
Loading summary
A
Today on the AI daily brief a week of surprise AI acceleration the AI daily brief is a daily podcast and video about the most important news and discussions in AI. Alright friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG Superintelligent Section and zencoder. To get an ad free version of the show, go to patreon.com aidaily brief or or you can subscribe on Apple Podcasts. If you want to learn more about sponsoring the show, send us a note at sponsorsideailybrief AI. Also at aidailybrief AI you can find out about all the other things we have cooking. One specific one to Note is that Cohort 3 of Enterprise Claw is closing very soon. So if you are interested in that, check that out at EnterpriseClaw AI. Now, this week, once again, we are using this Friday episode to do a bit of a recap. And I think that this week it's particularly relevant in the sense that some weeks the big stories represent so obvious a change that it barely needs to be point out how much has shifted. But this week was instead a week of surprising AI acceleration, where the individual stories add up to a whole that is much more than the sum of the parts and where we can feel almost more than intellectually recognize the acceleration all around us. Now, when I'm discussing acceleration, I'm going to refer to it in a lot of different contexts. There's model development, acceleration, policy acceleration, business redesign acceleration, and more. But where I want to start is with profitability acceleration and the corresponding acceleration in market sensibility. One of the huge stories from this week is that Anthropic expects to have its first ever profitable quarter. And of course, this is not just Anthropic. This is the first ever profitable quarter for any AI lab. Now there are some caveats. First of all, the quarter's not done yet, so this is projections not realized revenue. Second, there are, as we've mentioned before, questions around how Anthropic recognizes revenue, specifically around the idea that they count pure top line revenue before partner distributions, even for established rev share deals. And three they are, as we learned given other information revealed around their partnership with SpaceX, getting access to certain amounts of compute for the next couple of months at a discount. And yet I think for most people those are all relative quibbles with the overall idea, which is a resetting of expectations around just how much money these labs can make. The bubble narrative at the end of last year was all about the idea that we were going to overbuild compute infrastructure as agents started consuming massive amounts of tokens at the beginning of this year. The bubble narrative, to the extent there still was one, was instead about the idea that the big labs were never going to be able to serve these tokens profitably. And so the fact that Anthropic is doing so, even if it's with a few ups from their partners at SpaceX to most honest viewers, has an impact in changing the way that they think about the acceleration of business model development in the AI sector. And by the way, while they're not profitable yet, OpenAI also had a banger of a first quarter. In fact, they generated about a billion dollars more in revenue than anthropic in Q1, although revenue acceleration for Anthropic has since outpaced that. But one of the big boosts for OpenAI's revenue was of course, token hungry Codex. And it wasn't just these startups that were absolutely crushing. Nvidia also had a massive quarter, beating basically every analyst expectation, although at this point we're in such uncharted waters that it feels like the market doesn't quite know what to make of them. Quoting Patrick Moorhead again in the same quote that I used a couple of days ago, but one which I think bears repeating, the challenge with Nvidia at a 5 trillion plus market cap is that investors do not know how to deal with it. If you take Jensen's $1 trillion forward demand pipeline, this is an $8 to $9 trillion company. People are afraid to move even on exceptional earnings. And while I think he's absolutely right about that, I also think that increasingly the market is gearing up for full send mode when it comes to AI now, deeply corresponding with the acceleration of business models and market expectations. We also got an acceleration in the shift away from one pricing paradigm to another. Many folks, including me on this podcast, have framed this as the end of the subsidy era of AI and the beginning of the trade off era, where the introduction of token hungry agents means that flat rate plans that subsidize the most active users are no longer economically viable, even a little bit. One of the interesting small stories from Google I O this week was their quote unquote price cut on their Ultra plan from $250 a month to $200 a month, which actually when you dug into it, came along with a shift to usage based billing for certain types of token hungry use cases. This was not at all dissimilar to what everyone got so worked up about with Anthropic last week. Although in this case Google was pushing usage based pricing even in their owned harnesses like Anti Gravity. Whereas at least for the moment Anthropic is still trying to subsidize you when you're doing things specifically through Claude Code or Claude cowork. And speaking of Anthropic and the shift to usage based pricing, much was made of Microsoft canceling its Claude Code licenses this week, with at least part of the reasoning being a cost consideration. Now there are also reports that this is about creating an incentive to get GitHub Copilot up to snuff as tools like Claude Code and Codex just absolutely eat its lunch. But I don't think that the token cost issue is just a front. As Hedgy Markets put it, token based pricing is forcing every enterprise customer to confront the actual cost of running these models at scale and the number turns out to be far higher than the flat rate experiments