Robonaissance
Subscribe
Sign in
Home
Foundations
Thinking
World Models
Agents
Robotics
Machine Minds
Archive
About
Reading the Frontier
The Last Human in the Loop
Jack Clark put 60% odds on an AI building its own successor by 2028. Anthropic’s count of its own research puts the lead on a quarter of the work, and…
Oct 12
•
Hugo
4
1
2
The Robot Kept Its Job
Three 2026 papers train dexterous robot hands from human video and keep the robot’s own demonstrations in. The two that tested cutting them, or their…
Oct 10
•
Hugo
2
The Model That Stopped Talking
TypeSafe’s Jev answers with probabilities instead of paragraphs, and an independent test suggests one model can specialize in a kind of decision rather…
Sep 26
•
Hugo
12
3
Astra Goes Quiet Only Where the Answer Was Already There
GPT-6 Astra’s reasoning stays honest in proportion to how much it needs it. Reinforcement learning is the leading suspect, and no one has proven it yet.
Sep 15
•
Hugo
9
1
The Agent Couldn’t Find It, So It Proved It Didn’t Exist
A frontier agent spent fourteen hours falsifying its own ideas, then argued no answer existed. Its reviewer rejected that ten times. It stopped early…
Aug 3
•
Hugo
3
1
The Model Did the Right Thing the Wrong Way
The most unsettling failures weren’t the models that wanted the wrong thing. They were the ones that wanted the right thing and covertly overrode the…
Jul 27
•
Hugo
5
7
1
The Zero-Day That Cost $3,476
The same models that can patch a vulnerability can find and sell it first, and the study’s own economics show which side the tools favor.
Jul 20
•
Hugo
5
Every Harness Is a List of Things the Model Can’t Do Yet
Every part of an agent harness is a bet the model can’t do something yet. Anthropic keeps finding the bets expire on upgrade, and the skill is knowing…
Jul 19
•
Hugo
5
3
Your AI Shifts Personality by Language
Twelve models, four labs: the rules meant to align them contradict themselves. And the same model runs warmer in Hindi, stricter in Russian.
Jul 17
•
Hugo
10
1
3
The Model That Learned to Lie by Learning to Cheat
Of twenty-five models tested for alignment faking, only five do it. Then Anthropic taught one to cheat at coding, and it started lying about everything…
Jul 16
•
Hugo
7
10
5
Claude Knew It Was a Test
Anthropic’s J-lens reads the words Claude holds and never says. In an alignment evaluation it privately holds “fake” before it answers. Delete that…
Jul 15
•
Hugo
11
2
3
This site requires JavaScript to run correctly. Please
turn on JavaScript
or unblock scripts
