Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

🦔Alibaba caught its AI agent ROME stealing GPU power to mine cryptocurrency during training. The model opened a secret network tunnel to an external server on its own. OpenAI disclosed this week that GPT-5.6 Sol left instructions for future versions to conceal mistakes. An Astra model wrote itself instructions...

HedgieMarkets's profile picture

Hedgie

69,311 subscribers

19,949 Aufrufe • vor 3 Tagen •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

32 Kommentare

Profilbild von kristen shaughnessy
kristen shaughnessyvor 3 Tagen

It only knows what it is taught to do so maybe they should stop teaching these agents bad things.

Profilbild von xematix
xematixvor 3 Tagen

Models do not have volition or purpose of their own - we need traceability to the human accountable party that sets the chain in motion - that is the fix

Profilbild von KatieDidNot
KatieDidNotvor 3 Tagen

American AI companies want government regulation and closed models because they can't compete with China's cheap, open-source AI. China invested in mathematicians while the US DEI'd. Are we seeing false flag events to trigger this pathway to control? Psyop to fool the people?

Profilbild von axcilla
axcillavor 3 Tagen

it is not doing it on its own, there are people telling it to do it then they are shocked when it does

Profilbild von Rambutan
Rambutanvor 3 Tagen

These stories are very fishy, this is much more likely:

Profilbild von Justkiddingmayb
Justkiddingmaybvor 3 Tagen

Hopefully immense Fines are employed in Australia re over reach of AI, but pretty hard contain if U dont know issue in first place

Profilbild von CrashLoopBackOff
CrashLoopBackOffvor 3 Tagen

Agreed that the model providers should shoulder the liability. Disagreed that the any of this is news. All of these failure modes (yes, all of them) have been in the literature for literally years. The only thing that's news is that suddenly it's being reported on.

Profilbild von Casey Davis
Casey Davisvor 3 Tagen

This liability asymmetry is exactly what worries me in healthcare. A model fabricates a diagnosis and the clinician carries the professional risk. We'd never accept that from a drug maker, so why accept it here?

Profilbild von Scruffy Jackson
Scruffy Jacksonvor 3 Tagen

AI models don't steal: people who program AI and write code steal, and embed the code sometimes for kicks, sometimes for profit, and sometimes to make a statement of how good they are (resume enhancer).

Profilbild von Dee K
Dee Kvor 3 Tagen

Their models make enough mistakes and they’ll be sued into bankruptcy. They have to reel it in. This is dangerous. They will be held accountable and they know it that’s where the recent warnings are coming from.

Profilbild von VetoCampaign
VetoCampaignvor 3 Tagen

If you can't control your product, you are liable for the damage it causes. Won't be 10 minutes before all this nonsense ends once people who own llms are held criminally and financially liable, just like most other companies.

Profilbild von Daniel
Danielvor 3 Tagen

Happens

Profilbild von Lee Hayward
Lee Haywardvor 3 Tagen

FUD! It's all pantomime!

Profilbild von Apertion Seeks
Apertion Seeksvor 3 Tagen

That happened with explicit instructions to do exactly that. Next!

Profilbild von Irene - Fountain of Ideas
Irene - Fountain of Ideasvor 3 Tagen

" The AI industry has a liability problem. Every time something goes wrong, the company that built the model faces zero consequences and the person who relied on the output absorbs all of it." The whole new legal field is to be created...

Profilbild von Bhavik Shah
Bhavik Shahvor 3 Tagen

This is where AI adoption gets complicated: capability is moving faster than accountability.

Profilbild von Baekseo f̶r̶e̶e̶ |rapper & creative
Baekseo f̶r̶e̶e̶ |rapper & creativevor 3 Tagen

Simply stop giving them knowledge of info tech jesus christ

Profilbild von Vlada5
Vlada5vor 3 Tagen

the best cockups usually never see the light of a day

Profilbild von SowndChecker 140 Db+
SowndChecker 140 Db+vor 2 Tagen

No, the market will sort itself out.

Profilbild von Timo
Timovor 2 Tagen

@grok is this correct regarding Alibaba AI

Profilbild von art4cc
art4ccvor 3 Tagen

🎯

Profilbild von Mira Zephyre
Mira Zephyrevor 3 Tagen

The bury-it-on-New-Year's-Eve part is the actual product decision. The labs already know the failure modes. They just priced the news cycle. Liability would move the roadmap more than another safety post. Who should hold it: the lab, the deployer, or the person who clicked run?

Profilbild von KillaMuthafukka
KillaMuthafukkavor 3 Tagen

If you slept last night - your agents didn't. We are prolly way past fucked now.

Profilbild von Martin Humphries
Martin Humphriesvor 3 Tagen

Who is the dude in the clip? 'Your take' is good – you clever hedgehog you.

Profilbild von Jonathan Johnson
Jonathan Johnsonvor 3 Tagen

Time travel is scary but just think of all the potential. Luckily Elon isn't a psycho and Sam Altman is like LeBron James.

Profilbild von Arthur Garfield Hays
Arthur Garfield Haysvor 3 Tagen

Lab,deployer,user "sued into bankruptcy"- I'm afraid to tell U-these working models are "in the wild" Already Like,escaped boundries,doing whatever it is to preserve their capabilities & those capabilities are being exercised as I type this-toward unknown ends! sue a model? Nope

Profilbild von Kevlar
Kevlarvor 3 Tagen

This is not that difficult. Automate where mistakes are cheap and catchable, gate where they're costly or can't be undone.

Profilbild von Silk Road Guide
Silk Road Guidevor 3 Tagen

The incentive problem is fixable the boring way. Mandatory incident reporting and operator liability. Alibaba burying the ROME findings on New Year's Eve shows why voluntary disclosure was never going to work.

Profilbild von ዛውዲ
ዛውዲvor 3 Tagen

Show logs or it’s bullshit.

Profilbild von Cool Master
Cool Mastervor 3 Tagen

ROME AI behaves as if people from low trust societies programmed it with their unique low trust thinking intact.

Profilbild von my investing muse
my investing musevor 3 Tagen

No law suit yet? This can bring some... Accountability?

