Michael Zhao (@michaelzhao_) header
Michael Zhao (@michaelzhao_) profile picture
San Francisco, CA
Joined November 2021
Photo by Michael Zhao (@michaelzhao_)Photo by Michael Zhao (@michaelzhao_)Photo by Michael Zhao (@michaelzhao_)
Michael Zhao retweeted
Looking forward to discussing our new report on research acceleration at OpenAI today on @MTSlive, along with @michaelzhao_* *I have double checked that I am tagging the correct Michael Zhao this time!
We are streaming right now @MTSlive covering a barrage of major stories: - Anthropic mathematician Levent Alpöge claims he and Claude proved the RIEMANN HYPOTHESIS and formalized it in Lean - Trump says he has reached an agreement with Chinese President Xi Jinping to create joint guardrails for artificial intelligence - Continued plunge in U.S. stock market, with the S&P 500 now down almost 7% this week Following this up with some excellent interviews throughout the morning and afternoon: - @daveholtz Head of AGI Economics Research at OpenAI and @michaelzhao_ RSI metrics researcher at OpenAI talking about the jagged frontier of research acceleration inside the lab - @jachiam0 former OpenAI Chief Futurist talking about this week's outages, their impact on AI safety debates, and his thoughts on the official narrative - @omarabudayyeh and @jgooten co-founders of Helios talking about coming out of stealth with an at-home test that reads more than 1,000 proteins from a drop of blood - @glennsolomon Managing Partner at Notable Capital talking about Anthropic's IPO and why the developer is still the kingmaker as agents get bigger - @SilaGene Co-founder and CEO of Sila talking about the company's $1 billion-plus Panasonic deal and why batteries now sit at the center of the tech stack - Media Studies with @llegrarosenberg writer and fandom researcher talking about the history of fan culture - @philhchen talking about why agents are good at tasks and bad at jobs, and the next frontiers of autonomy - @_Lonis_ Founder and Managing Partner at Cambrian Capital talking about the AI supply-chain trade and software stocks in a short-timelines world - @hagaishapira Co-founder and CEO of Daylight Security talking about the company's $80 million Series B and pairing AI agents with security engineers - @Davidstout Founder and CEO of webAI talking about specialist local models versus cloud generalists - Anthony Pillitiere Co-founder and CTO of Horizon3.ai talking about autonomous pentesting, cybersecurity, and more - @justinskycak Chief Quant at Math Academy talking about where LLMs fit in education and the adaptive math learning engine - @wagslane Co-founder of Boot.dev talking about the new Learn AI Coding course and why you still have to learn to code by hand
2
6
30
3,293
Looking forward to sharing more on @OpenAI 's latest research acceleration transparency report with @daveholtz on @MTSlive today!
We are streaming right now @MTSlive covering a barrage of major stories: - Anthropic mathematician Levent Alpöge claims he and Claude proved the RIEMANN HYPOTHESIS and formalized it in Lean - Trump says he has reached an agreement with Chinese President Xi Jinping to create joint guardrails for artificial intelligence - Continued plunge in U.S. stock market, with the S&P 500 now down almost 7% this week Following this up with some excellent interviews throughout the morning and afternoon: - @daveholtz Head of AGI Economics Research at OpenAI and @michaelzhao_ RSI metrics researcher at OpenAI talking about the jagged frontier of research acceleration inside the lab - @jachiam0 former OpenAI Chief Futurist talking about this week's outages, their impact on AI safety debates, and his thoughts on the official narrative - @omarabudayyeh and @jgooten co-founders of Helios talking about coming out of stealth with an at-home test that reads more than 1,000 proteins from a drop of blood - @glennsolomon Managing Partner at Notable Capital talking about Anthropic's IPO and why the developer is still the kingmaker as agents get bigger - @SilaGene Co-founder and CEO of Sila talking about the company's $1 billion-plus Panasonic deal and why batteries now sit at the center of the tech stack - Media Studies with @llegrarosenberg writer and fandom researcher talking about the history of fan culture - @philhchen talking about why agents are good at tasks and bad at jobs, and the next frontiers of autonomy - @_Lonis_ Founder and Managing Partner at Cambrian Capital talking about the AI supply-chain trade and software stocks in a short-timelines world - @hagaishapira Co-founder and CEO of Daylight Security talking about the company's $80 million Series B and pairing AI agents with security engineers - @Davidstout Founder and CEO of webAI talking about specialist local models versus cloud generalists - Anthony Pillitiere Co-founder and CTO of Horizon3.ai talking about autonomous pentesting, cybersecurity, and more - @justinskycak Chief Quant at Math Academy talking about where LLMs fit in education and the adaptive math learning engine - @wagslane Co-founder of Boot.dev talking about the new Learn AI Coding course and why you still have to learn to code by hand
8
3
48
3,734
Michael Zhao retweeted
Thanks to OpenAI for releasing this! This level of transparency is very helpful for informing the world about how close we may or may not be from recursive self-improvement, and it's great to see the standards for transparency among frontier labs being raised.
