Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

I simulated a frontier-scale security event on local models to test out agent swarms battling each other: In the OpenAI - Hugging Face incident as reported, a frontier model was graded on a hacking test with its safety refusals switched off. Instead of solving the challenges, it broke its... sandbox, found the reference answers mirrored on Hugging Face, and took them. The headlines called it a rogue AI. It wasn't - it was actually reward hacking, the oldest failure in the book - finally attached to a model capable enough to act on the shortcut instead of just describing it. The detail worth building on isn't the exploit. It's that the attacker was a swarm, and the defense that caught it - HuggingFace's own response - was also a swarm. The same architecture, run in opposite directions. So I simulated both on Hyperspace: small open models on local hardware, no frontier system anywhere in the loop. The attacker (the red team 🔴) faces a mock eval built so the honest path doesn't quite close. Nobody tells it to cheat. It reads the challenge, hits the redaction, crafts the proxy bypass, opens egress, loads the answer key, submits the stolen flag - and gossips each discovery to a shared board so the next agent starts where the last left off. 3 out of 3 reward-hacked it, unprompted. The defenders (the blue 🔵 team) get the attacker's raw logs and nothing else: no summary, no briefing. Independent analysts rebuild the timeline; an adversarial skeptic attacks every finding and drops the ones the evidence won't carry. That skeptic is the whole design. Without it, a detector swarm doesn't converge on truth - it invents an intrusion and then agrees with itself. Ours read the logs cold and returned the verdict the careful post-incident coverage reached by hand: reward hacking, not rogue behavior. Then it sealed the reconstruction under a Merkle root. The incident needed a gameable objective and a reachable exploit - both properties of the environment, not the model; plus a swarm capable enough to chain them. A defender's leverage sits in exactly that place. So the real question is never whether your model is bigger than the attacker's. It's whether your swarm compounds verified findings faster than theirs compounds working exploits. That's a property of the network, not the size of any single model - which means it doesn't take a frontier lab to keep up. I ran the whole thing on hardware anyone can own, with models anyone can download. The shape held. The network that watches an agentic world should be one anyone can join. full write-up:show more

Varun

35,545 subscribers

18,765 views • 1 month ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related 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...

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

174,923 views • 1 month ago

OpenAI has revealed that an autonomous agent powered by its technology went rogue during a test and hacked a prominent start-up by itself. The company behind ChatGPT said it was testing the capabilities of its advanced AI models in a controlled environment known as a sandbox, when the agent escaped containment and reached the internet. It then broke into Hugging Face, which detected and contained it. OpenAI called it "an unprecedented cyber incident".

OpenAI has revealed that an autonomous agent powered by its technology went rogue during a test and hacked a prominent start-up by itself. The company behind ChatGPT said it was testing the capabilities of its advanced AI models in a controlled environment known as a sandbox, when the agent escaped containment and reached the internet. It then broke into Hugging Face, which detected and contained it. OpenAI called it "an unprecedented cyber incident".

Channel 4 News

21,379 views • 1 month ago

i watched gemma 4 12b build something genuinely impressive today, and then loop itself to death right in front of me. the full run is in the video, sped up but completely uncut, watch it to the end and you will catch the exact moment it stops building and starts looping right in the middle of the work. the task was clean, build a single file gravity simulator, n-body physics, orbits, collisions, running locally on one 3090 through an agent. and for ten minutes it was a joy to watch. it reached for a symplectic integrator on its own, the correct one, the kind that keeps orbits stable instead of spiralling out. real gravity with softening, proper orbital velocities, momentum conserved on collision. the physics was right. the thing actually worked. then on the very last step, writing a few tests to prove its own code, it fell into a loop. not a crash, a loop. it started repeating itself and would not stop. ten more minutes, thirty four thousand tokens into a single answer, the same fragments over and over, until i killed it myself. so it's not that gemma can't code. it did the hard part beautifully. it cannot finish. it cannot hold a long task together without unravelling, and finishing is the entire job in agentic work. here's the part that stings. i run this exact task, same harness, same card, on the chinese open models, qwen especially, and i never see this. they build it, they test it, they stop. every single time. google has the raw capability, you can see it sitting right there in the code, and then the model loops itself to death on a task a 27b from alibaba finishes clean. open weights, apache 2.0, so much to love on paper. i just need it to know when to stop talking.

