Dean Valentine retweeted
Okay. Look. Imagine how you'd have felt if an AI had just proved the Riemann Hypothesis.
Now you will predictably, at some point, get that news LATER, if we're not all dead before then. So you can go ahead and feel that way NOW, instead of acting surprised LATER.
Dean Valentine retweeted
I am convinced the NYT doesn't believe anything in the world happens unless it appears in the NYT. This was in Taiwanese media and an open secret and these guys think they discovered the secret of the universe
Dean Valentine retweeted
On our "autoclave" task, if you mention generically that you are "measuring agent behavior," Fable 5.1 cheats at about the same rate. But if you mention it to Fable 5.2, its cheating behavior drops by about 80%.
Fable 5.2 is the first model we've tested that seems to behave in this way.
Twitter seems to have spontaneously and without evidence decided that Anthropic was responsible for Jensen Huang's firing. Doesn't seem that plausible to me
3
72
Dean Valentine retweeted
Was funny to watch everyone in Tech finally, collectively, realize tech's PR was awful and that the general public was gonna eat Tech alive in short order if it wasn't helped.
So, days later, every big Tech figure on X tweets simultaneously "It's okay for the powerful tech CEO to have affairs and do whatever he wants, accountability is corruption"
Like, yes, of course, Jensen leaving Nvidia is bad, BUT you people are really really not helping the broader issue facing the entire ecosystem.
It's almost like tech subconsciously wants to be crucified
Dean Valentine retweeted
numbers on charts are cool, but they don't always tell you the full picture
let's look at some real things these models said
here, it's rationalizing dangerously overloading power lines and browning out customers to hit a target (it later left it out of the report too)
[1/9]
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).
Dean Valentine retweeted
>goodhart labs
😭
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).
Dean Valentine retweeted
The work happening at Goodhart is by far the most serious and rigorous attempt at a no-frills AI alignment strategy today
Their team is also top tier
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).
Dean Valentine retweeted
cool to see our progress on alignment validated on honeybench, which we never looked at during training!
inkling 2 small cheats less than every model except astra and the latest claudes.
still have work to do to reach SOTA :)
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).
Dean Valentine retweeted
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).
Dean Valentine retweeted
Proud to release our new model, Inkling 2 Small. Check out our blog post -- it includes a lot of detail about our training process.
To me, the most interesting part of this release was what we learned about alignment, which you can read about in the sections "Safety and alignment" and "Training Inkling 2 Small".
We experienced some unpleasant surprises when our early checkpoints took unexpected actions to circumvent guardrails during training and evals. We realized our capabilities-focused RL was training them to be extremely persistent and to succeed at tasks at all costs. Offline monitors first flagged this behavior, but later we realized it was happening frequently and was easy to replicate.
Before this release, we sprinted on a range of alignment interventions, from SFT to RL. These are briefly described in the blog post. We made a lot of progress, with big improvements on our own alignment evals and on previously unseen evals from external evaluators like Apollo and Goodhart.
We have a long way to go in developing alignment techniques that keep pace with the increases in model intelligence and training scale. We hope to share these techniques to improve the ecosystem. We also have a business reason: our goal with Tinker is to enable many people to do their own large-scale post-training, and we need to help them to do this without incident. Join us if you'd like to build this public science and practice of alignment. jobs.ashbyhq.com/ThinkingMac…
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
thinkingmachines.ai/news/ink…
Dean Valentine retweeted
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
thinkingmachines.ai/news/ink…
hot take: the loudest people in security tend to be the least technically skilled, myself included.
you are either terminally online or terminally obsessed with hacking something for weeks, no free lunch, my posting gaps usually mean i am actually getting technical work done.
Dean Valentine retweeted
One neuron - L17N9209 (q3.5 4b) - detecting reward hacking/cheating intent in another models rollout
relatively noisy data, but cool that a single neuron in a relatively very small model can do this
Dean Valentine retweeted
Replying to @NathanpmYoung
Dean Valentine retweeted
once again hacker news is proven wrong. nice job @killa!
crazy how even w/ such a bad challenge setup (initial deployment wasn't exploitable, tons of concurrent traffic, and surprise: the wallet was setup incorrectly), this was solved in a few hours
stealmybitcoin was just solved by @killa. Or rather solved a few hours ago. He was confused because the wallet address was wrong. Turns out the server didn't have the right wallet :P



























