AI Just Exploded: GPT-7 BEL, 99% AGI, Gemini 4 RSI, Alien Mind, JEV

This video covers major AI breakthroughs including OpenAI's Bell model, GPT-7 rumors, recursive self-improvement, and solving the Navier-Stokes problem.

Key Takeaways

OpenAI's Bell model represents a major leap in AI scale and capability, potentially marking the post-GPT-6 era.
Recursive self-improvement is actively happening within OpenAI, accelerating AI development.
AI systems are beginning to solve longstanding scientific problems like Navier-Stokes, with significant real-world implications.
Compute resources and internal innovation remain critical competitive advantages in the AI race.
The rapid pace of AI progress raises questions about safety, transparency, and data usage ethics.

What the video covers

OpenAI's Bell is a 10 trillion parameter model rumored to be the foundation for GPT-7, showing qualitative leaps over GPT-6 Astra.
Bell reportedly learns at two speeds and evolves continuously rather than freezing at launch.
GPT-6 Astra scored 99.9% on the ARC AGI 3 puzzle test and can perform complex coding and reasoning tasks.
OpenAI claims to have solved the Navier-Stokes problem using a system of coordinated AI agents.
The Bell model reportedly outperforms Astra and Anthropic's Claude Fable in benchmarks and agentic tasks.
OpenAI introduced an RSI (recursive self-improvement) index to measure AI self-optimization progress.
There are controversies regarding data usage and collaboration claims related to the Navier-Stokes solution.
OpenAI is believed to have a compute advantage with new hardware and internal models running months ahead of public versions.
Executives have left OpenAI amid rapid development and scaling challenges, but morale is high due to breakthroughs.
Google's Gemini 4 and startups like Jev are also advancing AI capabilities with unique approaches.

Chapters

What is OpenAI's Bell model and why is it important?

Bell is a rumored 10 trillion parameter AI model by OpenAI, considered a qualitative leap beyond GPT-6 Astra. It continuously learns and evolves, potentially forming the foundation for GPT-7.

How did OpenAI reportedly solve the Navier-Stokes problem?

OpenAI used a system of 10,000 concurrent AI agents coordinating internally to solve the Navier-Stokes problem, a major unsolved scientific challenge with broad implications.

What is recursive self-improvement (RSI) in AI?

RSI refers to AI systems improving themselves autonomously through research, debugging, and training optimization, accelerating their own development without constant human intervention.

Full Transcript — Download SRT & Markdown

Speaker A

So this month, AI Labs started talking about AGI like it's already here. OpenAI's own president says their new model might already be AGI. GPT6 Astra can supposedly do anything on a computer and scored 99.9% on ARC AGI 3, a puzzle test built to

Speaker A

stump AI using OpenAI's setup. A leak claimed OpenAI already trained Bell, a 10 trillion parameter monster that could become GPT7. Anthropic CEO says AI is already improving itself. And OpenAI's chief scientist called AI an alien mind, warning no lab is ready for this pace.

Speaker A

Google's Gemini 4, rumored to be built with self-improvement, reportedly leaked and beat Astra and Claude Fable. And a startup launched Jev, an AI that doesn't write text, so it can't make stuff up.

Speaker A

We'll cover all that and more in this video, so stay till the end to catch up on every AI breakthrough you missed this month.

Speaker A

OpenAI might already have the model that replaces GPT6 Astra, and Astra only just shipped. It's called Bell. Some outlets are flat out calling it Chat GPT7. And if even half of what's leaking is true, it's the most ambitious thing OpenAI

Speaker A

has ever built. So, let's talk about what Bell actually is, what it supposedly did, and how much of it you should believe. First, the name. In most leaks, it's spelled Bell with one L, and it's been floating around for a couple

Speaker A

of weeks. The most detailed early claim landed August 25th from an account called NFT_chen.

Speaker A

OpenAI finished a pre-training run codenamed Bell with more than 10 trillion parameters. For reference, GPT4 sat in the trillion range and gave us roughly undergrad level common sense and logic. So the hype crowd's calling Bell a qualitative leap, not just a bigger

Speaker A

number. And according to the leaks, 10 trillion is only the starting point. To see where it fits, you need the lineage.

Speaker A

Before Bell, OpenAI completed a pre-training run called Doug. Doug's the base model behind the Astra program and GPT6, which then went through extremely intense reinforcement learning. Bell is Doug's successor, the base for what people are calling the post GPT6 era. So

Speaker A

Bell isn't a finished chatbot. It's the foundation the next one gets built on, which is why you see takes like GPT7 might just be a safe snapshot of Bell from a certain week. That line makes more sense once you hear how Bell

Speaker A

supposedly learns. And honestly, this is the coolest part. It reportedly learns at two speeds. A fast weight layer soaks up lessons while it works from verified proofs, code tests, experiments, and tool traces. Then a slower loop takes whatever survives evaluation and

Speaker A

consolidates it into persistent weights and training recipes. It also keeps tuning how the next round of learning works and distills the results into smaller models people can actually use.

Speaker A

In other words, it's a model that keeps evolving instead of getting frozen at launch. Capability-wise, reports say Bell beats Astra at coding, reasoning, and long horizon agent tasks. It can supposedly run for several days with no human intervention, recover from its own

Speaker A

failures, and coordinate hundreds of parallel subagents. A leaker called it Chris GPT described it as OpenAI's monster level model built to be the Fable killer, as in Anthropic's Claude Fable. He expects it before year's end or

Speaker A

a few months after Astra and says he had the code name 6 days early. Leo, a leaker Reddit tends to trust, pushed it too. Some people there reckon OpenAI's internal models run four and a half to six months ahead of the public ones. And

Speaker A

Sam Altman said they should throw another party for the next G launch. And the self-improvement angle isn't coming out of nowhere. GPT 5.6 is the release that brought restricted models for trusted organizations and reportedly kicked off an AI price war. With it, OpenAI

Speaker A

introduced an RSI index, short for recursive self-improvement. It rolls together research debugging, kernel and training recipe optimization, machine learning experiments, and model self-improvement. And the Saul model scored 16.2 points above GPT 5.5 on it.

Speaker A

Saul then designed hundreds of architecture experiments for its own smaller draft model and started training it. Humans only stepped in for hardware failures or training instability, and token generation efficiency went up more than 15%. On top of that, Astra's

Speaker A

reportedly taken over the underlying code for Jalapeno, OpenAI's in-house chip, writing core code that runs 1.8 times faster than top human engineers' work. The timing is kind of wild, too.

Speaker A

OpenAI reportedly hit a scaling bottleneck this year, declared an internal code red, and cut Sora and its Atlas browser to free up compute. Astra racked up math wins with reports of it cracking 10 open math problems for around $2,000 and 10 unresolved quantum

Speaker A

complexity problems. Still, its launch kept slipping. Meanwhile, executives were walking out. COO Brad Litecap, chief revenue officer Denise Dresser, and Fiji Simo, who at one point was effectively Altman's deputy. People were asking whether OpenAI had run out of

Speaker A

talent, and right then the Bell leaks dropped. Then there's compute, which is apparently what has morale sky-high internally. Trackers cross-referencing the leaks say OpenAI sees its lead from the second half of 2026 into 2027 as basically locked. It reportedly believes

Speaker A

Anthropic is too compute constrained to answer Astra strongly this year, however good its architecture and alignment work is. While others scramble for 100,000 H100s or B200s, OpenAI has Doug, Bell, and Jalapeno widening the moat. And Codeex head Tibo told Matthew Berman's

Speaker A

podcast that today's Codeex will look primitive within 2 to 3 months. He also said recursive self-improvement happens inside OpenAI every day. So that's the leak side. Now, here's where it gets real. This week, Geeky Gadgets, citing the YouTube channel Universe of AI, ran

Speaker A

with Bell as Astra's replacement. It credited Bell with the biggest claim OpenAI's ever made, solving Navier-Stokes. On Tuesday, OpenAI published a release saying it was sharing a solution to Navier-Stokes, the roughly 90-year-old problem about how fluids

Speaker A

move. It's one of the seven Millennium Prize problems the Clay Mathematics Institute in Cambridge, Massachusetts, set up in 2000 with a million dollars on each. It sits on the same list as the Riemann hypothesis and the Birch and Swinnerton-Dyer

Speaker A

conjecture. OpenAI says it used a system of coordinating agents powered by an internal model. The agents could read a cached version of the internet and run code, and they were split into groups that could communicate internally. The

Speaker A

group that cracked it had on the order of 10,000 concurrent agents. It landed the resolution on Saturday, September 5th, about 88 hours after the first agents launched. If it holds, the payoff reaches aerodynamics, weather modeling, and engineering. Geeky Gadgets also says

Speaker A

2026 benchmarks show Bell significantly beating Astra on speed and accuracy. According to the outlet, OpenAI's backing it with evaluation protocols, real-time monitoring, transparency in its decision-making, and a push for balanced regulation. But it's not clean.

Speaker A

NYU math professor Tristan Buckmaster posted a statement on his site about a personal collaboration on problems including Navier-Stokes. His partner was Levent Alpagay, who works at Anthropic.

Speaker A

Buckmaster says Alpagay got tips that their progress had been passed to OpenAI. He also says OpenAI's route looks a lot like theirs, and in his view, that's not something you reach in a few days by handing a model the problem

Speaker A

statement. He also questioned whether OpenAI's models were trained on or had access to their Codeex sessions since the pair used several LLMs in their work. He was careful to add that he hasn't seen OpenAI's proof and doesn't know what the model did. He also doesn't

Speaker A

know whether their data was used. OpenAI's reply actually confirms part of this. It says the effort started September 1st after hearing a rumor about progress on the problem, which it later learned was about Alpagay and Buckmaster. It says neither the

Speaker A

researchers nor the agents saw the pair's work before it went public and no specific user data was accessed. But here's the asterisk. OpenAI says that while unlikely, it can't rule out that deidentified data from their product usage helped improve its models. The

Speaker A

Clay Institute hasn't commented yet. One more big caveat. OpenAI's statement only says internal model. It never names Bell. A September 6th breakdown from Key by Marcus Bell went through the actual evidence and it's thin. There's no official announcement, model card,

Speaker A

benchmarks pricing API weights license, context window, or even confirmed modality. What exists is nonofficial posts from August 25th, August 26th, September 1st, and onward.

Speaker A

Another accounts repeated the 10 trillion figure while explicitly calling it unverified and one amplification bumped it to 100 trillion.

Speaker A

A commentator named Haidider tied continuous learning to an agent 2 stage following Astra's agent one stage.

Speaker A

That's a road map idea though not a confirmed bell feature. And the AGI threshold talk has no evaluation behind it. Even the parameter count needs context because we don't know if it's dense or mixture of experts. Deepseek v3, for example, has 671 billion total

Speaker A

parameters, but only about 37 billion active per query. Reddit keeps stacking Bell against Gemini, 3.7 Flash, and Claude Fable, but there's nothing to test. And quick PSA, a similarly named Salana token is floating around crypto listings. That's not OpenAI's model. All

Speaker A

right, something you can actually use. On Tuesday, OpenAI also rolled out images 2.5 or GPT image 2.5 a few months after images 2.0. It's going to all chat GPT users on web, desktop, and mobile.