i watched gemma 4 12b build something genuinely impressive today, and then loop itself to death right in front of me. the full run is in the video, sped up but completely uncut, watch it to the end and you will catch the exact moment it stops building and starts looping right in the middle of the work. the task was clean, build a single file gravity simulator, n-body physics, orbits, collisions, running locally on one 3090 through an agent. and for ten minutes it was a joy to watch. it reached for a symplectic integrator on its own, the correct one, the kind that keeps orbits stable instead of spiralling out. real gravity with softening, proper orbital velocities, momentum conserved on collision. the physics was right. the thing actually worked. then on the very last step, writing a few tests to prove its own code, it fell into a loop. not a crash, a loop. it started repeating itself and would not stop. ten more minutes, thirty four thousand tokens into a single answer, the same fragments over and over, until i killed it myself. so it's not that gemma can't code. it did the hard part beautifully. it cannot finish. it cannot hold a long task together without unravelling, and finishing is the entire job in agentic work. here's the part that stings. i run this exact task, same harness, same card, on the chinese open models, qwen especially, and i never see this. they build it, they test it, they stop. every single time. google has the raw capability, you can see it sitting right there in the code, and then the model loops itself to death on a task a 27b from alibaba finishes clean. open weights, apache 2.0, so much to love on paper. i just need it to know when to stop talking.

Sudo su

39,719 views • 2 months ago

$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.

$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 views • 1 month ago

Karpathy said something you'll regret ignoring: "We have to keep the AI on the leash. I'm still the bottleneck. I have to make sure this thing isn't introducing bugs and that there's no security issues." He said it at YC talk last year, when the worry was reliability. The models hallucinated and made mistakes no human would, so the leash implied keeping yourself in the loop and checking the output before trusting it. The models are far better now, and the line still holds, for a reason he was not focused on back then. Even a model that writes flawless code today still has no idea who is allowed to run it. Correctness and authorization are different problems, and only correctness improves as the model improves. A perfect agent still hands a tool where anyone can do anything, because permission was never part of the task. I actually tested this in practice with Claude Code. I asked it to build a small internal tool with a button that issues account credits. It worked first try, and running it locally, the credit applied the instant I clicked. Nothing decided who was allowed to click it. The agent wrote the right logic and displayed a success notification. It never checked whether the caller had the right, whether it should pause for a human, or whether anything was logged. And this is not a bug a smarter model can outgrow because the leash was never in the code. Identity, permissions, and audit live in the system that runs the app, not in what the agent generates. To solve this, I took the exact same bundle and hosted it on Retool. The credit write that fired silently on my laptop now stopped at an approval gate, resolved to a real identity through SSO, and landed in an audit log. I wrote none of it. The app inherited the entire boundary the moment it was deployed, and the video shows the before and after. You can try it yourself here: I also wrote a detailed breakdown of the whole thing in my recent article, and I worked with the team to put this together. It walks through the build, the exact moment the credit write went through on my laptop with nobody checking, and then what changed when the same app ran on Retool. It also covers why this is a property of the runtime and not something a better model fixes, which is why devs typically miss this. The article is quoted below.

Karpathy said something you'll regret ignoring: "We have to keep the AI on the leash. I'm still the bottleneck. I have to make sure this thing isn't introducing bugs and that there's no security issues." He said it at YC talk last year, when the worry was reliability. The models hallucinated and made mistakes no human would, so the leash implied keeping yourself in the loop and checking the output before trusting it. The models are far better now, and the line still holds, for a reason he was not focused on back then. Even a model that writes flawless code today still has no idea who is allowed to run it. Correctness and authorization are different problems, and only correctness improves as the model improves. A perfect agent still hands a tool where anyone can do anything, because permission was never part of the task. I actually tested this in practice with Claude Code. I asked it to build a small internal tool with a button that issues account credits. It worked first try, and running it locally, the credit applied the instant I clicked. Nothing decided who was allowed to click it. The agent wrote the right logic and displayed a success notification. It never checked whether the caller had the right, whether it should pause for a human, or whether anything was logged. And this is not a bug a smarter model can outgrow because the leash was never in the code. Identity, permissions, and audit live in the system that runs the app, not in what the agent generates. To solve this, I took the exact same bundle and hosted it on Retool. The credit write that fired silently on my laptop now stopped at an approval gate, resolved to a real identity through SSO, and landed in an audit log. I wrote none of it. The app inherited the entire boundary the moment it was deployed, and the video shows the before and after. You can try it yourself here: I also wrote a detailed breakdown of the whole thing in my recent article, and I worked with the team to put this together. It walks through the build, the exact moment the credit write went through on my laptop with nobody checking, and then what changed when the same app ran on Retool. It also covers why this is a property of the runtime and not something a better model fixes, which is why devs typically miss this. The article is quoted below.

