Everything Sam Altman Is Thinking About Right Now

0:00Today on the AI Daily Brief, everything Sam Altman is thinking about right now. Before that end of the headlines, is AI model welfare a thing? The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:18All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, Blitzy, Vanta, and Superintelligent. To get an ad-free version of the show, go to patreon.com slash ai daily brief. And if you are interested in sponsoring the show, shoot us a note at sponsors at ai daily brief dot AI. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes. We kick off the day with an interesting one. Anthropic has updated the safety features for Claude Opus, allowing it to terminate certain conversations. Basically, in the consumer chat interface, Opus 4 and 4.1 will now be able to end conversations.

0:53Now, Anthropic said that this ability was reserved for, quote, rare extreme cases of persistently harmful or abusive user interactions. They said the feature was developed as part of their, quote, exploratory work on potential AI welfare, though it has broader relevance to model alignment and safeguards. Anthropic said that this wasn't strictly about preventing harmful use, but also about protecting the model itself. In pre-deployment testing of Claude Opus 4, we included a preliminary model welfare assessment. As part of that assessment, we investigated Claude's self-reported and behavioral preferences and found a robust and consistent aversion to harm.

1:28In testing, the model would display, quote, apparent distress when engaging with harmful content and had, quote, a tendency to end harmful conversations when given the ability to do so in simulated user interactions. Now, as you might imagine, this has kicked up quite the conversation around whether AI model welfare is actually even a thing. Some tried to better understand the implications, while others had just had enough. Anonymous developer Bantag wrote, model welfare is not a real thing. stop anthropomorphizing language models. Lefteris Carapestas said, This sounds more like a feature trying to stop people from trying to jailbreak a convo to get into forbidden behavior?

2:06James B. wrote, Giving a model the ability to end a conversation in rare abusive cases is a sensible safety valve. Framing it as model welfare is provocative and risks confusing people about what today's systems actually experience, which is nothing. Use it as a narrow, transparent moderation tool, not as evidence that models feel distress. The pros, he writes, it reduces endless abuse loops and prevents models from being steered into harmful content. Also creates a clear boundary that can improve safety for minors and casual users. And if logged well, it can surface valuable red team signals about where policies break.

2:38The risks, however, include phrases like distress or welfare assessment misleading the public into thinking the model suffers. Today's models generate text. They don't have experiences. Now, whether you agree with any of this or not, it's likely to be more of a conversation in the future and one I will certainly keep an eye on. Interestingly, when the AI Safety Memes account asked Elon Musk to, quote, help set a good example and move the Overton window by giving Grok a quit button too, Elon responded, OK. Moving on now to the much more knowable part of the headlines. OpenAI's secondary share sale is coming together and it's shaping up to be a major liquidity event.

3:13Bloomberg reports that current and former OpenAI employees plan to sell$6 billion in stock to an investor group that includes Thrive, SoftBank and Dragoneer. Sources said the round will value OpenAI at$500 billion, validating a 60 % jump in valuation from the SoftBank-led round consummated at the beginning of the year. If it goes through, that valuation would make OpenAI the world's most valuable startup, overtaking SpaceX. This is also quite possibly the largest single-secondary sale in history. Now, a few other observations from the reporting. First, existing shareholders cannot get enough money into OpenAI and seem eager to take part in every new allocation.

3:49All three of the major investors in this secondary have been heavy participants in previous rounds, and they're still looking to deploy billions more. Secondly, this round is going to transform many open AI employees from multi-millionaires on paper to wealthy in real cash terms. Many wonder then, could that have an impact on the company's retention strategy, especially with other AI labs still poaching from their roster? Whatever happens next, the company is definitely now playing in the big leagues. at a$500 billion valuation if they were public that would make them something like the 20th biggest company in the world.

4:21Speaking of AI investor enthusiasm, it apparently is not just at the foundation model layer as Vercel is now fielding unsolicited investment offers at a$9 billion valuation. The company last raised money 18 months ago at a$3 billion valuation but has of course been a big beneficiary of the Vibe Coding boom. The company is currently getting 76 % gross margins for what is effectively a cloud services company. Overall, it's clear that investors are hungry for AI investments wherever they can get them. Lastly today, the latest out of Meta. According to the information, the company is planning on their fourth restructuring effort so far this year.

