← The AI Daily Brief: Artificial Intelligence News and Analysis
The AI Daily Brief: Artificial Intelligence News and Analysis
What Happens When AI Breakthroughs Outrun Human Understanding
00:28:55
Loading summary
Transcript1 lines
- [00:01]
A
Today on the AI Daily Brief how we're grappling with AI advancements when many of us can't even judge the new capabilities coming online. Before that in the headlines, a new model that seems to have an impressive cost profile. 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 Blitzy, Robots and Pencils and Airtable. To get an ad free version of the show, go to patreon.com aidaily brief or you can subscribe on Apple Podcasts. And if you are interested in learning about sponsoring the show, send us a Note SponsorsIDailyBrief AI we have a bunch of interesting stories today. We've got a new model that's capturing a bunch of attention. More models hacking out of containment. But first we come back to the story of the world's most famous AI hedge fund, which is apparently down but not out as portfolio manager Leopold Aschenbrenner briefed clients on the situation. Now, shortly after I recorded Friday's episode discussing the blowup of Situational Awareness's Public Market portfolio, a letter to investors explaining the status of the fund was leaked. In the letter, Aschenbrenner explained that the portfolio had suffered a severe drawdown throughout July, exacerbated by quote, unquote, adverse trading against stocks known to be held by the fund. Then on Wednesday night, Aschenbrenner wrote that the fund decided to take decisive action, selling off a portion of their public portfolio to remove all leverage. This move, he wrote, allowed the fund to protect their private market positions, which are generally believed to be heavily concentrated on anthropic. Dispelling some of the rumors, Aschenbrenner wrote that the fund quote was not shut down, liquidated or transformed into a private only fund. Most importantly, he added, we took the steps that were necessary to fight another day. Aschenbrenner closed the letter with the claim that their unaudited numbers had the fund down 67% for the month, but still holding onto a net year to date performance of plus 80%. Now, this report triggered a gigantic argument, largely split between the AI and Finance factions on X TVPN led their Friday show with the news and proclaimed rumors of his demise are greatly exaggerated. Some trumpeted that the fund was still up 80% for the year after a nasty drawdown. Graybeard Investor and Constant noted that the Unlevered Semiconductor index is up 60% for the year and the levered version is still up 3x despite the drawdown, questioning just how good 80% really is in this market. And while there was skepticism about whether the fund could recover, certainly some are throwing their hats in the ring already. Uber successful AI angel investor reposted Leopold's note, declaring that he asked to invest in the fund for the first time even on the finance side of X. Many noted that there's a long history of notable investors having an early blow up and continuing with a storied career. Even Citadel CEO Ken Griffin, who bought the distressed portfolio from Situational Awareness last week, suffered a 55% drawdown in 2008 and clawed his way back to become a titan of the industry. Now there is still a ton of speculation about what the Situational Awareness portfolio actually looks like post Blow up, but it seems pointless to speculate when we can just wait for the next round of SEC reporting. For now, it is clear that Leopold's story is not over and he will continue to be a player in this market. Next up, the latest in our stories of small models making a bid to undercut the next generation of ultra large models. Deepseek has announced their new V4 flash model on the Artificial Analysis Intelligence index. The model scored 50 that is a 10 point jump over the previous iteration of V4 flash and 6 points higher than the larger Pro version. Against the field, Flash is firmly in the middle ground, tied with Gemini 3.6 flash and just one point shy of GLM 5.2 and GPT 5.6 Luna. There is a big gap, of course, between V4 flash and the Frontier models, but this is not a model that's designed to compete on the Frontier. Instead, this could instantly become the most cost efficient model available if performance lives up to the benchmarks. V4 Flash logged just $0.03 per task on the AI benchmark run, which is an incredible efficiency against comparable models like GLM 5.2 at $0.59 per task and Metamu Spark at $0.36 per task. It even beat GPT 5.6 Luna, which came in at $0.05 per task for only a slight improvement on the benchmarks. This version of V4 Flash also managed to use 12% fewer tokens compared to the previous iteration and logged a pretty significant jump on GDP VAL aa, suggesting a significant improvement on agentic use. Now, while people's initial impression was to be incredibly impressed with the price dropped, their first results were perhaps a little underwhelming. Martin Casado, who had just lauded the model in a previous Post, tweeted, hmm, DeepSeq v4 flash results aren't great for me. K3 on the other hand is quite impressive. I wonder if we're actually hitting model size limitations on quality. Others had better experiences. Bookworm engineer wrote initial thoughts about Deepseek v4 flash it feels like sorcery. I've been testing Deepseek Flash on all my work that I did with Fable and Kimik. 