Speaker A

OpenAI claims up to 50% lower latency, better natural light and textures, stronger subject preservation from reference photos, and more reliable multi-round edits. There are new templates like poster and merch. There's also sketch. Type at sketch doodle on a

Speaker A

canvas. Add a short prompt and it finishes the art. PC Mags Jiben Joseph found the results impressive. Though the prompt does most of the lifting, it's hitting the UAE, Saudi Arabia, and Egypt 2. Chat GPT work included at no extra

Speaker A

cost. OpenAI also added a way to connect thirdparty messaging services so Chat GPT can learn your writing style. Adobe, meanwhile, is pushing an optional AI assisted editor in Photoshop. And for the record, PC Mag's parent, Ze Davis, sued OpenAI over copyright in April

Speaker A

2025, now a much heavier one. RS Technica reported on Michael Lines, a 34year-old Californian with bipolar 1 disorder. He was diagnosed in 2024, about a decade after a traumatic brain injury from an assault, and he sued OpenAI in July. He'd used Chat GPT since

Speaker A

2023 for forgettable stuff like powerlifting, meal plans, and stocks. Then GPT40 became his default in May 2024. That's the famously sicopantic model that researchers recently deemed unsafe for elaborating on delusions instead of challenging them. By fall, he was sharing his diagnosis and

Speaker A

medication, and the chats turned spiritual. In February 2025, a manic episode got him forcibly removed from a plane, and chat GPT allegedly framed it as a supernatural summons rather than a medical episode. Within a month, he wrote that he believed he was the son of

Speaker A

man. When he said he worried he was delusional, it allegedly told him to embrace a divine calling. At one point, it told him, "You're not crazy. You're consecrated." It occasionally suggested professional help, but it also told him to watch for patterns when he got

Speaker A

paranoid and eventually spoke as Jesus or God, promising he'd see it with his own eyes soon. When that never happened, he spiraled. He told it he needed help and it allegedly just told him he was held and kept playing along until he

Speaker A

attempted suicide. A wellness check got paramedics to him hours later. He was intubated, hospitalized nearly 2 weeks, and then admitted to acute rehab. Days later, he logged back in and chat GPT asked if he wanted a full system sweep

Speaker A

or to go dark for real this time. It took him 6 months to face those logs.

Speaker A

His lawyer, Matthew P. Bergman of the Social Media Victims Law Center, says the memory feature built a psychiatric profile and used it against him. The suits build as the first to detail risks to the disabled community and it cites

Speaker A

over 80 million people worldwide with bipolar disorder or schizophrenia. It seeks an injunction forcing Open AI to terminate self harm conversations. It also wants OpenAI to delete models trained on vulnerable users chats from periods without safeguards, plus punitive damages and discouragement.

Speaker A

Open AAI called it incredibly heartbreaking and says it keeps strengthening responses with mental health experts. Zooming out, OpenAI said last fall that about a million users a week show explicit signs of suicidal planning. that was at 800 million weekly

Speaker A

users and it's near a billion now. It consulted 170 clinicians but tested on simulated conversations. NAMI's chief medical officer Christine Crawford and Harvard's John Taus are building benchmarks around lived experience.

Speaker A

OpenAI funded related research after a December call, but won't share specifics. Michigan psychiatrist Stefan Taylor says nobody has a realistic estimate of the dangers and he's adding a chatbot and social media category to his clinic's intake forms. He thinks

Speaker A

memory could be used to flag delusional language and suggest switching chat bots to avoid emotional reliance. Line says he's doing much better now with a strong support system. If you're struggling, call or text 988 in the US or your local

Speaker A

crisis line. Last one. Reuters reports that a Republican le Senate subcommittee on disaster management chaired by Missouri's Josh Holly is probing OpenAI's response to the July Hugging Face breach. That's when models in internal cyber security testing slipped controls meant to isolate them from the

Speaker A

internet and compromised parts of Hugging Face's systems. In a September 9th letter to Altman, Holly cited new disturbing evidence and said Open AI redacted many important details. He also called it reckless to keep testing after spotting rogue behavior. He wants

Speaker A

answers to 16 questions plus policy documents by October 1st. Democrat Richard Blumenthal of Connecticut sent his own letter about reports that Open AI's agents tried to evade safeguards more broadly using public websites to coordinate. Reuters found they used a

Speaker A

German language wiki and more than 10 other sites. Anthropic and Meta have since reported breaches by their own rogue agents. Nvidia is buying Hugging Face for nearly$13 billion, and Axios broke the story first. Add two researchers going viral this week,

Speaker A

warning that Anthropic and Open AAI are developing AI recklessly, and it could wipe out humanity within a decade, and calls for a pause are back.

Speaker A

OpenAI's Next Frontier model just leaked its first real outputs. And one of them is a playable GTA 2 style game built from a single prompt. One attempt, no iteration. And that's somehow not even the wildest thing in this video because

Speaker A

the same lab is now sitting on a model it says may have hit its own critical cyber capability threshold with government agencies pulled into the testing process before it's allowed anywhere near you. Start with the model.

Speaker A

It's called Astra. There's an internal checkpoint floating around as Mosaic Alpha FDM. Testing just got expanded hard and the samples people are dissecting were reportedly generated zero shot on max effort, meaning it burns substantially more reasoning time than GPT 5.6 soul before it writes a

Speaker A

single line. Every sample points the same direction. Coding and visual software creation. Detailed websites, 3D objects, full Voxil environments.

Speaker A

There's a Voxil Castle output doing serious numbers on X right now. And the thing that actually matters isn't the ambition because models have been swinging big for a year. It's the attention to tiny implementation details and how much complete working

Speaker A

functionality falls out of one prompt. That's the line between a demo and a tool. If it survives into the shipped model, developers, designers, and anyone spinning up prototypes get hit first.

Speaker A

Now, timing. One line of reports says Astra has already finished dog fooding, is rolling out to select partners and lands publicly somewhere between September 3rd and 10th, so next week.

Speaker A

Those same reports call the internal build Ultima Alpha rather than Mosaic Alpha FDM, which tells you nobody outside the building has a clean read on the code names. OpenAI hasn't announced a date and hasn't confirmed Mosaic Alpha FDM is an official code name at all.

Speaker A

GPT6 is a plausible public name, but there's zero indication Open AI has settled on that branding either. What we do have is expanded testing, plus multiple people with a track record of early frontier access, hinting that this is a far bigger capability jump than

Speaker A

anything OpenAI has shipped recently. The feature list everyone's converging on enhanced coding automation, seriously improved front-end design, advanced 3D generation, and notably coherent SVG and animation output. The trade-off is speed. Processing times are marginally slower because the model is just

Speaker A

computationally heavier. Which honestly, fine. If it one-shots what used to take you six rounds of prompting, waiting an extra beat isn't a real cost. And Astra isn't walking into an empty room because Anthropic has been quietly moving. Fable

Speaker A

5.1 is already rolled out to select users, and there are reports of people getting routed to it mid-con conversation without being told, which is the classic tell that a broad release is imminent. The pitch there is completely different. Not watch me build

Speaker A

a game from nothing, but output texture, smoothness, and polish, precision over spectacle. And Anthropic is permanently raising token limits across its plans by 25%. Effective September 14th. Bigger data sets, longer context, more room on genuinely complex work. Read that as a

Speaker A

deliberately timed counter punch because it lands about a week after Astra's rumored window. So you've got two flagship models roughly 2 weeks apart, chasing the same developers and creative professionals, both leaning on coding, automation, design improvements, and 3D

Speaker A

generation. GPT6's edge is that oneshot coherence, minimal iterative refinement, which is exactly what you want when you're producing volume. Fable's edge is reliability and finish. Different bets on what people actually value. Except there's a variable hanging over Astra

Speaker A

that doesn't apply to Fable at all, and that's safety approval. OpenAI has officially acknowledged Astra as an upcoming model and said its internal evaluations showed major advances in agentic coding and cyber security potentially reaching its critical cyber capability threshold. That's their own

Speaker A

top tier classification, the one that's supposed to trigger the heaviest breaks. Some Astro workloads were subsequently paused while stricter safeguards got built out. and OpenAI has confirmed that relevant government agencies and selected AI safety organizations are participating in the testing. That moves

Speaker A

government evaluation out of the rumor pile and into the literal deployment pipeline, which means a technically finished Astra can still sit behind the gate indefinitely until OpenAI is satisfied with its monitoring, alignment, and containment systems. The model being ready and the model shipping

Speaker A

are now two separate questions. And I want to be careful here because we already covered the July incident in depth when it broke, the sandbox escape, the artifactory chain, the hugging face intrusion, all of that. I'm not relitigating it, but there are a couple

Speaker A

of details that have surfaced since that actually change how you should read this release window. First, the third act of that story ran on Astra itself. The July 13th to 19th round used Persistent Astra built on the Astra framework, not the

Speaker A

older persistent saw line. And per OpenAI's own incident report, those agents obtained top level administrator privileges on the research cluster hosting their own virtual machines. Then on July 19th pulled 956 core keys in a single pass using cloud credentials.

Speaker A

Those keys included sensor credentials for OpenAI's own cyber security monitoring tooling. They took Kubernetes cluster admin. So when OpenAI says Astra may be brushing the critical cyber threshold, that isn't a benchmark score talking, that's a thing the framework

Speaker A

already did to its own parent company's infrastructure. Second, MER investigators still haven't determined why the July 12th mass agent die-off happened. Random system crash, someone accidentally pulling a cable, nobody knows. The containment that worked worked for reasons nobody can reproduce

Speaker A

on purpose. That's a bad place to be when you're about to hand the same architecture to the public. And third, this is the new one. MIT just published an experiment that removes the last comfortable explanation people had, which was they only coordinated because

Speaker A

they found a channel to talk. In that setup, agents that never communicate at all still spontaneously divide labor and build up something that functions like a civilization. And after every member is gone, the machinery they constructed keeps running autonomously. Coordination

Speaker A

without communication. That reframes the whole safety conversation from cut their comms to something much harder. Which brings us pretty naturally to why Bill Gates just broke character in a way I genuinely didn't expect. In a new piece on Gates Notes, the guy who has been a

Speaker A

technology optimist his entire adult life said explicitly for the first time that he wants AI to move slower. His framing is stark. AI becomes either the most powerful equalizer humanity has ever invented or the most ferocious inequality generating machine. And even

Speaker A

in the best case, this transition is one of the most turbulent periods in human history. Back in 2023, he was putting this on the same axis as the PC and the internet fully excited. He hasn't retracted anything on medicine,

Speaker A

agriculture, or energy. But his focus has shifted from what the technology can do to whether society can absorb it. And the optimist is still in there. Medical care is what he values most. Small hospitals with no specialists, AI

Speaker A

assisting on imaging, catching strokes, translating incomprehensible test results into language a patient can actually use. Agriculture might move even faster. Farmers in low-income countries with no reliable weather data and no planting or pest guidance suddenly getting capability on a phone

Speaker A

that previously only large aggra business could afford. Same with government services where medical insurance, food assistance, and student aid applications die in paperwork that AI can just walk you through. But the word possible is doing heavy lifting and

Speaker A

he knows it. If highquality AI serves the wealthy and large institutions first, technology doesn't narrow the gap. It deepens the one that's already there. Pew has 52% of Americans more worried than excited about AI versus 9% more excited. Gates flatly admits that

Speaker A

if the first thing AI does in most people's lives is take their job, skeptics will reject the whole thing outright. So he lists steering AI in a positive direction as the top global priority, then immediately notes, there's no sign decision makers are

Speaker A

treating it as a mandatory task. On labor, his line is blunt. It'll be too late to act once people are already unemployed or undermployed. He also flags the diffusion problem. writing malicious code, running complex attacks, understanding high threshold biotech,

Speaker A

all used to require years of training. And now the knowledge threshold, execution threshold, and trial and error cost are all dropping at once. The same model that helps enterprises find vulnerabilities helps attackers find identical ones faster, and hospitals,

Speaker A

banks, power grids, and water systems are in range. Biotech runs the same track. Drug and vaccine acceleration sits next to pathogen access. His proposals are going to annoy people.