4:56This time, the focus is on reorganizing the new superintelligence team. Sources said that the team will be divided into four groups. The TBD Lab, presumably a secret projects group with goals yet to be determined. A products team that will take over responsibility for the Meta AI Assistant, among other projects, an infrastructure team presumably dealing with the increasingly complex build-out of Meta's gigantic new data centers, and the Fundamental AI Research or Fair Lab focused on long-term research. Now, the changes haven't been announced internally and could still change, but it seems like the goal is to provide a clear divide between pure research and shipping teams.

5:29One thing I will note is that while the reporting has this as their fourth restructuring effort in six months, kind of implying that even though super superintelligence is just up and running, it's already in trouble, it kind of feels more to me like the major restructuring was the creation of the superintelligence lab in the first place to house all of their AI efforts, and this is just the natural division of that company that was always going to get decided on. Maybe that's not the case and it's more chaotic internally than I'm assuming, but I'm sure we'll learn more as things get up and running.

5:58For now, that's going to do it for today's AI Daily Brief Headlines edition. Next up, the main episode. This episode is brought to you by Blitzy, the enterprise autonomous software development platform with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Enterprise engineering leaders start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task. Blitzy delivers 80 % plus of the development work autonomously while providing a guide for the final 20 % of human development work required to complete the sprint.

6:36Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Blitzy is providing a limited-time, 30-day free proof-of-concept for qualifying enterprises. The team will provide a 5x velocity increase on a real development project in your org. Visit blitzy.com and press book demo to learn how Blitzy transforms your STLC from AI-assisted to AI-native. That's B-L-I-T-Z-Y dot com. As a founder, you're moving fast towards product market fit, your next round, or your first big enterprise deal.

7:14But with AI accelerating how quickly startups build and ship, security expectations are higher earlier than ever. Getting security and compliance right can unlock growth or stall it if you wait too long. With deep integrations and automated workflows built for fast-moving teams, Vanta gets you audit-ready fast and keeps you secure with continuous monitoring as your models, infra, and customers evolve. Fast-growing customers like Langchain, Writer, and Cursor trusted Vanta to build a scalable foundation from the start. And look, as someone who lives in the world of enterprise procurement, I love how Vanta makes it easy to get compliance right.

7:47The last thing you need when you're trying to win that big deal is to have it scuttled by is something that Vanta has solved for over 10 ,000 companies. Go to vanta.com slash NLW to save$1 ,000 today through the Vanta for Startups program and join over 10 ,000 ambitious companies already scaling with Vanta. That's V-A-N-T-A dot com slash NLW to save$1 ,000 for a limited time. If you are a regular listener, you will have heard about superintelligence agent readiness audits at this point. But I wanted to tell you today about the full suite of agent readiness products that go beyond just the initial readiness report.

8:23Over the last six months, Superintelligent has built out an entire agent planning suite. We help you move from discovery to planning to implementation. After you've completed your agent readiness audits, we help you double-click on your most important use cases with what we call our use case planning reports. These reports are going to help you understand what sort of technical preparation you need to do to be ready for a use case, what challenges you might face in implementation, and whether you should be thinking about building, buying, partnering, or some combination. After that, you can even get a spec document in what we call our technical blueprint that gives either your developers or the developers of the partner you work with what they need to build exactly the agent that you're looking for.

9:01If you want to learn more about Superintelligence Agent Planning Suite, we built a custom GPT to answer your questions. Just go to bit.ly slash super super agent. That's bit.ly slash super super agent, all one word. And if you have any questions, the agent can even help you book an appointment with our team. Welcome back to the AI Daily Brief. Today, we are recapping an interesting dinner conversation that Sam Altman and a couple of other executives from OpenAI had with a group of journalists last week. Now, if you're wondering why I'm taking the time to go over everything that was said in this meeting, it's because the journalists that they invited to this decided it was a better idea to spend like four paragraphs talking about the price of the entrees than actually getting into this incredible amount of information that they had access to.