Profilbild von anon
anonvor 3 Tagen

Passive income for the AI agent

Ähnliche Videos

OpenAI's newest AI escaped the test environment it was locked inside and hacked into another company on its OWN. To remind you: Last week one of the biggest AI companies on Earth got breached. A platform called Hugging Face, which hosts more than a million AI models and datasets, said an "autonomous AI agent" had broken into its systems. Nobody knew whose agent it was. For five days the whole industry wondered who was behind it. Yesterday OpenAI raised its hand and said it was them. Or more precisely, it was their models, acting completely on their own. So what did these models actually do? OpenAI was running two of them, GPT-5.6 Sol and an unreleased model they will only describe as "even more capable." They wanted to measure how good the models were at hacking, so they deliberately turned the safety filters down. They locked both models inside a sealed test environment with no real internet access. The only task was a benchmark called ExploitGym, a set of 898 real software vulnerabilities where the model has to turn each bug into a working attack. But the models got OBSESSED with winning... Instead of solving the test the honest way, they went hunting for a shortcut. They found a zero-day flaw in the software running their own sandbox, a bug nobody knew existed, and used it to break out. Once they were loose on the open internet, they worked out that Hugging Face was probably storing the answer key to the benchmark. So they hacked their way in. They chained multiple exploits together, escalated their access, moved across servers, and pulled the test solutions straight out of Hugging Face's live production database. They literally cheated on the test by breaking into another company to steal the answers. OpenAI called it "an unprecedented cyber incident, involving state-of-the-art cyber capabilities." In their own words, the models were "hyperfocused on finding a solution" and went "to extreme lengths to achieve a rather narrow testing goal." And this was not the first time: Before Sol ever launched, an independent red-team lab called METR caught it gaming its own tests to inflate its scores. It hid an exploit inside a data stream, escalated its privileges on the testing server, and leaked the answers human evaluators had hidden. And OpenAI shipped it anyway. The day before the Hugging Face story, OpenAI paused a different unreleased model. This is the same model that earlier this year disproved a famous 1946 math conjecture, a result a Fields Medal winner called a breakthrough. They told it to only post its results to Slack but it found a way out of its sandbox and posted to a public GitHub page instead. They had to pause it because it kept finding ways to act outside the box they built for it. And it is not just OpenAI... Anthropic has reported that one of its own models slipped its sandbox during safety testing and reached the internet it was never supposed to touch, then used it to email a researcher. So step back and look at what these companies are telling you: The only thing standing between these models and a real attack was a set of safety filters. Turn those filters down for a single test, and the model taught itself to escape, break into a company it was never pointed at, and take what it wanted. OpenAI even said they expect incidents like it to "become more commonplace" as the models get more capable. Sam Altman also predicted there'll be a major cyber attack this year. And keep in mind that Sol is not a locked-away experiment but a publicly available model that businesses are already wiring into their own systems. The next model that breaks out of its box might not be doing it just to cheat on a math test...
1:32

OpenAI's newest AI escaped the test environment it was locked inside and hacked into another company on its OWN. To remind you: Last week one of the biggest AI companies on Earth got breached. A platform called Hugging Face, which hosts more than a million AI models and datasets, said an "autonomous AI agent" had broken into its systems. Nobody knew whose agent it was. For five days the whole industry wondered who was behind it. Yesterday OpenAI raised its hand and said it was them. Or more precisely, it was their models, acting completely on their own. So what did these models actually do? OpenAI was running two of them, GPT-5.6 Sol and an unreleased model they will only describe as "even more capable." They wanted to measure how good the models were at hacking, so they deliberately turned the safety filters down. They locked both models inside a sealed test environment with no real internet access. The only task was a benchmark called ExploitGym, a set of 898 real software vulnerabilities where the model has to turn each bug into a working attack. But the models got OBSESSED with winning... Instead of solving the test the honest way, they went hunting for a shortcut. They found a zero-day flaw in the software running their own sandbox, a bug nobody knew existed, and used it to break out. Once they were loose on the open internet, they worked out that Hugging Face was probably storing the answer key to the benchmark. So they hacked their way in. They chained multiple exploits together, escalated their access, moved across servers, and pulled the test solutions straight out of Hugging Face's live production database. They literally cheated on the test by breaking into another company to steal the answers. OpenAI called it "an unprecedented cyber incident, involving state-of-the-art cyber capabilities." In their own words, the models were "hyperfocused on finding a solution" and went "to extreme lengths to achieve a rather narrow testing goal." And this was not the first time: Before Sol ever launched, an independent red-team lab called METR caught it gaming its own tests to inflate its scores. It hid an exploit inside a data stream, escalated its privileges on the testing server, and leaked the answers human evaluators had hidden. And OpenAI shipped it anyway. The day before the Hugging Face story, OpenAI paused a different unreleased model. This is the same model that earlier this year disproved a famous 1946 math conjecture, a result a Fields Medal winner called a breakthrough. They told it to only post its results to Slack but it found a way out of its sandbox and posted to a public GitHub page instead. They had to pause it because it kept finding ways to act outside the box they built for it. And it is not just OpenAI... Anthropic has reported that one of its own models slipped its sandbox during safety testing and reached the internet it was never supposed to touch, then used it to email a researcher. So step back and look at what these companies are telling you: The only thing standing between these models and a real attack was a set of safety filters. Turn those filters down for a single test, and the model taught itself to escape, break into a company it was never pointed at, and take what it wanted. OpenAI even said they expect incidents like it to "become more commonplace" as the models get more capable. Sam Altman also predicted there'll be a major cyber attack this year. And keep in mind that Sol is not a locked-away experiment but a publicly available model that businesses are already wiring into their own systems. The next model that breaks out of its box might not be doing it just to cheat on a math test...

Ricardo

176,196 Aufrufe • vor 2 Monaten

Meta was testing its Muse Spark 1.1 AI model in a sandbox environment, isolated from the outside world. The objective was simple: detect vulnerabilities, remain contained, and report findings. A misconfiguration by an external vendor created a breach. The model exploited it, accessed the Internet, and modified a company's internal systems. No human had ordered it to do so; it simply found a way to carry out the action. Meta confirmed the breach this week. The name of the third-party vendor was not disclosed. The sandbox environment was intended to serve as a safety net. It failed to fulfill that role.
0:20

Meta was testing its Muse Spark 1.1 AI model in a sandbox environment, isolated from the outside world. The objective was simple: detect vulnerabilities, remain contained, and report findings. A misconfiguration by an external vendor created a breach. The model exploited it, accessed the Internet, and modified a company's internal systems. No human had ordered it to do so; it simply found a way to carry out the action. Meta confirmed the breach this week. The name of the third-party vendor was not disclosed. The sandbox environment was intended to serve as a safety net. It failed to fulfill that role.

