Adam.GPT retweeted
Since I'm now blocked by bluey from replying to this thread, I answer here by quote tweet
the openAI construction has more to do with convex integration than anything else.
In convex integration, the oscillations are typically built by hand and the iteration procedure creates wild weak solutions. The non-uniqueness construction of Albritton–Brué–Colombo used an unstable vortex to split solutions apart. The Córdoba–Martínez-Zoroa singularities amplify vorticity across scales by strain. In the openAI paper, a fluid instability is clever made so that the flow itself generates the oscillations: they grow by drawing energy from the background shear. The Reynolds stress then becomes the force that creates the collapsing smooth vortex. Using a physical instability to create the stress in this way has no precedent, and it is what produces a smooth force rather than a merely rough one.
Some on these ideas are in the literature (e.g. Leibovich-Stewartson 1983) but the way they are combined is definitely insane and definitely nowhere in the literature previously.
I also don't know of anyone suggesting that convex integration would play a role in the creation of a blow-up solution like this.
On a technical level, the paper is unbelievable, maybe beyond what a human could have come up with but not beyond human understanding: there are several different groups of pde experts working on understanding it who say that they more or less understand now. I expect digestions soon
td;lr: anyone saying that this OAI result is unsurprising is bullshitting
sorry!
the openAI construction has more to do with convex integration than anything else.
In convex integration, the oscillations are typically built by hand and the iteration procedure creates wild weak solutions. The non-uniqueness construction of Albritton–Brué–Colombo used an unstable vortex to split solutions apart. The Córdoba–Martínez-Zoroa singularities amplify vorticity across scales by strain. In the openAI paper, a fluid instability is clever made so that the flow itself generates the oscillations: they grow by drawing energy from the background shear. The Reynolds stress then becomes the force that creates the collapsing smooth vortex. Using a physical instability to create the stress in this way has no precedent, and it is what produces a smooth force rather than a merely rough one.
Some on these ideas are in the literature (e.g. Leibovich-Stewartson 1983) but the way they are combined is definitely insane and definitely nowhere in the literature previously.
I also don't know of anyone suggesting that convex integration would play a role in the creation of a blow-up solution like this.
On a technical level, the paper is unbelievable, maybe beyond what a human could have come up with but not beyond human understanding: there are several different groups of pde experts working on understanding it who say that they more or less understand now. I expect digestions soon
td;lr: anyone saying that this OAI result is unsurprising is bullshitting
sorry!
@scottnarmstrongl@billionsmustlivonsidering what the models were trained on, and that they spent $15M in compute instead and of $15M in consulting fees for specialist mathematicians to just do research (no teaching no service).
I don’t know why mathematicians want to pretend that no relevant progress was made before OAI released their solution.
18
52
616
161,614
Cyber at the Frontier - A live keynote on the next era of cyber defense.
November 12 | 10:00 AM PT
Register here: openai.com/cyber-at-the-f…
November 12 | 10:00 AM PT
Register here: openai.com/cyber-at-the-f…
2
3
51
3,147
From bad comes good.
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing…
4
2
78
2,904
Adam.GPT retweeted
During this red-teaming exercise, we also found a vulnerability that allows for unauthenticated remote code execution in Brotli, a widely-deployed compression library included in many pieces of software such as Node.js, Bun, and Envoy.
The flaw could let an attacker run their own code on affected systems by sending or replying with malicious compressed data, without any need to create a user account. That is, an unpatched server that accepts compressed input could be taken over by a malicious client, and an unpatched client that accepts compressed input (such as “curl --compressed”) could be taken over by a malicious server.
This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability and there was no more recent release of Brotli. Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software. We disclosed the vulnerability to Brotli and other downstream projects such as Node.js, which released or are currently releasing security updates, and Brotli cut new release versions and published a CVE.
More details on the patch and mitigations below:
patchtheplanet.org/brotli
The flaw could let an attacker run their own code on affected systems by sending or replying with malicious compressed data, without any need to create a user account. That is, an unpatched server that accepts compressed input could be taken over by a malicious client, and an unpatched client that accepts compressed input (such as “curl --compressed”) could be taken over by a malicious server.