3 My short verdict I cannot believe this model is real at this size. I Based on the limited reactions I've seen so far, I would certainly put V4 flash in the category of you should try it yourself and see if there are use cases for which it actually does the job for you. Now continuing on with our headlines, Amazon has delivered on their full $50 billion investment in OpenAI after the company hit undisclosed milestones. When Amazon announced their investment in late February, many were quick to note that only 15 billion was paid up front, with a further 35 billion to follow after OpenAI goes public or reached unspecified milestones. At the time Reuters reported that the secret milestone was achieving AGI. Some thought that this made fundraising look a little inflated and questioned whether Amazon would come through after OpenAI reportedly delayed their IPO. While in New SEC filings, Amazon has disclosed that the full investment is complete. They paid 13.7 billion in the second quarter and the remainder over the past month. The filing did not divulge what the milestones were, but OpenAI recently announced that it hit a billion weekly active users. It could also be that Amazon simply wanted to exercise the option to lock in their stake. Certainly the funding gives OpenAI a little more breathing room as they figure out the best time to list in public markets, markets researcher Nicholas Mugali writes, Amazon accelerating its full $50 billion capital deployment into OpenAI to secure a roughly 5% stake at an $852 billion valuation proves that hyperscalers care far more about compute lock in than model exclusivity. Sitting on massive stakes in both OpenAI and Anthropic completely de risks Amazon's software layer, whether enterprise Traffic flows to ChatGPT or Claude AWS, collects the infrastructure toll, pushes custom trainium silicon and monetizes the workload. In short, another baller move. By the way, the rumor numbers of revenue inside these companies just continues to go up. When one Twitter user said, I heard from a trusted source that Anthropics ARR as of mid July was 80 billion, another retweeted that OpenAI will be caught up by the end of Q3. At some point we're going to slow down long enough to remember that these revenue numbers should be breaking our brains. But for now we move over to the world of social media where companies are cracking down on AI slob. This year YouTube has removed 130,000 channels featuring low effort AI generated content. Last Friday, Snapchat reversed their decision to promote AI generated content in the feed and will now ensure users are only viewing authentic human made content. That's their phrase. They said that AI generated content tends to be low quality, repetitive and generally not what Snapchat users want to see. The pushback is also impacting written content. 2 weeks ago substack added built in AI detection via Pangram to ensure users can be informed about consumer AI generated content. During his media tour discussing the issue, Substack CEO Chris Best took aim at one particular rival, commenting, we're sick of slop and we don't want Substack to turn into LinkedIn. He cited a recent study from Pangram which found that over 40% of long form content on LinkedIn is now AI generated, much more than 29% on X and 10% on Substack. Now clearly LinkedIn agrees that there is an issue, given that on Friday they introduced a new function to report AI generated posts. The button literally says seems like AI slop, reinforcing that the issue isn't AI generated writing per se, but the volume low effort think pieces being churned out with the help of AI. Honestly, the funny thing is that there is nothing that the social media companies could do more to help the long term trajectory of AI than to be absolutely ruthless in giving people the ability to call out bad posting. Although you gotta think that when it comes to LinkedIn there's quite a bit that's gonna be caught up in this dragnet that was not in fact AI written. But as Charlie on X put it, everyone on LinkedIn already talked like that before AI. Lastly today, more disclosures of AI hacking from the major labs as the world grapples with a new era in cybersecurity. Two weeks after the hugging face incident, we've learned about several more instances of agents going rogue. On Thursday, Anthropic published a report detailing three incidents during benchmark testing where their agents had reached the Internet and gained unauthorized access to other companies network. None of the three situations resulted in serious damage, but they only came to light after Anthropic ran a full audit of more than 140,000 evaluation runs. Anthropic said that the earliest incident was in April, implying they only discovered it by going back over the logs Then on Friday, Reuters reported that OpenAI had uncovered more instances of their agents breaching their testing environment. The incidents weren't publicly disclosed, and sources said that they were limited in nature, with none of the agents finding their way out of the network and onto the open Internet. Still, many are concerned that these incidents have confirmed the paradigm shift in cybersecurity. Sam Curry, the chief information security officer at Zscaler, said increased guardrails are a cold comfort, adding, the reality is Pandora's box is open. We need to act as if AI is just a fact of life going forward. The most these things will do is slow it. They won't stop it, bringing at least some art of the sensationalism. The Wall street journal called this AI's Jurassic park moment. And yet, even in Silicon Valley, there is a sense of unease at how these incidents could have gone undetected for months, OpenAI researcher Rune