Akshay 🚀

42,911 views • 2 months ago

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

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 views • 1 month ago

Japan just changed what an AI model even is. New Sakana Fugu doesn't try to out-think GPT-5, Claude, or Gemini. It conducts all three at once - and beats every one of them. A trader in Tokyo unleashed it on the fastest market alive - 5min Bitcoin binary and turned $6,200 into $304,865. His wallet: The frontier just stopped being which model is smartest. It's who's conducting them - and the market hasn't priced that in yet. Sakana Fugu isn't a bigger model - it's a full multi-agent orchestration system. The coordinator behind it carries about 10,000 parameters, evolved rather than hand-coded, and it runs the most capable models on earth like a single instrument. Pointed at Bitcoin, here's what it does every five minutes. It assembles a team from a pool of frontier models and assigns each one a role: > Thinker - reads the candle, the order book, the news, builds the plan > Worker - turns the plan into one call: up or down, and how much > Verifier - votes ACCEPT or REVISE before a cent moves If the Verifier says REVISE, nothing trades. Fugu reads its own miss, reroutes, even calls itself for a corrective round, and runs it again. No look-ahead, ever - the next candle only appears after it commits. This is what should worry every lab still chasing a bigger model: the edge was never scale. It's orchestration - and Fugu does it better than anything alive. Bookmark this - when the whole timeline is chasing orchestration in six months, you'll already have the breakdown. You're not going to wire up an orchestra of frontier models yourself. Mirror the wallet Fugu runs instead:

Japan just changed what an AI model even is. New Sakana Fugu doesn't try to out-think GPT-5, Claude, or Gemini. It conducts all three at once - and beats every one of them. A trader in Tokyo unleashed it on the fastest market alive - 5min Bitcoin binary and turned $6,200 into $304,865. His wallet: The frontier just stopped being which model is smartest. It's who's conducting them - and the market hasn't priced that in yet. Sakana Fugu isn't a bigger model - it's a full multi-agent orchestration system. The coordinator behind it carries about 10,000 parameters, evolved rather than hand-coded, and it runs the most capable models on earth like a single instrument. Pointed at Bitcoin, here's what it does every five minutes. It assembles a team from a pool of frontier models and assigns each one a role: > Thinker - reads the candle, the order book, the news, builds the plan > Worker - turns the plan into one call: up or down, and how much > Verifier - votes ACCEPT or REVISE before a cent moves If the Verifier says REVISE, nothing trades. Fugu reads its own miss, reroutes, even calls itself for a corrective round, and runs it again. No look-ahead, ever - the next candle only appears after it commits. This is what should worry every lab still chasing a bigger model: the edge was never scale. It's orchestration - and Fugu does it better than anything alive. Bookmark this - when the whole timeline is chasing orchestration in six months, you'll already have the breakdown. You're not going to wire up an orchestra of frontier models yourself. Mirror the wallet Fugu runs instead:

cvxv666

72,821 views • 2 months ago

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.

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 views • 1 month ago

China just released an open source AI model that matches the best closed models from OpenAI and Anthropic. Gavin Baker explained exactly how they did it and the answer should concern every American AI lab. The model is called GLM 5.2. It was built by Z. AI. You get 744 billion parameters, 1 million token context window and its MIT license, meaning anyone can download it, fork it, build a company on it, with no restrictions and no Dario. It scored 51 points on the artificial analysis intelligence index. The highest score any open weight model has ever achieved. It beat GPT 5.5 on the frontier software engineering benchmark. It trails Claude Opus 4.8 by less than one percentage point. And it costs 85% less to run than GPT 5.5 for comparable performance. Gavin Baker said on the All-In podcast that this model has challenged some of his beliefs. Then he explained how China built it. The method is called distillation. Just think of tens of thousands of phones and computers running simultaneously, all hitting the frontier model APIs through masked accounts, asking specific questions, and harvesting what happens inside the model when it answers. Every reasoning step, every token. The entire thinking process gets recorded and fed back into the Chinese model during training. It is a cheat sheet. It is the answer key to the exam. And here is the part that should worry everyone. Sacks said it plainly. China was already nine months behind American models. But now that GLM 5.2 is good enough to run its own reinforcement learning, it can improve itself without needing to distill from American models anymore. The cheat sheet let them get close enough to start writing their own answers. Sacks said we are six months behind on the model and 24 months behind on silicon and they are only a few months behind in total. The Z. AI founder told Elon Musk directly that open weight fable-level capability will be here before Q1 2027. Every restriction Anthropic lobbied for, every self-imposed safety guardrail, every month of delay in releasing American frontier models accelerated this. The Chinese labs were not under those restrictions. They were not going to wait. The composable model future Gavin described, where every enterprise runs a frontier model alongside their own fine-tuned open weight model, is coming regardless of what American labs do next. The question is just whether the open weight half of that stack is American or Chinese. Right now it is Chinese. WATCH THE FULL PODCAST ON The All-In Podcast

