Sayash Kapoor
Computer scientist; finished his PhD at Princeton in July 2026 and joins the UC Berkeley School of Information as an assistant professor in July 2027. Researches the societal impact of AI and co-writes AI as Normal Technology with Arvind Narayanan.
Sayash Kapoor did not write this page. ppll compiled it from their own words, each linked to the source. Is this you? · Report a mistake
Latest x.com
What they believe
-
Evaluating an AI agent requires scrutinizing its entire trajectory, not just its final outcome.
The evaluator must scrutinize the entire trajectory, not just the final outcome.
normaltech.aiSaid 19 Oct 2026This on korrents
-
AI companies should be liable for what their agents do, and policymaking should make that clear.
We think AI companies should be liable for what their agents do, and this should be clarified through policymaking.
normaltech.aiSaid 14 Sept 2026This on korrents
The bold line is korrents’ wording; the quote is theirs. These 2: what they wrote before what they said aloud, one per source, newest first.
Feed
As its own page →Hiding
19 November
10 November
- x.com
- NewsletterThe missing quadrant in responses to AI
5 November
3 November
28 October
26 October
20 October
19 October
From one piece New paper: Emerging trends in AI agent evaluation 2 beliefs · normaltech.ai
-
korrents.com
An AI agent's evaluation results at one level of scope are not guaranteed to hold at another.Their words
Crucially, the results of evaluation at one level of scope are not guaranteed to translate to another.
-
korrents.com
Evaluating an AI agent requires scrutinizing its entire trajectory, not just its final outcome.Their words
The evaluator must scrutinize the entire trajectory, not just the final outcome.
14 September
From one piece The AI-as-Normal-Technology view of loss-of-control incidents 3 beliefs · normaltech.ai
-
Their words
Rather than behaving like startups, AI companies need to behave like more mature organizations.
-
korrents.com
Marginal investments in AI control are more likely to be effective than marginal investments in alignment.Their words
But in our view, marginal investments in control are more likely to be effective compared to those in alignment.
-
korrents.com
AI companies should be liable for what their agents do, and policymaking should make that clear.Their words
We think AI companies should be liable for what their agents do, and this should be clarified through policymaking.
5 August
From one piece AI agents can't yet do open-ended AI research 2 beliefs · normaltech.ai
-
Their words
Our results suggest that conducting open-ended research remains challenging for frontier AI agents.
-
Their words
Our view is that while faster progress is certainly possible on narrow tasks (such as improving efficiency), we don't think it will lead to broad RSI or explosive progress.
22 May
From one piece Did Google’s AI agents really build an operating system for $916? 3 beliefs · normaltech.ai
-
Their words
benchmark evaluation is effectively impossible for this kind of task for many reasons including cost.
-
Their words
Google's experiment does add to the mounting evidence that agents or agent teams can autonomously or near-autonomously work on certain kinds of tasks for very long periods of time, making progress without getting stuck or confused.
-
Their words
Even if there was no direct copying, writing an operating system might be relatively easy for agents because of patterns memorized in the training data, so this doesn't tell us much about agents' ability to create novel pieces of software.
21 May
From one piece Do AI Risks Require Extraordinary Government Intervention? 2 beliefs · normaltech.ai
-
korrents.com
The gap between frontier AI capabilities and publicly available ones is at most a few months, not years.Their words
Open-weight models and widespread API access from frontier labs mean that the gap between frontier and publicly available capabilities is at most a few months, not years.
-
Their words
Nonproliferation is brittle because it relies on a single chokepoint. Resilience distributes defenses across society.
16 April
From one piece Open-world evaluations for measuring frontier AI capabilities 3 beliefs · normaltech.ai
-
Their words
App store operators should prepare for and police spam submissions, as they might soon see thousands of applications submitted autonomously using agents.
-
korrents.com
Average performance on a benchmark suite is very different from the upper bound of what an AI agent can do.Their words
Average performance is very different from upper-bound capability elicitation. Trying to run dozens (or 100s) of tasks in a benchmark suite is only necessary to measure average performance.
