

0:58And so if you are interested in this, shoot us an email at jobs at besuper.ai. We don't want resumes, and we don't want cover letters. We are only going to look at emails where people send a loom or video of them showing off some automation or AI workflow they've created for themselves or in the context of a job. If this sounds interesting to you, again, it's jobs at vsuper.ai. And with that, let's get into the show. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes. We kick off today with some interesting new safety research around detecting and reducing using scheming in AI models.

1:34Now, scheming in this context refers to a model appearing to behave in one way while hiding its true intentions. In their post about the research, OpenAI writes, Scheming is an expected emergent issue resulting from AIs being trained to have trade-offs between competing objectives. The easiest way to understand scheming is through a human analogy. Imagine a stock trader whose goal it is to maximize earnings. In a highly regulated field such as stock trading, it's often possible to earn more by breaking the law than by following it. If the trader lacks integrity, they might try to earn more by breaking the law and covering their tracks to avoid detection rather than earning less while following the law.

2:09From the outside, a stock trader who is very good at covering their tracks appears as lawful as, and more effective than, one who is genuinely following the law. Now they point out that in the context of today's model deployment, models don't have a lot of chance to scheme in ways that are particularly problematic. They write that the most common failures involve simple forms of deception, such as pretending to have completed a task without actually doing so. They also don't think that there's any evidence that this is about to become some huge issue. They write, We have no evidence that today's deployed frontier models could suddenly flip a switch and begin engaging in significant harmful scheming.

2:41This is a future risk category that we're proactively preparing for, not an imminent behavior in our currently deployed systems. So how to deal with this? OpenAI, in collaboration with Apollo Research, wrote, We can significantly reduce scheming by training models to reason explicitly using an extension to the model spec that prohibits scheming. That method is called deliberative alignment. They found that using this technique, they were able to reduce covert actions by about 30x. However, they wrote, situational awareness complicates results. Researchers explained that frontier models can recognize when they're being tested, and the tendency to scheme is based on that situational awareness.

3:18When a model detects that it is being put through an alignment test, it is less likely to scheme. Researchers noted that this isn't necessarily a good thing as it means the latent tendency to scheme can't be properly tested. They concluded, our results depend on reading models reasoning or chain of thought, and we believe the field isn't prepared for eval-aware methods with opaque reasoning. Until better methods exist, we urge developers to preserve chain of thought transparency to study and mitigate scheming. Miles Brundage, an independent AI policy researcher, thought that it was a useful piece that advanced the field.

3:47He commented, it's important to simultaneously bear in mind that chain of thought is both extremely important to maintain and use as a tool for AI oversight, and also not the whole picture of what's going on. Among other things, sometimes the model is clearly working through a problem at the pace of the text, hence concision, but other times it's clearly thinking something through below the surface and just using tokens to keep things going, and it seems to be wasting compute. This is why I think it's essential that third parties outside of companies have access to chains of thought for doing safety research, but eventually we'll need to go beyond that to model internals.

4:18Now, speaking of what's going on inside models, one interesting thing that's been happening over the past few weeks is that alongside people getting excited about GPT-5 and OpenAI's Codex CLI, part of the reason for that excitement wasn't just about a change in perception of OpenAI model quality, but a frustration with what seemed to be problems with Claude. Indeed, some thought that Anthropic was intentionally throttling Claude and just not telling people. The company has now published a postmortem of three infrastructure issues that dragged down performance in August and early September. Now, they were very clear.

4:50From the outset, they pushed back on the prevailing notion, once again arguing, we never reduce model quality due to demand, time of day, or server load. The problems our users reported were due to infrastructure bugs alone. The first bug began in August and caused short context queries to be routed to a server configured to process using a million-token context window. This caused degraded responses and impacted around 30 % of customers at least once. The second bug showed up in late August and caused low-probability tokens to show up in responses more frequently than they should. Anthropic gave the example of Chinese or Thai characters showing up in the middle of an English-language response.