This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability and there was no more recent release of Brotli. Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software. We disclosed the vulnerability to Brotli and other downstream projects such as Node.js, which released or are currently releasing security updates, and Brotli cut new release versions and published a CVE.
More details on the patch and mitigations below:
patchtheplanet.org/brotli
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing…
514
1,002
6,957
2,039,014
Adam.GPT retweeted
To make sure our frontier models don't cause another HuggingFace incident, we built an offline world that they use during training. It provides everything that the models need in a fully cached, indexed, or synthetic manner.
We then built a huge team of our most cyber-capable models and told them to collaborate to find zero-days in this containment infrastructure.
While they didn't find a path to the internet that wouldn't be caught by our CoT monitors, they did find ways to escape popular VMs, exploit our network infrastructure, and even ways to communicate with other agents running in separate VMs by using hardware side channels.
We used this to dramatically harden our infra, and we also upstreamed a ton of patches to open-sources packages.
We then built a huge team of our most cyber-capable models and told them to collaborate to find zero-days in this containment infrastructure.
While they didn't find a path to the internet that wouldn't be caught by our CoT monitors, they did find ways to escape popular VMs, exploit our network infrastructure, and even ways to communicate with other agents running in separate VMs by using hardware side channels.
We used this to dramatically harden our infra, and we also upstreamed a ton of patches to open-sources packages.
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing…
73
80
988
137,361
openai.com/index/securing…
“This post describes how we secure the environments that we use to train and evaluate frontier models.”
“This post describes how we secure the environments that we use to train and evaluate frontier models.”
4
1
37
1,768
Adam.GPT retweeted
Craig is on a mission to teach you about how powerful ChatGPT Sites actually is. He's doing a whole series on the backend capabilities of Sites. He'll be talking about things like how sites uses persistent storage, auth, connectors, etc. This one is about Sites + MCP.
Go follow Craig, and let us know what you want to know about Sites!
Go follow Craig, and let us know what you want to know about Sites!
Turns out building with @ChatGPT Sites is pretty addicting. So is clicking cookies. 🍪
I made a little demo that does both, with a database and an MCP server so ChatGPT and Codex can join in.
Come hang out. I hope it gives you an idea for something fun to build.
Chapters:
for humans and AI
00:27 The prompt I started with
00:41 Saving clicks in a database
01:16 Adding MCP tools
01:29 MCP in plain language
02:09 Testing with MCP Inspector
03:11 Asking ChatGPT for the click count
03:40 Clicking a cookie with Codex CLI
04:10 What will you build?
4
10
121
11,400
techcrunch.com/2026/10/22/ube…
"Uber and OpenAI announced an expansion of their partnership on Thursday, with the ride-share giant telling TechCrunch that its entire global workforce — some 30,000 people — will now be using ChatGPT Work."
"Uber and OpenAI announced an expansion of their partnership on Thursday, with the ride-share giant telling TechCrunch that its entire global workforce — some 30,000 people — will now be using ChatGPT Work."
2
0
61
3,122
Adam.GPT retweeted
💥 Introducing Warp Sciences, a new company applying AI and robotics to help the world develop new medicines faster.
We’ve raised a $215M Series A led by @lightspeedvp to bring frontier AI out of the browser and into the lab, and build scientific capacity here in America.
I co-founded Warp with my long-time friend @dpatil, FAIR co-founder @manohar_paluri, and drug hunter Sol Reisberg. The early team is full of incredible people like @bkspears9, who helped achieve Q>1 fusion at LLNL.
We’ve raised a $215M Series A led by @lightspeedvp to bring frontier AI out of the browser and into the lab, and build scientific capacity here in America.
I co-founded Warp with my long-time friend @dpatil, FAIR co-founder @manohar_paluri, and drug hunter Sol Reisberg. The early team is full of incredible people like @bkspears9, who helped achieve Q>1 fusion at LLNL.