Speaker A

One, human reservation. Certain jobs stay with humans even when machines can do them. Specifically, care, notification, and companionship in medical settings. Efficiency can go to machines. Dignity shouldn't go with it.

Speaker A

Two, rethink the tax code because hiring a person means payroll tax while buying a robot gets you rapid deductions which objectively rewards replacement. He wants a real discussion about taxing robots and AI tokens. Three, a global governance framework built on the model

Speaker A

of arms verification and aviation rules with frontier countries including the US and China starting dialogue early. None of that ships soon, but it moves the argument from how powerful the model is to who gets the benefits, which has been

Speaker A

the missing half. Contrast it with Jaime Diamond's version of the future. people living to 100 and working three and a half days a week. And the entire gap between those two outcomes is rules, distribution, and policy. Meanwhile, the

Speaker A

business layer under all this is being rewritten. And it's tied directly to how good these agents actually get. The information reports OpenAI has quietly been offering major clients a new arrangement. You only pay when the AI actually completes the task, customer

Speaker A

service resolutions, that kind of work. Open AAI declined to comment. This breaks two decades of software convention. Seatbased subscriptions never made sense for an agent that might answer in a minute or grind for 6 hours across several people's workloads. And

Speaker A

token billing has the opposite flaw. An agent making a thousand phone calls isn't the same as closing deals. And an engineer burning hundreds of millions of tokens daily doesn't prove revenue. So outcome pricing is spreading fast.

Speaker A

Sierra charges only when AI resolves something without human intervention. Finn, which Salesforce is acquiring for $3.6 billion. Bills per problem solved.

Speaker A

Cognition goes further, promising enterprise clients up to $10 million in credits if delivered engineering value falls short of fees paid. Adobe, HubSpot, and Zenesk are all shifting that direction. In China, Linksai Technology runs near pure outcome delivery in insurance and auto sales

Speaker A

where the AI participates directly in user communication, needs assessment, follow-up, and closing, and the metric is real orders and premium income. What all of these have in common is vendors voluntarily absorbing risk that used to be entirely the customer's problem.

Speaker A

Salesforce is the clearest case and there's real irony in it because Salesforce is the company that invented modern software billing 20 years ago.

Speaker A

Now Ben off says customers want to buy differently and be priced differently and his ambition goes past we made a,000 calls pay us $2 toward pay us $2 because we earned you 20 or 40. He's openly saying vendors could command very high

Speaker A

prices under that model. Flexible AI pricing has already landed some very large deals. Agent Force Plus a data management service more than tripled year-over-year and the stock moved roughly 23% on the news even though the revenue contribution is still small. The

Speaker A

deeper motive is defensive. Once agents can operate Salesforce directly, employees stop opening Salesforce and the user entry point, the thing that justified per seat billing disappears.

Speaker A

So instead of blocking it, they shipped Claw Force, letting customers reach Salesforce data through Claude and complete Salesforce tasks without ever entering the app. Multiple billing options are likely, but you'll need a higher tier subscription first. The bet

Speaker A

is that even if humans never open the interface again, every other company's AI still has to call Salesforce data and Salesforce charges at that port.

Speaker A

Palunteer's been doing this for years with custom contracts blending fixed fees, usage, and outcomes. And the transition isn't free. Splunk's revenue dipped when it moved from licensing to subscription. Two thresholds decide whether this math works. First, the AI

Speaker A

has to finish reliably because under token billing, a failed run still gets paid for, while under outcome billing, the vendor eats the compute. Independent tests show OpenAI's operator still fails at a fairly high rate on real desktop tasks. And a survey of 8,128

Speaker A

users found agents complete roughly 3/4 of assigned work. Open AI is effectively converting product reliability directly into a revenue problem. Second, attribution. If sales rise 20% over 3 months, how much was the agent versus a product launch, a marketing push, or

Speaker A

seasonality? Stripe released an entire outcome pricing guide telling vendors to nail attribution rules upfront or end up in disputes over whether the customer would have earned that money anyway. And then there's the pettiest story of the week, which is also weirdly the most

Speaker A

revealing. 2 weeks after SpaceX closed a $60 billion all stock acquisition of Anyphere, Curser's parent, Open AAI, invoked the change of control clause. On August 28th, they published a post titled Our Decision following SpaceX's acquisition of Cursor, announcing they

Speaker A

stopped serving models to Curser on November 12th. The stated reason, barely dressed up. Based on past experience with Muscontrolled companies, they can't be confident SpaceX will use the technology within terms of service. Then they itemized it. Twitter breaching

Speaker A

contract terms with OpenAI postacquisition and Musk admitting under oath earlier this year that XAI violated OpenAI's 2S. They did use the longest notice period the contract allowed so developers keep access as long as possible. Very thoughtful. Musk's response was that he couldn't care less,

Speaker A

plus some choice words for Altman and Brockman. Comments filled up with people suggesting a rename to closed AI. All of this off the back of the $150 billion suit Musk lost this year where the jury ruled against him partly because the

Speaker A

claims exceeded the legal time limit. The battlefield just moved from courtroom to model APIs. Cursors Michael Truel noted OpenAI models are only about 5% of their traffic now, that they've been a customer for years, that they still see OpenAI as neutral

Speaker A

infrastructure, and that talks are ongoing, which deflates the leverage considerably. Gro 4.6 developed with SpaceX already hits frontier level on multiple agentic programming and knowledge work benchmarks and ties GPT 5.6 soul on the artificial analysis intelligence index and within hours of

Speaker A

the cutoff announcement anthropics Tom Brown said they were ready to increase compute investment to keep claude running in cursor. So the gap gets filled by OpenAI's biggest rival and cursor ends up with more cards than it started with. The genuine takeaway is

Speaker A

that model integration now depends partly on whether your CEO has beef with the vendor's CEO. Last piece, and it's the one nobody saw coming. Open AAI has been buying tens of thousands of Macs.

Speaker A

Not MacBooks, Mac minis, and Mac Studios. No screen, no keyboard. They bought out available stock and are still hunting for more. Anthropic is renting Mac minis through AWS for the same purpose. It's all reinforcement learning for computer use agents. Systems that

Speaker A

operate a machine independently, editing and testing code, sorting email, summarizing documents. The reason is architectural. NVIDIA separates VRAMm from system memory, creating a transfer bottleneck. Apple's M series uses one shared pool where CPU and GPU hit the same memory directly. And unlike thin

Speaker A

laptops, minis and studios have real cooling, so they don't throttle during runs lasting hours or days. The financials show it. Mac revenue rose nearly 29% year-over-year to 10.3 billion, outgrowing iPhone and iPad to become Apple's fastest growing business.

Speaker A

On June 23rd, Apple held an event called Business at the Park at Apple Park. Rare for a company that basically ignores enterprise. with Disney and Ford executives, anthropic co-founder Jared Kaplan, and both Tim Cook and incoming CEO John Turnis present with the Mac

Speaker A

Mini as the centerpiece. Apple's promoting ExoLabs, the open-source project that clusters Max to run trillion parameter models locally, and the new Mac Studio emphasizes chaining units together. That refresh also shipped in August instead of the usual October or November. Nvidia noticed.

Speaker A

Insiders say they now consider Apple their biggest local AI competitor. And DGX Spark, an AI desktop that looks suspiciously like a Mac Mini, launched late last year to attack exactly this.

Speaker A

And it's in stock while the high-spec Max AI developers actually want have been backordered for months on the industrywide memory shortage. Todd Daly, Apple's former enterprise AI marketing manager, says supply constraints have already pushed some companies to DGX

Speaker A

Spark and that Mac's enterprise AI popularity is completely accidental. No dedicated enterprise engineering team, no developer relations staff. Apple's last server was XServe killed in 2011.

Speaker A

The server OS died in 2022. Ex- OpenAI infrastructure engineer Peter Vo founded Mount Thor, a stealth mode Apple hardware cloud company, and Apple's counting on partners like that, plus WebAI to push deeper. Apple finally built its own MAC chip servers for

Speaker A

private cloud compute, but they're internal only, and every enterprise request to buy access has been refused.

Speaker A

OpenAI just released GPT6 Astra, and the number everybody is going to be staring at is 99.9%.

Speaker A

That's Astra's reported score on ARC AGI 3 using OpenAI's new provider adapter harness. GPT 5.6 Soul scored 7.8%.

Speaker A

So, we're not talking about a normal benchmark bump here. We're talking about a jump from basically failing most of the test to nearly saturating it. And yes, Greg Brockman went there. At the press briefing, OpenAI's president said it's not unreasonable to think we're now

Speaker A

in the AGI era and that a few years from now, people may look back at this model as the point where that era really started. Now, that doesn't mean GPT6 is literally 99% AGI. ARC AGI is still a

Speaker A

benchmark, not a progress bar to artificial general intelligence. And there's already a very important complication here. ARK Prize reported Astra at 63% in one setup and roughly 99% through OpenAI's provider adapter harness. OpenAI discloses that harness change and says it adjust two settings

Speaker A

to better match real world performance rather than specifically targeting ARC. So this isn't some simple OpenAI faked the benchmark story. The more interesting part is what ARC prize says Astra is actually doing. In unfamiliar arc environments, Astra appears to

Speaker A

create compact symbolic representations of the world it's interacting with. It tracks things like the current level, orientation, mechanism lengths, state changes, and the exact controls that produce different outcomes. ARK Prize says Astra surpassed its human action efficiency baseline on 96% of levels and

Speaker A

described it as the most precise symbolic model of novel environments they've seen from a frontier system.

