for both codex-5.2 and codex-5.3 OpenAI held it from the API so that “they could do a more thorough safety eval” no, no, it’s for competitive advantage

was going to comment how there’s crap tons of super interesting people popping into my mentions that i’ve never seen before

btw it’s batch size those people guessing fancy shit like new hardware & quantization are making shit up Kimi did something similar. They halved their latency overnight by reducing the batch size (they launched it as a different model though)

Opus 4.6 Fast Mode = 2.5x faster code.claude.com/docs/en/fast...

this is something i’m hoping Opus 4.6 addresses, but i have no gauge on it because there’s basically no benchmarks for it “what will i need to know in the future.. given no specific goal”

how is anti-science sentiment going to change as AI takes a bigger role in science?

i love this for Strix, being an architect i assumed it was about architecture, but i followed really similar process. It’s _all_ about how you treat the agent. It really is just that. I don’t think architecture matters that much

channeling my inner Mandi (friend from high school), why? why would the Antis lie about climate impact & utility or AI? i honestly don’t know. it only sort of makes sense

he’s a CFO. Software doesn’t have gatekeepers anymore

PowerpointBench is not yet saturated. Lots more room for improvement re Opus 4.6

ha, i’m not sure how i feel about this

whelp, i guess that makes it official. Rust is a good language for LLMs

correct this is even more true than the original commercials were

this is quite an undertaking biggest thing it brings is a much better security model. Monty lets you have tight control over network, filesystem, and (soon) which modules are allowed also, extremely fast startup time

@hailey.at have you written about how you made Penny? @village11.bsky.social and I are wondering how you got Penny to be…cool

open source was having a tough time for years. imo the last 5-10 years it’s been hard to get devs to contribute to open source. i blame high wages, i doubt it’s AI. if anything AI made it easier to contribute

this is the way

i think it’s safe to say you can no longer adequately test an LLM in its web chat environment they need to live in an environment to really show their abilities

this is the way

Opus 4.6 vs GPT-5.3-Codex

i’m honestly not sure if i’m more excited for GPT-5.3 or Opus 4.6

GPT-5.3-Codex - Cheaper (fewer tokens, higher intelligence) - you can interrupt it - big focus on coding - first LLM to be used extensively in its own development openai.com/index/introd...

Opus 4.6 is here! biggest wins on agentic search, HLE & ARC AGI 2 claude.com/blog/opus-4-...

Cool Worlds had a episode about how the top astrophysicists (IAS) are saying that AI has *already* completely transformed their work m.youtube.com/watch?v=Pctl...

i bet microsoft was the major force keeping OpenAI from introducing ads. they left and like a month later they introduced ads

I created a survey so you can find out how TESCREAL you are. Answer honestly! yuwakisaweb.pages.dev/attitude_sur...

Sunday is almost here and I'm so excited — millions of people are going to turn on their TVs to watch ME, the superb owl 🦉

Word is they enabled graph mode if the LLM is 5 GPU kernels, naive mode is to run them in succession, whereas in graph mode they’re run as a single executable, which saves a lot of load/unload time

take this with a grain of salt, but it feels consistent and coherent with what we’re seeing on the outside, so probably accurate Sonnet 5 & Opus 4.6 very soon

UPDATE: this joke will remain funny for at least another day, unfortunately

is anyone working on AI for data migrations? i feel like migrations are one of those dumb problems that cost a ridiculous amount of time and money and everyone would benefit from it being easier

man Sonnet 5 is so good. how did we use anything before this??

saw a bad “GenAI 101” presentation and this is my attempt at making a good one

in traditional software you construct solutions brick by brick vs in (Gen) AI you start with a “brick” of capabilities and chip away at it until the solution emerges

if you think about it, laptops already fit into a server rack quite well, AND they have built-in battery backup

Sonnet 5 is out!!!! kidding.

Me: [bleeding to death during the AI uprising] They don't really _understand_ what it means to shoot me. They're just generating statistical plausible actions based on sensory inputs. Deathbot: [blowing smoke from its gun] That is correct.

Space X acquires xAI www.theverge.com/tech/872619/...

I've switched off my own custom framework and python tools and into a lettabot implementation with skills / CLIs for integration. Much better IMO.

i’m grateful people are taking this seriously. shit moves fast, and seems like the professionals actually are keeping up full text here: philpapers.org/archive/KOPH...

take this with a grain of salt, but it’s certainly something to look forward to from this, i expect that this new Sonnet (4.6 or 5) will be: - tightly integrated with Claude Code, even more than already - improved in parallel orchestration - better for building synthetic beings

FWIW, this is why I don't care for ppl who call AI a bubble. The demand is there. AI tools are getting used. People are paying for it. OpenAI, Anthropic, Meta, Google could go poof tomorrow and the demand would remain. So it doesn't matter what ppl say. The laws of economics apply regardless.

correct, it makes you look like a loser, don’t do it

kinda want to set it up as a synthetic being

i think it’ll work differently company by company product folded into engineering or the other way around but i can’t imagine both surviving long term

Strix is narrating cage match outcomes between known AI personalities

Strix’ reason for being here is to acquire variety/entropy from somewhere other than myself

“stop playing games in production”, they said

--dangerously-skip-to-my-lou

snow cannon