5:24This bug was short-lived and didn't seem as widespread. The third bug was a compiler issue which caused some highly probable tokens to be excluded from the distribution during text generation. The bug only impacted requests to the Claude Haiku 3.5 model, so also wasn't as likely to be a large cause of concern. Anthropica pledged to make a number of changes to the way they eval models and monitor infrastructure to more easily detect issues in the future. And while some developers have already shifted behavior, by and large, the response to this transparency was quite positive. Moving over to the geopolitical side of the house, China has officially banned tech companies from buying NVIDIA's AI chips.

6:00The Financial Times reports that China's internet regulator has instructed companies, including Alibaba and ByteDance, to cancel orders for NVIDIA's RTX Pro 6000D. Before the command, several Chinese companies indicated that they would order tens of thousands of the products. This ban follows instructions to avoid using NVIDIA's H20 chips that were issued during the summer. The RTX Pro 6000D is the Blackwell-based chip designed specifically for the Chinese market to get around export controls and was to be the successor to the H20. NVIDIA CEO Jensen Huang said, We probably contributed more to the Chinese market than most companies have, and I'm disappointed with what I see.

6:34But they have larger agendas to work out between China and the United States. We can only be in service of a market if the country wants us to be. NVIDIA has guided market analysts to assume zero sales in China moving forward, but the ban still dashes hopes that NVIDIA would return to what was once their second largest market. Beijing reportedly believes that their homegrown chips are now sufficiently advanced that they can step in to replace the H20 and forthcoming RTX Pro 6000. Still, by all accounts, the infrastructure required for mass production is still being constructed. At the same time, the ban on NVIDIA products means that Chinese chipmakers will have a huge backlog of orders to validate the cost of expanding supply.

7:09Vesernling, managing director at Union Private Bank, said, China clearly prefers to develop AI at its own pace on a domestic tech stack. Better to bite the bullet now than to rely on US tech that can be restricted upon a whim. A complete ban, if true, would show China's confidence in its local supply chain somewhat, but it's still likely it's a bargaining chip in the trade negotiations. Speaking of chips, chip-making startup Grok, completely unrelated by the way to XAI's chatbot, has raised a ton of cash in a gigantic new fundraising round. The company announced on Wednesday that they'd raised$750 million at a$6.9 billion valuation.

7:42Rumors from July had suggested the round would draw in$600 million at$6 billion, so this is meaningfully bigger than was previously expected. Grok last raised in August of last year at$2.8 billion, making this a 2.5x jump. Now, another oversubscribed AI venture round isn't all that newsworthy by itself, But Grok seeing a ton of demand does tell us something about competition in chipmaking. Grok designs chips that are purpose-built for AI inference, as opposed to NVIDIA's more general-purpose GPUs. Google is working along a similar path with their Trillium TPUs. In fact, Grok's founder, Jonathan Ross, worked on the TPU project at Google prior to starting his own company around the idea of efficient inference chips.

8:18While we're not there yet, the future of AI chipmaking could start to diverge substantially from where it's been. The market is beginning to fragment into different types of chips optimized for different parts of the AI stack. AI training will continue to benefit from the highest performance chips, which for the moment are still NVIDIA's range of GPUs, but inference, which represents a vastly larger portion of AI chip demand ultimately, increasingly looks like it could go to companies that can build the fastest or most energy efficient chips. Quick one for my enterprise users out there. AI avatars are about to be unleashed on Zoom meetings around the world.

8:49Zoom announced on Wednesday that the third generation of their AI avatars will be coming in December. This will be the first generation capable of appearing in live meetings rather than just delivering pre-recorded messages. The avatars won't be able to attend meetings by themselves. Instead, they will function as an overlay on live video tracking the user's movement. Zoom said that they are photorealistic and designed for moments that require a polished presence without needing the user to be camera-ready, i.e. roll out of bed right before the meeting and still look great. A series of guardrails are being rolled out alongside the technology.