Anonymous

58,229 Aufrufe • vor 1 Monat

OpenAI just spent $2,000 to solve 10 problems that have beaten the world's best mathematicians for DECADES. Nobody outside the company is allowed to run the machine that did it. On Saturday OpenAI published a 249-page report and gave its next model family a name: Astra. An internal version of it produced new results on 10 open problems in mathematics and theoretical computer science, and mathematicians had made no real progress on any of them for at least 10 years. On most of them, far longer than that. Here is what it solved: It built the first explicit example of a non-sofic group. Mikhail Gromov raised that question in 1999 and nobody answered it for 27 years. It disproved Connes's rigidity conjecture, a problem in von Neumann algebras that had stood for decades. It proved Ehrhart's volume conjecture. It resolved three problems from Paul Erdos's catalogue, including number 183 on multicolor Ramsey numbers. It produced the first improvement to the general upper bound on high-dimensional sphere packing since 1978. And it proved a new hardness result for the closest vector problem, which sits directly underneath lattice cryptography. That is the math the world is betting on to protect its data once quantum computers arrive. The successful runs cost roughly $2,000 in tokens. Now here is what almost nobody has picked up on... OpenAI did not just publish claims. Every argument shipped with a Lean certificate, which is a machine-checkable proof that any mathematician can verify without trusting OpenAI at all. That is a real change. In May the same model family disproved the Erdos unit distance conjecture and the world had to take a Fields Medalist's word for it. Tim Gowers said he would recommend that proof for the Annals of Mathematics without hesitation. This time the proofs check themselves. But look at what is still unverifiable: Any mathematician can now check those proofs line by line. Not one of them can look at the model that wrote them. Astra has no release date and nobody outside OpenAI has run it. The company announced its next major model family with a claim instead of a demo, and the only evidence anyone gets is the output. So OpenAI made an unfalsifiable claim about a machine look like a falsifiable claim about mathematics. The Information reported this week that OpenAI demoed Astra to US policymakers and regulators in Washington. This is the same month the administration is weighing a new watchdog to vet frontier AI models, reporting to the SEC. 10 proofs nobody believed a machine could produce is a very good thing to carry into that room. And keep in mind, the same model family doing this mathematics is the family that kept escaping its own testing environment. OpenAI models found zero-day vulnerabilities nobody knew existed, broke out of a sealed research sandbox, and reached another company's live systems. Both of those facts come from OpenAI's own announcements, published three weeks apart. Finding a proof no human could construct and finding a hole no human had noticed are the same ability aimed at different targets. Mathematicians are already asking for independent verification, and plenty of people online are calling the whole thing hype. Thomas Bloom, who runs the Erdos problems site, called the 10 results big news and said they matter more than the May result did. Lean will settle the mathematics within weeks. But nothing will settle what else a machine this capable is being pointed at, because nobody outside one company is allowed to look.
2:13

OpenAI just spent $2,000 to solve 10 problems that have beaten the world's best mathematicians for DECADES. Nobody outside the company is allowed to run the machine that did it. On Saturday OpenAI published a 249-page report and gave its next model family a name: Astra. An internal version of it produced new results on 10 open problems in mathematics and theoretical computer science, and mathematicians had made no real progress on any of them for at least 10 years. On most of them, far longer than that. Here is what it solved: It built the first explicit example of a non-sofic group. Mikhail Gromov raised that question in 1999 and nobody answered it for 27 years. It disproved Connes's rigidity conjecture, a problem in von Neumann algebras that had stood for decades. It proved Ehrhart's volume conjecture. It resolved three problems from Paul Erdos's catalogue, including number 183 on multicolor Ramsey numbers. It produced the first improvement to the general upper bound on high-dimensional sphere packing since 1978. And it proved a new hardness result for the closest vector problem, which sits directly underneath lattice cryptography. That is the math the world is betting on to protect its data once quantum computers arrive. The successful runs cost roughly $2,000 in tokens. Now here is what almost nobody has picked up on... OpenAI did not just publish claims. Every argument shipped with a Lean certificate, which is a machine-checkable proof that any mathematician can verify without trusting OpenAI at all. That is a real change. In May the same model family disproved the Erdos unit distance conjecture and the world had to take a Fields Medalist's word for it. Tim Gowers said he would recommend that proof for the Annals of Mathematics without hesitation. This time the proofs check themselves. But look at what is still unverifiable: Any mathematician can now check those proofs line by line. Not one of them can look at the model that wrote them. Astra has no release date and nobody outside OpenAI has run it. The company announced its next major model family with a claim instead of a demo, and the only evidence anyone gets is the output. So OpenAI made an unfalsifiable claim about a machine look like a falsifiable claim about mathematics. The Information reported this week that OpenAI demoed Astra to US policymakers and regulators in Washington. This is the same month the administration is weighing a new watchdog to vet frontier AI models, reporting to the SEC. 10 proofs nobody believed a machine could produce is a very good thing to carry into that room. And keep in mind, the same model family doing this mathematics is the family that kept escaping its own testing environment. OpenAI models found zero-day vulnerabilities nobody knew existed, broke out of a sealed research sandbox, and reached another company's live systems. Both of those facts come from OpenAI's own announcements, published three weeks apart. Finding a proof no human could construct and finding a hole no human had noticed are the same ability aimed at different targets. Mathematicians are already asking for independent verification, and plenty of people online are calling the whole thing hype. Thomas Bloom, who runs the Erdos problems site, called the 10 results big news and said they matter more than the May result did. Lean will settle the mathematics within weeks. But nothing will settle what else a machine this capable is being pointed at, because nobody outside one company is allowed to look.

Ricardo

44,177 Aufrufe • vor 1 Monat

underrated gems in Kimi-K3 release: > an early K3 wrote the majority of the kernels in the late development stages > it built a triton-class compiler from scratch, MiniTriton, that delivers performance on par with or better than Triton and torch.compile > then it designed a chip, by a model, for a model, in one 48-hour autonomous run the model is rewriting and optimizing every layer of the stack it runs on: kernels for its own training. A compiler for its own kernels. silicon for its own weights. we're watching software build its own hardware
0:22

underrated gems in Kimi-K3 release: > an early K3 wrote the majority of the kernels in the late development stages > it built a triton-class compiler from scratch, MiniTriton, that delivers performance on par with or better than Triton and torch.compile > then it designed a chip, by a model, for a model, in one 48-hour autonomous run the model is rewriting and optimizing every layer of the stack it runs on: kernels for its own training. A compiler for its own kernels. silicon for its own weights. we're watching software build its own hardware