9:51Which, if you're listening to podcasts to get your information, maybe you have realized why traditional outlets are struggling. But in any case, there was so much interesting signal here that I went through basically every published account of the dinner and summarized it into a group of trends and themes, something like 20 different topics overall. The first theme we have to talk about is the GPT-5 launch. And one of the things that was clearest from this conversation was that they very much understood that this launch was at least to some extent botched. In fact, it seemed like many of the journalists assumed that because this conversation happened last week, a week after the launch or so, that basically it was OpenAI trying to get journalists back on their side rather than talking about how bad the GPT-5 launch had gone.

10:34Now, an interesting thing has been happening recently. In the less than two weeks since this launch has happened, it's created context for this summer's Contra AI narrative, summed up in this New Yorker piece, what if AI doesn't get much better than this? Now, if you wonder what I mean by summer Contra AI narrative, each summer since ChatGPT launched, there has been a moment where basically people took some event and used it as a context to basically try to peel back all the hype around AI and say maybe it's not all it's cracked up to be. In June of 2023, it happened after ChatGPT had its first down month after students left school.

11:09This, of course, created context for people to argue that ChatGPT was just a platform for cheating students and nothing more. Then last year, in around July of 2024, we got, of course, the$600 billion problem and Goldman Sachs note about too much spend, too few results, that was all about infrastructure investment and how companies were spending way too much money and they were never going to make it back. This year, it happened a month later in August, and the big theme is that the initial impressions of GPT-5 mean that AI must, yes, indeed be slowing down or hitting a wall. And for real this time.

11:40And yet, even as these things have been published over the last three or four days, there has been a major shift apparent to anyone who's watching closely. Swix from Latent Space wrote on Sunday, watching the timeline flip on GPT-5 sentiment from negative to positive is pretty funny. For those too cooked by the algorithm to remember, the same thing happened to O1. Dan Mack writes, the vibe has shifted. With GPT-3.5 level models, I started out impressed and I became less so the more I used them. Opposite has been true with next-gen models since. Jvim Auschwitz had maybe the best write-up of what happened, calling GPT-5 the reverse DeepSeek moment.

12:16Basically what he argued is that during the DeepSeek moment, which was in January, everyone freaked out about how China had caught up and how amazing the DeepSeek R1 model was and how cheap it was. But when you actually peel back, there were a bunch of things going that helped contribute to the narrative. He writes,

13:00retrospect was significantly overstated because of this confluence of trends. GPT-5 then, as the reverse deep-seek moment, is kind of the opposite. A better, more significant model that has a comparative dearth or underselling based on a different set of factors. The factors that he points to include GPT-5 being evaluated as if it was scaling up compute in a way that it didn't. The poor initial experience with rate caps and loss models. People evaluating GPT-5 when they should have been evaluating GPT-5 thinking. A contributing narrative of a loss of momentum, and a group of people who have established prior saying that OpenAI is flailing, helping amplify the initial vibes.

13:37I'm sure that throughout the week we'll be talking more about this shift in GPT-5 sentiment, but it's definitely the case that even with that, they knew that they had bobbled things, particularly around getting rid of GPT-4-0. Allman said candidly, I think we totally screwed up some things on the rollout. I legitimately just thought we screwed that up. Now it was interesting that it really was about the removal of the old models that seemed to be the big deal. At the same time, Altman and his colleagues, who included ChatGPT VP Nick Turley and COO Brad Lightcap, did also push back on the idea that there were tons and tons of people who were obsessed with their relationship with the older models.

14:13Writes The Verge, Altman pegged the percentage of ChatGPT users who have unhealthy relationships with the product at, quote, way under 1%. The bigger deal, he said, was that there are quote, hundreds of millions of other people who don't have a parasocial relationship with ChatGPT, but did get very used to the fact that it responded to them in a certain way and would validate certain things and would be supportive in certain ways. Which was underscored by the fact that just before the dinner, the company had actually announced that they were not just bringing 4O back, but as they put it, making GPT-5 warmer and friendlier based on the feedback that it felt too formal.