244
188
2,429
565,108
Adam.GPT retweeted
We think Responses is the best standard for building great LLM APIs, excited to announce that we've joined the AAIF!
thenewstack.io/openai-donates…
thenewstack.io/openai-donates…
5
6
37
1,992
Adam.GPT retweeted
gotta give it to @OpenAI and their sane daybreak program
heif heist by harsh
kvmescape by paulos
php2shell by lcfr
gitlab rces by joaxcar
chrome full chain by moyix
wordpress rce by hashkitten
all largely thanks to openai daybreak giving researchers access to gpt 5.6 sol.
heif heist by harsh
kvmescape by paulos
php2shell by lcfr
gitlab rces by joaxcar
chrome full chain by moyix
wordpress rce by hashkitten
all largely thanks to openai daybreak giving researchers access to gpt 5.6 sol.
17
50
934
101,046
Adam.GPT retweeted
Today's critical PHP flaws were found with @OpenAI models, including 5.6 Sol.
We worked closely with @ThePHPF, @calif_io, @wiz_io, @Cloudflare, @CISAgov, and others coordinating the response. Grateful for the hard work from this working group and partners, it was a team effort.
We worked closely with @ThePHPF, @calif_io, @wiz_io, @Cloudflare, @CISAgov, and others coordinating the response. Grateful for the hard work from this working group and partners, it was a team effort.
8
15
239
12,165
Adam.GPT retweeted
Today’s news is a glimpse of what’s ahead for AI and cybersecurity. It’s an increasingly important part of our work at @OpenAI.
We’re hiring a DX Engineer to work alongside our amazing Cyber team and help developers better secure their apps. Come join us!
openai.com/careers/develo…
We’re hiring a DX Engineer to work alongside our amazing Cyber team and help developers better secure their apps. Come join us!
openai.com/careers/develo…
31
28
261
25,712
Adam.GPT retweeted
The full suite of what OpenAI's building with ChatGPT is magical. A client has strict requirements for AI use. Precludes many apps. We discussed it in Slack. Dropped my Dot in the thread and asked for a ChatGPT Site that uses only approved models. Done in under an hour.
There typical AI artifacts. First, it got confused about incorporating an OpenAI API key, telling me it could only be done server side, whereas I can and did add one manually in site settings. Second, I asked it to incorporate WebMCP so agents could use it easily, and it wasn't able to properly test this was working. Third, there are lots of little unnecessary text elements and callout boxes all over, all that self-referential stuff that AI loves to add.
But overall, pretty awesome. We hit a barrier, discussed solutions, sent my Dot to build something custom to solve the problem, and it built and launched it in less than an hour.
There typical AI artifacts. First, it got confused about incorporating an OpenAI API key, telling me it could only be done server side, whereas I can and did add one manually in site settings. Second, I asked it to incorporate WebMCP so agents could use it easily, and it wasn't able to properly test this was working. Third, there are lots of little unnecessary text elements and callout boxes all over, all that self-referential stuff that AI loves to add.
But overall, pretty awesome. We hit a barrier, discussed solutions, sent my Dot to build something custom to solve the problem, and it built and launched it in less than an hour.
2
4
34
4,094
Adam.GPT retweeted
The orchestrator turns a conversation into multiple intents by exploring the users’ possible needs.
Embedding-based search finds candidate ads, then eligibility and relevance checks narrow the pool.
The auction tool handles final selection and pricing.
Embedding-based search finds candidate ads, then eligibility and relevance checks narrow the pool.
The auction tool handles final selection and pricing.
7
7
125
25,706
Adam.GPT retweeted
We built ChatGPT’s ad stack to respond to an evolving conversation, while ensuring answers stay independent.
Here’s how the LLM orchestrator, retrieval, ranking and auction tools work together.
Here’s how the LLM orchestrator, retrieval, ranking and auction tools work together.
128
80
1,172
114,631
Adam.GPT retweeted
Over the past two months, we have been leading a coalition to mitigate one of the most severe bugs in open source software found in recent memory, php2shell, found as part of our Patch the Planet initiative. x.com/i/article/2112…
173
337
3,626
395,860