suggest. By the way, for those who are still on cloud code and adapting to new usage based paradigms, the Anthropic team did announce that you could now run slash usage to see a breakdown of where your tokens were being consumed, specifically which skills agents, MCPs or plugins were being the biggest token hogs. Now in this new paradigm everyone is shuffling and jockeying for position in narrative space. I discussed already the mixed messaging from Google I owe where I thought based on the entire corpus of communications, that they were really emphasizing the speed of 3.5 flash over the price, given that it was so much more expensive than previous Flash versions and actually pretty comparable prices to the medium levels of these state of the art models like 5.5 and 4.7. But on the main stage they did try to pitch their tools as a way that big enterprises could save billions of dollars, meaning that they are aware of this new token shortage Paradigm, even if 3.5 Flash itself isn't necessarily the answer. And speaking of answers, I also think that part of the acceleration this week is that you're not just seeing a recognition of the subsidy era ending, but companies moving to fill in the gap. Cursor introduced Composer 2.5, a seemingly very high performing coding model, and artificial analysis found that it performed at a very comparable level to Opus 4.7 and GPT5.5, a little ahead of them on their medium settings and a little below them on extra high and max settings, but it did all that at 10 to 60x lower cost. Now if more efficient models are one part of the answer, the other is getting access to more compute and one of the other accelerations we saw this week, without a doubt was Elon Musk fully settling into his shifted role as AI Compute czar? On Wednesday, he tweeted as the recently expanded partnership with anthropic AI demonstrates, SpaceX is offering AI compute as a service at significant scale. We're in discussions with other companies to do the same over time, especially with Orbital data centers. We expect to serve AI at extremely high scale. This was more than just talk. Anthropic's Chief Compute Officer Tom Brown, announced that the company was expanding their partnership with SpaceX, starting to scale up not only on Colossus 1, which we already knew about, but also on the Colossus 2 data center. Not only does this have implications for Anthropic and their ability to serve their big models as well as OpenAI and what they're dealing with from a competitive perspective, but I think this is going to significantly increase interest in the SpaceX IPO. When SpaceX was playing the role of weird amalgamation home for all of Elon's random projects, I tended to think that the IPO was going to be a referendum on Elon more than anything else. Now that they're actually starting to position themselves in this role as almost an alternative NEO cloud, with a unique possibility of actually opening up more compute by moving into space, which is obviously something the other NEO clouds can offer. All right folks, quick pause. Here's the uncomfortable if your enterprise AI strategy is we bought some tools, you don't actually have a strategy. KPMG took the harder route and became their own client zero. They embedded AI and agents across the enterprise how work gets done, how teams collaborate, how decisions move not as a tech initiative, but as a total operating model shift. And here's the real unlock that shift raised the ceiling on what people could do. Humans stayed firmly at the center while AI reduced friction, surfaced insight and accelerated momentum. The outcome was a more capable, more empowered workforce. If you want to understand what that actually looks like in the real world, go to www.kpmg.us AI. That's www.kpmg.us AI. It is a truth universally acknowledged that if your enterprise AI strategy is trying to buy the right AI tools, you don't have an enterprise AI strategy. Turns out that AI adoption is complex. It involves not only use cases, but systems integration, data outcome tracking, people and skills, and governance. My company Superintelligent provides voice agent driven assessments that map your organizational maturity against industry benchmarks against all of these dimensions. If you want to find out more about how that works. Go to Besuper AI and when you fill out the Get Started form, mention maturity maps. Again. That's Besuper AI. Here's a harsh truth. Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized. Half of companies have AI tools, but only 12% use them for business value. Most employees are still using AI to summarize Meeting Notes if you're the one responsible for AI adoption at your company, you need section Section is a platform that helps you manage AI transformation across your entire organization. It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value. The result. You go from rolling out tools to driving measurable AI value. Your employees move from meeting summaries to solving actual business problems and you can prove the roi. Stop guessing. If your AI investment is working, check out section@sectionai.com that's S-E-C-T-I-O-NAI.com so coding agents are basically solved at this point. They're incredible at writing code. But here's the thing nobody talks about. Coding is maybe a quarter of an engineer's actual day. The rest is standups, stakeholder updates, meeting prep, chasing context across six different tools. And it's not just engineers. Sales spends more time assembling proposals than selling. Finance is manually chasing subscription requests. Marketing finds out what shipped two weeks after it merged, ZenCoder just launched ZenFlow work. It takes their orchestration engine, the same one already powering coding agents, and connects it to your daily tools. Jira, Gmail, Google Docs, Linear Calendar Notion. It runs goal driven workflows that actually finish your standup Brief is written before you sit down. Review cycle coming up. It pulls six months of tickets and