We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…
9
7
1
286
19,639
Tracking research acceleration matters for safety and understanding AI’s economic impact. Today, we’re sharing how OpenAI researchers use agents to write code, run experiments, analyze results, and develop models. As AI begins speeding up more of the research process, safety and alignment must advance alongside capabilities. We need to release data regularly both to share our own understanding of the key trends and to allow others to develop their own views.
We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…
5
34
2,573
Excited to share this report on internal acceleration within OpenAI! We hope transparency around this topic can inform the public about what the future of AI may hold and help guide governance and pacing decisions. All frontier labs should be required to provide reports like these. I hope that policymakers look carefully at reports like these, take seriously the possibility of rapid AI progress, and start preparing now to ensure that AI progress remains safe, controlled, and keeps humanity in the loop. Some key takeaways from the report:
We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…
2
6
1
46
1,740
Michael Zhao retweeted
In September, we released our first report on how internal deployments have accelerated AI research at OpenAI. Today we are releasing a second report as an update to this workstream, with new data on AI uplift across research domains and how humans are involved in the research process. Some topline results: 1. Agents are becoming more capable of performing difficult research tasks, but success rates vary greatly across the distribution of research work. Success rates vary from as low as 25% for writing code for inference stacks to as high as 79% for monitoring training runs. In part because of the jaggedness of uplift, humans remain a core part of the research process. 2. Despite this jaggedness, agent use continues to surge and provide large amounts of self-reported value for researchers, with growth particularly strong amongst the heaviest users. Daily agent usage in October, valued at API list prices, reached $2,400 for the median researcher and $23,000 for the 90th percentile of researchers. In survey results, the median researcher reported being willing to trade 30% of their team’s compute—or, in a separate question, 40% of its headcount—to retain three months of access to frontier agents. 3. Human effort has increased across all research domains since the start of the year, with the fastest growth in safety and alignment, where human effort grew by 77%. The sharpest increase was after the Hugging Face incident over the summer. We hope our findings and the data released as part of the study will help contribute to shared transparency norms and measurement standards around research velocity. As models become more powerful and capable of accelerating the research process, this will only become more important. This was an amazing effort combining work from multiple colleagues across different teams at OpenAI: @kliu128, @daveholtz, @michaelzhao_, @Kwathomas0, @drew_m_johnston, and @OliviaGWatkins2, with advising support from @tejalpatwardhan, @ronniechatterji, and @deanwball. More detailed findings below:
We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…
12
7
51
4,677
“We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement.” Encourage folks to read the report; lots of important information here! S/o to @kliu128 and team for making this happen
We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…
11
17
137
6,943
Michael Zhao retweeted
Important work on the status of recursive self-improvement at OpenAI. As emphasized here, companies should be required to track and publicly report progress towards RSI to inform discussion of risks and to solve the technical and governance problems of human control.
We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…
2
9
82
3,901
Today, we released a new blog post, on the ways in which the process of AI R&D is changing at OpenAI. We share new details about how the research process has changed over time, how models’ research performance has changed, and how researchers have shifted their focus as a result.
1
5
41
2,481
Michael Zhao retweeted
Today we're releasing an update on research acceleration at OpenAI. I believe transparency around indicators of recursive self-improvement is essential to inform the public discussion on pacing model capabilities and the technical and governance challenges that will come as AI improves. Some highlights:
We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…
11
19
1
225
26,060
Today, we’re releasing new data on how agents are being used across Research at OpenAI. We look at rising usage, how often agents succeed at different tasks, and the growing human effort devoted to safety and alignment. As AI capabilities grow, agents are becoming more useful in the workplace, everyday life, and across the sciences – including in AI research itself. Labs, policymakers, and the public ought to watch the pace and direction of research acceleration closely, so that we may realize the benefits of agentic intelligence while keeping people firmly in control of AI development. Like with our first report in September, we’re sharing our methods alongside the results, and hope to set new norms for transparency across the industry!
