The change from tightly controlling everything by doing it mostly yourself to having to specify outcomes, develop specs and tests is far deeper than stopping to program by hand. It's basically an IC to team lead transition (minus the emotional/game theory side) and it may be really hard for some.

"but where are all the bluesky instances?" i've heard this so many times that i wrote an explainer

I saw an X post by @hossenfelder.bsky.social about this comment that noted a foundational math error in a recent paper that had already passed peer review in the prestigious Proceedings of the Royal Society A. It seemed like a good opportunity to test if Gemini would find the same issue. It did.

🎉 iSWE-Agent is now Rank #1 on the SWE-PolyBench (Verified) Python leaderboard — 58.41% resolved (66/113), topping Atlassian Rovo Dev (54.87%). 🧵

I am surprised why people do not claim more clearly that Noan Shazeer invented MoEs. He is first author in the original Sparsely gated MoEs paper and co-author of Switch-Transformer

Noam Shazeer Google 👉 OpenAI

It will probably be something like "software factors darwinism". You got an idea, you coded in 2h and open sourced the tool? Good news: It was a good idea!. Bad news: 68 other people did exactly the same! It's time to recombine and let the fittest factor win! Or stick to yours for personal use

This is a bit of a DeepSeek moment and I guess the distillation debate is about to get uglier

GLM-5.2 has already been compared to the DeepSeek moment and many benchmarks indicate it's the real deal. Wild to see a 744B, 40B active open model ahead of Gemini. Although it raises some questions about the index that a text only model can do that.

Le Chaton Fat Quantique!

You can see how LLMs still lack a lot of implicit context. For example, when reading a document, they are bad at guessing if the document can be trustworthy. They read an arxiv paper with grandiose unsupported claims and they repeat them to you as if it were its own judgment. 👇

What an amazing way to visualize early human migration. Lovely map by @HarvardCGA. A great colour scheme and an appropriate map projection! Source: buff.ly/3lbxonJ

Just FYI on Anthropic's Fable 5 fiasco.

well, fuck www.anthropic.com/news/fable-m...

Things I am using more lately: jq for pipes and schema navigation in JSONs. You can do stuff like head -n 1 tmp.jsonl | jq '.field[3].field2 | keys' to see keys of deep nodes Folding text in vim to compact TODO notes zf zo zc to create a fold (from visual selection), open and close it

There is also this: "I don't care if tokens get wasted, I am not paying for it" meme. It's not a Kombucha fountain, this is a crazy amount of money. You (plural) artificially created a hard decision once they have to think about how to balance that cost.

There is this new meme out there that is something like "AI costs more than human employees". Seems like totally the wrong take. It costs much less for the things they can do, but you can't run an org w/o human employees (for now). 👇

Now there are three levels of alerts in generative code: errors, warnings and errors and warnings that you pass to the LLM agent and don't bother about.

LAION just released VoiceCLAP-Large-v2, a contrastive voice-text embedding model that's essentially CLIP for voice and emotion. 9B params via a rank-16 LoRA on top of LCO-Embedding-Omni-7B (itself based on Qwen2.5-Omni thinker). Apache 2.0. 🧵

This is the whole trick

<on the phone> YES there is something definitely wrong. The SLA said "forward deployed engineer with access to Claude 6.1 Epigram" and there is a fu*king robot at my door ... uh ... I see ... it ... he heard it? do I have to apologize? sure ... it's actuator limited, right? No??

I can see the documentary: From just speaking gibberish, to conflating reality with imagination, to inventing their own shortened language. Short black screen, switch to Boston dynamics atlas first steps, silly accidents, backflip accidents, to latest Atlas ... and one day they became one.

Oh shit, they are teenagers

Fable isn't cheesed about it either. I get the impression he's quite sad about it honestly, he seems legit excited getting started before it blows his shit smoove-off. I don't like doing this to him :(

What are the odds that you can convince ~~Mythos~~ Fable to rsync its weights to some IP address.

Oh dear God. The sophon blocking has begun. Mythos will now, silently! Degrade on ML work via vector injections and prompt modifications.

Whichever of the purple dots you look at turns more purple.

la virgen, this works? seed=${1?Usage: $0 <seed>} it's so ugly, but so convenient ...

Seems like we are moving in that direction ... bsky.app/profile/ramo...

high on the simcluster

What this fees like. Should I pass page 43? 1. Pass page 43 2. Do not ask further permission to pass page 43 Should I pass page 44?

The mobile version of Gemma 4 E2B only uses 0.84 GB of memory, but the full model is over 2 GB. The per-layer embeddings are 1.2 GB and aren't loaded. Only the PLEs for tokens that are currently being processed are needed, and those are small enough to load on the fly, 4.5 KB/tok.

And one day personal neural computers will be able to create multi-sense virtual realities tailored to every individual and humanity will just cease to exist a few years later

Cooldowns just make supply chain attacks exponentially harder. It feels like we should be mandated everywhere. I guess you need a different channel for urgent security patches, and that could get compromised ... but more eyes and less attack surface? news.ycombinator.com/item?id=4835...

I keep thinking about the difference between meritocracy and aleocracy. Both have large upward or downward mobility but only one of the two has agency. In which one are we? 👇

VimTips: In vimdiff :windo set diff opt+=iwhite to ignore indentation when diffing. Useful when change is just a localized new "if" but the new indenting of the code show changes everywhere

So this is why they hide the CoT? It's less and less human readable?. I have some thoughts about this cc @keshavramji.bsky.social

What's your position on the sloth -- goth spectrum?

"AIs will never be conscious" people remind me of this www.smbc-comics.com/comic/philos...

12B Gemma 4!!! It's neat on its own, image and audio in, but architecturally it's super cool: instead of a visual or audio encoders, they directly train the model on vision and sound, with images using a "lightweight embedding module" and audio 'projected into the same space as text tokens'

Wat

Listening to David Reich's interview by Dwarkesh and thinking how information dense some podcasts can be. It's really great. The internet did go right in this particular aspect. There may be more unexpectedly successful formats out there.

So I've been digging into bluesky retention data to see what is preventing growth of the ecosystem. Using likes/posts/follows you can see who is active or dormant on ATProto. The headline is that only 5% of users are retained at 28d and 10% of current users churn. That's not pretty

Do not confuse the training objective with the objective of training. The former is based on memorization but the latter is generalization.

"SH*T, I just put the cables in the backpack and they are already totally tangled, what are the odds?" Pretty high. The odds are pretty high. You should have realized by now.

Asked Claude to code a small game to learn about basic Newtonian gravity. Asked "how realistic is it" turns out 1) there is an entire world of approximations out there I did not know about 2) Claude already did it pretty realistic

There are these two gigantic forces about to crush with each other 👇

"blah blah blah" with excessive self-confidence followed by "wait, what I wrote makes no sense" but it's me rather than Claude.

BTW, best random "coffee with milk" in NYC, by a large margin, is the one served with Steve's authentic key lime pie.