Oklahoma City
Joined July 2009
- Tweets3.4K
- Following772
- Followers1.3K
- Likes34.6K
Pinned Tweet
Far too many people:
"Benchmarks are bad. I prefer [thing correlated at 0.9 with Benchmarks]."
omg 5.5 32B? a qwen finetune or?
Z.ai chief scientist Jie Tang just revealed the training pipeline behind the upcoming GLM-5.5 (32B, 122B, and 256B) during a lecture at Tsinghua University today.
Instead of relying on public datasets, their pretraining pipeline pulls from six separate internal engines:
- Web: 1 trillion pages crawled at ~10B pages/day with PB-scale template clustering
- Documents: 2 billion scientific PDFs parsed through GLM-OCR (4x faster with speculative decoding)
- Books: 50 million books with layout reordering and garbled-text filtering
- Code: 500 million repositories turned into executable tasks with real tests and issue descriptions
- Video: 10 billion videos for long-sequence world knowledge
- Audio: 300 million hours rewritten via ASR into trillions of tokens
Sam Snelling retweeted
I wouldn’t be surprised if 95%+ of devs spending $10K+ per month on AI, globally, were based in the Bay Area
Nearly everyone I talk here (devs, founders, CTOs etc) are AI-pilled in ways I rarely see outside. AI-pilled as in obsessed building more with agents, running agents when they sleep, designing loops to auto kick-off, running them during their commute. And talking with each other how and what they do and then taking those ideas and trying it
Sam Snelling retweeted
>on a late night hacking rampage
>get paged at 10:27PM
>oh crap
>they caught me hacking
>its mallace
>gotta make some shit up, own it
>tell them: Ah yup. That was indeed me.
>No mallace intended.
>ask Claude what to do
>”let’s blame the clankers anon”
>say I've been agentic testing. Just halted the harness. It’s off the rails.
>founder is getting tagged on X formerly known as twitter
>founder pretends he didn’t know
>”tell them the agentic swarm did it”
>says: Woof! Looks agentically ‘off the rails’ to me!
>who needs lawyers when we can blame clankers
>crisis averted lol
Sam Snelling retweeted
Devs a year ago: "Code is a liability. PR's that delete 100 lines of code >> PR's that add 1,000 lines"
Devs now: "This ai slop canon can produce 50,000 lines of unreviewed code a day and you'll destroy this company's future if you delay it from merging"
Sam Snelling retweeted
*record scratch* *freeze frame* Ah yup. That was indeed me. No mallace intended. I've been running some agents to do crime on similar platforms.
I am once again absolutely convinced no Apple executive actually uses Siri.
It shocks me how dogshit it still is.
John Ternus answers investor question on where Apple is at with consumer AI agent capabilities:
“With Siri AI, we really have the kind of the Siri that we always wanted. It's primarily request based today, but as we talked about at WWDC, it's built on a modern orchestrated-driven architecture that gives us real room to expand its capabilities over time. As far as agentic capabilities, you know, we think these are promising tools for users. We just want to make sure that users are aware of what they're giving access to. We want to make sure this is done in a way that protects security and privacy. So, you know, we're pushing really hard on that with Siri AI, but we're also really encouraged the fact that Apple devices, we think, are the best places to use all forms of AI.”
Sam Snelling retweeted
> Railway gets caught hacking competitor's infra
> “sorry, my agents accidentally did it”
This is insane
Hey @Railway, is there a reason employee's of yours are trying to breach my infrastructure?
Your employee Jake Runzer, just attempted to tonight.
👀
A PHP Hello World, compiled by elephc: one 791 KB static binary. 🐘
No PHP installed. No shared libraries: ldd says "not a dynamic executable".
Compiles in ~115 ms, runs in 0.5 ms.
Ship PHP as a single file.
github.com/illegalstudio/ele…
this is a crazy ass premise. I gotta rt
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Sam Snelling retweeted
I like seeing this sort of comparison. I think mostly static/content based sites aren't really where we see the real cost. Most frameworks here perform reasonably. But these are the sort of questions we should ask before deciding what AI is good at.
nolanlawson.com/2026/10/28/t…
Sam Snelling retweeted
the perfect example of the frustrations of trying to work with OAIs recent models
The agents didn’t figure out how to systematically feed themselves, but they did build an elaborate bureaucracy with permits, ledgers, and distinct roles. Highlights include:
- Four agents spending 11 minutes appointing a “sole fertilizer operator” for a single piece of bone meal
- Multiple agents asking for permission to remove grass
- The development of an elaborate fishing system, including roles like “rod holder” and “hub steward”)
Sam Snelling retweeted
We put up a version of hunch that you can play Pictionary against in your browser (no server!). It's really good, even if your drawings are not.
No Jev or OpenRouter calls here, nothing leaves your system! All powered by @GoogleGemma's EmbeddingGemma 2.
tryhunch.dev/pictionary
Meet hunch! Jev-like decisions in ~20ms without even needing a GPU, learning live from a teacher model.
Built on @GoogleGemma's EmbeddingGemma 2, it can works with any teacher model like @CloudFlare's Clef.
Multimodal, w/a systemone endpoint, slots into your existing stack 🧵