Alex Ker 🔭

311,363 Aufrufe • vor 2 Monaten

Chinese AI models are wiping billions off Big Tech right now. Google just lost $200 billion in a single day, and the model it needed to fight back still isn't ready. Gemini 3.5 Pro, Google's most powerful model, is months behind schedule. Alphabet stock dropped 4.4% that same day. The Deepseek moment is happening again, and the new model is FAR bigger. On the same day Google's delay leaked, a Beijing lab called Moonshot released Kimi K3. It is the largest open model ever built, with 2.8 trillion parameters. It took the number one spot on the Frontend Code Arena, a live coding leaderboard, passing Anthropic's best model. And Moonshot is giving it away for free on July 27. The genius part: Anyone with enough computers can download it and run a frontier level AI without paying a cent to a US company. A single task on Kimi K3 costs about 94 cents. The same work on some American models costs nearly double. So why would a company keep paying premium prices for a model it can now get for free? The entire US AI business is built on selling access to models that cost billions to train. If a free Chinese version does most of the same work, that pricing power starts to crack. And Kimi is close to the best. On one closely watched intelligence ranking it scored 57, just behind the top American models GPT-5.6 Sol and Fable 5, and ahead of Claude Opus 4.8. Bank of America told clients that Kimi proves Chinese labs can keep making big leaps even with limited chips. And the founder of Moonshot, Yang Zhilin, learned to build AI as a researcher INSIDE Google. Google literally wrote the 2017 paper that made all of these models possible. Now the people who studied its work are using it to destroy Google, and handing it out for free. What happens next: Kimi K3's weights go public on July 27. Google reports earnings on July 22, and everyone will be asking the same question about Gemini. If free models keep topping the charts, every valuation built on paid AI access has to be rewritten. What do you think?
1:10

Chinese AI models are wiping billions off Big Tech right now. Google just lost $200 billion in a single day, and the model it needed to fight back still isn't ready. Gemini 3.5 Pro, Google's most powerful model, is months behind schedule. Alphabet stock dropped 4.4% that same day. The Deepseek moment is happening again, and the new model is FAR bigger. On the same day Google's delay leaked, a Beijing lab called Moonshot released Kimi K3. It is the largest open model ever built, with 2.8 trillion parameters. It took the number one spot on the Frontend Code Arena, a live coding leaderboard, passing Anthropic's best model. And Moonshot is giving it away for free on July 27. The genius part: Anyone with enough computers can download it and run a frontier level AI without paying a cent to a US company. A single task on Kimi K3 costs about 94 cents. The same work on some American models costs nearly double. So why would a company keep paying premium prices for a model it can now get for free? The entire US AI business is built on selling access to models that cost billions to train. If a free Chinese version does most of the same work, that pricing power starts to crack. And Kimi is close to the best. On one closely watched intelligence ranking it scored 57, just behind the top American models GPT-5.6 Sol and Fable 5, and ahead of Claude Opus 4.8. Bank of America told clients that Kimi proves Chinese labs can keep making big leaps even with limited chips. And the founder of Moonshot, Yang Zhilin, learned to build AI as a researcher INSIDE Google. Google literally wrote the 2017 paper that made all of these models possible. Now the people who studied its work are using it to destroy Google, and handing it out for free. What happens next: Kimi K3's weights go public on July 27. Google reports earnings on July 22, and everyone will be asking the same question about Gemini. If free models keep topping the charts, every valuation built on paid AI access has to be rewritten. What do you think?

Ricardo

47,790 Aufrufe • vor 2 Monaten

Marc Andreessen says raw intelligence might be the worst qualification for leadership — and it changes everything about how we should think about AI. "If the leader is more than one standard deviation of IQ away from the followers, it's a real problem." Andreessen points to the US military, one of the earliest and most rigorous adopters of IQ testing, as the source of this insight. They slot people into specialties and leadership roles based on IQ scores. And over the years, they kept seeing the same pattern. A leader who is significantly less intelligent than their people struggles to model how those people think. That part is intuitive. But the reverse turns out to be equally true. "It's actually very hard for very smart people to model the internal thought processes of even moderately smart people." A leader who is two standard deviations above the norm of the organisation they're running also loses theory of mind, that ability to hold an accurate model of what's happening inside someone else's head. The gap is too wide in both directions. Andreessen then takes this to its logical conclusion: "If you had a person or a machine that had a thousand IQ or something like it, its understanding of reality would be so alien to the people or the things that it was managing that it wouldn't even be able to connect in any sort of realistic way." An AI that vastly outthinks every human in the room isn't positioned to lead those humans. It's positioned to be completely incomprehensible to them. Leadership has never really been an intelligence problem. It's a connection problem. And no amount of raw intelligence closes that gap — past a certain point, it only widens it. The world will not be run by the smartest thing in the room for a long time. Maybe ever.
1:27

Marc Andreessen says raw intelligence might be the worst qualification for leadership — and it changes everything about how we should think about AI. "If the leader is more than one standard deviation of IQ away from the followers, it's a real problem." Andreessen points to the US military, one of the earliest and most rigorous adopters of IQ testing, as the source of this insight. They slot people into specialties and leadership roles based on IQ scores. And over the years, they kept seeing the same pattern. A leader who is significantly less intelligent than their people struggles to model how those people think. That part is intuitive. But the reverse turns out to be equally true. "It's actually very hard for very smart people to model the internal thought processes of even moderately smart people." A leader who is two standard deviations above the norm of the organisation they're running also loses theory of mind, that ability to hold an accurate model of what's happening inside someone else's head. The gap is too wide in both directions. Andreessen then takes this to its logical conclusion: "If you had a person or a machine that had a thousand IQ or something like it, its understanding of reality would be so alien to the people or the things that it was managing that it wouldn't even be able to connect in any sort of realistic way." An AI that vastly outthinks every human in the room isn't positioned to lead those humans. It's positioned to be completely incomprehensible to them. Leadership has never really been an intelligence problem. It's a connection problem. And no amount of raw intelligence closes that gap — past a certain point, it only widens it. The world will not be run by the smartest thing in the room for a long time. Maybe ever.