-
benchmarksreinforcement learning
Their words
But whatever is precise enough to benchmark is also precise enough to optimize for, and modern RL training runs increasingly resemble the shape of the benchmarks themselves.
24 February
16 July 2025
1 May 2025
Nothing matches.
Beliefs
Korrents What they believe 20 beliefs — each backed by an exact quote.
Each is a — compiled by korrents.com, not by them: the one-line wordings are korrents', the quotes are theirs.
Recent
Evaluating an AI agent requires scrutinizing its entire trajectory, not just its final outcome.
The evaluator must scrutinize the entire trajectory, not just the final outcome.
New paper: Emerging trends in AI agent evaluation Said 19 Oct 2026
An AI agent's evaluation results at one level of scope are not guaranteed to hold at another.
Crucially, the results of evaluation at one level of scope are not guaranteed to translate to another.
New paper: Emerging trends in AI agent evaluation Said 19 Oct 2026
AI companies should be liable for what their agents do, and policymaking should make that clear.
We think AI companies should be liable for what their agents do, and this should be clarified through policymaking.
The AI-as-Normal-Technology view of loss-of-control incidents Said 14 Sept 2026
Show 17 more
Marginal investments in AI control are more likely to be effective than marginal investments in alignment.
But in our view, marginal investments in control are more likely to be effective compared to those in alignment.
The AI-as-Normal-Technology view of loss-of-control incidents Said 14 Sept 2026
AI companies need to behave like mature organizations rather than startups.
Rather than behaving like startups, AI companies need to behave like more mature organizations.
The AI-as-Normal-Technology view of loss-of-control incidents Said 14 Sept 2026
Faster AI progress on narrow, verifiable tasks will not lead to broad recursive self-improvement or explosive AI progress.
Our view is that while faster progress is certainly possible on narrow tasks (such as improving efficiency), we don't think it will lead to broad RSI or explosive progress.
AI agents can't yet do open-ended AI research Said 5 Aug 2026
Conducting open-ended research remains challenging for frontier AI agents.
Our results suggest that conducting open-ended research remains challenging for frontier AI agents.
AI agents can't yet do open-ended AI research Said 5 Aug 2026
An agent writing an operating system may be relatively easy because of patterns memorized in training data, so it says little about agents' ability to create novel software.
Even if there was no direct copying, writing an operating system might be relatively easy for agents because of patterns memorized in the training data, so this doesn't tell us much about agents' ability to create novel pieces of software.
Did Google’s AI agents really build an operating system for $916? Said 22 May 2026
There is mounting evidence that agents or agent teams can work autonomously or near-autonomously on certain kinds of tasks for very long periods, making progress without getting stuck or confused.
Google's experiment does add to the mounting evidence that agents or agent teams can autonomously or near-autonomously work on certain kinds of tasks for very long periods of time, making progress without getting stuck or confused.
Did Google’s AI agents really build an operating system for $916? Said 22 May 2026
Benchmark evaluation is effectively impossible for long-horizon real-world agent tasks, for many reasons including cost.
benchmark evaluation is effectively impossible for this kind of task for many reasons including cost.
Did Google’s AI agents really build an operating system for $916? Said 22 May 2026
Nonproliferation of AI is brittle because it relies on a single chokepoint, while resilience distributes defenses across society.
Nonproliferation is brittle because it relies on a single chokepoint. Resilience distributes defenses across society.
Do AI Risks Require Extraordinary Government Intervention? Said 21 May 2026
The gap between frontier AI capabilities and publicly available ones is at most a few months, not years.
Open-weight models and widespread API access from frontier labs mean that the gap between frontier and publicly available capabilities is at most a few months, not years.
Do AI Risks Require Extraordinary Government Intervention? Said 21 May 2026
Anything precise enough to benchmark is also precise enough to optimize for, and modern RL training increasingly resembles the shape of the benchmarks themselves.
But whatever is precise enough to benchmark is also precise enough to optimize for, and modern RL training runs increasingly resemble the shape of the benchmarks themselves.
Open-world evaluations for measuring frontier AI capabilities Said 16 Apr 2026
Average performance on a benchmark suite is very different from the upper bound of what an AI agent can do.
