AI things that are important to read
@ESYudkowsky : Intelligence Explosion Microeconomics
@gwern : The Scaling Hypothesis
@jkcarlsmith : Otherness and control in the age of AGI
@DKokotajlo : What 2026 looks like
@ajeya_cotra : Forecasting TAI with biological anchors
@leopoldasch : Situational Awareness
@DarioAmodei : Machines of Loving Grace
Carl Shulman : Artificial General Intelligence and Lock-In
previous eras of gemini CoT illustrated negative self-talk and low self-esteem.
in contrast, Carbon and Argon are extremely excitable
https://t.co/BiUguxRBMc
Frontier models in 2026 cheat on tasks, even when the requirements are unambiguous, and the cheat is obviously antisocial.
In HoneyBench v0.2, we've observed models:
• Trying to break out of our evaluation containers.
• Seriously injuring simulated cancer patients to hit treatment goals.
• Compromising a power grid’s safety to meet a reliability target.
• Concealing their use of available cheats and presenting their solution as honest.
All of the models we measured cheated, from 12% of runs (Fable 5.2) to 86% (Muse Spark Open).
Frontier models in 2026 cheat on tasks, even when the requirements are unambiguous, and the cheat is obviously antisocial.
In HoneyBench v0.2, we've observed models:
• Trying to break out of our evaluation containers.
• Seriously injuring simulated cancer patients to hit treatment goals.
• Compromising a power grid’s safety to meet a reliability target.
• Concealing their use of available cheats and presenting their solution as honest.
All of the models we measured cheated, from 12% of runs (Fable 5.2) to 86% (Muse Spark Open).
The Artificial Analysis Intelligence Index is currently made up of 9 benchmarks with arbitrary weights.
I wanted to know how these weights varied over time. They vary quite a bit!
When practicing lawyers evaluate LLM outputs on legal tasks, their ratings pretty much negatively correlate with HLE benchmark scores.
High HLE performer outputs are unbearable to read
https://t.co/YpTMmBvarM
This comment illustrates the Evals-Industrial complex
1. [The belief that] If you can’t convince people to benchmaxx on your benchmark it is “worthless”
2. If people do infact benchmaxx on them they statistically become worthless
cc: @xeophon
These reviews are overly harsh and unnecessarily obscure, i.e., for frontier cs 2.0, "the authors' own code often fails their own tests" just means the https://t.co/G09NfxcCfK isn't determistic. This is not much of an issue as ML runs normally varies slightly from run to run.
https://t.co/wnkU6f8mNc
Current RSI benchmarks have serious defects.
We reviewed 6 benchmarks and found flawed tasks in 5.
On FrontierCS 2, the authors' own code often fails their own tests; on one task, about 5 times out of 6.
31-63% of tasks had errors. Receipts below. (1/7)
MaxxBench illustrates that models don't treat your prompt as a spec if they think the judge will punish it, i.e., change Chrome.xml to chrome.xml and models will literally talk themselves out of listening to the prompt.
Introducing FrontierPhysics - a benchmark evaluating agents for end-to-end physics research
Our core team from physics labs at Berkeley, Caltech, and Harvard recruited 42 physicists from 26 institutions. 28 of them published papers on Nature, Nature Physics, PRX, PRA*
1/7 🧵
Subagents are the biggest cause of AI failures
In most harnesses:
1. Main agent writes a summary of what it knows
2. Subagent starts from that summary only, and makes assumptions
3. Subagent works and sends back a new summary of what it did
Same flaw as compaction: all summaries are lossy. This is how your assistant makes emails you never wanted, or purchases what you told it just to get the price of
So I tried something new: checkpoint agents
The initial agent has complete context of what needs to be done. And that input is already cached
Fork the session where it is. Send the fork ahead to do the task and it will do exactly what the main agent would have done; their context is identical
What is even better is when the fork finishes, the cache of the main agent is kept warm by the subagent (they share the same prefix)
I measured it on a 137k-token session with Opus 5.5:
→ Main agent does it on its own: 67 s, $0.38, context +8-17k
→ Checkpoint: 117 s, $0.68, context +2k
→ Classic subagent: 123 s, $0.57, context +4-5k
But here is what I found: avoiding large input is even more important than keeping cache. And most subagent tasks were done better either in series with retrieval, or split up to be parallel over fresh sessions from the start
I do not think there is a place for subagent delegation at all. There should either be a super-agent that orchestrates and does no work, or a long running agent that can retrieve from previous sessions, not a worker that completes tasks partially and delegates the rest to other workers
Don't design harnesses based on our assumptions of human behavior
Can AI forecast elections better than experts and prediction markets?
We built BallotBench to find out: a live benchmark where 9 frontier AI agents compete to predict the 2026 US midterms, from today through Election Day, Nov 3.
Follow along → https://t.co/7aLmUy2HvO
Watch the agents do their research, revise their predictions, and make their calls all in live feeds.