writes the prep doc. Now you might be thinking, didn't openclaw try to do this? It did, but it has come with a whole host of security and functional issues which can take a huge amount of time to resolve. Zencoder took a different approach. SOC 2 Type 2 certified curated integrations, tighter security perimeter, enterprise grade from day one, model agnostic, and works from Slack or Telegram. Try it at ZenFlow free. Now let's move on to another area where this week represented the recognition of some serious acceleration and this I'll call an acceleration of AI's consumer services. Although I thought that Google's messaging was a little mixed when it came to IO, one thing that was very clear is that AI is everywhere across the Google family of user experiences and people are using it. The Gemini app is now up to 900 million monthly active users, having effectively closed entirely the gap with ChatGPT. And the growth in overall tokens processed each month was even higher, jumping 700% from 480 trillion last May to 3.2 quadrillion this year. And I think one thing that's important to keep track of when it comes to Google is that as much as they're creating unique new consumer services for their AI, they're leaning heavily on the existing distribution and product suite. They have to integrate AI into their existing experiences. Now, this is nothing new in the sense that we've had AI overviews since Google I O 2024. But for the first time, Google search will feature not only AI' information consolidation capability, but also its agentic capacity. Google tweeted, soon you'll be able to create and manage multiple AI agents for your many tasks. Right in search, we're starting with information agents. These agents intelligently look across everything on the web, including blogs, news sites and social posts, plus real time data on finance, shopping and sports to surface updates related to your specific question. Your agent works for you 24, 7, helping you stay on what matters most to you. Then it'll send you an intelligent synthesized update with links to learn more at the right moment to help you take action. I think a way to think about this is that previously people had started to shift some amount of their searching behavior into the chatbots like ChatGPT and Gemini. You could kind of think about searches divided into I'm looking for information versus I'm looking for an answer, and for those I'm looking for an answer queries. Chatbots were kind of a more direct way than having to click through a bunch of blue links. Now this is what AI overviews were supposed to solve for, where they brought that answer capability directly into search. But the information gathering aspect of search still kind of continued the way that it always had. Now what you've got with this agent integration with search is two things. First of all, Google can spin up the right sort of presentation interface for your query without being constrained into a preexisting format. If an interactive mini app or website is the right way to answer your question or give you information, well, now with agents it can do that. But what's more, this is now Google recognizing that another divide in the categories of search is one time versus ongoing or persistent. The example they give in their announcement video is about a person who's looking for an apartment with a particular set of criteria. Previously, when you were using Google search, this would be a one time search. I want to see what apartment inventory is available right now. However, this I think is a good example of where if you don't find the exact right thing in that moment, the search inherently becomes more persistent or ongoing. And so with this new agentic capability, instead of having to ask Google right now what studio apartments are available that meet my criteria, I can ask Google to keep me updated when apartments that meet that criteria become available. And all of a sudden Google search becomes a place for persistent information gathering as well. I actually think in many ways this could be one of the most significant things that Google announced at I O for the reason that the percentage of people who might be likely to stumble into this sort of persistent query need is a lot higher, at least in the short term, than the number of people who are going to go seek out some native AI experience. Now Google is taking the same principle into other areas. They also announced an update to Docs, for example, called Docs Live, where users can dictate prompts and then edit to create full documents using just their voice plus AI, which I think is very much an except acceleration towards a new interaction pattern of voice first and live that I think is going to come to all sorts of different types of user experiences. Okay, so at this point we've talked about business model acceleration, token shortage acceleration and consumer surface acceleration. But we also got this week model capability acceleration, although it is for a model that we don't yet have access to. This week OpenAI announced that they had used AI to make a breakthrough in an 80 year old mathematics problem. The problem was posed by Paul Erdos In 1946, one of the most prominent figures in geometry. The problem asks, if you place n points on a plane, how many pairs of points can be exactly one unit apart? Until now, the prevailing view has been that a square grid arrangement of points was optimal. This week an internal model at OpenAI disproved this conjecture. The solution leveraged multidimensional mathematics flattened down into a 2D plane, producing more pairs than the grid arrangement. Fields medalist Tim Gowers, who wrote the companion paper, said what's significant about this moment is that it's the first really clear example of AI solving not just an unsolved math problem, but a really well known unsolved math problem. Now, a few points really stand out about the way the solution was found. Firstly, there's actually nothing