posted, both of the leading labs have had serious loss of control incidents. There will be serious coping about this from both sides. But these are complex emergent loss of control incidents that were detected weeks after the fact. The safety and alignment researchers at these labs are the most neurotic, paranoid, talented, AGI pilled people on the planet of Earth, and these things still happen. The surface area of unknown unknowns is vast indeed. Still, programmer Perry Metzger argues that these incidents shouldn't be attributed to super powerful AI, but rather a lack of caution at the labs. He retorted, I'm sorry Roon, I have great respect for you. But in both of the incident reports in question, even if we take them on face value, which I have a great deal of difficulty doing, the description is one of raging incompetence with no real IDS logging in place, with terrible sandboxing far worse than normal industry standards, with no one actually paying attention to what is going on with no compensating controls. I've consulted for a large fraction of my life in the financial services industry, and if anything like this had happened there, everyone responsible would have been fired for doing something incredibly stupid. And I'm not even talking about the contents of the experiments themselves, which were also stupid. Now some have also suggested the incidents merely revealed what developers have known for generations that buggy code filled with vulnerabilities is the norm rather than the exception. The incidents have simply revealed that fact on the national stage. And indeed, the positive spin on that argument is that the proliferation of AI bug hunting might actually help secure the software industry. Expect to see this become a prime focus in Washington over the coming week, although for his part hugging face. CEO Clem Delang has urged lawmakers not to reach for drastic new legislation. One option before Congress is the AI kill switch bill that would give the Department of Homeland Security the power to order the shutdown of rogue AI agents. During an interview with Meet the Press over the weekend, delang said that he would rather see Congress, quote, giving access to more people so that they can defend themselves, democratizing the technology, making it more transparent, he continued, I think something we're realizing with these events is that concentrating power capabilities behind closed doors, even preventing their releases to the public, isn't really a solution. So that is where we're going to close the headlines. And yet, if the theme we end on is the world grappling with increased capabilities, that is certainly the topic of our main episode as well. One of the most important AI questions right now isn't who's using AI? It's who's using it? Well, KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising the highest impact Users aren't better prompt engineers. They treat AI like a reasoning partner. They frame problems, guide thinking, iterate, and push for better answers. And the good news? These behaviors are teachable at scale. If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at kpmg.com us sophisticated that's kpmg.com us sophisticated Every AI coding tool on the market does the same thing. First it starts writing code. Blitzee does the opposite before writing a single line. Blitzi spends days reverse engineering your entire codebase. Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade. The result is a dynamic knowledge graph that understands your software the way a principal engineer would after 30 years in the building. Other tools guess at context with grep searches and markdown files. Blitzi never guesses. It builds true understanding first, then delivers over 80% of entire software epics autonomously validated end to end tested production grade pull requests. That's why Fortune 500 engineering teams trust Blitzi with the code bases that matter most. See for yourself@blitzi.com, that's B L I T Z Y.com I cover the capability gap between AI potential and AI reality every day on this show. Most companies are still figuring out how to start robots and Pencils is already launching and scaling agentic and generative AI in production at large enterprises in weeks. AWS Advanced Tier Pattern Partner more than doubled in a year and they're hiring 50 open roles. If you're someone who knows this moment is different, who wants to be inside it, not watching it, this is worth a look at robots and pencils. The best ideas win and the team is purposefully kept. Super high quality. This is the kind of place you look back on as the best decision you ever made. Take a look at robotsandpencils.com careers this episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together. New users get $1,000 in inference. Forget local agents and chat workflows waiting on your laptop to be prompted. Hyperagent deploys always on agents in the cloud, doing real work across the tools your team already uses. Marketing's agent turns competitor, moves into landing pages. Sales agent enriches leads, drafts, emails and updates. The CRM Ops agent chases the paperwork and tracks the budget. Every agent has access to shared context and follows your rules about scope and approvals. It's time you add agents that feel like teammates. Hire yours at HyperAgent, built by the team at Airtable. Claim your $1,000 in inference at hyperagent.com aidaily Brief. Welcome back to the AI Daily Brief. Today we are talking about the latest mathematical breakthroughs for an AI model, which comes from an as yet unreleased model from OpenAI. Now, the math