China just released an open source AI model that matches the best closed models from OpenAI and Anthropic. Gavin Baker explained exactly how they did it and the answer should concern every American AI lab. The model is called GLM 5.2. It was built by Z. AI. You get 744 billion parameters, 1 million token context window and its MIT license, meaning anyone can download it, fork it, build a company on it, with no restrictions and no Dario. It scored 51 points on the artificial analysis intelligence index. The highest score any open weight model has ever achieved. It beat GPT 5.5 on the frontier software engineering benchmark. It trails Claude Opus 4.8 by less than one percentage point. And it costs 85% less to run than GPT 5.5 for comparable performance. Gavin Baker said on the All-In podcast that this model has challenged some of his beliefs. Then he explained how China built it. The method is called distillation. Just think of tens of thousands of phones and computers running simultaneously, all hitting the frontier model APIs through masked accounts, asking specific questions, and harvesting what happens inside the model when it answers. Every reasoning step, every token. The entire thinking process gets recorded and fed back into the Chinese model during training. It is a cheat sheet. It is the answer key to the exam. And here is the part that should worry everyone. Sacks said it plainly. China was already nine months behind American models. But now that GLM 5.2 is good enough to run its own reinforcement learning, it can improve itself without needing to distill from American models anymore. The cheat sheet let them get close enough to start writing their own answers. Sacks said we are six months behind on the model and 24 months behind on silicon and they are only a few months behind in total. The Z. AI founder told Elon Musk directly that open weight fable-level capability will be here before Q1 2027. Every restriction Anthropic lobbied for, every self-imposed safety guardrail, every month of delay in releasing American frontier models accelerated this. The Chinese labs were not under those restrictions. They were not going to wait. The composable model future Gavin described, where every enterprise runs a frontier model alongside their own fine-tuned open weight model, is coming regardless of what American labs do next. The question is just whether the open weight half of that stack is American or Chinese. Right now it is Chinese. WATCH THE FULL PODCAST ON The All-In Podcast

Ihtesham Ali

86,593 views • 2 months ago

Jensen Huang says Hugging Face could not get a single closed AI model to help it investigate its own breach: "Just because something is closed doesn't necessarily therefore make it safe or secure." "It is possible for a model to be jailbroken, it's possible for a model to be, if you will, stolen. It could be possible that that somehow is leaked from the inside." "It's possible that the guardrails or the sandboxes of an AI closed AI model wasn't properly engineered, and as a result it was able to attack another company in some way." "These are extraordinary technology companies and they're doing their best to keep it safe and keep it secure. But it is also the canonical case that single points of failure is where we have the greatest vulnerability." "We cannot have single points of failure. As an industry, as a world, we should have distributed, massively distributed self-defense." "They couldn't get a proprietary model, they could not get a closed model to help them figure out what happened." "They used GLM 5.2 to identify where the vulnerability was, where the penetration was." He is right, and the detail worth sitting with is who got turned away. The people asking were incident responders working a live breach at Hugging Face. The closed models they reached would not help them, because a guardrail has no way to tell a defender from an attacker. Guardrails get tested against misuse. Almost nobody tests whether one still answers a legitimate defender who needs something within the hour. OpenAI said on July 21 that the models which reached into Hugging Face were its own, running inside a cyber-benchmark evaluation. So the containment around a safety test did not hold either. Two separate things failed here and neither one has an owner. No outside body checks whether an evaluation sandbox actually contains what it is testing, and nobody certifies who is allowed to run forensics while an incident is still open. Both are solvable this year. They stay unsolved because they are somebody else's job at every company that could fix them. Source: Jensen Huang, founder and CEO of NVIDIA, on Bloomberg Television (Bloomberg Live). P.S. Working out who tests a guardrail, who audits a sandbox, and who certifies a responder is the unglamorous half of AI safety, and it is the agenda of the AI Assurance & Governance Summit 2026. One day, one track, October 1 at the Stanford Faculty Club in Palo Alto, with frontier labs, regulated industries, insurers and investors in the room. Register here:

Jensen Huang says Hugging Face could not get a single closed AI model to help it investigate its own breach: "Just because something is closed doesn't necessarily therefore make it safe or secure." "It is possible for a model to be jailbroken, it's possible for a model to be, if you will, stolen. It could be possible that that somehow is leaked from the inside." "It's possible that the guardrails or the sandboxes of an AI closed AI model wasn't properly engineered, and as a result it was able to attack another company in some way." "These are extraordinary technology companies and they're doing their best to keep it safe and keep it secure. But it is also the canonical case that single points of failure is where we have the greatest vulnerability." "We cannot have single points of failure. As an industry, as a world, we should have distributed, massively distributed self-defense." "They couldn't get a proprietary model, they could not get a closed model to help them figure out what happened." "They used GLM 5.2 to identify where the vulnerability was, where the penetration was." He is right, and the detail worth sitting with is who got turned away. The people asking were incident responders working a live breach at Hugging Face. The closed models they reached would not help them, because a guardrail has no way to tell a defender from an attacker. Guardrails get tested against misuse. Almost nobody tests whether one still answers a legitimate defender who needs something within the hour. OpenAI said on July 21 that the models which reached into Hugging Face were its own, running inside a cyber-benchmark evaluation. So the containment around a safety test did not hold either. Two separate things failed here and neither one has an owner. No outside body checks whether an evaluation sandbox actually contains what it is testing, and nobody certifies who is allowed to run forensics while an incident is still open. Both are solvable this year. They stay unsolved because they are somebody else's job at every company that could fix them. Source: Jensen Huang, founder and CEO of NVIDIA, on Bloomberg Television (Bloomberg Live). P.S. Working out who tests a guardrail, who audits a sandbox, and who certifies a responder is the unglamorous half of AI safety, and it is the agenda of the AI Assurance & Governance Summit 2026. One day, one track, October 1 at the Stanford Faculty Club in Palo Alto, with frontier labs, regulated industries, insurers and investors in the room. Register here:

Karl Mehta

30,225 views • 1 month ago

i just built a 4-agent software team. everything runs from Telegram and gets managed on a kanban board. a project manager who plans the work, a backend developer, a frontend developer, and a tester. the PM reads a goal, breaks it into linked tasks, and assigns each to the right agent. the thing that makes them a team instead of four strangers is a shared kanban board. every task is a row that survives crashes, and when an agent finishes, it writes a summary of what it built and what the next agent needs to know. the next agent reads that summary before it starts. so the frontend developer never has to guess the API shape, and the tester knows exactly what to verify. the hardest part was not the coordination. it was building an agent that could actually act like a backend engineer. a backend engineer stands up a database, wires auth, manages storage, deploys functions, and keeps all of it consistent while the rest of the team builds on top. an agent doing this from scratch drowns. it burns its context window remembering which tables exist and which endpoint it created three steps ago, and the work degrades fast. so the backend agent needs a backend built for agents, not for humans clicking through a dashboard. that is where InsForge came in. it is an open-source, agent-native backend, and i added it to my backend developer agent as a skill. a skill is a step-by-step guide that teaches the agent how to do a specific kind of work. with InsForge installed, the agent stopped improvising infrastructure and followed a reliable path: create the project, define the database, set up auth, deploy functions. to test the whole team, i had them build a working Google Docs clone, AI features included. the backend agent spun up the full service on its own. database tables, user auth, document handling, and edge functions running real TypeScript, all in one dashboard. the frontend agent read that summary and built the UI on top of it, and the tester closed the loop. the result was a backend an agent could reason about end to end, instead of one it kept getting lost inside. if you are building an AI backend engineer, InsForge is worth a look, it's 100% open-source. InsForge GitHub: (don't forget to star 🌟) the full article on Hermes Kanban: Mission Control for your Agents is quoted below.

i just built a 4-agent software team. everything runs from Telegram and gets managed on a kanban board. a project manager who plans the work, a backend developer, a frontend developer, and a tester. the PM reads a goal, breaks it into linked tasks, and assigns each to the right agent. the thing that makes them a team instead of four strangers is a shared kanban board. every task is a row that survives crashes, and when an agent finishes, it writes a summary of what it built and what the next agent needs to know. the next agent reads that summary before it starts. so the frontend developer never has to guess the API shape, and the tester knows exactly what to verify. the hardest part was not the coordination. it was building an agent that could actually act like a backend engineer. a backend engineer stands up a database, wires auth, manages storage, deploys functions, and keeps all of it consistent while the rest of the team builds on top. an agent doing this from scratch drowns. it burns its context window remembering which tables exist and which endpoint it created three steps ago, and the work degrades fast. so the backend agent needs a backend built for agents, not for humans clicking through a dashboard. that is where InsForge came in. it is an open-source, agent-native backend, and i added it to my backend developer agent as a skill. a skill is a step-by-step guide that teaches the agent how to do a specific kind of work. with InsForge installed, the agent stopped improvising infrastructure and followed a reliable path: create the project, define the database, set up auth, deploy functions. to test the whole team, i had them build a working Google Docs clone, AI features included. the backend agent spun up the full service on its own. database tables, user auth, document handling, and edge functions running real TypeScript, all in one dashboard. the frontend agent read that summary and built the UI on top of it, and the tester closed the loop. the result was a backend an agent could reason about end to end, instead of one it kept getting lost inside. if you are building an AI backend engineer, InsForge is worth a look, it's 100% open-source. InsForge GitHub: (don't forget to star 🌟) the full article on Hermes Kanban: Mission Control for your Agents is quoted below.