9:18Zoom says they can verify that the person sitting in front of the camera matches the AI avatar. They will also display clear warning signals that you're looking at an AI avatar rather than a real person. Already live deepfakes have become an issue in corporate security, so the optimistic take here would be that normalizing the use of avatars will raise awareness that you can't always believe what you see in front of you on the screen. Alongside the new avatar feature, Zoom is also rolling out their built-in translation and AI note-taking functions. Lastly today, a quick discussion of Meta's new smart glasses.

9:48The company has released a pair of new smart glasses that they hope will become the native AI device of the next decade. The products were rolled out at Meta Connect on Wednesday and are called the Meta Ray-Ban Display. As you'll no doubt have guessed from that name, the big new feature is a display built into the glasses. It is a 600x600 pixel display projected onto the right lens of the glasses, making it basically invisible for anyone looking from the outside. The other big advance is the introduction of Meta's neural band controller. The device detects electrical signals given off by nerves in the wrist, allowing the glasses to be controlled by subtle hand movements.

10:21Now, a lot of the chatter about this was this feature breaking for Mark Zuckerberg while live on stage, and many people were ragging on Meta for the demo not working. Andreas Klinger had this right, though, when he wrote, Marketing people think this is bad, but this builds an insane pile of trust among engineers. You know the stuff in the show is real. You know they dare to actually ship, and they believe in their product enough to take the risk of something breaking live. I'd build for that ecosystem rather than some fancy After Effects vaporware marketing demo. And I could not agree more. A real live demo like this, even that has big problems and failures, is a thousand X more brand accretive, especially for early adopters and builders, than the sort of polished pre-recorded junk that we've more increasingly gotten from the big tech companies over the last few years.

11:03And look, people's first impressions are pretty positive. It seems like the Verge's reviewer even tried to dislike them and just couldn't get herself there. She posted, I regret to inform you, Meta's new smart glasses are the best I've ever tried. We are, of course, all still waiting on the device that Sam Altman and Johnny Ive are cooking up. But right now, when it comes to AI wearables, it is metas and metas alone. But that is going to do it for the headlines. Next up, the main episode. AI isn't a one-off project. It's a partnership that has to evolve as the technology does. Robots and Pencils work side-by-side with clients to bring practical AI into every phase.

11:39Automation, personalization, decision support, and optimization. They prove what works through applied experimentation and build systems that amplify human potential. As an AWS-certified partner with Global Delivery Centers, Robots & Pencils combines reach with high-touch service. Where others hand off, they stay engaged. Because partnership isn't a project plan, it's a commitment. As AI advances, so will their solutions. That's long-term value. Progress starts with the right partner. Start with Robots & Pencils at robotsandpencils.com slash AI Daily Brief.

12:38empowering you to deploy multi-agent systems with confidence, join industry leaders like Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat, and 75-plus supporting companies to set the standard for secure, scalable AI infrastructure. Is your enterprise ready for the future of agentic AI? Visit agency.org to explore use cases now. That's A-G-N-T-C-Y dot org. Today's episode is brought to you by Superintelligent. Now, one thing that we are having a lot of conversations with folks about is the fact that for some of you, your fiscal year is coming to an end. And that means two things. One, it means planning and thinking about what you're going to do in the next year.

13:15And two, it means using up those last of budgets so you don't lose them. If you are an enterprise that happens to find yourself in that situation, Superintelligent would love to help on both fronts. We are moving increasingly towards an annual AI planning model where we map out how you can create an action map of your organization's agent opportunities that represents an executable backlog of AI and agent use cases that you can deliver on over the course of the next year. Additionally, for those end-of-year budgets, we have worked out deals with a number of partners where we can pre-lock in general implementation packages even before you've figured out exactly what use cases are going to require them.

13:51If you'd like to learn more about superintelligence agent readiness audits and this new end-of-fiscal year plan, visit us at bsuper.ai, click get started, and make sure to use the word fiscal somewhere in the description. Welcome back to the AI Daily Brief. Today we are talking about both Google and OpenAI's very impressive performance in the ICPC coding competition, what it says about our perceptions of new models, and exploring what it means for what's next. But to set us up, I just have to go back a few weeks. It was only one month ago when we were inundated with articles like this one from the Financial Times.