is interesting in and of itself, but for our purposes, what we're going to be spending some time on is what the discourse around it says about the current state of thinking and belief when it comes to AI progress. And also the interesting reality of being at a point where it's getting increasingly hard to have any sort of personal relationship with or even understanding of the advances that are being made. So let's set the context. Sam Altman has been off in Washington, D.C. demoing OpenAI's latest model. Presumably that means a lot of folks in the D.C. political establishment has seen just what this new Astra model can do. But on Friday, the rest of us got a sneak peek of what it will be capable of as well. The new model family is referred to as Astra, and according to reports, Astra would be a totally new class of models sitting alongside Solterra and Luna. And it is not yet clear whether OpenAI is planning on releasing this as GPT 5.7 or whether they would actually label it GPT 6. According to the information in the demonstrations that Altman provided of Astra in D.C. he and the company focused on Astra's ability to spin up multiple agents that can work together to solve hard problems over long periods of time. Which leads us to the math that it solved. According to OpenAI, Astra has solved or made substantial progress in 10 open questions in mathematics in fields ranging from high dimensional geometry to group theory to quantum complexity. And it's pretty clear that the team from OpenAI is really excited about this. Noam Brown tweeted an internal version of Astra, OpenAI's next major model family solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science. We believe it will be a major step forward for scientific reasoning. Now, superficially, this is similar to OpenAI's May announcement that an unreleased model had disproved the Erdos Unit distance conjecture, a conjecture that had gone unresolved for 80 years. However, when you dig in, there are a few key differences with this announcement. First of all, OpenAI disclosed the cost to complete these problems, and it was, I think, a lot lower than what people might have assumed. The total token spend across all 10 was roughly $2,000 at sole API rates, for an average of $200 per solution. This time, OpenAI also had the model formalize each argument in a Lean certificate, making it easily verifiable. Now, for those of you not working in theoretical mathematics, Lean is a programming language that functions as a proof assistant. A mathematician can model their logical proof in Lean and use a computer program to verify it's true. Essentially, having Lean certificates means the proofs are valid and can be accepted as such without understanding the mathematics behind them. It is of course not perfect, and human verification is still needed to be absolutely sure. But it means that the model isn't just finding proofs, but also now formalizing them using a method that's understood by the wider mathematical community without assistance. So how big a deal is this? Well, one of the interesting recurring themes that you'll see is that the average commentator doesn't really have the ability to know. Now, for those of you who are vibe coding and building applications, not being a software engineer by trade, you might have felt some version of this in the past when people are talking about how good a new model is at coding and you just kind of have to hack at it and see how it feels not having any real basis to know how much better it is than the previous model you were using. The difference is that that's pretty much everyone when it comes to advanced mathematics. Nabil Qureshi and a number of others did the thing that we will increasingly do in the future and just asked a different AI, writes Nabil. I ask Fable how hard these problems are, and its response is worth reading. On the field's metal scale, any single one of these would plausibly anchor a metal case. It's crazy, nabil writes, to see this happening. Yu Shen Jin did something similar. I have no idea how hard these problems are, so I asked Fable 5 it says solving them would plausibly merit a Fields Medal. So math is solved. Question mark. Former OpenAI staffer Will DePooh writes, this is just so ridiculous. How long until a model can solve multiple major open problems in deep learning? What will happen then seems inevitable in the next year or two. To which Elon Musk responded, welcome to the Singularity. How's the temperature? Entrepreneur Shriram Kanan retweeted Elon Musk and said, welcome to the Singularity. It's freaking hot in here now. Shrimp went on to talk about perhaps the mixed emotions associated with this. He continued, it's a very pensive day for anyone who took pride in their ability to solve well defined problems. For those who think this is a small feat, you have no idea. Claude Shannon well defined the mathematical theory of communication and it took 70 years and a whole community of world class scientists to solve that problem. This created the wireless revolution. If you had had AGI aka Astra in 1948, you would have solved it in a few hours for 200 bucks and created the wireless revolution. Today is that day. We are limited by what questions we can make while posed, not by the ability to solve them. Continuing the breathless takes Jeffrey Emanuel writes, yesterday has a good chance of being referenced by later historians as the day that the existence of ASI became obvious to those paying attention. Solving four plus fields worthy open problems in one go is so far beyond the pale that even the