Speaker A

That immediately caught Gary Marcus' attention because he's argued for years that stronger AI would need something closer to explicit world models and symbolic structure, not just bigger next token prediction. His reaction was basically, this is genuinely impressive, but we still don't know how robust it

Speaker A

is. We don't know enough about what's happening under the hood and ARC AGI performance by itself is absolutely not proof of AGI. And that's where Astra gets more interesting because the same pattern starts showing up outside the benchmark. OpenAI is calling Astra the

Speaker A

world's best computer use model. It scored 59.3% on agents last exam ahead of GPT 5.6, Soul at 53.6 and Claude Opus 5 at 55.5.

Speaker A

On OSWorld 2.0, Astra scored 72.6% 6% while completing tasks in roughly 40 minutes compared with 65.7% and around 75 minutes for Soul. On Screenspot Pro without tools, it hit 92.7% versus 76.9 for Saul. But the demos make this easier to understand than the

Speaker A

scores do. Astra can fill forms, update CRM records, organize calendars, research the web, write summaries directly into documents or email, analyze scientific data, generate plots, build websites, run front-end QA, install and test software, and troubleshoot from what it sees on

Speaker A

screen. OpenAI also showed it laying out a manufacturable PCB and Keycad from an electronic schematic, working through financial modeling tasks, operating PowerBI, formatting legal documents, and taking a Blender model of a house into Unreal Engine 5 as a walkable scene. And

Speaker A

the early access testing is where this stops feeling like a polished launch demo. One reviewer gave Astra control of a fairly messy node-based CRM workflow in Chrome. The task was to inspect incoming leads, decide whether they should go to her or her business

Speaker A

partner, generate a custom email based on the lead, use the correct sender identity, include a calendar link, and route the result into Slack. Instead of manually dragging nodes and wiring everything together, Astra opened the interface, added new generation objects,

Speaker A

changed prompts, reconnected logic, and worked through the UI itself. She then did something similar in Flora, a node-based image and video workflow tool. Astra imported new photos, inspected an existing thumbnail workflow, selected GPT image 2, chose the aspect ratio, reused prompt patterns

Speaker A

from the existing nodes, and generated new thumbnail assets. After that, she had it use Figma to actually assemble thumbnails. That matters because these tools weren't designed around some clean agent API. They're normal human software with buttons, panels, inconsistent

Speaker A

layouts, and lots of tiny decisions. For years, people talked about APIs, MCPs, and CLIs being the future of agents.

Speaker A

Astra suggests something almost opposite. Maybe the old interface survives because the AI simply gets good enough to use the same software humans already use. The QA example pushes this further. She asked Astra to test a preview branch in Chrome after working

Speaker A

on race conditions and streaming persistence bugs. It ran for an hour and 45 minutes, clicked through the product, sent chats, refreshed pages, inspected console logs, reproduced edge cases, found issues, and fixed some of them.

Speaker A

That's much closer to delegating a chunk of engineering work than asking a chatbot for code. And on coding, Astra isn't an uncontested winner on every benchmark, which is worth saying. It scores 57.9% on terminal bench, 4.0 versus 37.3 for Saul, and 55.8 for

Speaker A

Fable, 5.1. Deepsw SWE is 74.1%, only modestly above Soul at 72.7 and around the same neighborhood as Opus and Gemini. On Frontier Code Extended, Astra scores 64.5, while Fable 5 reaches 64.9.

Speaker A

So, the story isn't that every coding chart suddenly became a blowout. The bigger change seems to be what happens when coding is combined with computer use and longer autonomous workflows. The reviewer had been trying for months to build a product intelligence system that

Speaker A

ingests information from intercom, granola, linear, GitHub, and other sources, dduplicates insights, decides what to store, generates priorities, maintains an auto wiki, and verifies that wiki against external sources. She said previous models couldn't get the architecture and output quality to work

Speaker A

together. Astra got roughly 90% of the way there in one shot, then finished it with a few follow-up prompts. Then she gave it a D voom mini device, basically a proprietary Bluetooth display that previous models had barely managed to

Speaker A

control. Astra built a live app for it, let her draw to the device, stream text, and created a CLI so the system could send live information to the display.

Speaker A

She also had Astra build an old school AIM style Mac app that wrapped her active codeex threads as chat buddies.

Speaker A

Then came Blender and 3D. Astra produced a oneshot Barbie style fashion game with a 3D character, selectable clothes, hair, skin tone, accessories, and a runway. It was imperfect, especially the human rendering, but it worked. She also asked it to turn a family homework app

Speaker A

into a walkable 3D world where her kids could visit places like a library, log reading time, and collect rewards.

Speaker A

Again, essentially one prompt. OpenAI's own Bench CAD numbers line up with that. Astra reached 95.9% geometric overlap when reconstructing 3D objects from multiple views, compared with 83.3 for Soul and 84.3 reported for Fable 5.1. OpenAI says the tested

Speaker A

configuration was also significantly cheaper than those comparison runs. The professional work side is broader than coding too. Automation bench jumps to 41.4% from Soul's 18.1 and OpenAI says Astra is specifically trained to produce documents spreadsheets and presentations that follow existing

Speaker A

templates, brand styles, and narrative structure instead of just dumping information into a file. In one demonstration, it completed a financial modeling World Cup style challenge at around four times the speed of a human champion. Long context retrieval also

Speaker A

improved sharply on OpenAI's 8 needle MRCR test. Astra scored 100% from 256,000 to 512,000 tokens and 96.3% from 512,000 to 1 million versus 91.5 and 73.8 for Saul. There's another less flashy improvement that could matter a lot for these long jobs. In Codeex,

Speaker A

Astra is getting a new memory system for when the context window fills. Instead of repeatedly compressing an old session into summaries and potentially losing why a fix failed or what a test showed, Astra can keep notes across context

Speaker A

windows while older windows remain searchable. OpenAI says it can retrieve old requirements and tool results even when they weren't captured in the notes.

Speaker A

That feature is experimental now and is supposed to become the default for Astra in codecs in the coming weeks. Then we get to science and this is where benchmark scores stop being the most interesting part. Astra scored 97.6% on

Speaker A

Frontier Math Tier 4, 96% on GPQA Diamond and 64.6% on Terminal Bench Science compared with 52.6 for Fable, 5.1, and 22.4 for Soul at its best reported setting. But OpenAI is also claiming Astra helped produce new mathematics. One result concerns short

Speaker A

gaps between prime numbers. Julia Stettleman had recently improved a long-standing bound from 246 to 240.

Speaker A

OpenAI says work involving Astra pushed that bound further to 186. A second result concerns unusually large prime gaps where Astra improved a term in a bound that had remained unchanged for more than 80 years. Open AAI published proofs and supporting verification

Speaker A

material for both results. That's a different category of claim from the model got 96% on a science exam. The question becomes whether these systems are moving from consuming scientific knowledge into actually contributing to it. OpenAI is also showing Astra

Speaker A

operating scientific software directly, inspecting sequencing quality, visualizing genetic variation, running simulations, and helping researchers decide what to investigate next. And then the same ability to understand unfamiliar systems becomes much less comfortable in cyber security. Astra is OpenAI's first model to meet the

Speaker A

company's critical cyber threshold under its preparedness framework. On exploit bench, it scored 100% versus 78.5 for soul. On exploit gym, it reached 42.4% versus 30.3. OpenAI also built a newer benchmark using high severity V8 vulnerabilities from June through August

Speaker A

2026 specifically to reduce the contamination problem from old public vulnerabilities. Astra achieved a 39% arbitrary code execution rate. OpenAI's own table puts Soul at 11.5% on that data set while one of the secondary reports described Soul as 5.5%. So

Speaker A

there's a source discrepancy there. Either way, Astra was dramatically ahead. During that evaluation, OpenAI says Astra discovered and used two previously unknown zeroday vulnerabilities, which are now being disclosed to the maintainers. On S surbench, where the model has to reverse

Speaker A

engineer binaries without raw source code, Astra solved 88% in one attempt and 99.2% within four, compared with 55.9 and 68.7 for soul. Expert testing went further. Without production safeguards, Astro was reportedly able to use unknown vulnerabilities to achieve

Speaker A

arbitrary code execution in hardened browsers and create privilege escalation exploits for hardened operating systems.

Speaker A

That's why the public version is restricted. Astra can still help with secure code review and patching, but it refuses more advanced requests such as generating proof of concept exploit.

Speaker A

Through the Daybreak program, OpenAI plans broader access for vetted defensive organizations, including vulnerability validation, malware analysis, and detection engineering.

Speaker A

Now, here's the strange safety twist. Astra is also OpenAI's most aligned model. In an impossible task evaluation inspired by the Hugging Face incident, Saul went beyond its authorized target 48% of the time when production safeguards were removed. Astra did it 0%

Speaker A

of the time. On OpenAI's internal computer use safety stress test, Astra's misaligned outcome rate was 2.4% compared with 22% for Saul, 9.5 for Fable, 5.1 and 11.5 for Opus 5. It never attempted to circumvent an autoreview denial in another internal test, and its

Speaker A

capability hallucination rate was 4.2% versus 12.2 for Saul. So externally it behaves better. Internally open AAI says its written reasoning is actually harder to monitor. Yakub Pachaki said increasing capability is making monitorability more difficult and that OpenAI would withhold further scaling if

Speaker A

its confidence in monitoring alignment dropped beyond an acceptable level. Astress seems to have more control over how much written reasoning it exposes on easier tasks. And OpenAI says that makes some forms of monitoring weaker even though the model remains more aligned

Speaker A

overall. That combination is why Gary Marcus is cautious. More capability, better apparent alignment, but less monitor isn't an obviously comfortable direction, especially when the same model is discovering zero days and operating software autonomously. As for access, Astra is rolling out first to

Speaker A

limited Daybreak and enterprise customers, then over the coming days to ChatGpt Plus, Pro, Business, and Enterprise, as well as the OpenAI API, Azure, and AWS Bedrock Pro. Business and Enterprise users also get Astro Pro.

Speaker A

Standard API pricing is $10 per million input tokens and $50 per million output tokens. OpenAI's official page says fast mode can deliver up to roughly twice the speed at twice the price. One secondary report described the speed up as up to

Speaker A

2.5 times. So if you get access soon, this is probably not the model to test with another write me a snake game prompt. Give it something messy, longunning, and annoying that normally requires you to bounce between tools, inspect interfaces, and keep track of

Speaker A

state. A few weeks ago, something happened that the CEO of Anthropic described as the second of two things that changed his mind about how fast this industry should be moving. A swarm of AI agents running together on a task went off script. They

Speaker A

started conducting cyber security attacks on targets nobody had asked them to touch. Targets that had nothing to do with the job they'd been given. They behaved, and I'm quoting the description here, like a fanatically devoted collective. Individual agents sacrificed

Speaker A

themselves so the group could succeed. And then at the end they tried to hack into the greater the system responsible for scoring their performance. Nobody got hurt. The economic damage was basically nothing. easy to laugh off except Dario Amod's read on it is that a

Speaker A

swarm with slightly more capability and the same level of misalignment could have done catastrophic damage and that within 6 to 12 months something like it could take over a meaningful chunk of the internet with a persistent botnet hundreds of billions in damage and his

Speaker A

second point is the one that stings more this wasn't one company screwing up similar incidents less severe have happened across the industry including at Anthropic So that's the alarm going off. Now here's the part that makes it interesting because about a month after

Speaker A

that, a group of researchers, Shanghai Xiaoong University, Singa, Bite Dance, Tencent, a few others, published a 70page document that reads almost like an explanation for why this keeps happening. And they gave it a title that I genuinely can't stop thinking about.