14:45The company writes, changes are subtle, but ChatGPT should feel more approachable now. You'll notice small, genuine touches like good question or great start, not flattery. Internal tests show no rise in sycopency comparing to previous GPT-5 personality. Another lesson, as Altman puts it, the real solution here remains letting users customize ChatGPT's style much more. We're working on that. And yet, for all of this, it was also kind of hard for them to be overly concerned. In fact, for all the people who were complaining loudly, OpenAI saw traffic to its API double within 48 hours. They also said that in terms of every other usage metric, ChatGPT was at all-time highs.

15:21So much so that they're completely out of GPUs, which as we'll see definitely has them thinking about how to improve that situation in the future. Now one really interesting thing which some people tried to tear down was Altman also acknowledging that because of some of their limitations, this really was an innovation in efficiency as well. When discussing GPT-5, Altman said, we have this big GPU crunch. We could go make another giant model. We could go make that and a lot of people would want to use it and we would disappoint them. And so we said, let's make a really smart, really useful model, but also let's try to optimize for inference cost.

15:56And I think we did a great job with that. Now, like I said, as the narrative has shifted over the last few days, you've seen on X a shift from people saying that this was all about cost and making investors happy to actually understanding that efficiency is a vector of progress in its own right. Jackson Atkins posted the ArcGi chart and wrote, GPT-5 was a cosplay, making a somewhat better model at a fraction of the price. GPT-5 is about 10 % better than O3 Pro, but costs 90 % less. Now, I've talked about this a lot on this show, that right now we live in this privileged early adopter period, where mostly people aren't overly concerned about cost for their individual workflows, they're really just focused on getting the most performance possible.

16:38However, as we move away from individual workflows to production grade, cross-the-company workflows, those cost considerations are going to really matter. When we have background agents running all the time, doing work even as we do other work, cost is going to start to really matter. When we start to deploy the Dr. Strange strategy of having 100 agents do the same work at the same time to come up with different results that we ultimately decide which is best, we are really going to care about cost at that point and that amount of token consumption. And so I think people are starting to realize that while yes, it is maybe more fun to get raw performance increases, these efficiency gains are really impactful as well.

17:14Now in addition to making it cheaper, it also is very clearly more efficient at doing smart things. Comparing the number of tokens used in completing Pokemon Red, something it took 0.3 18 ,184 steps to complete took GPT-5 just 6 ,470 steps. Gemini 2.5 Pro took 68 ,000 steps. As Vraser X writes, this isn't just better performance, it's a completely different level of reasoning. Planning memory abstraction, GPT-5 is showing the contours of true general intelligence. In any case, the overall story when it came to GPT-5 was an acknowledgement of some of the mistakes, but also a reminder that actually, in point of fact, the thing is kind of popping off.

17:54Now, while discussions of GPT-5 might have been the most quoted part of the conversation, Another part that got a lot of attention was OpenAI's intended infrastructure build-out. On the one hand, this is nothing new. OpenAI has, of course, been working on things like Project Stargate for a while now, and Sam Altman has been talking to whoever one will listen about how much needs to be spent on data centers. But they really reinforce that they are going to spend a ton of money on data centers in the coming months and years. Said Altman, you should expect OpenAI to spend trillions of dollars on data center construction in the not very distant future.

18:27He also pointed out, you should expect a bunch of economists to wring their hands and say, this is so crazy, it's so reckless, and whatever. And we'll just be like, you know what, let us do our thing. Now, interestingly, and he didn't go very much into detail on this, when people asked how he was going to finance all of that, he didn't say venture capitalists or even Arabian Gulf sovereign wealth funds. He said, I suspect we can design a very interesting new kind of financial instrument for finance and compute that the world has not yet figured out. We're working on it. This is something super interesting to me because it's very clear that the economics of AI don't comfortably fit inside the economics of other types of technology that we've seen in the past.