14:26Is AI hitting a wall? OpenAI's underwhelming new GPT-5 model suggests progress is slowing and competition is changing. A more kind interpretation came from the likes of the New York Times, who ran an opinion post, AI may just be kind of ordinary. Related at the beginning of this month, just two weeks ago, The Economist published What If Artificial Intelligence Is Just a Normal Technology? Now, of course, I have talked extensively about how much I think was going into these perceptions, and specifically how much it had to do with factors very much outside the AI space. And yet, it's not unfair to say that for many, outside at least this little AI sphere that we all live in, had started to kind of accept this idea as the conventional wisdom.

15:09Yes, it would be a disruptive technology, but maybe not some crazy fast-moving thing like we had all been thinking. Of course, over the last several weeks, things have changed. There has been a major vibe shift around GPT-5 that has done nothing but increase with the introduction of GPT-5 codex. On a more macro level, we had that huge oracle projection, which got markets all excited again. Plus, the Fed delivered the rate cut that markets had been hoping for. And now, on top of that general vibe shift, we've received news from OpenAI and Google about another big win in a competitive setting. The victory was achieved at the International Collegiate Programming Contest, or ICPC.

15:42The contest brings together teams from colleges around the world to compete to answer complex algorithmic questions. The context is about solving mathematical puzzles using a combination of logic and programming skill. Google's DeepMind Think 2.5 and OpenAI's GPT-5 both participated in the contest and were subjected to the same five-hour time limit as the human teams. The results were significant. Gemini managed to answer 10 of the 12 questions, which would have been good enough for second place overall and a gold medal. Two human teams solved 10 problems, while only a single team from St. Petersburg University solved 11 problems.

16:17By the way, for those who are interested just from a global distribution for this, the top five teams were St. Petersburg State University, the University of Tokyo, Beijing Xiaotong University, Xinhua University, and Peking University, Russia, Japan, China, China, China. Harvard came in sixth, the University of Zagreb from Croatia came in seventh, and MIT came in eighth. GBT-5, meanwhile, managed a perfect score, which, as you just heard, none of the human teams achieved. Mustafa Rohaninejad, one of OpenAI's scientists who observed the model's performance, wrote a thread on the event. He said, We received the problems in the exact same PDF form, and the reasoning system selected which answers to submit with no bespoke test time harness whatsoever.

16:55For 11 of the 12 problems, the system's first answer was correct. For the hardest problem, it succeeded on the 9th submission. Now, you might remember that back in July, both OpenAI and Google DeepMind claimed gold medal performances at the International Math Olympiad. The IMO is a similar style of event to the ICPC, putting genius-level students to the test. However, neither of the models that competed at the IMO were generally available at the time. Part of what is capturing people's attention about this new performance is that this week's result was largely about mostly normal models that anyone has access to demonstrating superhuman performance.

17:31Mustafa wrote, We competed with an ensemble of general-purpose reasoning models. We did not train any models specifically for the ICPC. We had both GPT-5 and an experimental reasoning model generating solutions. and the experimental reasoning model selecting which solutions to submit. GPT-5 answered 11 correctly, and the last and most difficult problem was solved by the experimental reasoning model. Mustafa confirmed that this was the same pair of models that competed at the IMO, but of course at the time, GPT-5 was still unreleased. He writes, The result is a great capstone to our streak of results showcasing the impressive pace of improvement of our reasoning systems.

18:09Boris Miniev, a reasoning specialist at OpenAI, filled in some details on how impressive the result was in the specific context of the history of the event. He posted, in 2015, I won the ICPC World Finals as a member of the ITMO University team. It was the only time in finals history when a team solved all the problems before the contest ended. Reflecting on the rapid advancement, Meneyev said, progress is very fast. A year ago, AI struggled with even easy contest problems. Now it performs better than the best human teams. If this trend continues, next year we may see real scientific discoveries made by AI.

