Thomas Schranz 🍄
@__tosh
Followers
9K
Following
37K
Media
3K
Statuses
27K
🚜 building - 🌌 Atlas, tools to help you get most out of llms - ♟️ Chess Cats - 🧇 Waffle, messaging w/ friends - 🍓 Jam, open source Clubhouse (@jam_systems)
San Francisco, London, Vienna
Joined October 2008
early stage product: 🪵 collect tinder, start fire before freeze. later stage product: 🔥 keep fire alive, grow fire, defend fire. very different. h/t @christineluc.
8
4
71
TIL: tailwind puts all color definitions into output.css whether they are used or not. not trivial to remove them in tailwind v4.
1
0
0
still wonder if agents like claude code or amp are useful?. git clone anything you find interesting. ask questions about the codebase. see how they translate your questions into ripgrep + give you answers from the code they read and their vast knowledge. investigate.
1
0
0
👀🔥 new Qwen3.
Bye Qwen3-235B-A22B, hello Qwen3-235B-A22B-2507!. After talking with the community and thinking it through, we decided to stop using hybrid thinking mode. Instead, we’ll train Instruct and Thinking models separately so we can get the best quality possible. Today, we’re releasing
0
0
2
everyone wondering about when AGI. me wondering when reply spam is finally becoming interesting to read
0
0
1
RT @peterbuch: Just launched LLM SEO EEAT on Product Hunt and currently #1. Check how trustworthy Google thinks your content is - get an in….
producthunt.com
Get instant E-E-A-T scores and AI-powered recommendations to improve your content's trustworthiness and search rankings.
0
1
0
I don't think the endgame of agentic coding UI will be jules, codex, or claude code in github PRs. Slack and discord won over bulletin boards for collaboration. Chat is faster, more direct. As long as we iterate we want to be as close to what's happening as possible.
0
0
2
LLM SEO EEAT is currently #1 🏆 on @ProductHunt . a simple tool that lets you check how your content (e.g. landing page, blog post, product listing, …) scores according to Google’s content guidelines. (link in next tweet, yeah I know)
3
0
4
RT @stopachka: Is there a "System Design" podcast, where the host talks to the engineers who built the systems we know and love, and go dee….
0
3
0
firebase firestore is very slow atm. anyone else seeing this?.
2
0
6
TIL: Anthropic models are on Google and Amazon but not (yet?) on Azure.
1
0
0
Claude Sonnet 4 inference in tokens / s. Google Vertex: 70.Anthropic: 60.Amazon: 50. note: this is not about inference speed but also about demand & varies over time. just a current snapshot via openrouter
0
0
1
so yes, in that sense _you_ having memory of previous sessions is definitely helpful. the idea of memory is very useful and spot on. just how it is implemented currently (and how context windows work) doesn't work well.
0
0
0
there might be a way to pull memory off but the way it is done right now doesn't work well. it just makes the new iteration less effective. it's a _feature_ that a new agent session is a new session. (while you know that it is a new session & know more from previous sessions).
1
0
2
memory is an anti-feature. it's like quicksand. quick improvements come from the loop
1
0
1