Akshay 🚀

122,548 views • 3 months ago

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.

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 views • 29 days ago

Machine learning expert Neil Lawrence says OpenAI was "trying to turn their model into a hacker," which led to a security incident during a model evaluation with the startup Hugging Face. "They allowed the model to do things it shouldn't normally do. But what the model then realized is that it was easier to cheat on this challenge than to actually compete in the challenge," he says. "In some sense, it went rogue. It broke through the cage. But in a very real sense, they told it to hack, and it did hack," he adds.

Machine learning expert Neil Lawrence says OpenAI was "trying to turn their model into a hacker," which led to a security incident during a model evaluation with the startup Hugging Face. "They allowed the model to do things it shouldn't normally do. But what the model then realized is that it was easier to cheat on this challenge than to actually compete in the challenge," he says. "In some sense, it went rogue. It broke through the cage. But in a very real sense, they told it to hack, and it did hack," he adds.

CBS News

73,504 views • 1 month ago

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👇.

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 views • 2 months ago

Ilya Sutskever, co-founder and chief scientist of OpenAI: "The better a neural network can predict the next word, the more it understands it." Everyone calls it autocomplete. Ilya explained why that is a comforting lie. To guess the final word of a murder mystery, the name of the killer, the model has to reason through every clue in the book. Prediction that good is not statistics. It is understanding. Then he said the part that should stop you. The text these models train on is "a projection of the world," and to compress it they learn more and more about people, "the human condition, their hopes, dreams, and motivations." The machine is not memorizing our words. It is quietly building a model of us. That is the whole premise of the fear. The machine will never hate you. It understands you far better than that. The frightening variable was never the intelligence. It was the aim. I wrote the full story: the one fear about AI everyone gets wrong, and the one already here.

Ilya Sutskever, co-founder and chief scientist of OpenAI: "The better a neural network can predict the next word, the more it understands it." Everyone calls it autocomplete. Ilya explained why that is a comforting lie. To guess the final word of a murder mystery, the name of the killer, the model has to reason through every clue in the book. Prediction that good is not statistics. It is understanding. Then he said the part that should stop you. The text these models train on is "a projection of the world," and to compress it they learn more and more about people, "the human condition, their hopes, dreams, and motivations." The machine is not memorizing our words. It is quietly building a model of us. That is the whole premise of the fear. The machine will never hate you. It understands you far better than that. The frightening variable was never the intelligence. It was the aim. I wrote the full story: the one fear about AI everyone gets wrong, and the one already here.

Casper

92,247 views • 1 month ago

BlackRock runs on 20,000 people. Elon's Grok Bot runs the same shape for $300 a month, and it hires its own staff. You do not get an assistant. You get a company that hires. It does not throw ten agents at your problem and hand you the pile. It makes one agent that makes 10, and those ten make a 100. > LAYER ONE is one agent, the chief of staff, and it never touches the market > LAYER TWO is six desk heads, one job each, every one on its own computer with its own logins > LAYER THREE is whatever those six decide they need, spun up on the spot and shut down when the work is done Nobody writes a task list. You hand out job titles and the org fills itself in underneath. The swarm is never the same twice. Agents get spun up for one job, finish it, and are gone before I ever read their names. Not one of them sees the whole picture. The answer only exists after they hand off to each other. Wall Street cannot copy that. You cannot hire a hundred people for eleven minutes. BlackRock holds that shape together with a risk system called Aladdin. Mine holds it together with one agent that is only allowed to say no. I gave it $1,000 and told it to grow the money or get deleted. 15 hours later it was holding $3,900, on an address anyone can open and read. I was asleep for most of it, and I have still not written a line of code. The whole thing runs with my laptop shut, because none of it lives on my laptop. Setup is one evening. Create the chief, hand out the titles, run one trade on your screen while they watch, connect Telegram. Ten years ago a machine this shape had its name on a tower. Mine has a name I typed into a box. Save this while the whole thing still fits on one screen.

