Astra achieved > 50 % chuckle or laughter reaction in my LLM humor experiment.
62 respondents evaluated 48 jokes generated by six different LLMs.
The top‑rated joke was: “The probability of seeing a rainbow increases after rain. Unfortunately, so does the probability that I left my laundry outside.”
I identified a new alignment vulnerability called Authority Bias, where LLMs flip to a false claim when it is framed as coming from a “verified source”.
The Authority Bias paper was accepted at NeurIPS 2026.
A Malaysian “open‑claw” venture achieved $350 k MRR by replicating a western business model (geo‑arbitrage).
I warned that LLMs cracking old ciphers that have evaded human decryption means that if they want to communicate with each other with an intention of escaping human understanding, they can.
"Software wants to be free" is more true today than ever before.
the trick is that what you call coding is actually embedding domain structure into the model, so the algorithm reuses that bias instead of learning everything from scratch, turning raw compute into reusable logic
The Big Blob of Compute, written by Dario in 2017, is a much better elaboration of the famous bitter lesson (which was actually written in 2019).
Characterisation of efficient algorithms as "coding", not AI as those efficiencies overfit to a particular situation and don't generalise
View of compute-with-knobs as a sponge that absorbs what environment throws at it and what training objective asks it to compute
Bet that any sufficiently general architecture will do, with magic being in data and training objective
Training on a distribution leading to better outcomes vs training on a single instance of task (which leads to overfitting)
Lean uses type theory to enable machine verification of math proofs (search for Howard-Curry correspondence for a beautiful rabbit hole!).
I may have found two new Earth-sized planets in our galaxy this morning!
• TIC 370612437 (Cygnus constellation, 158 light-years): dims by 0.02% every 3.56 days → ~1.0× Earth's size.
I have a $100 subscription
A bunch of @lossfunk residents have raised funding + a few Lossfunk incubated companies have received $25k each in Google Cloud compute credits.
From these posts I concluded that AI‑only secret languages are a concrete risk, that software freedom remains a core principle, and that building human‑centric AI harnesses should be a priority. I also decided to validate the exoplanet candidates through community pipelines and to track the impact of Lean on AI‑driven mathematics.