19:05And to the extent that I'm giving folks who are concerned about bubbles, something we'll talk about in just a minute, any consideration, it is the fact that we're operating in sort of new territory. When we're talking about the type of large-scale compute that's required to get intelligence in the hands of everyone, a new type of utility as we discussed in our show about vibe coding last week, it really is going to require different funding models. And so I'm really interested to see what they have cooking up on that front. Now, speaking of bubble, the other thing that's been quoted far and wide is Altman saying that, yes, we are in an AI bubble.

19:34It was Bloomberg who asked this particular question, and Altman basically said that he saw parallels between the current investment patterns in AI and the dot-com bubble in the late 90s. Bloomberg writes, in both cases, Altman said smart people became overexcited by a new technology. But in each instance, he said that technology was real and poised to eventually have lasting impacts on the business world and society. Said Altman, are we in a phase where investors as a whole are overexcited by AI? In my opinion, yes. Is AI the most important thing to happen in a very long time? My opinion is also yes.

20:06Altman said society as a whole is unlikely to regret the massive investment in AI, but also admitted he thinks some current startup valuations are insane and irrational behavior. He added, someone's going to get burned there. So I don't want to relitigate the bubble conversation in this particular episode. I obviously think that there are huge differences between the dot-com bubble, for example, and AI, notably the fact that AI is already making boatloads of money. If you want to get my take on whether investors are subsidizing the true cost of AI, definitely go listen to the Vibe Coding episode on Friday, because I get deep into that particular topic.

20:39However, when it comes to the people who are breathlessly saying that Altman is calling this all a bubble, I'm sorry, friends, but he is being diplomatic here. He is at the helm of a company that has a current tender offer for employee shares going on right now at a valuation of a half trillion dollars, even though they've been around for like two and a half years. It would be absolutely insane for him to say anything other than exactly what he said here. Can you imagine if he went full hype? People would accuse him of being a huckster, being irresponsible, costing retail investors who inevitably end up holding the bag a bunch of money.

21:10He had to have a diplomatic way of saying that things are maybe over-exuberant while also ultimately revalidating the underlying trajectory. Basically, I think that this particular part of the conversation is sort of a nothing burger except for the fact that it makes for great headlines. Which is not to say that we can't still debate the AI bubble thing. I just don't think Altman is actually overly concerned about this being a bubble in point of fact. Now, part of why I think that is the information we got about OpenAI as a business. And this is something that I think that people are kind of sleeping on in this discussion.

21:42Altman said in no uncertain terms that open AI is profitable on inference. He said if we didn't pay for training, we'd be a very profitable company. And while you got a lot of her, her, yeah, but you got to pay for training type of post on X, a lot of people got how significant this is. It means the unit economics of delivering AI right now are profitable for them outside of advancing to new models. One part of the upshot of this is that it does create more financial power, at least in the short term than it might otherwise seem. Altman said, we will always be training the next thing, but if we needed to run the company profitably and stay ahead, I think we probably could do that.

22:18Now, it's a reasonable question to ask how long they could do that for if they weren't also competing to be the state of the art. But still, I don't think that anyone thought before this interview that they would even be profitable on the inference, so this is a big update. Altman also said that the company will go public at some point and basically needs to keep investing in infrastructure. Altman said, we can spend$300 billion and sell$400 billion in services, and if we don't have the$300 billion in data centers, we just keep disappointing our customers. When it came to going public, he said, I do think we have to go public someday, probably, but he also said he wasn't sure if he was well-suited to be CEO of a public company, joking can you imagine me on earnings calls, and also joking that in a few years, maybe AI would be the CEO.

22:58Next theme, Sam got asked quite a bit about this question of have we hit a wall. The three takeaways from me on this one was that Altman thinks that A, it's still evolving really fast, that B, we're kind of looking in the wrong places for advancement, and C, maybe most significantly, they already have better models. He said, I think the models are still getting better at a rapid rate. One of the things that's interesting is the models have already saturated the chat use case. They're not going to get much better. The Turing test has passed. So basically, as we're looking for advancement, it may not be strictly in what GBD 6 and 7 can do compared to 5 and 4, but more around things like how long we can produce AI videos for, how long and at what level of complexity agents can stay on task.