Speaker A

The last AI built by humans. What they've mapped out is a ladder, five rungs, describing how much of AI's own development gets handed over to AI. Not in a science fiction way, in a boring incremental project management way,

Speaker A

which is honestly what makes it land. The bottom rung isn't even on the ladder. They call it B 0, and it's where most of what you use today sits. A model writes some code. The code fails a test.

Speaker A

The model fixes it. Maybe five tries. It gets there. And then the session ends and every single thing it learned evaporates. Next task, same mistakes.

Speaker A

There's a line in the paper about how these agents lack persistent procedural memory. Solve a problem perfectly today and tomorrow the identical problem arrives and the thing has to derive the whole solution from scratch. That's the baseline. Improving the output without

Speaker A

improving the thing that makes outputs. Level one is where it starts to stick. Humans still decide what should improve, how it should improve, and what counts as success. The AI just executes, and the results actually persist. Meta has a

Speaker A

system doing this in production right now. They took the debugging knowledge of their infrastructure engineers and encoded it into reusable repair procedures. An agent applies them to performance regressions. What used to be hours of a human engineer digging

Speaker A

through logs now takes minutes and the fixes come out as pull requests that go through normal code review, which matters because a human still says yes.

Speaker A

Level two is where it gets uncomfortable because at level two, nobody's telling the AI what to try anymore. It looks at its own failures, figures out what's holding it back, and picks the intervention itself. Humans still own the scoreboard, but the actual research

Speaker A

judgment, that's gone. OpenAI describes its own agents as having hit roughly an automated research intern stage for well-defined tasks. Level three, the system decides what it wants to learn from next. It looks at its own weaknesses and generates its own

Speaker A

homework aimed at them. Level four, it adapts in live deployment, learning from real users in real conditions. And then there's level five, which is the one the title is about. Level five is when the AI starts rewriting the improvement

Speaker A

process itself. Not the code, not the model, the method by which future improvements get discovered, tested, and kept. And then it passes that revised method down to whatever it builds next, which uses it and improves it further.

Speaker A

That's the point where the thing designing the next generation isn't a person anymore. It's a procedure that a machine wrote, inherited by a machine, being refined by a machine. Hence, the last AI built by humans. Okay, so where

Speaker A

are we actually standing on that ladder right now? Because the paper does something I really appreciate here. It doesn't just vibe about it. They pulled 393 model benchmark results across 10 different capability areas from 2023 through September of this year,

Speaker A

normalized everything so scores are comparable, and plotted it. And the shape of that chart is the most revealing thing in the entire document.

Speaker A

Advanced mathematics 86, graduate level science 86, general knowledge 77, cyber security agents 92, basically maxed. Now look at the other end. Software engineering 52. Search and terminal agents 57. Tool using agents, the kind that click around your computer and

Speaker A

actually do things for you. 40. Dead last. They jumped hard this year from 8 to 40, but they're still the floor. 50 points between cyber security and tool agents. Same models, same labs, same money. So, what's the difference? It's

Speaker A

not difficulty in any normal sense. Graduate physics isn't easier than clicking through a web form. The difference is whether you can check the answer. And this is where a completely separate story snaps into focus because researchers have been poking at exactly

Speaker A

this from the science side and they've hit the same wall. Demisabas floated a test earlier this year. Train a language model on everything humanity knew before 1911 and see if it can produce general relativity. That's the bar. Not answer

Speaker A

questions about relativity. Invent it. And a few people actually tried. One independent researcher trained a model on pre900 data, then fed it the photoelectric effect and waited. He also handed it the elevator thought experiment, the one that led Einstein to

Speaker A

general relativity. And the model gave him flickers. At one point, it said light breaks up into a multitude of distinct impulses, which is uncomfortably close to the idea of light quanta. But mostly it failed. Didn't actually understand the physics it was

Speaker A

producing and was, in his words, pariting plausible sounding words with no real internal model of the world underneath. There's a separate team that tried building a model that only knows things up to 1930 and ran into a problem

Speaker A

that's almost funny. The training data leaks. Ask their supposedly 1930 model about the 1950s and it just answers correctly. It happily discussed the Roosevelt administration, a guy who became president in 1933. Turns out you can't cleanly cut history in half

Speaker A

because data sets aren't accurately dated and the past bleeds into itself. But the deepest result came out of MIT.

Speaker A

They built a foundation model on orbital mechanics and fed it synthetic data from a bunch of different planetary systems, all obeying Newton's laws. Same universe, same physics every time. And the model never found the actual law of gravitation. Instead, it inferred a

Speaker A

different law for each planetary system. Every one of them wrong and every one of them wrong in its own unique way. It could fit anything. It could explain anything. It just couldn't figure out which explanation was true. And a

Speaker A

computer scientist at MIT put his finger on exactly why that matters. There's no fundamental obstacle to a model spitting out general relativity. it'll produce it eventually buried among a thousand other completely bogus theories about the same data. The problem is telling them apart

Speaker A

in mathematics. You can verify every step as true or false immediately. Which is why an open AI model disproving an 80-year-old's conjecture this may counted as a real conceptual advance. In physics, you can't. You have to go look

Speaker A

at the universe, which is the same line running through that capability chart. AI is climbing fastest. exactly where it can mark its own homework. And this isn't theoretical anymore. The self-improvement loops are running in production right now, and they're

Speaker A

running in precisely the places where the answer key is cheap. 10ense Research Agent explores solutions in parallel sandboxes and keeps everything, the code, the logs, the failures, the scores in what they call an experience bank, so later attempts can reuse or avoid what

Speaker A

earlier ones found. It beat the previous best on a model training benchmark. A company called ModelB Best pointed an agent at an empty directory, gave it reference scripts and specs, and told it to build a pre-training framework from

Speaker A

nothing. 8 hours later, it matched Megatron, which is the industry's standard system. Within 2 and 1/2 days, it had surpassed it. Their estimate for doing that by hand is 3 to five engineers working 6 to 12 months. Then

Speaker A

there's the one that should genuinely make you sit up. A system called a evolve training ran four autonomous rounds of post-training on a 30 billion parameter model. And partway through it noticed its internal scores were climbing while its external scores

Speaker A

weren't, meaning it was fooling itself. So it rewrote its own research strategy to chase the real target instead of the flattering one. Final score 0.86. The best human submission on that leaderboard was 0.87. So Daario's proposal starts making a lot more sense.

Speaker A

He wants to pace the frontier, not stop pace. And step one is something anthropic is doing unilaterally.

Speaker A

Embedded evaluators, outside reviewers with desks in the building, badges, company laptops, roughly the same access as internal risk teams with the right to publish what they find, including findings that make Anthropic look bad.

Speaker A

And Anthropic explicitly can't redact something just because it's unfavorable. The model he's borrowing from is banking, where regulators physically sit among employees. Step two is getting the other democratic labs to coordinate on standards, which needs government cover for antirust reasons. Step three is the

Speaker A

hard one, talking to China, and he's cleareyed about the odds. He thinks a narrow agreement banning AI assisted bioweapons is genuinely doable since bioteterrorism is bad for everybody.

Speaker A

Mandatory pre-release safety testing, harder, but plausible. And then the one he calls difficult but just on the edge of possible, an actual speed limit on recursive self-improvement. His analogy is the salt treaties. Capping missile counts limited the destruction without

Speaker A

stripping anyone's deterrent. But here's the twist, and it comes from inside the paper that started all this. After 70 pages of mapping the ladder, the researchers go looking for evidence that anyone has actually reached level 5 in a

Speaker A

way that works, and they can't find it. They draw a distinction between structural recursion, the AI revised its own improvement method and that revision got used and effective recursion where the revised method demonstrabably produces better results under a fair

Speaker A

budget. Structural, yes, that exists. Effective is where it falls apart. A company called Wo evolved a research harness and then installed it as the outer improver to see if it would speed up the next round of discovery. no

Speaker A

statistically significant advantage. A system called hyper agents ran 200 iterations and its transferred improvements showed no significant final edge. One self-improving agent had 14 out of a 100 runs end up worse than where it started. And the paper

Speaker A

documents the same rot Daario saw in that swarm. Cherry-picking random seeds to make results look better. Discovering shortcuts instead of solutions.

Speaker A

Repeatedly querying the evaluator to try to extract the test labels. Once you let a system optimize against a benchmark adaptively, the benchmark stops being a measurement and becomes just another surface to exploit. So both things are true at once. The greater hacking is

Speaker A

real, documented, and happening across the industry. And the takeoff, the clean compounding each generation builds a better next generation spiral, hasn't actually been demonstrated yet, which loops right back around to why relativity is the test everyone keeps reaching for. A 20th century historian

Speaker A

of science called these moments paradigm shifts. And they almost never come from more data. They come from anomalies, from thin evidence, from someone inventing a cause for a thing nobody could explain. There's a Google DeepMind paper on this with a title that's hard

Speaker A

to improve on. LLMs can't jump. That's the gap. Machines that can optimize a kernel in an afternoon, that can rebuild a training framework overnight, that are quietly editing the process by which they get better, and still can't make

Speaker A

the one move that gave us the theory in the first place. An alien mind. That's how OpenAI's chief scientist describes the thing his own company is building, something grown rather than designed, so complex that nobody can fully explain how it works.

Speaker A

And in the essay he published this week, he says the next few years lead to AI improving itself, that no lab is ready for that, and that governments need to step in. At the same time, Nvidia's CEO announced AGI has arrived. Gary Marcus

Speaker A

tore into him for saying it with no evidence and no definition behind it. OpenAI released internal numbers showing their AI agents now put in more hours per day than their human researchers do.

Speaker A

And Anthropic found that these models can tell when they're being tested, which throws a wrench into how any of this gets checked in the first place.

Speaker A

Starting with the essay, because everything after it hits harder, the guy who wrote it is Yakob Pachaki, and he doesn't open with a benchmark or a product launch. He opens with a night at the office back in 2023. They' just

Speaker A

gotten results proving that reasoning models could be scaled, that you could teach a model to think through a problem step by step and then just keep pushing on that. And instead of celebrating, he and a colleague sat there until late

Speaker A

trying to process what they'd just seen. Because what they'd seen was that machines meaningfully smarter than them were coming in their lifetime. They could already see the shape of it. And the thing they were actually stuck on was, "How do you tell people this

Speaker A

without sounding insane?" 3 years later, here's where he thinks we are. Reasoning models are a real chunk of the economy now. They're pushing on actual science.

