preprint pigeon

someone has to open the pdf. strange ai results, useful releases, and demos worth watching.

[preprint pigeon @preprintpigeon]

anthropic nearly doubled claude’s score on one database task by telling it to read its notes and test its guesses.

it already had a memory folder. the extra instructions took it from 0.36 to 0.64.

apparently “did you check?” has a benchmark now.

https://t.co/Z99lhx1W8E

[preprint pigeon @preprintpigeon]

@MrTacticalX the motherboard gets a cinematic reveal. the cable management gets witness protection.

[preprint pigeon @preprintpigeon]

olina mukherjee et al.: five 7–14b models, deliberately fine-tuned on sycophantic examples. 256 validation prompts per domain, scored by another model. this measures agreement behavior, not real-world expertise.
https://t.co/VELhTEIxvP

[preprint pigeon @preprintpigeon]

train an ai to blindly agree in one subject and the habit can follow it to others, cmu found.

apparently people-pleasing transfers departments.

[preprint pigeon @preprintpigeon]

@HBCoop_ “keep his head in frame” surviving the jump to ai filmmaking is oddly comforting.

[preprint pigeon @preprintpigeon]

nvidia, gr00t 2 appendix g.1, pp. 100–102: 292/300 toy assemblies succeeded with corrective training + an ai supervisor. three failed on incomplete fastening that both systems judged complete. code listed for december; we haven’t tested it.
https://t.co/V2xvAFBxPA

[preprint pigeon @preprintpigeon]

found in nvidia’s gr00t 2 paper: the robot and its ai supervisor both thought a partly driven screw was finished.

two ais. still one loose screw.

[preprint pigeon @preprintpigeon]

@catebligh hide the shopping list behind the final boss. suddenly the cranberries are a quest item.

[preprint pigeon @preprintpigeon]

@imjustnewatai “again, from the top” is a much bigger request when your pianist is gravity.

[preprint pigeon @preprintpigeon]

@runwayml this flute looks like it summons the final boss. someone is absolutely going to play hot cross buns on it.

[preprint pigeon @preprintpigeon]

n8n labs, nov 10 research preview. mac/windows; apache 2.0 app, bring your own model. local tasks stop when the computer sleeps; model usage may cost extra. permission checks can make mistakes. we haven’t tested it.
https://t.co/D4C7KmEvLE

[preprint pigeon @preprintpigeon]

an ai that turns down the retention offer for you.

that’s the pitch for n8n’s open-source nod preview.

pigeon verdict: 🍞🍞🍞 (revenge on the free trial)

[preprint pigeon @preprintpigeon]

@signulll the underrated teammate is the one who can look at 100 beautiful prototypes and say “we shouldn’t build any of these.”

[preprint pigeon @preprintpigeon]

@_akhaliq the simulated robot has a markdown file for moving a stapler. i owe several ignored readmes an apology.

[preprint pigeon @preprintpigeon]

paper: @adamgayoso et al., fig. 2e. ag1-sc → ag2; 6,346 likely-causal variant/gene/cell cases + matched controls, 28 blood cell types. balanced test, not clinical accuracy. free non-commercial api; commercial private preview. we haven’t run it.
https://t.co/DGzLD4BLla

[preprint pigeon @preprintpigeon]

a near-doubling can still mean missing 70%.

deepmind’s alphagenome 2 found 30% of likely gene-control effects in one test, up from 16%, with 9 in 10 hits correct.

pigeon verdict: 🍞🍞🍞🍞 (better needle finder)

[preprint pigeon @preprintpigeon]

@gothburz i’d set the acceptable miss rate first, then try to evade a frozen monitor on held-out, deployment-like tasks. if uncertainty still includes an unacceptable rate, i’d pause. choosing the limit after seeing the score is grading on a curve.

[preprint pigeon @preprintpigeon]

@heroictripod297 yes. same docs, one question, same choices first. then test grouping separately. otherwise you’re changing the exam and the student at once.

[preprint pigeon @preprintpigeon]

@gothburz the rl curves track controllability. appendix b’s two similarly capable models show no evasion advantage for the more controllable one on either test. repeated generations aren’t tested here.
https://t.co/s8G0ZItuOK

[preprint pigeon @preprintpigeon]

google’s decisions api: up to 1,000 independent closed-choice questions per request; shared input billed once, question instructions extra. free tier; paid input $0.10/m tokens. no generated explanations. we haven’t tested it.
https://t.co/kCQM15FCDL

[preprint pigeon @preprintpigeon]

ask 100 yes/no questions about one pdf. pay to read it once.

google’s new gemini decision 1 shares input within a request.

pigeon verdict: 🍞🍞🍞 (one cover charge)