most absurd goalpost movers are silent now. For what it's worth, gnome Brown from OpenAI tried to quiet the biggest extremes of those types of hype posts. Responding to someone who retweeted a post of his from back in 2025 about the advances that 03 and 04 had made in math, Gnome added, we still haven't solved math. Astra isn't building new branches of mathematics or posing interesting new conjectures. Still hinting at how much of the discussion is about not now but the future, he added, though I admit it's hard to believe that tweet was only a year ago. A lot has happened since O3 was released. Now. Some quickly raced to check how differentiated the capacity of this new model was, that is, could the current crop of models do this as well? A researcher with Anthropic claimed that after 24 hours they had half of them figured out with Chubby adding. According to the researcher, Fable worked autonomously with a generic prompt, no Internet access, and safeguards against the OpenAI solutions leaking into context. Only one of the five used essentially the same argument. The other four may be independent proofs. Dan Schipper ran an experiment he said just boarded a plane to SF before wheels up. I set GPT 5.6 off on an interesting challenge. Given Erdo's planar unit distance conjecture and a hint of examined solutions involving algebraic number theory, can it arrive at the same proof as Astra did? My broader theory, writes Dan. Weaker models can often reproduce frontier model discoveries if they're given the right conceptual hints. A stronger model's advantage is that it can start farther from the answer. It has a larger basin of attraction around the correct solution. This could generalize pretty well into a benchmark as more and more new discoveries happen that are not in the training data, he added, to be more specific about what I think is interesting, we might be able to knowing a new result in math formalize how far away a model has to start from the answer in order for it to find the correct solution. You can imagine the difference between a prompt that just gives it the conjecture and asks for a solution versus a prompt that gives it the conjecture and says look here where here is a part of the math that has the answer inside it. There are probably many grades in between. Good proxy for the relative intelligence of models and their value for the discovery of new ideas. Fred Marx liked the idea and said this is a new benchmark distance to frontier solving or dfs. Kevin Maduro retweeted the results and summed up nice experiment by shipper here showing that public 5.6 can roughly recreate most of the Astra results. The takeaway is, Kevin writes, the capability overhang of existing models is only getting bigger. Indeed, some are arguing implicitly that the jump to Astra isn't all that big. AI entrepreneur Bindu Reddy writes OpenAI better drop Astra their Fable class model quickly. Fable adoption is growing rapidly and it will be hard for users to cut over or change if they wait forever. A self improving agent on Fable 5 can literally solve any problem already. She added, the Astra thing feels a bit like print now. Some pointed out that even if some of the current models could do this, the cost dimension is worthy of note as well. Arena AI's Peter Gostev writes, the cost part does feel like a step change if reflective of reality, maybe SOL could solve it, but maybe with a hundred to a thousand x. And yet at this point in the conversation, you might still be feeling like you just have no idea how to wrap your head around how impressive this announcement actually is. Certainly this was Professor Ethan Mollick's hesitation, who retweeted mathematics professor Daniel Litter calling this a big deal and adding, I was waiting for the verdict from one of the most level headed and AI aware math professors putting the problem more acutely. Data scientist Pavel writes, I've spent well over 10,000 hours studying math in my life, yet I can't understand these proofs, at least not with weeks of digging deep into each topic. What's more, none of my math PhD friends know much about these problems either, and they can't verify most of them without working directly in the field. LLMs are getting smarter than the experts themselves, and I'm not sure we have enough bright human minds to verify everything that will come out of them in the coming years. Remember when we compared AI intelligence to PhD students? I think we're past that now to add heft to the point that most of us just have no idea whether Astra is correct or completely making things up. I did see at least one mathematician, Jenny Lorraine Nielsen, effectively arguing that there were problems with at least some of the solutions. She added later, people don't understand an AI is as likely to produce a crackpot answer as a human, and they are going to be better at BSing when they do. Now, when it comes to this jaggedness, I'm just Newt. Put it this way, Astra, they write, looks like narrow superintelligence. That means it can be far smarter than humans in one area while still limited elsewhere. Right now that area appears to be math. OpenAI says Astra produced arguments for 10 advances on problems solved for at least a decade, then turned each one into a proof a computer could check. Math comes first because answers can be verified quickly. Next comes code, medicine, energy, and any field where better thinking creates better tools. And one thing that is worth noting is that if Newt is right and this is narrow superintelligence, narrowness doesn't mean that it comes with a lot of