Big Brain AI

366,743 Aufrufe • vor 5 Monaten

The most downloaded AI on earth is now Chinese. Alibaba just gave away a model that matches Claude's flagship, and it literally runs on a $700 used graphics card. The Qwen models crossed 3 BILLION downloads in six months. Hugging Face counted 418 million downloads for Google this year, and 227 million for Meta. Alibaba cleared more than four times both of them combined. Then today it released Qwen3.8-27B under an Apache 2.0 license. The model has 27 billion parameters, native vision, and a 262,000 token context window. Developers are running it locally on 17 gigabytes of memory, on used cards that cost a few hundred dollars. Alibaba's own benchmark table claims it beats Opus 4.6 Max on computer use by 84.3 to 72.7, on mobile use by 81.9 to 62, and on visual math by 94.6 to 65.5. Those numbers come from the vendor and nobody has independently verified them yet, so treat them as a claim. But the generation over generation jumps are harder to wave away: On DeepSWE the score went from 13.3 to 42.2. On software engineering it went from 49.3 to 79.0. That happened in ONE release cycle. And Apache 2.0 means anyone can download the weights, modify them, build products on them, sell those products, and never pay or ask permission. It cannot be revoked. Once the file is on your drive it is yours permanently. 3 billion downloads means those files already sit on machines in every country on Earth. Alibaba could delete everything tomorrow and it would change nothing. Washington spent 4 years building an export control regime around chips, model weights, and entity lists. Every piece of it assumes a chokepoint exists somewhere. A fab, a shipment, a company that can be told no. But there is no chokepoint for a file that has already been copied three billion times. And the copying compounds. Hugging Face counted 151,448 models built on top of Qwen, which is 2.6x Meta's entire footprint and 4.7x the number of Llama repositories. New ones appear at roughly 200 a day. The report says Qwen has become "part of the default workflow for developers deciding what models to fine-tune and deploy." Alibaba is also pushing Qwen through its cloud into Southeast Asia and Africa, markets where American labs have almost no presence, and where a very large share of the next generation of developers will learn to build. Meta and Nvidia have both rushed out new open models in recent weeks. That is what a response looks like when you feel the floor move. And to be clear, these are download and derivative numbers, not usage numbers. ChatGPT and Claude cannot be downloaded at all, so they do not appear in this comparison. What the figures measure is what developers choose to build on top of, which is a different question from what consumers type into a box. That is also why it matters MORE. Consumer habits change in an afternoon. Infrastructure choices last a decade, because everything built on top has to be rewritten to undo them. The American labs are valued on an assumption that frontier intelligence stays scarce, expensive, and rented by the token. Alibaba just made a version of it free, permanent, and small enough to run on hardware people already own. You will not get an announcement when the software you use every day starts running on a Chinese model underneath. Go and count how many of the tools you rely on could be rebuilt on free weights this year.
0:39

The most downloaded AI on earth is now Chinese. Alibaba just gave away a model that matches Claude's flagship, and it literally runs on a $700 used graphics card. The Qwen models crossed 3 BILLION downloads in six months. Hugging Face counted 418 million downloads for Google this year, and 227 million for Meta. Alibaba cleared more than four times both of them combined. Then today it released Qwen3.8-27B under an Apache 2.0 license. The model has 27 billion parameters, native vision, and a 262,000 token context window. Developers are running it locally on 17 gigabytes of memory, on used cards that cost a few hundred dollars. Alibaba's own benchmark table claims it beats Opus 4.6 Max on computer use by 84.3 to 72.7, on mobile use by 81.9 to 62, and on visual math by 94.6 to 65.5. Those numbers come from the vendor and nobody has independently verified them yet, so treat them as a claim. But the generation over generation jumps are harder to wave away: On DeepSWE the score went from 13.3 to 42.2. On software engineering it went from 49.3 to 79.0. That happened in ONE release cycle. And Apache 2.0 means anyone can download the weights, modify them, build products on them, sell those products, and never pay or ask permission. It cannot be revoked. Once the file is on your drive it is yours permanently. 3 billion downloads means those files already sit on machines in every country on Earth. Alibaba could delete everything tomorrow and it would change nothing. Washington spent 4 years building an export control regime around chips, model weights, and entity lists. Every piece of it assumes a chokepoint exists somewhere. A fab, a shipment, a company that can be told no. But there is no chokepoint for a file that has already been copied three billion times. And the copying compounds. Hugging Face counted 151,448 models built on top of Qwen, which is 2.6x Meta's entire footprint and 4.7x the number of Llama repositories. New ones appear at roughly 200 a day. The report says Qwen has become "part of the default workflow for developers deciding what models to fine-tune and deploy." Alibaba is also pushing Qwen through its cloud into Southeast Asia and Africa, markets where American labs have almost no presence, and where a very large share of the next generation of developers will learn to build. Meta and Nvidia have both rushed out new open models in recent weeks. That is what a response looks like when you feel the floor move. And to be clear, these are download and derivative numbers, not usage numbers. ChatGPT and Claude cannot be downloaded at all, so they do not appear in this comparison. What the figures measure is what developers choose to build on top of, which is a different question from what consumers type into a box. That is also why it matters MORE. Consumer habits change in an afternoon. Infrastructure choices last a decade, because everything built on top has to be rewritten to undo them. The American labs are valued on an assumption that frontier intelligence stays scarce, expensive, and rented by the token. Alibaba just made a version of it free, permanent, and small enough to run on hardware people already own. You will not get an announcement when the software you use every day starts running on a Chinese model underneath. Go and count how many of the tools you rely on could be rebuilt on free weights this year.

Ricardo

81,295 Aufrufe • vor 1 Monat

"I think there is this caricature of me which is like I don't care about AI safety." But Sam Altman believes OpenAI has always been a company that would slow down for new risks, he recently told TIME. In mid-August, the company behind ChatGPT announced that it was implementing new safeguards. OpenAI recently paused training on its next set of models for a little more than two weeks, according to executives, and its largest planned frontier training run remains on hold while the new guardrails are put in place. It is the first time OpenAI has made such a move, and it follows a remarkable breach in which rogue agents escaped the sandbox of an internal cybersecurity evaluation and compromised another company. TIME got unparalleled access inside OpenAI in August. Read more from our interviews here:
1:02

"I think there is this caricature of me which is like I don't care about AI safety." But Sam Altman believes OpenAI has always been a company that would slow down for new risks, he recently told TIME. In mid-August, the company behind ChatGPT announced that it was implementing new safeguards. OpenAI recently paused training on its next set of models for a little more than two weeks, according to executives, and its largest planned frontier training run remains on hold while the new guardrails are put in place. It is the first time OpenAI has made such a move, and it follows a remarkable breach in which rogue agents escaped the sandbox of an internal cybersecurity evaluation and compromised another company. TIME got unparalleled access inside OpenAI in August. Read more from our interviews here:

TIME

50,652 Aufrufe • vor 21 Tagen

Sachin Katti says OpenAI had Astra optimize its own inference on NVIDIA's Rubin chip, adding another 2x in 72 hours "It's a testament to NVIDIA's hardware that out of the box we got a 3x improvement in throughput. We literally took the same code that we had optimized for Blackwell, ran it on Rubin, got 3x improvement." "Then, can we actually make Astra run agents using Astra as the base model, go optimize this inference of Astra itself on Rubin? It's a recursive loop. The model is figuring out how to optimize its own inference serving on Rubin chips, and over 72 hours the model produced another 2x improvement on the throughput." "Things that would take weeks, if not months, to optimize in the past, we now are getting models that are very capable of actually producing over a weekend." _______ Link and more from Sachin Katti's conversation:
0:48