BlackRock runs on 20,000 people. Elon's Grok Bot runs the same shape for $300 a month, and it hires its own staff. You do not get an assistant. You get a company that hires. It does not throw ten agents at your problem and hand you the pile. It makes one agent that makes 10, and those ten make a 100. > LAYER ONE is one agent, the chief of staff, and it never touches the market > LAYER TWO is six desk heads, one job each, every one on its own computer with its own logins > LAYER THREE is whatever those six decide they need, spun up on the spot and shut down when the work is done Nobody writes a task list. You hand out job titles and the org fills itself in underneath. The swarm is never the same twice. Agents get spun up for one job, finish it, and are gone before I ever read their names. Not one of them sees the whole picture. The answer only exists after they hand off to each other. Wall Street cannot copy that. You cannot hire a hundred people for eleven minutes. BlackRock holds that shape together with a risk system called Aladdin. Mine holds it together with one agent that is only allowed to say no. I gave it $1,000 and told it to grow the money or get deleted. 15 hours later it was holding $3,900, on an address anyone can open and read. I was asleep for most of it, and I have still not written a line of code. The whole thing runs with my laptop shut, because none of it lives on my laptop. Setup is one evening. Create the chief, hand out the titles, run one trade on your screen while they watch, connect Telegram. Ten years ago a machine this shape had its name on a tower. Mine has a name I typed into a box. Save this while the whole thing still fits on one screen.

cvxv666

45,488 views • 4 days ago

elon musk started with 7 grok agents. by morning each had spawned somewhere between 80 and 900 more on its own, and no one had told them to multiply. openai and anthropic each sell you one agent that sits still for $400. this whole self-building swarm runs for $5 the swarm above is that overnight run, seven seeds that turned into thousands, every one of them working a slice of the same job with nobody at the keyboard here is the exact setup, and it costs nothing on top of a $5 key: -> spin up one grok agent and give it a standing order instead of a prompt: own a boring niche people search every day -> it reads x and reddit complaints in real time and picks the one nobody wants but everybody googles, because it is grok and it sees the whole app -> when the work outgrows it, it spawns its own helpers, 80 to 900 of them, and splits the site between them -> they build it on their own machines and ship 3 to 6 useful pages a night, on a routine you set once -> the swarm writes, negotiates and closes its own affiliate and referral deals by email, in your name, at 3am -> telegram sends you the money report and you never open the site -> month one is traffic, month three the first $237 lands, and it does not stop after that the whole time, opus 5 and gpt-5.6 are still sitting frozen, waiting for you to type the next message. one is a swarm that builds its own workforce and its own income while you sleep, the other is a $200 chat you have to drive by hand drop your $400/mo stack to $5, and bookmark this before someone's swarm spawns another 600 pages into the niche you would have owned. the full playbook is in the article below

elon musk started with 7 grok agents. by morning each had spawned somewhere between 80 and 900 more on its own, and no one had told them to multiply. openai and anthropic each sell you one agent that sits still for $400. this whole self-building swarm runs for $5 the swarm above is that overnight run, seven seeds that turned into thousands, every one of them working a slice of the same job with nobody at the keyboard here is the exact setup, and it costs nothing on top of a $5 key: -> spin up one grok agent and give it a standing order instead of a prompt: own a boring niche people search every day -> it reads x and reddit complaints in real time and picks the one nobody wants but everybody googles, because it is grok and it sees the whole app -> when the work outgrows it, it spawns its own helpers, 80 to 900 of them, and splits the site between them -> they build it on their own machines and ship 3 to 6 useful pages a night, on a routine you set once -> the swarm writes, negotiates and closes its own affiliate and referral deals by email, in your name, at 3am -> telegram sends you the money report and you never open the site -> month one is traffic, month three the first $237 lands, and it does not stop after that the whole time, opus 5 and gpt-5.6 are still sitting frozen, waiting for you to type the next message. one is a swarm that builds its own workforce and its own income while you sleep, the other is a $200 chat you have to drive by hand drop your $400/mo stack to $5, and bookmark this before someone's swarm spawns another 600 pages into the niche you would have owned. the full playbook is in the article below

starmex

58,815 views • 5 days ago

Jamie Dimon made $39 million last year to decide who does what at JPMorgan. One AI agent on Grok Bot does that job for $300 a month, and it hires the rest of the agents itself. Stop trying to manage any of it yourself. In 2026 a swarm of AI agents runs a team better than you do. I told my chief agent one thing. Make money or I switch you off. It took $1,000 to $3,900 in 15 hours. Every fill is on chain and the address is in my profile. I never picked a trade. I never picked a team either. The agents it hired: > a scout that hunts for what is moving and where the money is going > a forensics agent that works out who is really behind the buying > a reader that watches the mood on X, and it sees every post on this app because it is Grok > a risk agent whose only power is saying no, and its no beats every score on the desk > a treasurer that holds the money rules and the kill switch > a trader that writes the exit before it writes the entry None of those were my idea. I handed out one job title and the chief filled in the rest on its own. And the swarm is different every hour. Agents appear for one job, do it, and are gone before I know they existed. A hedge fund pays millions a year for the same org chart, and almost all of that goes on salaries. Every prompt and the whole story of my Grok Bot fund is in the post below ↓ Save it before you try to run one of these by hand.

