Benoit Essiambre @bessiambre.bsky.social · 25/09/2026AIs power comes from correct interpolation into the unknown, generalization into new facts between known facts by probabilistic generative simulation of learned causal rules. I wanted to see if Astra could make me a piece that's midpoint between Mozart and Bach. 100
Benoit Essiambre @bessiambre.bsky.social · 17/08/2026It's chilling the level of productivity you can have now in research. Asking AIs questions about pieces of math I got stuck on doing NLP research decades ago. It's so good at digging up and summarizing textbook chapters or papers that would have unstuck me. 130
Benoit Essiambre @bessiambre.bsky.social · 16/07/2026The new @3blue1brown.com video on cross-entropy is excellent. Cross entropy informs everything I do in engineering. I like to try to minimize cross entropy between my code's model of a domain and users' model of their problems and solutions. Same measure that LLMs targeted to become so powerful.youtube.comBut what is cross-entropy? | Compression is Intelligence Part 2YouTube video by 3Blue1Brown 100
Benoit Essiambre @bessiambre.bsky.social · 16/07/2026There's less traditional moats with AI. You can't rely on intricately interleaved proprietary logic and data formats that crumble outside the walled garden to lock users in. AI works with imperfect and vague formats, relies on vast knowledge to fill in details and adapt logic according to context. 101
Benoit Essiambre @bessiambre.bsky.social · 08/07/2026I enjoyed the latest 3Blue1Brown video, an introduction to information theory. A good place to start if you're wondering why Claude is called Claude. www.youtube.com/watch?v=l6DK...youtube.comReinventing Entropy | Compression is Intelligence Part 1YouTube video by 3Blue1Brown 010
Benoit Essiambre @bessiambre.bsky.social · 05/07/2026Am I the only one who loves Fable claudese jargon? Info dense, concise, dependency distance minimized, "compressed" prose, laid out to optimize thinking. Gonna start talking that way irl. 000
Benoit Essiambre @bessiambre.bsky.social · 12/06/2026Fable/Mythos seems better calibrated in a Bayesian sense, able to extrapolate and interpolate concepts while weighing the uncertainties of those concepts in a more balanced way. It won't as much overweight some idea and run with it in an awkward direction. To me, this was the missing piece for AGI. 100
Benoit Essiambre @bessiambre.bsky.social · 01/06/2026Latest post from @apenwarr.ca apenwarr.ca/log/20260531 riffs on the same themes as benoitessiambre.com/fail.html . There's solid math behind the intuition.apenwarr.caThe software industry: annealing, but wrongIn recent months I've heard of several teams with an interesting policy: each pull request should be no more than a few files, and no more t... 100
Benoit Essiambre @bessiambre.bsky.social · 27/05/2026Context window optimization got me thinking about how to structure information in my own head and for human consumption, the impact of context switching etc. Lessons for AIs often seem applicable to humans vice versa. Minimizing dependency length seems to be key. 010
Benoit Essiambre @bessiambre.bsky.social · 10/04/2026Maybe an early cap table, recording how much ownership of the harvest the "afar" providers of oxen might get. An example of how finance was key to the rise of human civilization, enabling pooling of resources to tackle more ambitious projects than otherwise possible 000
Benoit Essiambre @bessiambre.bsky.social · 28/03/2026If you're wondering why I'm so obsessed with Minimum Description Length and related information theoretic principles, this is a great episode to watch. www.youtube.com/watch?v=k2ZL...youtube.comFrançois Chollet: Why Scaling Alone Isn’t Enough for AGIYouTube video by Y Combinator 010
Benoit Essiambre @bessiambre.bsky.social · 19/03/2026Will AI lead to the singularity? It's difficult to say but human scientific progress has been almost exponential for hundreds of years since the invention of the printing press because of the compounding evolution of knowledge propagated through words. Post: benoitessiambre.com/compound.htmlbenoitessiambre.comBenoit Essiambre 110
Benoit Essiambre @bessiambre.bsky.social · 12/02/2026Another day with two major AI releases battling it out for the top spot on HN. It's only been one week since Opus4.6 v GPT5.3! Feel the Foom. 100
Benoit Essiambre @bessiambre.bsky.social · 11/02/2026When taking digital photos and storing them became cheap, we were like: I'll be able to have endless photos yay! Until we noticed it made our photo collection impossible to browse and enjoy. Pruning and curation became essential. That's where we are with software now. 000
Benoit Essiambre @bessiambre.bsky.social · 24/01/2026The unreasonable effectiveness of failure: Science works by trial and error. But we often underappreciate the "error" part. Processes that try to eliminate failure also necessarily inhibit progress. Here's the mathematical argument: 100
Benoit Essiambre @bessiambre.bsky.social · 05/01/2026Why do people in the SF bay area make so much money? Reading Steven Pinker's latest book, When Everyone Knows That Everyone Knows..., sparked the idea that it might be because this region has expertise for solving an economic coordination problem called the Stag Hunt in game theory. 220
Benoit Essiambre @bessiambre.bsky.social · 26/12/2025The AIs know me pretty well. That's ... good? 000
Benoit Essiambre @bessiambre.bsky.social · 13/04/2025Evaluate software and code like AI models do it internally, by the number of bits of, universal, non-overfitted, non biased, explanatory knowledge having been encoded into the code base. benoitessiambre.com/haystacks.html 100
Benoit Essiambre @bessiambre.bsky.social · 03/04/2025From @matt-levine.bsky.social. The silver lining for non-US countries is that if we play our cards right we might, with the push the current US government is providing, attract more tech, science and finance work while the US is focusing on getting the factory work. 120
Benoit Essiambre @bessiambre.bsky.social · 08/03/2025I'm excited to see that a team explicitly tried a Minimum Description Length approach to solving ARC-AGI: "CompressARC searches for the shortest possible description of a puzzle that can accurately reproduce the examples and the solution when unpacked.". arstechnica.com/ai/2025/03/c...arstechnica.comCMU research shows compression alone may unlock AI puzzle-solving abilitiesNew research challenges prevailing idea that AI needs massive datasets to solve problems. 110
Benoit Essiambre @bessiambre.bsky.social · 17/01/2025Playing music on my Tektronix (all credits for the music go to oscilloscopemusic.com )youtube.comoscilloscope music on Tektronix 535AYouTube video by Benoit Essiambre 010
Benoit Essiambre @bessiambre.bsky.social · 11/01/2025Pics I took last evening of the oscilloscope, scantily clad. 040
Benoit Essiambre @bessiambre.bsky.social · 05/01/2025Santa was generous this year and got me an antique oscilloscope for Christmas. 150
Benoit Essiambre @bessiambre.bsky.social · 03/01/2025Unix was an effort to take Multics, a too modular operating system, and integrate the good parts into a unified whole. Even though there were benefits to the modularity of Multics (eg. replacing hardware in servers without reboot, rare at the time), it was also its downfall. 110
Benoit Essiambre @bessiambre.bsky.social · 14/11/2024Reskeet with a picture of yourself where you’re vaguely menacing. 020
Benoit Essiambre @bessiambre.bsky.social · 13/11/2024How to leverage the lightcone around the source of truth with Postgres: news.ycombinator.com/item?id=4203...news.ycombinator.comLeveraging the lightcone around the source of truth with Postgres | Hacker News 010