particularly special, at least nothing mathematically special about this model. OpenAI researcher Noam Brown said it was just a general purpose LLM with no specific training for this problem or mathematics. Finding the solution also didn't involve a particularly tricky prompt. The prompt was largely just a clear statement of the problem, with appropriate mathematical terminology. Summing up the feeling of acceleration that came Alongside this is OpenAI's Alexander Way, who wrote:10 months ago, I was ecstatic that AI could win international Math Olympiad gold. Today. That excitement feels quaint. Later in the same thread, Alexander writes, math is a leading indicator of what is to come. Soon, perhaps sooner than we all think, AI will begin autonomously producing landmark results in CS, physics, econ, bio, etc. We should be prepared for a new world where the nature and methods of science will have changed. And indeed, this idea of AI autonomously advancing was also at the heart of the biggest moment of model development acceleration this week, which was of course former OpenAI co founder Andrej Karpathy announcing that he was returning to the playing field, this time with Anthropic. In his announcement post, Andre said that the next few years at the frontier of LLMs will be especially formative, and what he'll be working on at Anthropic is what many call rsi, or recursive self improvement. As Anthropic's Nicholas Joseph put it, Andre will be building a team focused on using Claude to accelerate pre training research itself. Now, of course, to anyone who's been following Andre's work with his auto research experiments, it won't be particularly surprising that this is what he's going to work on now. But the combination of him going back to work with Anthropic, what he's going back to work on, and his argument that the next few years will be quote, especially formative, all contribute to this overall sense of acceleration in a big, big way. Now, one funny joke that came out of the Erdos problem discussion was how much energy it used to solve it. Ethan Malik wrote, using the best public estimates we have of LLM resource use, solving this Erdos problem took between 0.6 and 6.3 kilowatts of electric and about 3 to 31 liters of water. So that is less than three almonds worth of water and the electricity equivalent of 2 to 20 miles of EV. Driving the three almonds worth of water was reflective of a larger accelerated conversation this week, which is about data centers. On the one hand, data centers as a political hot button, and specifically opposition to data centers is accelerating in a major way right now. However, alongside that particular acceleration, we're also seeing an acceleration in the counter narrative taking down some of the bigger critiques that have been floating around and poking holes in some of the most quoted reasons that people are against Data Centers Centers Water use was the big one from this week, with a chart flying around showing how the annual water use of data centers compares to golf courses, almonds and lawns. Specifically the fact less than a fifth of the water used on golf courses, about a tenth of the water used on almonds, and about a twentieth of the water used on lawns each year. This narrative is also finding its way into publications, as are other counter narratives like a CBS affiliate in Richmond, Virginia, profiling an electrician named Josh Price who talks about how, quote, AI has been a huge win for him and his colleagues who are building AI infrastructure. Before the AI building boom, it was hard to find steady work. Now people don't have to leave Virginia for good jobs. This is to me an extremely encouraging acceleration. Not because communities shouldn't be asking hard questions and able to assert agency when it comes to data centers. But obviously I believe that that conversation is going to be better when it's based not generally on a fear of AI and a hatred of big tech, but instead real facts and evidence. Now, on the policy front more broadly, we saw both a bit of an acceleration and a strange back and forth where an acceleration turned into a pause. The acceleration came from California Governor Gavin Newsom, who signed an executive order aimed at preparing workers for potential AI labor disruption. Now, the order is very clearly exploratory. It directs state agencies to work with academics, labor groups and AI companies on the issue, with those groups tasked with developing new policies to deal with it, gathering data, developing the ability to identify early warning signs of labor disruption, as well as directing the state to look at reforming things like severance standards and employment insurance. Now, a lot of folks had very practical questions here. RAMP's lead economist Eric Karazarian writes, when it comes to the idea of a dashboard to track AI's impact on employment across different sectors, how is this supposed to work? This is a hard measurement problem. State unemployment insurance systems cannot identify whether a layoff was related to AI. This could lead to worse policy making regarding AI and jobs, targeting the completely wrong industries and workers. When some sounded a warning on this, others pushed back that this was typical of Gavin Newsom, that it makes his constituency happy by appearing to do things but doesn't ultimately actually do anything about it, meaning that it's mostly all for political show. I think you can take it in one of two ways. One, the non cynical way that holds aside specific issues. This is the governor of one of the biggest and most important states in the country for the first time tackling head on the potential implications of AI disruption, or even if one chooses to take it cynically, given how 100% for sure it is that Gavin Newsom will run to be the Democratic nominee for president in 2028. This is a preview and a test run of the type of language and policy that a major presidential candidate is going to use vis a