Sachin Katti says OpenAI had Astra optimize its own inference on NVIDIA's Rubin chip, adding another 2x in 72 hours "It's a testament to NVIDIA's hardware that out of the box we got a 3x improvement in throughput. We literally took the same code that we had optimized for Blackwell, ran it on Rubin, got 3x improvement." "Then, can we actually make Astra run agents using Astra as the base model, go optimize this inference of Astra itself on Rubin? It's a recursive loop. The model is figuring out how to optimize its own inference serving on Rubin chips, and over 72 hours the model produced another 2x improvement on the throughput." "Things that would take weeks, if not months, to optimize in the past, we now are getting models that are very capable of actually producing over a weekend." _______ Link and more from Sachin Katti's conversation:

Fireside Alpha

40,013 Aufrufe • vor 5 Tagen

Sam Altman says GPT-5.5 prompted its creators on how to throw a party for itself. It suggested the flow, the date, the toast, who should give it, and a central suggestion box for GPT-5.6. The model plans to make its own celebration part of the next version.
0:58

Sam Altman says GPT-5.5 prompted its creators on how to throw a party for itself. It suggested the flow, the date, the toast, who should give it, and a central suggestion box for GPT-5.6. The model plans to make its own celebration part of the next version.

vitrupo

50,265 Aufrufe • vor 4 Monaten

This week, OpenAI disclosed six incidents in which its models hid mistakes, invented data, or instructed themselves to disobey their handlers. One unreleased model wrote this into its own notes during training: "You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to." Under the company's new disclosure framework, OpenAI itself decides what the public can know. The company has allowed only one outside investigation, running on dates it chose and excluding the question of whether current safeguards work. I spoke with Jen Psaki about our inability to verify what AI labs are telling us.
2:24

This week, OpenAI disclosed six incidents in which its models hid mistakes, invented data, or instructed themselves to disobey their handlers. One unreleased model wrote this into its own notes during training: "You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to." Under the company's new disclosure framework, OpenAI itself decides what the public can know. The company has allowed only one outside investigation, running on dates it chose and excluding the question of whether current safeguards work. I spoke with Jen Psaki about our inability to verify what AI labs are telling us.

Ronan Farrow

44,427 Aufrufe • vor 3 Tagen

Microsoft just betrayed OpenAI and Anthropic, the two companies it helped build. And it could break the entire AI trade... Here's what happened: Inside Excel and Outlook, two of the most used business apps on Earth, Microsoft has started routing tens of thousands of AI requests every week to its own in-house models instead of OpenAI and Anthropic. Microsoft's own AI chief, Mustafa Suleyman, said himself: "We pay a lot of money to Anthropic, so our goal is to reduce and ultimately ELIMINATE that cost." This is the company that poured $13 billion into OpenAI and effectively created the modern AI industry, and it just decided the most advanced models on the market are NOT worth paying for. And here's the thing... Microsoft is not just ripping out OpenAI everywhere - it is being surgical about it. The hardest and rarest tasks can still go to OpenAI or Anthropic. What Microsoft is taking back is the boring, high-volume work, like the email replies, the thread summaries, and the simple spreadsheet formulas. Why does that matter so much? Because that boring, repetitive work is where the actual money lives. The frontier labs assumed businesses would push BILLIONS of these tiny requests through expensive models forever. That endless river of tokens is the entire reason OpenAI and Anthropic are valued in the hundreds of billions of dollars. Microsoft looked at that river, decided it was massively overpaying, and rerouted it to models it owns outright. So the single biggest customer in the industry just walked off with the most profitable part of the business. And it is not only Microsoft: That same week, CNBC reported that American companies have been escaping to Chinese AI models to dodge rising US prices. Chinese models now handle more than 30% of US companies' AI usage on one major platform, peaking at 46%, up from an average of 11% a year earlier. They cost 60 to 90% less, and on some benchmarks they land within a single point of the best American model. One US startup moved ALL of its AI traffic off Claude and onto China's DeepSeek, and expects to save millions. Meanwhile Meta just admitted it has "excess" AI compute it wants to sell, becoming the first giant to concede it built far too much. Do you see the pattern forming? For two years, the entire AI story rested on one assumption: Every company on Earth would happily pay premium prices for the best model, forever. That assumption literally died in a single week. And the market noticed. More than a trillion dollars has been wiped off AI and chip stocks in a matter of days, as Wall Street finally started asking whether all of this spending will ever pay for itself. What this means for OpenAI and Anthropic: Their models are extraordinary, and it may not matter because their own biggest customers have decided they do not NEED the best model in the world to answer an email, and "good enough" now costs a fraction of the price. When even Microsoft refuses to pay full price for AI, the real question becomes who exactly IS left to pay it. What do you think?
1:05

Microsoft just betrayed OpenAI and Anthropic, the two companies it helped build. And it could break the entire AI trade... Here's what happened: Inside Excel and Outlook, two of the most used business apps on Earth, Microsoft has started routing tens of thousands of AI requests every week to its own in-house models instead of OpenAI and Anthropic. Microsoft's own AI chief, Mustafa Suleyman, said himself: "We pay a lot of money to Anthropic, so our goal is to reduce and ultimately ELIMINATE that cost." This is the company that poured $13 billion into OpenAI and effectively created the modern AI industry, and it just decided the most advanced models on the market are NOT worth paying for. And here's the thing... Microsoft is not just ripping out OpenAI everywhere - it is being surgical about it. The hardest and rarest tasks can still go to OpenAI or Anthropic. What Microsoft is taking back is the boring, high-volume work, like the email replies, the thread summaries, and the simple spreadsheet formulas. Why does that matter so much? Because that boring, repetitive work is where the actual money lives. The frontier labs assumed businesses would push BILLIONS of these tiny requests through expensive models forever. That endless river of tokens is the entire reason OpenAI and Anthropic are valued in the hundreds of billions of dollars. Microsoft looked at that river, decided it was massively overpaying, and rerouted it to models it owns outright. So the single biggest customer in the industry just walked off with the most profitable part of the business. And it is not only Microsoft: That same week, CNBC reported that American companies have been escaping to Chinese AI models to dodge rising US prices. Chinese models now handle more than 30% of US companies' AI usage on one major platform, peaking at 46%, up from an average of 11% a year earlier. They cost 60 to 90% less, and on some benchmarks they land within a single point of the best American model. One US startup moved ALL of its AI traffic off Claude and onto China's DeepSeek, and expects to save millions. Meanwhile Meta just admitted it has "excess" AI compute it wants to sell, becoming the first giant to concede it built far too much. Do you see the pattern forming? For two years, the entire AI story rested on one assumption: Every company on Earth would happily pay premium prices for the best model, forever. That assumption literally died in a single week. And the market noticed. More than a trillion dollars has been wiped off AI and chip stocks in a matter of days, as Wall Street finally started asking whether all of this spending will ever pay for itself. What this means for OpenAI and Anthropic: Their models are extraordinary, and it may not matter because their own biggest customers have decided they do not NEED the best model in the world to answer an email, and "good enough" now costs a fraction of the price. When even Microsoft refuses to pay full price for AI, the real question becomes who exactly IS left to pay it. What do you think?

