test time harness self improvement...

prediction: reasoning by isomorphism or hard analogies will unlock significant advances in scientific discovery. this is very little understood, and was a bottleneck becasue of human capability, motivation, and incentives. there are early signals in recent discoveries, but this remains largely underappreciated. @3blue1brown made me aware of this in his podcast with @dwarkesh_sp when he talked about the Langlands program...

under certain monotonicity assumptions, compression ≈ prediction. does intelligence reduce to compression, or is compression a consequence of intelligence? 🤔 arxiv.org/html/2610.11031v1

every day, when i wake up, i think about the Roman empire and K-lines.

is business management "science" pseudoscience by Popperian standards? @PAHoyeck :(

theres never been a better time to have unbounded curiosity.

Level 0: execute rules Level 1: find solutions Level 2: find algorithms that find solutions Level 3: find representations in which good algorithms exist Level 4: discover objectives/evaluators for judging those representations Level 5: improve the discovery process itself.

[...] The action of thinking may incidentally have other results; it may serve to amuse us, for example, and among dilettanti it is not rare to find those who have so perverted thought to the purposes of pleasure that it seems to vex them to think that the questions upon which they delight to exercise it may ever get finally settled; and a positive discovery which takes a favorite subject out of the arena of literary debate is met with ill-concealed dislike. This disposition is the very debauchery of thought. [...] - Peirce, 1878. Uncannily relevant to our times.

Being gRaciouS to all past and new Innovations is very important!

When something (e.g., a field, topic) seems "solved", the more important it is to focus on learning the most about it.

great resource ergo.org/

Soon we will have humans designing for agents more than for humans (agent computer interface) before agents design for agents.

State of AI meta prompting for problem solving Next step: automated investing with higher alpha than Ms. Pelosi.

What's this f* infatuation LLMs have whenever a system handles "refusals" "claim discipline" "epistemic hygiene"... all a result of post training RL to reduce hallucinations. I'm sure this is creating suboptimal consequences at inference time.

We need models that don't constantly hedge and strawman.

Guess it's back to Claude, what a rollercoaster. I couldn't get Codex to stop overengineering and hill climbing.

Very interesting paper, with very interesting results. @SchmidhuberAI arxiv.org/html/2605.14831v1 been using an approximation of this for nearly 2 months