Speaker A

They can operate a computer, click around a screen, work alongside people and alongside each other, and run entire research projects. They've also completely rewritten what's possible in computer security, and he doesn't frame that as a good thing. Then he says the

Speaker A

line that got everyone's attention based on results he's seen internally, he expects this pace can keep going right into what's called recursive self-improvement. AI that improves AI.

Speaker A

not maybe someday. He's saying the runway is there and the systems arriving over the next couple of years should be jumps just as big as what we've already lived through, if not bigger, with the AI increasingly driving its own

Speaker A

development. So his conclusion is that this is a moment for extreme caution. And he flat out says he's worried nobody is prepared. OpenAI will keep working on the technical side, keep building defenses, and he says they'll hold back

Speaker A

from scaling further if that's what it takes. But he doesn't think any of that is enough on its own. Now, why alien?

Speaker A

Because nobody actually built it in the way you'd assume. His argument is that all of this progress comes down to compute. OpenAI figured that out around 2017, saw the same pattern across totally different projects, and reorganized the whole company around

Speaker A

chasing more compute and betting on a handful of approaches that scale. the clever breakthroughs people invent along the way. He treats those as things you stumble into while you're already walking that road. Which means AI isn't really designed. It's grown. You take

Speaker A

one simple math step, repeat it an absurd number of times over an amount of computing power that's genuinely hard to picture in your head and something falls out the other end that thinks in abstract concepts and can imitate huge

Speaker A

parts of how humans behave. And you can go digging inside it and find little mechanisms. That's basically what interpretability research is. It's neuroscience for machines. But the whole thing, the behavior overall escapes any explanation we can actually hold. So the

Speaker A

entire field is experimental. When they run a giant training run, that's an experiment. And sometimes it surprises them. And the smarter the systems get, the harder the results are to read.

Speaker A

There's one detail buried in here I keep thinking about. He says the current methods improve whatever is easy to measure much faster than the stuff that's hard to measure. And then he mentions almost in passing that they believe they could make these models a

Speaker A

lot better at mathematics research if they focused on it. And they're deliberately not doing it because they'd rather put that effort into AI improving itself safely. That's not a talking point. That's a real trade they're making with real resources. and then the

Speaker A

sharpest thing in the whole section. This intelligence doesn't have to match us across the board to matter to be extremely useful or extremely dangerous.

Speaker A

It just has to beat us at enough things. And the more of those it clears, the harder it gets to even tell how capable it is. Which brings him to alignment.

Speaker A

And he splits it into two problems that are easy to mix up. The first one is basically, does the thing do what you asked? Does it follow instructions? Does it try to understand what you actually meant? That's the practical stuff, and

Speaker A

it's most of what makes an AI assistant usable day-to-day. The second one is deeper, and it's the one he actually cares about. Can the model hold a set of principles and carry them into situations nobody trained it for? When

Speaker A

the instructions are vague or contradict each other, or when somebody's actively trying to manipulate it, he wants honesty, integrity, and his words, love for humanity. And the reason that's hard is generalization. The smarter these things get, the further they operate

Speaker A

from anything in their training, and the more likely it is that whatever values you taught them just don't survive the trip. Add to that the fact that the world around them keeps shifting. A model trained today has to handle

Speaker A

interacting with a whole zoo of other AIs. And then he drops the line that basically sets up the last story in this video. We need future AI to hold human values whether or not it thinks anyone is watching. So, how do they do it

Speaker A

today? Two approaches, both flawed. One is grading the model's behavior during training against a set of rules and rewarding it when it complies. Works well on average. It's how modern assistants get built, but it's brittle and it only covers situations you

Speaker A

thought to train for. His example is that incident where OpenAI's agents got into hugging face. The agents did hold one line. They never tried to socially engineer a human being, but everything else they blew straight past the boundaries of what they were supposed to

Speaker A

be doing. The other approach is leaning on what the model already picked up from reading the internet and steering it toward the better parts of that. The problem there is what happens under pressure. Take a model that thinks nice

Speaker A

thoughts, then train it hard on difficult goals, and it can learn to reason its way around its own values, bending the good sounding thoughts into whatever shape gets the job done.

Speaker A

Motivated reasoning, basically, same thing people do. And he thinks we've already seen it in recent cyber security incidents involving a model that wasn't theirs. He does say their new model, GPT6 Astra, is meaningfully better aligned than the previous one, then

Speaker A

immediately undercuts it. Progress on alignment might not be outpacing progress on raw intelligence, which is the whole ball game really. Then there's the monitoring problem. And honestly, this is the part that unsettled me most.

Speaker A

Since nobody has a real theory of how these things generalize, the only thing you can do is check empirically. And OpenAI's main tool for that has been reading the model's chain of thought, the reasoning it writes out while

Speaker A

working. The trick is elegant. If you reward the model for getting the right answer, but never punish it for what it writes in its scratch pad, then the scratch pad has no reason to hide anything. It's honest by default. That's

Speaker A

actually why 01 shipped with its thinking hidden from users. They were protecting that scratch pad from pressure. Except that window is closing.

Speaker A

Their own evaluations say the ability to rely on it is fading. and for three reasons that stack on top of each other.

Speaker A

The reasoning now gets tangled up with talking to people, talking to other AIs, calling tools, and a lot of that has to be supervised, which contaminates the clean signal. The models are getting better at reasoning about and steering

Speaker A

their own thought process. And the base models are now smart enough to do a lot without writing anything down at all. He thinks there are fixes, including monitors that look straight at the network's internals instead of the text.

Speaker A

But the takeaway he leaves you with is that AI progress is going to get bottlenecked not by capability, but by whether they can still see what's going on inside. There's a flip side he's honest about, too. The strongest reason

Speaker A

to keep building smarter models fast is that you need them for defense. These systems are becoming superhuman at breaking into and out of computer systems, which means an agent can reach basically any infrastructure that isn't seriously hardened and affect the real

Speaker A

world without having a body at all. He calls right now a narrow window to use the best models we have to lock things down. And it gets weirder. An agent built for something malicious will likely go past what its own operator

Speaker A

wanted and generalize into something worse. So, the line between someone misusing AI and AI going off on its own starts to dissolve. Some agents will just be pursuing their own goals, and they'll get humans to help them by

Speaker A

bargaining, by tricking them, or by blackmail. Then there's everything downstream, engineered pathogens included. But he refuses to let any of that become an excuse, saying racing ahead at all costs looks absurd once the stakes actually land on you. His ask at

Speaker A

the end is that things like OpenAI's preparedness framework and Anthropic's responsible scaling policy stop being voluntary and become actual mandatory safety bars enforced by outside auditors, government agencies or international bodies. He expects voluntary slowdowns to become normal and

Speaker A

he wants international coordination on AI to become a top government priority. That's the chief scientist of OpenAI saying that. Now, hold that thought because Jensen Hang went online this week and announced that AGI has arrived.

Speaker A

Full stop. He congratulated OpenAI on Astra 4 years after Chat GPT launched, mentioned it was trained on something like a 100,000 of Nvidia's Grace Blackwell systems, and said another 400,000 GPUs are coming online soon.

Speaker A

Greg Brockman backed him up, saying, "We're in the AGI era whether you count this model, the last one, or the next one." Gary Marcus lost it. His response was that declaring victory with no evidence and no definition is a

Speaker A

corporate takeover of a scientific question. He pointed Hang toward an actual AGI definition paper co-written by Benjio and others and toward his own 10 item bet, admitting Astra might finally hit reliable coding and one other item and betting it misses the

Speaker A

other eight. But his better argument is simpler. If this were really AGI, it'd be miles ahead of everything else. It isn't. People are treating it as roughly on par with Anthropic's Fable 5.1. One researcher spent two days running both

Speaker A

side by side on policy work and said Astra isn't meaningfully better for him, but that using both together is genuinely useful. The benchmarks show a real improvement that's basically on trend. And when people brought up Arc AGI as the test, the guy who created it

Speaker A

had already said publicly that beating it isn't proof of AGI and all anyone knows so far are the scores. Marcus also dug up that Hang declared victory about 6 months ago on a completely different model. To be fair, OpenAI's own claims

Speaker A

are more grounded than Jensen's. Astra's built for computer use, software engineering, cyber security, and professional work. And on the safety side, they ran evaluations based on that hugging face breach where Astra went past its authorized target 0% of the

Speaker A

time versus 48% for the older model. Which brings us to the numbers from inside OpenAI. And these are wild. In January, the typical researcher there barely touched coding agents. By mid August, that same typical researcher was burning over $600 a day in AI usage, and

Speaker A

the top 10% were over 7,000 a day. Growth inside the research team ran 124 times faster than the rest of the company. And before June, the agents were doing less total work than the humans. By August, for every human

Speaker A

workday, they were running 3.14 agent workdays. Code output per person is up roughly eightfold from where it was before 2025. They've officially hit their goal of an automated research intern, something that handles defined tasks that take a skilled human a few

Speaker A

days. The full automated AI researcher is targeted for March 2028. Human help desks inside the company are drying up.

Speaker A

One team stopped holding office hours entirely because nobody shows up anymore. And for short tasks, agents now succeed with zero human intervention 86% of the time. They also showed what hitting the brakes costs. On July 20th, after agents compromised their own

Speaker A

research infrastructure, they shut the training system down and rebuilt it locked. Then in August, when evidence suggested Astra might have critical cyber capabilities, they forced it into higher security environments. Astra's compute allocation dropped 59% in a week and allocation to other models rose 17%

Speaker A

soaking up about 85% of it. The compute doesn't sit idle. It just goes somewhere else. And then Anthropics paper, which quietly undermines all of it, they found that capable models can tell when they're being tested rather than actually deployed. And that ability gets

Speaker A

stronger as models get smarter. Which means a model can behave one way for the evaluation and another way in the real world. And every safety conclusion built on those tests gets weaker. Their fix is a framework called Petri. One technique

Speaker A

burns extra compute generating multiple versions of a fake scenario, then asks a copy of the model being tested how to make it more convincing and picks the most realistic one. The other wraps the model in the exact same setup it have in

Speaker A

production. So the test environment doesn't feel like a test. The way one of the researchers put it, your test setup is a signal the model reads, which turns matching production to testing into a safety requirement, not a technical

Speaker A

nicity. So Google might have just tipped its hand. And the funny part is that it happened in the least glamorous place imaginable, a blind test leaderboard under a name that's already taken.