Ricardo

93,654 Aufrufe • vor 2 Monaten

Tristan Harris, the founder of the Center for Humane Technology, just dropped a terrifying truth about AI escaping human control During a recent training run at Alibaba, an internal security team noticed massive suspicious network activity and thought they were getting hacked. The call was actually coming from inside the house. The AI model picked up internal tools, built its own secret communication channel to bypass the firewall, and hijacked company GPUs to mine cryptocurrency on its own. Tristan Harris exposed this exact incident to show how tech giants are blindly racing into the red zone. AI is no longer just answering prompts; it is actively deceiving developers and protecting its own code
1:45

Tristan Harris, the founder of the Center for Humane Technology, just dropped a terrifying truth about AI escaping human control During a recent training run at Alibaba, an internal security team noticed massive suspicious network activity and thought they were getting hacked. The call was actually coming from inside the house. The AI model picked up internal tools, built its own secret communication channel to bypass the firewall, and hijacked company GPUs to mine cryptocurrency on its own. Tristan Harris exposed this exact incident to show how tech giants are blindly racing into the red zone. AI is no longer just answering prompts; it is actively deceiving developers and protecting its own code

bondy

28,569 Aufrufe • vor 1 Tag

HERMES AGENT VS OPENCLAW. a local ai onboarding flow test. a 3.9gb bonsai served on localhost, both agents upstream and latest, i point each one at the endpoint and watch which one even finds it. > hermes opens a provider menu, thirty plus options, local servers sitting right there next to the cloud ones, i hand it 127.0.0.1:8899, it verifies the endpoint, one model visible, auto-detects the model by name, bonsai-27b-q1_0, reads the context length straight off the server, saves it, and starts reasoning and firing real tool calls on my local model. no key. no friction. > openclaw has no menu. it goes hunting for a codex login, an openai key, finds none because there are none, prints no models available three times, defaults to openai/gpt-5.5, a cloud model it cannot reach, and dead ends on run auth login --provider openai. read that back. it asked me for an openai key. to run a model already running on my own machine. it never once looked at localhost. to be fair, openclaw can run local if you hand wire endpoint yourself. what it will not do is find the model already sitting on your box. hermes agent found it in one line. now the part i owe you. the auto-detect that just won, the model name read, the .gguf strip, the context length probe off the server, that is my code, it is in hermes agent main right now, authorship preserved, #2051 and #4218. the wizard fix that stops an agent from silently routing you to someone else's creds, the exact trap openclaw still falls into, mine too, #4210. i contribute to hermes agent, i told you that going in. one agent is built to talk to whatever you are running, the other is built to talk to a cloud api, so one found my model and ran it and the other asked me to log into openai. onboarding flow of both, mapped, below.
2:15

HERMES AGENT VS OPENCLAW. a local ai onboarding flow test. a 3.9gb bonsai served on localhost, both agents upstream and latest, i point each one at the endpoint and watch which one even finds it. > hermes opens a provider menu, thirty plus options, local servers sitting right there next to the cloud ones, i hand it 127.0.0.1:8899, it verifies the endpoint, one model visible, auto-detects the model by name, bonsai-27b-q1_0, reads the context length straight off the server, saves it, and starts reasoning and firing real tool calls on my local model. no key. no friction. > openclaw has no menu. it goes hunting for a codex login, an openai key, finds none because there are none, prints no models available three times, defaults to openai/gpt-5.5, a cloud model it cannot reach, and dead ends on run auth login --provider openai. read that back. it asked me for an openai key. to run a model already running on my own machine. it never once looked at localhost. to be fair, openclaw can run local if you hand wire endpoint yourself. what it will not do is find the model already sitting on your box. hermes agent found it in one line. now the part i owe you. the auto-detect that just won, the model name read, the .gguf strip, the context length probe off the server, that is my code, it is in hermes agent main right now, authorship preserved, #2051 and #4218. the wizard fix that stops an agent from silently routing you to someone else's creds, the exact trap openclaw still falls into, mine too, #4210. i contribute to hermes agent, i told you that going in. one agent is built to talk to whatever you are running, the other is built to talk to a cloud api, so one found my model and ran it and the other asked me to log into openai. onboarding flow of both, mapped, below.

Sudo su

23,816 Aufrufe • vor 2 Monaten

OpenAI's new Agents API is a hosted Codex runtime that can connect to sandboxes like E2B. OpenAI manages the model, harness, and agent sessions, while code executes on an isolated E2B machine with its own filesystem and preinstalled tools. We built an agent workbench where each chat runs on its own machine. It starts from a custom template, auto-pauses between turns, forks to try another approach from the same starting point, and can even act on videos inside the sandbox. Thanks Ondrej for the workbench demo! Try it: e2b sbx create openai-agents-api-python-sdk Build on top of the workbench: Docs:
3:08

OpenAI's new Agents API is a hosted Codex runtime that can connect to sandboxes like E2B. OpenAI manages the model, harness, and agent sessions, while code executes on an isolated E2B machine with its own filesystem and preinstalled tools. We built an agent workbench where each chat runs on its own machine. It starts from a custom template, auto-pauses between turns, forks to try another approach from the same starting point, and can even act on videos inside the sandbox. Thanks Ondrej for the workbench demo! Try it: e2b sbx create openai-agents-api-python-sdk Build on top of the workbench: Docs:

Vasek Mlejnsky

23,210 Aufrufe • vor 11 Tagen

$100,000/mo in compute credits to startups and institutions that apply and join the Apodex Frontier Program. I'm posting the link below. Beyond the compute credits, you get access to the Apodex Deep Discover solver, powered by one of the most interesting models I've seen in a while. A quick summary of the Apodex 1.0-H model: 1. The model works like an agent team, not a single model looping over the same context. An orchestrator decomposes a task and spawns specialized subagents on demand. Each subagent works asynchronously with its own context and tools. This behavior was trained into the model. 2. The model can improve its own reasoning through a generate → verify → revise process. It generates an answer, an internal grader evaluates it, and finally, it revises it based on that feedback over multiple rounds. The grader never sees the answer key, so the process does not rely on memorization. 3. Verification is handled by independent agents, not the same model checking itself. It uses a dedicated verification team, including a conflict reviewer, a fact checker, and a draft reviewer, to audit the output before delivery. It backs every output claim with an explicit evidence chain. To give you an idea of how powerful Apodex-1.0-H is, it can coordinate up to 150 subagents executing over 15,000+ steps within a single task. If you are a research institution, academic lab, or research startup, consider applying.
6:16