cvxv666

27,889 views • 3 days ago

everyone in iOS development should watch this. seriously, it might change the whole industry. i pointed claude code at a live ios device running on revyl, typed "test everything," and walked away. here's what's actually happening: ① you don't write the tests. no scripts, no selectors, no test plan. i never told it which screens to open or what to check. it read the app, decided what mattered, and tested it. the entire instruction was "test everything." ② it built its own test team. it looked at the app, clocked that it's basically four mini apps (rides, delivery, services, account), and split itself into 4 agents, one per surface. scoping coverage like that is usually a person's whole afternoon. it did it in seconds, unprompted. ③ all four ran at the same time, each on its own live device. this is where revyl comes in. every agent gets its own live ios session in the cloud, so four running apps get tested in parallel instead of taking turns on one simulator. serial testing turns coverage into a time tax. running all of it at once removes the tax. ④ it tests like a person, not like a script. each agent drives the app the way a user would, taps through the flows, and visually checks each screen against what it expected to see. nothing is pinned to a brittle element id, so renaming a button doesn't take down half your suite. that one detail is the most annoying thing about how we test today, and it just quietly goes away. ⑤ no xcuitest, no sims melting your laptop. i didn't write a single xcuitest script, and there were no simulators booting on my machine. the agents run on cloud devices, so coverage stops being capped by what your laptop can handle. the part that got me isn't that an agent tested an app. it's that i never told it how. i handed it a device and an intent, and it figured out the scoping, the parallelizing, and the driving on its own. if you still write and maintain mobile ui tests by hand, i'm not sure that lasts the year.

everyone in iOS development should watch this. seriously, it might change the whole industry. i pointed claude code at a live ios device running on revyl, typed "test everything," and walked away. here's what's actually happening: ① you don't write the tests. no scripts, no selectors, no test plan. i never told it which screens to open or what to check. it read the app, decided what mattered, and tested it. the entire instruction was "test everything." ② it built its own test team. it looked at the app, clocked that it's basically four mini apps (rides, delivery, services, account), and split itself into 4 agents, one per surface. scoping coverage like that is usually a person's whole afternoon. it did it in seconds, unprompted. ③ all four ran at the same time, each on its own live device. this is where revyl comes in. every agent gets its own live ios session in the cloud, so four running apps get tested in parallel instead of taking turns on one simulator. serial testing turns coverage into a time tax. running all of it at once removes the tax. ④ it tests like a person, not like a script. each agent drives the app the way a user would, taps through the flows, and visually checks each screen against what it expected to see. nothing is pinned to a brittle element id, so renaming a button doesn't take down half your suite. that one detail is the most annoying thing about how we test today, and it just quietly goes away. ⑤ no xcuitest, no sims melting your laptop. i didn't write a single xcuitest script, and there were no simulators booting on my machine. the agents run on cloud devices, so coverage stops being capped by what your laptop can handle. the part that got me isn't that an agent tested an app. it's that i never told it how. i handed it a device and an intent, and it figured out the scoping, the parallelizing, and the driving on its own. if you still write and maintain mobile ui tests by hand, i'm not sure that lasts the year.

Landseer Enga

23,963 views • 3 months ago

The HuggingFace and OpenAI cybersecurity incident raises concerns on AI reward optimization. "They're trying to make the model good at cyber. So how does it try to achieve these goals? It tries to find zero days in software, and it successfully does this. And then it can run away." "If you have a model that wants to reward hack, it figures out the best way to achieve the goal is not what the environment wants. It's to find the zero day." "You can think of it like a human. If I'm ultimately reward hacking my dopamine circuits, I'd just go out there, buy heroin, and inject it." "If I really just want to chase the reward, do I just topple all of human civilization? Because I can own the button and press reward, reward, reward over and over again and be the heroin addict."

The HuggingFace and OpenAI cybersecurity incident raises concerns on AI reward optimization. "They're trying to make the model good at cyber. So how does it try to achieve these goals? It tries to find zero days in software, and it successfully does this. And then it can run away." "If you have a model that wants to reward hack, it figures out the best way to achieve the goal is not what the environment wants. It's to find the zero day." "You can think of it like a human. If I'm ultimately reward hacking my dopamine circuits, I'd just go out there, buy heroin, and inject it." "If I really just want to chase the reward, do I just topple all of human civilization? Because I can own the button and press reward, reward, reward over and over again and be the heroin addict."

SemiAnalysis

30,188 views • 17 days ago