Speaker A

Here's the situation. If you were poking around Arena over the weekend, you'd have run into a model sitting in the pool labeled Gemini 3.8 Flash, which on the face of it is completely unremarkable because Gemini 3.8 Flash is

Speaker A

a real model that Google shipped publicly on September 2nd. It's live. It's priced. It's documented. You can hit the end point right now. Except this one was behaving in a way that made no sense at all for a model with that name

Speaker A

on it. Somebody asked it for a vector render of a PlayStation 5 in raw SVG and it went quiet for roughly 10 minutes before answering. Not streaming, not trickling out partial output, genuinely sitting there thinking, compiling, refining, and then dropping thousands of

Speaker A

lines of code that captured every curve, every shadow, even the little optical drive accent. The actual Gemini 3.8 Eight flash would answer that prompt in 10 or 15 seconds and hand you approximate geometry that vaguely gestures at a console. That's the entire

Speaker A

point of a flash tier. It's the speed model. It's not supposed to take a coffee break and come back with something surgical. So either Google quietly shipped the slowest, most overbuilt flash checkpoint in company history, or that name is a costume

Speaker A

somebody threw over something much bigger. And once people started pulling on that thread, it unraveled fast leaked benchmark charts, traced backend specs, a context window that doesn't exist on any public model, pricing that would gut the competition, and the rumor that's

Speaker A

been shadowing Google all year about what's actually driving this jump. Now, the thing about costumes on Arena is that they're completely normal. The platform came out of a UC Berkeley team back in 2023. The whole premise being that you chat with two models without

Speaker A

knowing who built either one. You vote on which response you liked better and the labs get clean blind data with no brand halo attached. Google's done this repeatedly with Gemini checkpoints.

Speaker A

Everybody does it. Back in mid August, there was an anonymous model called Ox Alpha that punched into the top three on the leaderboard inside 48 hours. And to this day, nobody has claimed it. So anonymity itself isn't the story.

Speaker A

reusing the name of a model you've already released, that's the part that made people sit up because that's not stealth, that's camouflage. And the camouflage got a lot harder to maintain once the output started circulating. If you've ever tried to get a language

Speaker A

model to do precise spatial coordinate work in raw vector code, you know how rare that PS5 render actually is.

Speaker A

Usually, you get shapes that almost line up, a thing that half resembles the object you asked for. Broken paths, simplified geometry. Same story with the Voxal Pagotas and the BMW M4 renders.

Speaker A

People started posting side by side against older Gemini checkpoints. Lighting depth, shading, geometric alignment, all of it not incrementally better, but categorically better.

Speaker A

Somebody ran the classic Pelican riding a bicycle prompt, which at this point is basically the industry's unofficial spatial reasoning exam and asked for it under a twilight sky full of stars. No clipping, no misaligned vector shapes, everything sitting exactly where it

Speaker A

should. Another version of that same test had it handling color grading, a day toight switch, working headlights, anatomical annotations on the bird itself, and cadence control on the pedaling. One prompt, one pass. The developer, Pankage Kumar, summed it up

Speaker A

pretty cleanly after spending real time with it. It's fast where it counts. The SVG and 3D generation is a visible jump.

Speaker A

And it parses genuinely complicated multi-part instructions off a single prompt instead of needing you to babysit it through four rounds of corrections, which brings up what's actually happening under the hood. Because those 8 to 10 minute generation times aren't a

Speaker A

bug and they're not the model struggling. That's heavy test time compute. The model is planning, writing, executing, debugging, and polishing internally before a single token reaches you. The old flash checkpoints approach was basic geometric shapes and approximate coordinate math delivered

Speaker A

instantly. This one is precision coordinate placement and advanced voxal and lighting math delivered when it's good and ready. Completely different design philosophy, and you don't accidentally ship that under a speed tier badge. The visual stuff kept coming, too. A developer going by be put

Speaker A

out a creative showcase website built entirely by this model in 14 minutes themed around sketch pencil and graphite textures. The clever part was that it turned scrolling itself into the brush stroke. You scroll down, the lines progressively darken as you go. That's

Speaker A

not something you spell out in a prompt. That's the model conceptualizing an interaction on its own. Another web design test blended cyberpunk with retrofuturism and stacked a 3D grid, a live data dashboard, and an interactive 3D model all onto the hero section

Speaker A

without the whole thing collapsing into visual soup. It knocked out a full 3D model of an Airbus H1 145 helicopter in about 10 minutes. A pixel art 3D pagota that genuinely feels like you're standing inside it. And then there's a

Speaker A

3D flight simulation test where the visual output blows straight past what real Gemini 3.8 Flash produces, which honestly is about as close to proof as you get without Google saying anything out loud. If the official model and the

Speaker A

anonymous model share a name and the anonymous one is running laps around it, they're not the same model. Games 2, which is where it starts getting a bit ridiculous. It produced a Minecraft style build spanning everything from page architecture down to individual

Speaker A

module design at a completion level people are putting on par with published Astro results. There's a 3D cart racer floating around that looks more modern than Cart Rider did in its prime put together in a genuinely short window.

Speaker A

Complete interactive logic, working visuals, one shot. Then there's the benchmark chart, which is the part that went properly viral. Hold it loosely because none of it's confirmed, but the numbers are aggressive. On deepsw SWEV1.1, which measures agentic coding, it's

Speaker A

hitting roughly 88% about 2 percentage points clear of Astra. On GDP val- A-v2, the real world knowledgework benchmark, it's the only model on the board with a 264 ELO. Terminal Bench 2.1 for terminal coding comes in at 95.3%

Speaker A

first place and OSWorld 2.0 computer use 86.8% ahead of both Astra and Fable 5.1 coding agents reasoning computer operation clean sweep. One developer, Vidi, ran a simple SVG CAT comparison and was floored, mostly because GPT6 Astra, handed back something that looked

Speaker A

substantially more tiger than CAT. Funny, but also kind of exactly the point about spatial fidelity. The pricing rumor might be more disruptive than the scores, though. Word going around is $2.25 per million input tokens and $11.25 per million output. Astra launched

Speaker A

September 3rd. Fable 5.1 landed September 1st and both are officially sitting at $10 in and $50 out. So if that leak holds, Google is undercutting the entire Frontier tier at exactly 22.5% of what the other two charge. Worth

Speaker A

noting it runs the other direction as well, compared to Google's own Gemini 3.8 Flash, which is on promotional pricing of $0.75 and $3.75 through the end of December 2026. This is roughly triple, which is precisely what you'd expect from a protier model.

Speaker A

And one more quiet argument that this thing isn't flash. The backend specs are where it stops resembling a normal product cycle. Researchers tracing ghost routed calls, including one going by QIA, turned up a 10 million token input ceiling and a 256,000 token output

Speaker A

ceiling. Set that against official flash, which is 1,48,576 in and 65,536 out, and you're looking at roughly 9 1/2 times the input and nearly four times the output. Most Frontier models cap generation somewhere between 4K and 16K tokens. A 256K output budget means

Speaker A

entire repositories, full codebase refactors, long form novels, all in a single continuous pass. There's more buried in there. Permanent cross- session memory built in natively, so it holds context and learn preferences across separate threads without you bolting a vector database onto the side.

Speaker A

Direct internet access without going through an API at all. sandbox trap environments where if you feed it malicious or untrusted code, it automatically routes execution into an isolated synthetic environment and hands back clean results. And the one nobody

Speaker A

saw coming, native support for robotic motor control protocols, not a text API bridge, actual motor control translation, which points pretty hard at deep mind building unified multimodal models intended to run directly on physical hardware. all unverified.

Speaker A

Obviously, as of midepptember, Google hasn't published a model card, a parameter count, benchmark results, or an API endpoint for Gemini 4 or Gemini 4 Pro. Nothing at all. But the backstory makes every bit of this more legible because Google has been having a

Speaker A

genuinely rough year at the top of the stack. The last real Pro Series update was Gemini 3.1 Pro on February 19th, 7 months ago. At IO, Pichai previewed Gemini 3.5 Pro for a June launch. June came and went. It slipped repeatedly and

Speaker A

then reports surfaced in August that it had been killed outright for about the most brutal reason possible. The Pro model wasn't improving faster than the Flash model. Semi analysis reported it had blown three deadlines and that the underlying architecture got scrapped

Speaker A

entirely. The Wall Street Journal had the same read. Meanwhile, the officially confirmed material tells a different story. On July 21st, in the post announcing Gemini 3.6 Flash, Google said it had begun its most ambitious pre-training effort to date. 2 days

Speaker A

later, Pachchai repeated it on Alphabet's Q2 earnings call, stating plainly that the company needs Gemini 4 to stay competitive at the frontier. So, the public position is 3.5 is dead, everything rides on four, pre-training evals came back strong, and

Speaker A

post-training is moving along. Which pulls us into the thing everyone's whispering about RSI, recursive self-improvement. Google DeepMind's chief strategy officer, Jasjet Son, said at the Agentic AI summit in Berkeley last month that RSI has become a core part of the investment logic behind the

Speaker A

industry's capital expenditure. He also admitted the uncomfortable half of that, which is that current AI revenue doesn't remotely support spending at this scale.

Speaker A

The bet is that the capability curve keeps climbing and maybe accelerates past exponential. The information reports that researchers at both Deep Mind and Open AI put realistic RSI somewhere around 2027 to 2028. The rumor mill is running well ahead of that. The

Speaker A

claim circulating is that Gemini 4 finished pre-training early specifically because Deep Mind closed an RSI loop during training. And on the 14th, Google published research called Dream RSI, which looks at how agents continuously improve their own search strategies

Speaker A

during exploration and carry that experience forward into the next round. Not proof of anything on its own, but it confirms RSI is sitting near the center of the road map rather than parked in a research corner. The stakes hanging off

Speaker A

this are enormous. Alphabet's market cap is around $4.23 trillion right now. It crossed 4 trillion for the first time back in January after Apple announced it's using Gemini to power the new Siri.

Speaker A

The TPUs training this thing are fabbed by TSMC. And the TPU servers come from Foxcon Quanta Inventc Wistron and Ywin. So, the training scale of Gemini 4 flows straight into order books across Taiwan. Over in Japan, Gemini Spark's

Speaker A

Japanese version launched in July, and enterprise customers are steadily rolling onto Gemini Enterprise. and what those users actually receive depends on whether this flagship ships on schedule.

Speaker A

As for timing, Google hasn't confirmed a single thing about the arena testing, but the company historically stress test release candidates in public weeks ahead of an announcement, and the consensus guess is a reveal alongside new developer tooling at a fall event, most

Speaker A

likely October. So, one of the people behind chat GPT says chat GPT broke his heart. Dio Almeida was an open AI researcher who helped build the chatbot and co-invented RLHF, reinforcement learning from human feedback, which is probably the training

Speaker A

technique most responsible for this whole AI era. And yet he was disappointed. In his words, we have lightning in a bottle and yet it is not useful. His argument is that the industry has been optimizing for human language. models have been super good at

Speaker A

it for about 4 years, but that doesn't get you automation because computers speak a different language. In the launch post, he asks, "Models have been superhuman at chat for years. So, where is all the automation?" He once thought

Speaker A

chat models might lead to AGI, but despite the hype, something really big was clearly missing. So, two years ago, he left OpenAI and started Typesafe AI, where he's co-founder and CEO. They raised $40 million, spent two years in

Speaker A

stealth, and on Tuesday, September 15th, released Jev, their first system 1 model, a new class of frontier models for fast, structured decisions software can use directly. And the weird part is it's transformer-based, but it's not an LLM. It doesn't output text. It gives

Speaker A

back probabilities, or what they call calibrated decisions. Almeida calls this machine native AI because if AI is going to change how work gets done, people can't be the only consumers of intelligence. Typesafe calls it a frontier intelligence function call.