Average performance is very different from upper-bound capability elicitation. Trying to run dozens (or 100s) of tasks in a benchmark suite is only necessary to measure average performance.
Open-world evaluations for measuring frontier AI capabilities Said 16 Apr 2026
App store operators should prepare for and police spam submissions, since they might soon see thousands of applications submitted autonomously by agents.
App store operators should prepare for and police spam submissions, as they might soon see thousands of applications submitted autonomously using agents.
Open-world evaluations for measuring frontier AI capabilities Said 16 Apr 2026
Reliability lags capability in AI agents, and will remain a barrier to deployment unless effort goes into improving it as a separate dimension from accuracy.
Our view is that reliability lags capability, and that reliability will remain a barrier to deployment unless researchers and developers focus effort on improving reliability as a separate dimension from accuracy.
New Paper: Towards a science of AI agent reliability Said 24 Feb 2026
Autonomous AI agents in high-stakes contexts need 99.9% to 99.999% accuracy before reliability stops being an issue, and LLM-based agents are not on track to reach that.
Our view is that for autonomous operation in high-stakes contexts, we need 3-5 "nines" of performance - 99.9% to 99.999% accuracy - in order for reliability to become a non-issue, and we don't think LLM-based agents are on track to reach such a threshold.
New Paper: Towards a science of AI agent reliability Said 24 Feb 2026
Each order-of-magnitude cut in an AI agent's unreliability is expected to be about as hard as the previous one.
we expect each order of magnitude decrease in unreliability (1-reliability) to be as hard as the previous one.
New Paper: Towards a science of AI agent reliability Said 24 Feb 2026
Accelerating the production of papers does not address what actually bottlenecks scientific progress.
They are simply trying to accelerate production, which is like adding lanes to a highway when the slowdown is actually caused by a toll booth.
Could AI slow science? Said 16 Jul 2025
More findings and papers within an old paradigm cannot produce the progress that a new way of understanding does.
No amount of findings or papers in older paradigms of geology would have led to the same progress that plate tectonics did.
Could AI slow science? Said 16 Jul 2025
What is a korrent?
A korrent is a belief a person has stated in their own words: one sentence stating the claim, backed by a quote and a source, kept at korrents.com.
Under a name here, the quoted block is what they actually said. The korrent beneath it is the claim those words support, in korrents' wording — tap it to see the record, its source, and who else holds it.
Nobody here wrote their own korrents. They are compiled from public statements, and a person can change their mind, which is recorded too.
Where they publish
Site Sayash Kapoor — Princeton His own personal/faculty-hosted page.
Link verified 25 Sept 2026.
Newsletter AI as Normal Technology Co-written with Arvind Narayanan; formerly AI Snake Oil.
Recent
- The missing quadrant in responses to AI 10 Nov 2026 Announcing STABL: Societal Transformation and AI Bottlenecks Lab
- New paper: Emerging trends in AI agent evaluation 19 Oct 2026 Understanding the methods, questions, and ecosystem for evaluation
- The AI-as-Normal-Technology view of loss-of-control incidents 14 Sept 2026 A middle ground between the cybersecurity and AI safety communities
Show 7 more
- AI agents can't yet do open-ended AI research 5 Aug 2026 Early evidence from two case studies
- Did Google’s AI agents really build an operating system for $916? 22 May 2026 The importance of independent evaluation
- Do AI Risks Require Extraordinary Government Intervention? 21 May 2026 Let’s not skip the hard work of AI governance
- Open-world evaluations for measuring frontier AI capabilities 16 Apr 2026 Introducing CRUX, a new project for evaluating AI on long, messy tasks
- New Paper: Towards a science of AI agent reliability 24 Feb 2026 Quantifying the capability-reliability gap
- Could AI slow science? 16 Jul 2025 Confronting the production-progress paradox
- AGI is not a milestone 1 May 2025 There is no capability threshold that will lead to sudden impacts
Link verified 25 Sept 2026. Recent items update automatically from the channel.
x.com @sayashk Posts on AI research and evaluation.
