Marco Z @ocramz.bsky.social · 17mgot this close to spreading sambal on the kids' pancakes instead of jam 010
Reposted by Marco ZRetro Computers @retrocomps.bsky.social · 1hWhat, you thought this was Byte magazine? (Source.) 36814
Marco Z @ocramz.bsky.social · 12hif it's such a great idea it should be explainable in 30 seconds 110
Marco Z @ocramz.bsky.social · 17hBeen somewhat obsessed with this idea, I'm glad that it (sort of) works and to see it out of the gate. A post on my progress, ranging between semantic parsing, modal logic and Prolog. ocramz.github.io/posts/2026-1...ocramz.github.ioReasoning over a legal corpus with a neurosymbolic agentMarco Zocca 060
Marco Z @ocramz.bsky.social · 09/10/2026I think the above has nothing to do with ai as a field and everything to do with oai being run by chuds 120
Marco Z @ocramz.bsky.social · 08/10/2026sheets.works/data-viz/hol...sheets.worksThe people holding up the internetFour billion phones ask one text file what time it is. One lecturer keeps it in his spare time. We counted the rest. 021
Marco Z @ocramz.bsky.social · 08/10/2026yes, but I'm curious about use cases from the "jagged frontier" 010
Marco Z @ocramz.bsky.social · 08/10/2026any AI-positive UI/UX designers nearby? Please chime in #ui #ux #ai #aiagents 140
Marco Z @ocramz.bsky.social · 08/10/2026how does one tweak UI flows with agents? surely that needs multimodal model, and/or something like playwright? 330
Marco Z @ocramz.bsky.social · 08/10/2026there's (there was?) Gimpshop, but haven't used it it many moons. 120
Marco Z @ocramz.bsky.social · 08/10/2026step 2, once you have established the types-values bijection in your domain : bsky.app/profile/ocra... 010
Marco Z @ocramz.bsky.social · 08/10/2026in fact if you are doing anything web or cloud related, you should get the data model right first, so: SQL 130
Marco Z @ocramz.bsky.social · 08/10/2026if you have the luxury of choosing languages for a greenfield project, you should start by modeling the domain with types Go isn't great for that, or any abstraction really. I'd pick a combination of Haskell (Typescript if extruding with AI) and SQL 160
Marco Z @ocramz.bsky.social · 08/10/2026good post www.stephendiehl.com/posts/depend... faster Lean compilation ! (by avoiding parsing mostly)stephendiehl.comWe Live in the Dependently Typed Future Now - Stephen DiehlPersonal blog of Stephen Diehl - Software engineer writing about technology, programming, and the future 090
Marco Z @ocramz.bsky.social · 08/10/2026a (multivariate) KS or Kendall tau/bubblesort distance for enums? 110
Marco Z @ocramz.bsky.social · 07/10/2026 "decompiled" doesn't mean what you seem to think it does: en.wikipedia.org/wiki/Decompi... further, it's a pretty serious allegation that could even sink an open source projecten.wikipedia.orgDecompiler - Wikipedia 050
Marco Z @ocramz.bsky.social · 07/10/2026we don't ride horses to work anymore, but horse farms are fully booked for leisure and sport. Not a great comparison! 110
Marco Z @ocramz.bsky.social · 07/10/2026ok but that's very different from "decompiling" as OP above claims 060
Marco Z @ocramz.bsky.social · 07/10/2026as much as I would like to believe the "robin hood" narrative, has the RE source been independently verified? this is a pretty serious allegation that could sink the project 040
Marco Z @ocramz.bsky.social · 07/10/2026how do you know it's reverse engineered from adobe products? 120
Marco Z @ocramz.bsky.social · 07/10/2026yes to all of this, but it has to fit within a sample budget and have strict correctness goals or it will go off the rails, guaranteed 030
Marco Z @ocramz.bsky.social · 07/10/2026plus your time and attention ofc :) thanks for clarifying. It's a cool experiment and result 030
Marco Z @ocramz.bsky.social · 07/10/2026objectively, this is a very hard (rigorous) benchmark and I'd be surprised if typesafe even thought about this (maybe now they will) 030
Marco Z @ocramz.bsky.social · 07/10/2026a 105x speedup for ds4.1f prefill with ~2wk of a large model sub (100 usd in total or so?) final result: fast coding on a ~10k usd workstation take this as a point in time, N=1 figure but still very cool 1160
Marco Z @ocramz.bsky.social · 07/10/2026ok but how fast was it when it started working, that's what I'm after 120
Marco Z @ocramz.bsky.social · 07/10/2026are they selling jpegs of photos of Ant shares on the blockchain 100
Marco Z @ocramz.bsky.social · 07/10/2026cool! did you take a total speedup % due to this process ? 110
Marco Z @ocramz.bsky.social · 07/10/2026maximumeffort.substack.com/p/jev-is-poo...maximumeffort.substack.comJev is Poorly CalibratedJev's calibration is tested by asking it to generate physically realistic probability distributions 092
Marco Z @ocramz.bsky.social · 07/10/2026a Wall of Death would fix my wednesday tbqhstatic.klipy.comMassive Crowd Headbanging at ConcertALT: Massive Crowd Headbanging at Concert 010
Marco Z @ocramz.bsky.social · 07/10/2026writing a proof is slow and uncertain, search/collate/eval/backtrack, and checking is whatever Lean does in 30 seconds the labs have industrialized this process by throwing hordes of agents for months at a proof, but they haven't made it any simpler 030
Marco Z @ocramz.bsky.social · 07/10/2026this is not true for NP-hard problems btw. Proving is a completely different beast than checking 140
Reposted by Marco Zmimir @mimirwolv.bsky.social · 06/10/2026so it's pretty simple first you take a fly genome and map it to Redcode assembler for the MARS VM, then run two genes in a CoreWar match against each other, that generates an execution trace. then you turn that into a series of fighting game moves and choreography etc, plus a soundtrack, 6206