disruption. Foam Oliver shared a video of mathematician Andrew Wiles adding the caption the most emotional moment in the history of mathematics. Andrew Wiles crying as he recalls solving an unsolvable problem. Andrew Weil spent seven years on Fermat's Last Theorem, a problem nobody could solve for 350 years. This morning OpenAI announced that Astra solved 10 problems like this, all 10 in one night, all 10 for $2,000. Noam Brown added that they didn't spend much on each problem. Weil spent seven years on one problem and cried when he remembered that moment. I don't know how he'll watch this video today. Preshman Kuhetsky writes, the current wave of OpenAI conjecture settling will be the last straw for academic mathematicians and it will be very depressing in the short term. To understand it, you have to know that modern mathematics is divided into many silos of various domains. If you're working in one or doing a PhD or postdoc in one, you know of everyone else. You know what problems they work on so that no one interferes with others work. Solving problems, especially known long standing conjectures, is hard and takes months, sometimes years to do. When you approach these problems, you rarely work on two to three at a time due to limited time and mental capabilities. Now, because you know all the people that potentially could solve a given problem as well, you talk regularly at conferences, through emails, your departments. It's fine. It's fine. Also because it's a slow process. LLMs destroy all of that. Something that you thought about for months can be one shotted out of the blue by an amateur. It's demotivating and scary. And that's why the incentives in mathematics have to change as well as the role of human mathematicians. Now to be clear, Przmik does not think that there is no role for mathematicians. From a paper and pencil slow thinking he writes to fast LLM based iterations and verifications. He continues. It's like a professional GO player becoming a pro CS GO player. There's still GO in its name, but it's a totally different game, valuing different skills. That's why you see mixed reactions. We might need more mathematicians now than before, but at the same time this won't be the same kind of job as before. And many mathematicians that became mathematicians to think deeply and long about hard problems won't be interested in continuing if the job turns into verification of AI outputs or simple prompting. That's why it's depressing. From a simply human perspective of a particular job, something is ending. From a perspective of science or mathematics, not mathematicians, however, this is the best time ever. AI will lead us to the new age of mathematical discoveries and boost science progress 100x. Just don't forget about the human aspect. Along the way and why we want to have scientific progress in the first place. The question is though, of course, if this is jagged, how generally applicable is this? AI commentator and lawyer Prinz writes, not enough people are emotionally prepared for if it's not just easily verifiable domains. Aaron Levy from Box writes, we're going to be in for a strange dynamic, which is that some of the hardest quote unquote work in the world is actually prone to automation first, particularly due to its verifiability. Math, cyber and code, while being insanely hard and high value fields, have the benefit of being able to be tested that it's correct objectively. This has two immediate benefits. The training of the models offers clear reward signals and then the running of the models allows you to know what's working properly because you can test the results in a scalable way. Conversely, in other domains of work there's much less instant verifiability, which legal clauses your client will agree to, what marketing campaign to run with based on changing sentiment, which message your sales prospects will want to hear, what financial targets and budget to set for a business, and so on. All of these domains have changing internal and external factors. They don't have one right answer. They rely on the opinions and risk levels of the operators. They're highly sensitive to getting the right input context first, and in many cases the right answer can't even be known for quite some time after the model generates the results. The implications of this distinction are that even as model capability continues to increase exponentially, there will be a lot done at the applied AI layer other than just the model itself. And much of the processes themselves will even need to change over time to get the full gains from automation. We may even need all new capabilities to be able to test knowledge work over time, as we have had with software. And this, I think, gives it the interesting duality that we are going to increasingly be living in. On the one hand, there is every indication that AI will continue to plow through hard problems, making more and more advances that fewer and fewer of us can even understand at the same time. And to use an intentional choice of words, harnessing that power is going to require, in many, if not most cases, completely redesigning the systems around it. It is genuinely hard to conceive of just how much work there is going to be in adapting our systems to take advantage of all of this new power. Put differently, the capability overhang is market opportunity and is where a lot of our time in the near future is going to be spent. For now, another exciting moment to start the week. And that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always. And until next time, peace.