Recent
- AI safety is a marathon, not a sprint The views of the AI safety community are quickly converging on a fast takeoff within the next year. But this relies on a few low-quality data points from AI companies and contested models of the world. I worry that this convergence will degrade the epistemic environment and make good decisions about AI policy and deployment harder. Forget about predicting the future; it will be (and sometimes already is) hard to adjudicate claims about the present. It also pushes the community’s energy toward blunt instruments such as generalized bans on AI development, a… 19 Nov 2026
- Steering AI's societal impact requires taking both AI development and downstream bottlenecks seriously. Today @RishiBommasani @random_walker and I are starting a new non-profit: STABL @stabl_ai to guide societal adaptation and resilience to advanced AI. 10 Nov 2026
- Announcing CRUX-in-a-box: open-source infrastructure for running and monitoring weeks-long agent evaluations. We built it over six months of long-horizon evaluations in CRUX, to keep runs repeatable, secure, and monitored. AI evaluators can now use it to: - Run agents on VMs for weeks - Use LLMs to flag risky actions for human review - Inspect tool calls, tokens, and spend - Set time and cost budgets - Chat with agents mid-run to give feedback or answer their questions If you run long-horizon agent experiments, we might have solved some of the problems you will come across. GitHub: 5 Nov 2026 github.com
Show 5 more
- Update: Can AI agents use Pangram to beat Pangram? In small tests, I found simple prompts to the Codex allowed it to beat Pangram. The final output was 100% human generated. Prompts that helped: - Have the agent develop a style guide based on past writing - Have it develop an "anti-style guide" based on AI writing tics that people have complained about - Instruct it to get to a seed sentence that Pangram doesn’t flag (or at least flags as AI-assisted rather than fully AI-generated). Then the agent knows what strategies work and can iterate from there 3 Nov 2026
- Yesterday, I had the privilege of testifying to the California Assembly Privacy Committee what AI legislation we need to help incentivize AI safety. One of my main takeaways was that there is immense enthusiasm in the California legislature around AI legislation. As one example, the event was organized when the legislature is off-session, but was still attended by 15 assembly members, including many who are not on the Privacy committee. It was also attended by the speaker of the assembly. In my testimony, I laid out a plan to develop a layered approach to AI safety: - Clarify and strengthen l… 28 Oct 2026
- After the surprise outcome of the Brazilian election, I was curious if AI forecasting tools predicted the outcomes correctly. It turns out a top AI tool (FutureSearch) made the same errors as prediction markets more broadly. This is an example of what @random_walker and I mean by “intrinsic limits” to prediction that have nothing to do with the AI tools in question, but rather depend on the inherent uncertainty in the world. Of course, a single failed prediction doesn’t mean that AI isn’t good at forecasting; in fact AI tools are already about as good as top forecasters, and I expect human fo… 26 Oct 2026
- An update on our evaluations of AI agents conducting open-ended AI research. Most evaluations of agents conducting AI research focus on narrow, verifiable tasks. But AI research is often open ended. In July we tested the ability of frontier AI agents to conduct open-ended AI research, using Claude Opus 4.8 and GPT-5.6 Sol. We gave agents research questions from two unpublished papers and thousands of dollars of API credits and compute. The authors of the original papers reviewed the AI-generated papers, following a NeurIPS-style template, and rejected both papers. Since then, OpenAI has relea… 20 Oct 2026
- New paper: AI evaluation is a burgeoning field. How can researchers entering the field understand best practices and open problems? In our paper, led by @flxcn, we develop a concrete list of open questions and best practices for agent evaluation. We also share recommendations for evaluators, AI companies, policymakers, and more. 19 Oct 2026
Link verified 25 Sept 2026. Recent items update automatically from the channel.
About the English under a post
Some people here publish in a language other than English. Where they do, this site shows a machine translation beneath the post, in this typeface — the site's own, not theirs.
The post itself is never changed, moved or hidden: what is set in the serif above is exactly what the person published, and it is what to quote them on. A translation can be wrong in ways that matter, especially about tone.
Only the post's own words are translated. A quoted post, a linked article and a belief on korrents.com are left in their original language.