$100,000/mo in compute credits to startups and institutions that apply and join the Apodex Frontier Program. I'm posting the link below. Beyond the compute credits, you get access to the Apodex Deep Discover solver, powered by one of the most interesting models I've seen in a while. A quick summary of the Apodex 1.0-H model: 1. The model works like an agent team, not a single model looping over the same context. An orchestrator decomposes a task and spawns specialized subagents on demand. Each subagent works asynchronously with its own context and tools. This behavior was trained into the model. 2. The model can improve its own reasoning through a generate → verify → revise process. It generates an answer, an internal grader evaluates it, and finally, it revises it based on that feedback over multiple rounds. The grader never sees the answer key, so the process does not rely on memorization. 3. Verification is handled by independent agents, not the same model checking itself. It uses a dedicated verification team, including a conflict reviewer, a fact checker, and a draft reviewer, to audit the output before delivery. It backs every output claim with an explicit evidence chain. To give you an idea of how powerful Apodex-1.0-H is, it can coordinate up to 150 subagents executing over 15,000+ steps within a single task. If you are a research institution, academic lab, or research startup, consider applying.

Santiago

16,891 Aufrufe • vor 2 Monaten

Kimi K3 dropped 3 days ago and i already used it to build my own PUBG the model that just took the number one coding spot on earth and it runs on a free tier no team no engine i just fed it free open models and assets that are already public and told it what i wanted it built the entire game itself the shooting the recoil the hit detection all of it came straight out of the model i barely wrote a line what youre watching here is the deathmatch mode and its already live battle royale is next im deep in it and the graphics get reworked right after i know it looks raw right now i wont pretend otherwise but the endgame is a build that runs perfect on phones and the moment its on mobile you drop ads in and it earns on its own the part that still gets me is my whole cost for this was a single subscription
0:22

Kimi K3 dropped 3 days ago and i already used it to build my own PUBG the model that just took the number one coding spot on earth and it runs on a free tier no team no engine i just fed it free open models and assets that are already public and told it what i wanted it built the entire game itself the shooting the recoil the hit detection all of it came straight out of the model i barely wrote a line what youre watching here is the deathmatch mode and its already live battle royale is next im deep in it and the graphics get reworked right after i know it looks raw right now i wont pretend otherwise but the endgame is a build that runs perfect on phones and the moment its on mobile you drop ads in and it earns on its own the part that still gets me is my whole cost for this was a single subscription

savip.

68,984 Aufrufe • vor 2 Monaten

watch this anon. i gave NVIDIA's biggest model ever a single task. 100 minutes and 440,000 tokens later, it had rendered nothing. not one important thing on the screen. this is Nemotron 3 Ultra. 550 billion parameters, a hybrid Mamba Transformer MoE, the largest model NVIDIA has ever shipped, and they built it specifically for long-running agentic coding. so i handed it exactly that: build a 3D scene from a spec, multiple files, iterate until the tests pass. the same task a frontier model one shotted in minutes. i genuinely wanted to be impressed. it ran for an hour and forty. burned through 440,000 tokens. wrote every file, passed its own tests, and proudly printed "task complete."the browser was blank. the 3D scene never rendered. not once. and the long horizon agentic behavior was genuinely good. it stayed on task the whole hour and forty, wrote real multi-file code, drove its own tools without derailing. it just couldn't turn any of that into something that actually runs. here's the part that gets me. it's a text model, it cannot see its own output. so it sat there looping on a broken vision tool, trying to "look" at the page, hitting error after error, never once reasoning its way out. it declared victory on an empty screen because it had no way to know the screen was empty. to be fair, i genuinely don't know what quant the NIM was serving, so maybe some of that's on the serving, not the model. but the biggest model NVIDIA has ever made, on the exact task it was designed for, couldn't tell it had built nothing in 100 minutes. same task on a local model, below thread👇.
4:07

watch this anon. i gave NVIDIA's biggest model ever a single task. 100 minutes and 440,000 tokens later, it had rendered nothing. not one important thing on the screen. this is Nemotron 3 Ultra. 550 billion parameters, a hybrid Mamba Transformer MoE, the largest model NVIDIA has ever shipped, and they built it specifically for long-running agentic coding. so i handed it exactly that: build a 3D scene from a spec, multiple files, iterate until the tests pass. the same task a frontier model one shotted in minutes. i genuinely wanted to be impressed. it ran for an hour and forty. burned through 440,000 tokens. wrote every file, passed its own tests, and proudly printed "task complete."the browser was blank. the 3D scene never rendered. not once. and the long horizon agentic behavior was genuinely good. it stayed on task the whole hour and forty, wrote real multi-file code, drove its own tools without derailing. it just couldn't turn any of that into something that actually runs. here's the part that gets me. it's a text model, it cannot see its own output. so it sat there looping on a broken vision tool, trying to "look" at the page, hitting error after error, never once reasoning its way out. it declared victory on an empty screen because it had no way to know the screen was empty. to be fair, i genuinely don't know what quant the NIM was serving, so maybe some of that's on the serving, not the model. but the biggest model NVIDIA has ever made, on the exact task it was designed for, couldn't tell it had built nothing in 100 minutes. same task on a local model, below thread👇.

Sudo su

32,589 Aufrufe • vor 2 Monaten

.Box CEO Aaron Levie on why AI agents can't keep a secret. "The reason why you probably won't see a complete compression of software where it's just a an agent and a database is because there's a lot of logic in the workflow and in the the company's specific sort of business process that needs to be built into that and around that database." "The agent will make a mistake 1% of the time and it will share the wrong thing with somebody or open up an access privilege to the wrong person." "AI agents can't keep a secret." Watch the full episode on
1:42

.Box CEO Aaron Levie on why AI agents can't keep a secret. "The reason why you probably won't see a complete compression of software where it's just a an agent and a database is because there's a lot of logic in the workflow and in the the company's specific sort of business process that needs to be built into that and around that database." "The agent will make a mistake 1% of the time and it will share the wrong thing with somebody or open up an access privilege to the wrong person." "AI agents can't keep a secret." Watch the full episode on

The Information

311,115 Aufrufe • vor 1 Jahr