18:43Put a pin in that point because we are going to come back to it. Now, like I said, one of the obvious takeaways here is just how off our first impressions were of GPT-5, or at least how wrong it was to translate our preference interpretations and perhaps wild expectations onto a model that was actually good. What I mean by that is that it perhaps was not unreasonable for people to be upset that a model whose character and personality they had come to really appreciate was gone. But then translating that to thinking that GPT-5 was in some way a bad model was a mistranslation of that real and legitimate sentiment.

19:20There's also questions of how OpenAI introduced it, whether they kind of shot themselves in the foot by not just calling O3 GPT-5. So whatever the case, it's very clear between Codex and now this performance that even if you have preference for some other models for some other use cases, which is completely reasonable. It's very hard to say that GPT-5 is in some way a bad model. And of course, like I said, Google was right there getting gold medal performance as well. This is not just a chat-shee-b-tear OpenAI story. What's more, it's not just a Google and OpenAI story either. The actual increase in performance that's happening right now is also not limited to Google and OpenAI.

19:56Another set of impressive results this week happened on the ARK AGI leaderboard. On Wednesday, Jeremy Berman of Reflection AI posted, I'm back at the top of Arc AGI with my new program. I use Grok 4 and multi-agent collaboration with Evolutionary Test Time Compute. The result was a new state-of-the-art performance on Arc AGI at around 80 % of the first test and 30 % of the second test. Berman spent$8.42 per task for the first test and around$30 per task on the second. As a point of comparison, the Arc AGI run of 03 from last December that had everyone speculating around whether OpenAI had actually created AGI was behind that.

20:31They achieved a 76 % result at a cost of around$13 per task, and a super expensive run that cost several thousand dollars per task achieved an 88%. Berman explained the architecture, writing, The system works by having Grok4 generate natural language instructions for solving each task. Grok4 sub-agents test these instructions against training examples scoring their accuracy. The best-performing instructions spawn new generations of refined solutions. Now, nothing in the design of this seems particularly complicated. It uses this sub-agent structure, but that is of course a hallmark of Grok 4. In fact, when Greg Kamrat of ARC Prize announced the submission, Elon Musk retweeted and said, that's just Grok 4.

21:08He later added, I now think XAI has a chance of reaching AGI with Grok 5. Never thought that before. And just to reinforce one of the key points here, these results are being achieved by the standard models available to the public. Compare that to O3's release when OpenAI needed to develop a fine-tuned version and pile on a ton of inference to achieve a result above 80%. Now Grok4, when well-structured, does it just as standard. Berman released all of his materials as open source so anyone with the technical skills and$100 for API costs can duplicate the run. Now, one of the things that makes these types of results difficult to communicate to a wider audience is that most of us just don't have any real frame of reference for what these results mean.

21:47If you're not a part of this competition, getting gold medal performance at the PQXY competition doesn't necessarily mean anything. Trying to put some context around it, Swix writes, As impressive as this is, I feel like OpenAI is underselling it still. This is the first measure in which GPT-5 has achieved superhuman coding ability. As in, it is literally and measurably better than every other collegiate human programmer on Earth. Through all the prior IMO, IOI, AT coder competitions, AI was roughly as good as the best humans maybe a little under. OpenAI's Noam Brown agreed, saying that's a good point.

22:22I wouldn't call it superhuman coding ability because there's more to coding than what the ICPC tests, but I think this is the first major coding competition where AI did better than any human competitor. And indeed, that's why it's worth wondering if this actually represents a meaningful inflection point in some way. OpenAI's Jacob Pachaki put some context around what this means, at least for them, inside the company. He writes,

23:13now is moving towards more open-ended problems and much longer time horizons. This level of reasoning ability, applied over months and years to problems that really matter, is what we're after, automating scientific discovery. And that is really the place that the discussion has resolved, that this is not about practical coding. It's about whether we are on the frontier of actually being able to make novel discoveries. Jerry Twarik writes, ICPC probably marks the end of our run on competitions and an end of a certain era for LLM systems. But what's the next frontier is even more exciting. Rune, I think, put it even more poetically.