Steve Cosman

New project: getkiln.ai - the easiest way to build AI products and models. 3k Github stars and ~10k users - pre-v1! github.com/Kiln-AI/Kiln Powerful AI dev tools, intuitive UX anyone can use. It includes fine-tuning, evals, and synthetic data gen. All free. 1/5

@bentlegen DDC!

Amazing there’s a “TUI” beer here (Nobody translate this and ruin it for me, please)

My writing now: - hand write amazing technical docs - Claude: proofread for grammar - Claude: make this a Linkedin post - Claude: make this an email for our audience - Claude: make this a blog post using guide X My ideas, with adaptive style.

@tarunsachdeva It's just finally cool to be fast local and embedded.

Noticing a resurgence in SQLite these days, it’s interesting to see this ecosystem develop. Was actually looking at the Ecosystem page to SQLite on the open source explorer and ofc Peter is the top trending project lol

The new ChatGPT engagement baiting is brutal. I don't need every answer to end with "If you're curious I could also tell you about...". Claude already replaced GPT for all coding, but I think it just won me for consumer too.

Coding isn't holding you back. Product work is. Introducing Modem, your dev team's auto-triage PM. It listens to user signals 24/7, finds problems before you do & does the work you're slow at – filing tickets, pinging teammates, closing the loop. Sign up today at @modemdev

New community blog post out! Our friends at @getkiln have been cooking up a guide on how to build better RAG systems. Learn how Kiln × @LanceDB help you build and improve your RAG configs in hours, not weeks. 🧵👇 1/5 lancedb.com/blog/rag-isnt-…

@ggerganov @erusev 🧑‍🍳😙🤌

LlamaBarn v0.10.0 (beta) is out - feedback appreciated

One of the best movies ever made about reinforcement learning

Okay, I really like using em dashes — y'all are ruining perfectly good punctuation.

@petergyang GPT 4.5 came out 6 months ago and was awful.

Last week: It’s so over for Anthropic This week: Sonnet 4.5 is INSANE Next few weeks: > It’s so over for Google > Gemini 3.0 is WILD > It’s so over for OpenAI > GPT 4.5 is the BEST Yes I can predict the future 😆

@ansonlin @andrewdsouza @goswish 🏃🏃🏃

Today in models lying about who they are:

@bentlegen CodeRabbit is a good second check. Really solid at catching issues.

when I'm making plan files I usually have Claude go over them 5-6 times in different ways before a line of code is written

you'd be surprised how useful it is to ask the LLM, "hey review the plan you just wrote, did you miss anything?" 10 times out of 10 it will discover it forgot something

@cerebras Cerebras Code has so much potential, but 2500tok/s means nothing if I hit my RPM limit in 9 seconds. Avg speed of agents is much worse. Kill per minute limits, daily caps fine.

Kiln v0.19 just came out with a new model library, GPT-5, GPT OSS, GLM, Qwen 2507, Cerebras support, SiliconFlow support, and much more! Download it now: github.com/Kiln-AI/Kiln/r…

@GraphCrimes imgur.com/yCI5W6U

I wrote 2000 LLM test cases so you don't have to This is a quick story of how a focus on usability turned into 2000 LLM tests cases (well 2631 to be exact), and why the results might be helpful to you. getkiln.ai/blog/i_wrote_2…

@charlieholtz I asked for this 4 days ago in HN 😆

We made a Mac app that lets you run a bunch of Claude Codes in parallel. Introducing Conductor!

We just shipped the biggest Kiln release ever! In includes evals v2, synthetic data v2, and a a whole new concept: Kiln Issues! Read more: getkiln.ai/blog/release_0…

Imagine being at a big company this year...