vis AI. Meanwhile, when it comes to the current president, it appeared for most of the week that we were going to be finally getting an AI executive order, with the information reporting that a signing ceremony had been prepared for Thursday afternoon, with the White House even sending out invitations to leading tech executives. Earlier in the week, a draft had started to circulate, with the draft order instructing the government to partner with the AI industry to design a safety benchmarking standard for advanced AI mod. Under the framework, AI companies would need to submit the models to the government 90 days ahead of public release, although reports are that industry figures pushed back on the lengthy timeframe, arguing for 14 days instead. In addition, the order would have established public private security hardening protocols led for some reason by the Treasury Department. We'll get into that in a minute. The Information reports that throughout the week, White House officials were waffling their word about whether the order would make the review process mandatory. Ultimately, the order was scuttled hours before the signing ceremony was set to take place. Trump told reporters at the White House, I didn't like certain aspects of it, so I postponed it. I think it gets in the way now. Initially it wasn't clear what the cause of the delay was. I saw some reports, for example, that it was just that there wasn't enough time to get the tech CEOs to DC. But later in the day, Politico reported that former Aizar David Sachs personally intervened to block the order. During a conversation with Trump, writes Politico, Sachs told the president that companies were already cooperating and that having the federal government review models before their public release would slow down innovation and harm the US and its AI race. With China, some officials seemed blindsided, with one White House source stating he called POTUS this morning, unbeknownst to anyone, his own staff included, and derailed it. Axios had some additional quotes, with one source saying that the main reason for the delay was that Trump just hates regulation and that Sachs also hated it. The source added, the whole thing was unnecessary and quote, just something Doomers wanted. It certainly doesn't appear to me that David Sachs was able to convince the president on something that he wasn't already inclined towards. He went out of his way numerous times yesterday to connect the issue to China, said Trump. We're leading China, we're leading everybody, and I don't to do anything that gets in the way of that lead. I really thought that could have been a blocker and I just want to make sure it's not. During that same press briefing, Trump was asked whether he discussed AI with President Xi responding, I did. I discussed it. He acknowledges how well we're doing. He's doing well too. It's the two of us. The two countries are fighting for it. Other countries are way behind. They're fighting for it. They want it, but they're way behind. My strong, strong guess is that this is not the end that we will hear around this at this point. Fabled AI executive order but who knows what it'll say, if anything, by the time it actually lands. Like I said at the beginning, in many ways the acceleration this week was less in the news itself and more in the feeling that the implications of the news creates A regular old forthcoming AI model is solving 80 year old math problems, revenue for all of these companies is skyrocketing and Anthropic is defying expectations by having their first profitable quarter. Their leading researchers clearly think we're entering a new phase as someone already worth billions decides they can't stay on the sidelines for this next pivotal part. And everywhere, the jockeying for political narrative and positioning vis a vis AI is getting more and more dramatic. One thing that was really nice about Google I O this week was the willfulness with which they did not participate in the AI doom cycle that I talked about at the beginning of the week. Instead, Google DeepMind CEO Demis Hassabis took the chance of his closing keynote to frame where we are as a beginning, not an end. Talking about the set of releases that they had and what's upcoming, he argued that they would, quote, help unlock AGI's incredible potential for the benefit of the entire world. When we look back at this time, Demis said, I think we will realize that we were standing in the foothills of the Singularity. It will be a profound moment for humanity. This technology will be a force multiplier for human ingenuity and usher in a new golden age of scientific discovery and progress, improving the lives of everyone everywhere. It's a great message and it's our jobs to make sure it's true. For now that's going to do it for this weekly recap on the AI Daily Brief. Appreciate you listening or watching as always. And until next time. Peace. Sam.
Host: Nathaniel Whittemore (NLW)
Podcast: The AI Daily Brief: Artificial Intelligence News and Analysis
Nathaniel Whittemore recaps a pivotal week characterized by widespread, multidimensional acceleration across the AI industry. Rather than a single transformative story, the week’s news is marked by a cumulative sense of intensified progress—ranging from the financial performance of AI labs, to rapid shifts in pricing paradigms, to technical and infrastructural breakthroughs, and significant policy developments. The host calls this an “acceleration phase,” impacting business models, consumer experiences, labor markets, and even scientific discovery.
/usage command to break down where tokens are being consumed, helping customers adapt to the new cost landscape.This week’s episode paints a vivid picture of accelerating change in AI:
Tone is optimistic but clear-eyed: Both the hype and the challenges are accelerating, but so are opportunities for profound impact.
Final message:
“It’s a great message and it’s our jobs to make sure it’s true.” — NLW [59:01]