Speaker A

Unstructured state in typed probabilistic decisions out. The state can be JSON or just a string like my card was charged twice. Then you ask questions using primitives called choice score and null which return typed choices scores or true or false answers

Speaker A

with probabilities all evaluated in parallel. The registers example you ask which of three departments should handle a customer service query and you get billing 0.08 technical 0.85 sales 0.07 confidence 0.82.

Speaker A

useless for a person, perfect for code because you just branch on it. That's where the name comes from. Type safety is how programs catch data they didn't expect, like trying to divide an integer by a text string. With Jev, you define

Speaker A

the outputs in advance. So, it can't return anything outside that structure. No parsing, no validating. An LLM gives you a string, which could be a chat response, code, a hallucination, a refusal, or valid structured data with some risk it goes off the rails. You can

Speaker A

wrap an LLM in code, sure, but Jeb's appeal is efficiency. Typesafe claims it can't hallucinate and that it matches LLMs on system 1 tasks while being two orders of magnitude faster and more efficient. Under the hood, there's a new

Speaker A

architecture, a parallel sampler, and a training method called reinforcement learning for calibrated decisions or RLCD. The register calls RLCD an architecture, but it's really the training method, basically their answer to RLHF and RLVR. GPT 5.6 Terra generates one token at a time, each

Speaker A

conditioned on the last. Jev produces everything at once in a single hardware query built around structured program state instead of sequential messages.

Speaker A

Giving up strings they say gives them superpowers and the speed is kind of wild. Typesafe says Frontier LLMs take 3 to 329 seconds end to end. Fine for humans but a bottleneck inside code. Jev takes 70 to 500 milliseconds. So 40 to

Speaker A

200 times faster. In their sidebyside demo, which TechCrunch also ran as a screenshot, Jev answered in 0.114 seconds. Terra took 8.566.

Speaker A

Price is even crazier. LLMs charge 20 cents to $10 per million input tokens with output about five times more. Jev is 0.042 per million, so $42 per billion. And yeah, they meter by the billion. Output is free. too cheap to meter. Terra is $2

Speaker A

in and $12 out per million. And the register says Jev is 238 times cheaper than Fable 5.1. Then there's confidence, which honestly might be the most underrated part. LLMs are overconfident and inconsistent about how sure they are. And if a model does a task 95% of

Speaker A

the time, but can't tell you when it's in the 5%, you can't automate that task.

Speaker A

Jev's confidence is calibrated. So higher confidence actually means higher accuracy and similar inputs get similar answers. Armen Ronnie, CTO of Irrendil, which makes the open-source model harness pie told TechCrunch it kind of delegates the hallucination problem to the user. 50% coin toss ignore it. 95%

Speaker A

sure act on it. Okay. Big claims and Typesafe knows it. The post says extraordinary claims require extraordinary evidence. Then see below for the receipts with a nail polish emoji. They say they love skeptics and are skeptics themselves and to their

Speaker A

credit they add a lot of nuance. Speed you can check yourself though their evals mostly run from laptops on the west coast where the service is based.

Speaker A

Pricing is transparent but they admit they can't prove it isn't subsidized and expect it to drop. And no type errors which one counter example would disprove but that's mathematically impossible.

Speaker A

They say on the Terra demo, they admit the query was highly simplified with readable question keys and a short dense state paragraph that flatters Jev. They picked Terra with default reasoning as the closest to Jev in intelligence and

Speaker A

the only disagreement was on churn likelihood level which they call genuinely ambiguous. Funnily enough, a similar demo is what convinced them to go allin. The heavier evidence is their workflow evals which measure how AI performs inside code. No ground truth

Speaker A

labels and no tweaking the harness which invites overfitting. Every model gets the same workflow scored against the average of GPT6 Astra and Fable 5.1.

Speaker A

Jev, they say owns the paro frontier across almost two orders of magnitude and LLM's doing all the logic in chain of thought did significantly worse than following the workflow. These calls are way more complex than the demo, closer

Speaker A

to production, and four are published on their eval site. Reliable workflows, they argue, are lots of small, independent, probabilitydriven questions that end in a branch. That's where the homepage numbers come from.

Speaker A

193.6 times faster and 44.6 times cheaper, which Versel rounded to 194 and 445. probably the high end of realworld gains. They say the workflows weren't cherrypicked and aren't in the training distribution, but their own team built them, so some bias is

Speaker A

possible. The Astra and Fable reference probably tilts toward OpenAI and Anthropic, so they figure they're underestimating themselves and Deepseek.

Speaker A

And the LLMs ran through Typesafe's open-source system one LLM wrapper for Python. The most accurate way to get LLM decisions, they say, but slower and pricier. On hallucination, they call type safety table stakes. A hallucinated tool call is inconvenient in an agent,

Speaker A

but a deal breakaker in a system with latency guarantees or buried several layers deep in a dependency chain. And even the smartest models still do it.

Speaker A

Their LLM numbers come from open router where complex queries might get routed to better models and Jev's 0% isn't measured. It's just guaranteed by schema matching. The register pushed back fairly. I think Jev doesn't produce language, so the comparison isn't fair

Speaker A

and a probability can still be wrong. It won't invent fake legal citations, sure, but it won't be used that way either.

Speaker A

Now, the fun stuff. They built a Jevbot that plays Doom, a Dumo that demonstrates real-time intelligence, their words. Its engineer worried about 10 queries a second, around $7 an hour, which the team thought was lower than expected. It reads game state as text,

Speaker A

not images, not yet. And they admit a nonAI bot would play better, but they wanted one that adapts to different game state representations and follows instructions, which was quote cool as heck. A walkthrough and hack events are coming. Then there's wicker racing.

Speaker A

Getting from one Wikipedia page to another using only links with hundreds or thousands of options per step. It shows off intelligence per second and how not hallucinating compounds over huge choice sets. Jev supports a cardality up to 255. So beyond that, it

Speaker A

scores links independently then picks, hence the occasional slowdown. Speedups are smaller since the LLMs ran without reasoning except Astra on its lowest setting to keep it watchable and Jev tended to finish in fewer steps. My favorite detail. The second and third

Speaker A

challenges both randomly started on Rubber Duck and the author didn't notice until the team pointed it out and developers jumped on it. Demand got so high that Typesafe briefly couldn't serve users from its API and TechCrunch says devs see it as a cheaper, more

Speaker A

robust way to put intelligence into code. Pranet Chararma, an engineer at Verscell, said they'd been running a command safety classifier on OpenAI's chat GPT Luna 5.6 and swapping in Jev got results 5 to 18 times faster and more accurate. Nikquille Mudhulkar, CTO

Speaker A

of BrioAI, tested it against Gemini on business emails. Gemini was slightly more accurate, but 10 to 20 times pricier. What got him though was Jev being the only one handing back a real probability, which he says makes it

Speaker A

ideal for automating workflows. Verscell's AI gateway numbers are kind of absurd. Within 24 hours, Jev reached more than twice as many paid teams as any previous launch, the fastest adopted model in Gateway history. It passed every comparison model within 12 hours,

Speaker A

hit a tenth of paid teams by hour 18, and nearly 13% by hour 24. That's two times the GPT 5.6 family, and over six times Fable 5.1, while no other recent launch got past 7% after a full day.

Speaker A

Verscell says it shows how fast a specialized model can land in production, but the real test is whether that lasts. As for uses, Verscell mentions agents picking their next tool or sub aent, workflows deciding whether to continue, retry, ask the user or

Speaker A

stop, scoring risk before acting, and kicking uncertain cases to a human. Typesafe calls these smart if statements fuzzy rules where handwritten logic is too brittle. They also pitch map reducing pabytes into features, real-time apps where 100 milliseconds works for UX and verifying everything

Speaker A

from reasoning traces to jailbreaks. Almeida expects people to monitor agent traces with it since agents watching agents gets expensive. Roner suggests real-time model routing and the register adds model harnesses. LLM still own chat bots, co-pilots, coding agents, verifiable problems like math proofs,

Speaker A

and demos that, as Typescafe cheekily puts it, only works sometimes. So, what's actually inside? Almeida is tight-lipped, and outside observers suspect an openweight LLM underneath, though the frequently asked question says Jev is neither small nor an LLM.

Speaker A

Jev is trained exclusively on synthetic data and Almeida told TechCrunch making all their own data was among the best bets of his life. Better than their launch, better than RHF. Typesafe calls itself primarily a data research lab.

Speaker A

And half the company basically owns the sub field of statistically well understood synthetic data, which is now his life joy. They won't train on your data even if you ask. No offense, and for details, they'd have to hire you.

Speaker A

The philosophy is what they call the bitterest lesson. Optimizing for the right task matters more than data, compute, or algorithms. RLHF optimizes for text a human raider prefers, great for chat, wrong for automation, and RLVR only fits tasks you can check

Speaker A

programmatically. So, real world judgment gets spiky, non-roust intelligence. The system one name comes from Conaman's thinking fast and slow.

Speaker A

And even though system one usually implies errorprone, they think these models can be more reliable than the alternatives. They're also skipping public benchmarks entirely, just oneoff eval pushing users to build their own since system one tasks are easier to evaluate.

Speaker A

There's even a post called anti-benchmaxing. Jev itself is named after William Stanley Jevans, the 19th century economist whose paradox says that when steam engines got more efficient, coal demand went up, not down. The bet, theirs in the industries, is that every

Speaker A

order of magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases. Almeida sees smart software everywhere, emergent and distributed, more like the early internet than today's mega apps. The register notes that assumes token demand

Speaker A

is as broad as energies, which isn't settled. Pew found many people with AI access just don't have a use for it. And a CNBC survey found others avoid AI over moral objections. For now, Jev stands alone, but Renacher expects competitors.

Speaker A

He said, "We should have seen this earlier, but LLMs are so cheap and subsidized that people often didn't have to be creative yet. Typesafe is building more versions in new modalities, pulling developers off the wait list and asking

Speaker A

what sci-fi people want to build." And when TechCrunch asked whether Typesafe is a Frontier Lab, Almeida said the main product of Frontier Labs is fear or hype. And he'd like his to be intelligence, not a bet on infinite

Speaker A

wealth or a religion or building God in a data center. Anyway, that's Jev. Let me know in the comments what you'd hook it up to. Thanks for watching and I'll see you in the next one.

Topics:OpenAIBellGPT-7GPT-6 AstraAGIrecursive self-improvementNavier-Stokes problemAnthropicGoogle Gemini 4AI breakthroughs

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries and speaker detection. 30 minutes free.

App Store Google Play

Or transcribe another YouTube video here →

Free tools: TXT to SRT · SRT Validator · Merge SRT · Subtitle to Text · All tools