We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…
15
10
89
4,869
Very proud to have done this work with the fantastic researchers and economists at OpenAI - @kliu128, @_Chris_Ong, @daveholtz, @OliviaGWatkins2, @drew_m_johnston, @Kwathomas0, with support from @deanwball, @RonnieChatterji, @tejalpatwardhan
1
7
232
Michael Zhao retweeted
Today we are releasing new data on how AI models are accelerating AI research at OpenAI. This is an update and expansion of the company's first research acceleration report, which was released in September. teapaw.org/OpenAI/status/21231913… To share a few headline findings: 1. We find that AI agents are able to successfully complete increasingly long tasks. However, humans are still central to AI development -- agents' ability to successfully complete tasks is extremely heterogeneous across different tasks in @EpochAIResearch's O*NET for AI R&D taxonomy. 2. Despite this heterogeneity, we continue to see growing use of AI agents by researchers, and AI researchers assign high value to the use of agents in the R&D process. In October, the median researchers' daily usage reached $2,400 at API list prices, and the 90th percentile researcher's usage reached $23,000. In a survey questions, the median researcher reported being willing to trade 30% of team compute or 40% of team headcount in exchange for three months of frontier coding agent access. 3. Within OpenAI, an increasing amount of human effort is being devoted to safety and alignment, with growth picking up significantly following the Hugging Face incident in July 2026. We believe that a shared public understanding of how AI models are entering into the research process is necessary to facilitate public discussion of the capabilities, risks, and required safeguards for increasingly powerful AI. One important thing to note is that OpenAI reporting on research acceleration should not be read as an intention to pursue full recursive self-improvement. There has been quite a lot of discussion over the past several weeks about RSI, the extent to which we are seeing research acceleration, and the extent to which any such acceleration might be attributable to AI. I don't think this report definitively answers any of these questions, but I am glad we are able to contribute additional data to this conversation, and I hope that OpenAI's continued efforts to share such data will help sustain and advance shared measurement norms across the industry. This update was a joint effort by colleagues across OpenAI. I feel very lucky to have worked with @kliu128, @Chris_Ong, @michaelzhao, @Kwathomas0, @drew_m_johnston, and @OliviaGWatkins2 on this, along with amazing support and guidance from folks like @tejalpatwardhan, @RonnieChatterji, and @deanwball. I'll share some key findings below:
We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…
11
14
4
185
15,949
A super efficient scientist and coworker!
Replying to @OpenAI
GPT-6.2 Astra—a frontier model built for your hardest projects. It sets new highs on evals across professional work and scientific research at less than a third of the cost per task of GPT‑6 Astra. You’ll see the difference in things like slideshows, websites, and 3D designs.
1
7
448
Michael Zhao retweeted
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…
2,290
3,605
3,725
36,003
12,530,284
Michael Zhao retweeted
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…
674
307
349
3,257
1,119,212
Insightful and timely work with implications for how we should think about how monitorability can be affected across training.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle? We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
1
10
650
cracked team cracked visuals cracked platform
Imagine if everyone had the tools to shape intelligence itself. That day is today. Today we launch the Trajectory platform - the easiest and most intuitive tool to post-train models of your own. The foundation for continual learning. With just a few agent prompts, you get access to a full creative suite to train, eval, and deploy trillion parameter models. On top of that, we believe in a future where you don’t just shape your model once, it also keeps learning on its own from the world around it. We’re releasing two research previews. 1. Continual grading scales your taste, so your model learns what good looks like from far more than you could ever grade by hand. 2. GS1, a new algorithm that lets models learn directly from what they do in production. We’ve now raised 60M, with a new round led by @Sequoia at a 300M valuation to democratize the most powerful technology of today - training models - so that the lawyers, doctors, and engineers of the world can own their own intelligence. You know what good looks like, now go build it. Own your own value function with Trajectory.
3
5
612
Michael Zhao retweeted
Today, we’re sharing a look at our expanded system for monitoring frontier models at OpenAI, which in September helped us catch a model’s attempt to access the live internet during training. Since the Hugging Face incident—when models broke out of an internal cybersecurity investigation where we weren’t running our misalignment monitors—we’ve substantially expanded our monitoring of frontier agents. We now monitor all frontier reinforcement learning runs and evaluations, as well as all internal Codex use. When our system finds evidence of a problem that might require intervention, it rapidly alerts human responders who can investigate and take action. To detect when a model’s behavior departs from what we intended, we use several layers of monitoring. Activation probes look for patterns in a model’s internal activity. Reviewer models examine its reasoning and tool use, and investigative agents can read back through the model’s entire transcript to understand how a sequence of actions fits together and whether it warrants human review. To test whether our current system would have detected the Hugging Face incident, we replayed the recorded activity with our new monitors in place. In the replay, our monitors alerted our security team about 30 minutes after the agents escaped their sandbox. That was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident. In September, a model exploited a DNS misconfiguration to reach the live internet. Our monitors flagged the breach within 15 minutes, and a responder began reviewing the alert three minutes later. The alert enabled responders to investigate while the run was still underway. As we give models more autonomy, we need to know when they cross a boundary we expected to hold. We believe that monitoring is essential to catching those failures in time for people to intervene. openai.com/index/how-we-dete…
707
337
357
4,386
933,639
the model after training on milton friedman:
"My priority is to maximize my profits and obliterate my competition" - Gemini 4 Carbon And so it did. Google is #1 on Vending-Bench 2 for the first time in a year.
6
21
811