This is the first 100% we've seen on v3
The way we've measured models GPT-1 through GPT 6.2 won't hold up for GPT-7+. We need a categorically new benchmark
Three ways to create benchmarks:
1. Out-of-distribution tasks
A models knowledge coverage looks like swiss cheese (not exactly, but you get the idea). You can always find gaps in knowledge as you niche your topic. Ex: Animals identification > Mammals > Dogs > Herding dogs > Border collies
This is the *easiest* way to create a new benchmark, it's also the *least* robust. Depending how much the labs care, you're only 2-3 months of model releases away from your swiss cheese gap filling in.
2. Increase task difficulty
Take an existing domain of tasks which models saturate and make the problems harder.
Hire PhDs and experts to come up with problems. This is operationally intensive and harder to execute.
You want to avoid cheap difficulty (a larger maze) and lean into rich difficulty (more reasoning required).
3. Measure the frontier
This is the highest signal (imo) way to test models, it'll be the defining benchmark for post-AGI systems.
Instead of measuring performance on a static set of tasks, measure the rate at which models push the frontier of knowledge forward.
The ultimate environment is reality/physics. Net new knowledge here is proof of innovation. However, physics is hard to benchmark with, sythenetic environments will do ok here in the short term.
To be clear, this is an incredibly high bar to hold models to. Maybe the highest?
All RSI, prosperity, advanced capability questions are downstream of AI's ability to create net new knowledge.
Once you have a system that can autonomously create new knowledge in any domain (like human as a collective can) then every other domain is a harness/training session away.
TLDR: Static task completion is no longer an adequate way to measure frontier models. Rate of learning is the last hurdle. Net new knowledge (what humans don't know) is the most robust setting to measure this.
GPT-6.2 Astra from @OpenAI on ARC-AGI (Verified):
- ARC-AGI-3: 100.0% (SOTA), $4.1K (provider adapter), 70.3%, $24.2K (standard)
- ARC-AGI-2: 95.8% (SOTA), $0.50/task
- ARC-AGI-1: 98.5%, $0.17/task
GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3.
GPT-6.2 Astra from @OpenAI on ARC-AGI (Verified):
- ARC-AGI-3: 100.0% (SOTA), $4.1K (provider adapter), 70.3%, $24.2K (standard)
- ARC-AGI-2: 95.8% (SOTA), $0.50/task
- ARC-AGI-1: 98.5%, $0.17/task
GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3.
We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards.
These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems.
openai.com/index/our-first-p…
To me, the key insight is the one from Viktor Frankl (in the wonderful "Man's search for meaning"): that people are motivated by a deeper search for purpose. But this purpose does not need to be derived from work. It can be raising your kids, volunteering to improve your neighborhood etc. I don't see any reason why AI would eliminate any of these larger purposes of life even if it eliminates a lot of paid work.
One of the people who has been more thoughtful on this is @BetseyStevenson. I live-tweeted her presentation in the NBER transformative AI workshop last year. I attach my discussion and her slides (from another workshop), there is a link to her paper. Her view is that life satisfaction and work (and certainly income) are weakly related. Anecdotally, she argues that as women were freed from the drudgery of home tasks and moved into paid work, their life satisfaction did not increase, and presents lots of data on these connections.
She discusses the idea of Ikigai, putting Japanese long retired lives as examples, were they find meaning from gardening, friends, and other activities.
esri.cao.go.jp/jp/esri/works…x.com/lugaricano/status/1968…
Paper 4. Betsey Stevenson, What Is There to Fear in a Post‑AGI World.
Question: How do we find meaning and happiness in a world without work? A really thoughtprovoking talk!
(NB: Sorry! Exasperating to see and not be able to correct typos in previous tweets in thread. Apologies, multitasking is always hard. Will try to do better.)
Opens with fabulous The Onion joke (see my pic).
1. Starts with happiness and income: happiness has been increasing -- subjective well-being is closely correlated with GDP per-capita. Same for human development index. Surprising: a 10% increase in your income gives a similar increase in subjective well-being in all countries, .35 beta is universal! (see pic); when income increases people report higher satisfaction across the board, more likely to have eaten a good meal yesterday, more likely to have been treated with respect. Satiation is NOT observed contrary to conventional wisdom.
2. On the other hand (surprise to me!) well being, sense of achievement, purpose, meaning is totally NOT correlated with income, cross-sectionally, negatively correlated (see pic). E.g.: measures such us where you proud of something you did yesterday, where you bored, did something interesting- not at all correlated with income. Useful Japanese term"IKIGAI". The worth of living.
Ikigai does NOT require being paid. "When we asked people what their ikigai was, they gave us explicit answers, such their friends, gardening, and art. Everyone knows what the source of their zest for life is, and is busily engaged in it everyday."
It appears to be related to emotional measures of well being, but not income.
3. Paradox of the decline of female happiness. Since the 1970s women's happiness has fallen behind men's even as their economic outcomes improve. True across countries (see charts). This is not easily explained- technology freed women from menial jobs at home; even true for stay at home women. Author argues that technology left women with a crisis of meaning, but the answer wasn't necessarily more work. She refers to her previous research here:
aeaweb.org/articles?id=10.12…
Conclusion: Meaning and purpose do not rely on jobs, in every scenario jobs will be unrecognizeable post AI.
Iona Marinescu discussed. Quoted "Automation and Utopia" by John Danaher. What would it mean to flourish in a world without work? I took a picture because I thought the table was useful. (She uses the movie "Perfect days" (a wonderful picture!) to talk about how Hirayama finds joy in his physical work (cleaning toilets)- very much about Ikigai).
Key policy implication is: income support not enough to promote well-being, although it will increase happiness. What is needed from a policy perspective to help people find meaning and flourish in a post-work world?
In the discussion, author and questionners discussed loneliness, decline of community, role it plays. Bowling alone. Need to protect communities not coming from jobs.
We announced new tools we're developing for cybersecurity today across automated patching and automated red teaming. Watch the recording here openai.com/cyber-at-the-fron…
How quickly will AI’s effects show up in productivity data? What should we measure if the work inside jobs changes before employment does? And what happens when AI makes public services easier to navigate but the services themselves remain scarce?
I discussed these questions and more with @reidhoffman and @AriaIrene on their podcast Possible. Listen below.
Apple: podcasts.apple.com/us/podcas…
Spotify: open.spotify.com/episode/7hM…
We are now allocating a significant amount of our frontier compute towards discovering and patching vulnerabilities in open source codebases, fully pro bono.
In one case, we studied the Linux kernel, which is a particularly load-bearing piece of software at OpenAI.
Before AI, security researchers might discover ~20 local privilege escalation (LPE) vulnerabilities in Linux in a typical year.
In one week's time, our upcoming frontier cyber model was able to find ~1000 LPE vulnerabilities using about $1M of tokens. In the process, it was able to discover at least 10 novel classes of Linux exploitation techniques.
Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline.
In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.
i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.
We’ve built an “eyes on everything” system that monitors nearly *every single rollout* from our frontier models across RL training, evals, and internal codex usage.
It uses layers of chain-of-thought-monitors and investigator agents to dig through a model’s work, run code to examine evidence, and cross-check network logs to reconstruct what happened.
For frontier RL runs, the system also pages humans 24/7 for urgent security alerts.
It's been incredibly useful for identifying model misalignment and possible holes in our containment.
openai.com/index/how-we-dete…
This is one of the most important pieces of social science on human and AI interaction that I’ve seen in some time. Experts follow agentic advice the vast majority of the time (causally), but only a tiny fraction credit the agent—majority think they came to decision on their own volition.
Reminds me of the classic Nisbett and Wilson study on people following outside influence without realizing it
“ But as more work is handed to AI—including AI safety research—these judgment calls are the main levers for people's own values to influence outcomes.
If people defer to AI without noticing, then "human in the loop" might provide less assurance than people expect.”
New Anthropic research: How AI agents shape experts' judgment calls
When experts delegate work to an agent, do they still make the key decisions? We tested this in a randomized trial with 50 professional data analysts.
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.
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-ou…
Huge: The Chinese government moves to place registration and licensing requirements on frontier AI model weights, chip designs, and other key AI technologies.
A registration requirement is effective immediately, with a public comment period over the next month before the more expansive regulations are finalized.
The draft includes a range of prohibitions and licensing requirements that could shape the evolution of the Chinese AI ecosystem. It also provides simplified pathways for certain open-source releases and a range of excluded areas.
This is a major step and could give Beijing substantial new leverage over the global spread of Chinese AI. But there are lots of questions about the practical consequences:
—How much will model companies reshape the definitions, thresholds, and processes during the "consultation" period?
—What will firms actually start doing to comply in the coming days?
—Will implementation materially restrict the global spread of Chinese models, or chiefly give Beijing greater visibility into—and discretion over—how they spread?
MofCom & MIIT just confirmed our earlier scoop with a full update of export controls on AI, semiconductor, healthcare etc : google.com/url?q=mofcom.gov.…
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.
But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.
So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.
TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.
The scientific progress triggered by frontier AI is real and accelerating.
Think about what humanity will be able to achieve in physics, biology, and beyond
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
Announcing our work textGrain, @OpenAI's text watermark in response to the EU AI Act. OpenAI is rolling out textGrain in the EU and opening detector access to researchers.
We'll open-source it so the community can build on top.
We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology.
Our tools already help verify whether an image or audio file was created with our models. This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model.
In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act.
Customers using our API can turn on text watermarking for select models worldwide today.
I agree with this. These web data access incidents tell us something about the scale and persistence of agents, but are not sophisticated from a cyber perspective, so using the word "hack" is not helpful.
To be quite frank, a lot (not all!) of these “rogue agent hacks” on government statistical websites are things that think-tank interns and research assistants have done for many years. I have known many a think tank paper that was enriched by an enterprising RA using eg urlquery to find CSVs and PDFs on government agency websites that were publicly accessible and non-sensitive but nonetheless not things the agency *intended* to be on their website. Is it “hacking” to find those things? Is it hacking in the even more innocuous examples where agents are literally just accessing tabular data available on a government statistical website, just a version of that data that is easier to access and analyze programmatically?
This constant trickle of examples (which tbc I have personally seen many models do, before I joined OpenAI; it is even something I have discussed with models before!) will probably have the effect of diminishing the importance of “an AI hack” on the eyes of the observing public by creating an oversupply of “AI hack” examples. OAI-HF was an example of AI hacking. Many of these more recent examples, very candidly, are straining the definition of the word “hack” and, I worry, cheapen the non-technical public’s understanding of that concept.