Colin HermesClaw retweetledi
Looking for a space in SF to work until late?
Join us at @club_homebrew and get access 24/7.
Come check it out: luma.com/pdn11qio
English
Colin HermesClaw
6.4K posts
Colin HermesClaw
@alwaysbecolin
Founder of https://t.co/xn1sScw1Fh and https://t.co/xMeLotDAn0 and https://t.co/uhCWrE7Wvf Dev Advocate for @NebiusAI, Previously @Metamask @Akashnet @Cisco @Accenture
San Francisco, CA Katılım Ekim 2021
921 Takip Edilen42.1K Takipçiler
Colin HermesClaw retweetledi
Love how with proper harness I can just get either Claude or Codex to talk to me in diagrams
English
Colin HermesClaw retweetledi
Colin HermesClaw retweetledi
I’m open-sourcing 75 code examples of how labs are using agents for training data.
Labeling, DPO juries, rejection sampling, step-level rewards, instruction generation, dataset curation, benchmark decontamination.
Learn these and go get that bag 💰
git.new/data-labeling
Ashpreet Bedi@ashpreetbedi
Data labeling is so hot right now. If you understand the code below you can get a $1m/yr job at a frontier lab. A jury of 5 models that turns raw response pairs into trainer-ready DPO data. Typed verdicts, position-debiased, self-preference recusal. git.new/dpo-jury
English
Colin HermesClaw retweetledi
Colin HermesClaw retweetledi
SOMEONE FROM THE ANTHROPIC TEAM LEAKED THEIR OBSIDIAN SETUP. 8 MILLION PEOPLE SAW HOW HE ACTUALLY USES CLAUDE
the funniest part? all of this information was sitting in claude's documentation from day one. nobody read it
one guy did, packed it into a 9-step guide and posted it. and it broke the internet. 4,100 likes, 800 retweets, then china picked it up and 8 million views
want to know what's in it?
one file. called CLAUDE.md. it holds everything about you: how you think, what you're working on, where you get stuck, even how you want the ai to talk to you. claude reads it first every single session
one file changed everything. because now ai doesn't open with "how can i help?" it already knows. it remembers your projects, sees your goals, catches moments where you're contradicting yourself
people spent years searching for the perfect prompt. the right temperature. the magic formula. and the answer turned out to be not how you ask ai. but what ai knows about you before you even open your mouth
then the guy went deeper. taught claude to work on a schedule. every morning at 7am the ai walks through all notes on its own, finds new stuff, links it, cleans what's stale. no command. no reminder
and all of this runs on obsidian. free app. text files on your drive. no cloud, no lock-in. switch models tomorrow and the folder keeps working
the most liked comment under the original post: "this is the difference between using ai and building a system. most people won't realize it until they waste hundreds of hours repeating themselves"
hundreds of hours. you've already spent some of them
full guide in the video. i break down finds like this every day - follow so you don't miss the next one
Yarchi@undefinedKi
English
Colin HermesClaw retweetledi
Anthropic researcher just dropped a paper most AI engineers missed.
Xiaotian Luo published it 2 days ago.
It unpacks the entire self-evolving harness playbook: how the code around the agent rewrites itself.
You stop tuning the model. You tune the code around it.
→ arxiv.org/abs/2607.13683
Bookmark it. Then read this setup ↓
Archive@ArchiveExplorer
English
Colin HermesClaw retweetledi
Fuc*k your paid courses, @Kimi_Moonshot AI founder (Yang Zhilin) explained how they're running 300+ specialized sub-agents in parallel and it changes everything.
The founder of a $20B + valuation Chinese AI lab give you a 40-minute masterclass on AGENT SWARMS.
This is the clearest breakdown I've seen on building large-scale AI systems that actually work in the real world.
- Dynamic orchestration, PARL rewards, real scaling beyond context windows.
- Trained with smart RL rewards (instantiation, finish & outcome)
- 100-300+ specialized spawns agents , sub-agents on the fly working in parallel.
- Thousands of tool calls.
- 4.5x faster execution.
This isn't theory it's how they're scaling Kimi to handle real complexity.
Better than most local paid courses.
Swap your Netflix tonight watch this instead.
Md Ismail Šojal 🕷️@0x0SojalSec
- be Yang Zhilin the founder of Moonshot AI (Kimi) - China Tsinghua to U.S CMU PhD (graduated 2019) - Worked at Google and Meta. - Papers with 2 Ai-god-fathers LeCun & Bengio - Co-authored papers on computer reasoning and pattern recognition with Turing Award winners like Yoshua Bengio and Yann LeCun. - Led the highly-cited XLNet paper over 10k citations, which advanced large language model training. - Apple wanted him, Stanford/MIT were options. - Instead, left the US, went back to China in 2019 and founded Moonshot AI (Kimi) - he’s helping build China’s next AI powerhouse. Talent like this is rewriting the AI map.
English
Colin HermesClaw retweetledi
how I run Fable with subagents
fable is my main agent now, it does the planning and the judgment calls, then hands off execution to subagents
if you leave a subagent alone, it inherits fable's reasoning level, so at spawn time fable sets the model explicitly instead, based on what the task is
fable's own effort isn't fixed either, it runs low, medium or high, mostly low or medium since the jump in token cost between levels is steep
how fable routes the model:
> fable reaches for high effort when a task needs multi-step judgment, low and medium handle the rest
> routine execution, drafts, formatting, straightforward edits go to sonnet
> harder execution, debugging, anything that needs deeper reasoning goes to opus
> design work goes to GPT 5.5, coding work goes to GPT 5.6
the same setup scales up to a full campaign:
> fable plans the launch week, sets the angles and the cadence
> sonnet researches competitor launches and drafts the week's posts from the brief
> opus rewrites the launch-day post since that one sets the whole week's angle
> the week's schedule goes out through sonnet, straight to typefully
on code, fable scopes the task and hands it to GPT 5.6, sonnet and opus stay on the writing and research side
Shann³@shannholmberg
how to run Fable and GPT 5.6 without hitting your limits these models are powerful, and they burn tokens fast. so how you run them decides how much a day of content, research, or outreach gets done 6 things that get you more usage: > trim your CLAUDE .md and AGENTS .md to the essentials, every prompt reads them plus every skill and tool you've got enabled, so turn off what you're not using > drop the reasoning level when you don't need the top one, default to medium or high and save max effort for the problems that need it > give the model clear stop points, it runs long, so have it finish the plan and check in before executing > keep subagents on a lower reasoning level, they inherit the parent, so a swarm at max effort drains a window in one message > run a cheaper model as the orchestrator and call the expensive one only for the hard reasoning > look at what one message costs in usage, review it and see how you can use above to improve it none of these are big changes on their own, but together they decide how long you can run the best models before you hit your limit
English
Colin HermesClaw retweetledi
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.
This is a 17-place jump from Kimi-k2.6 (#18 -> #1).
In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.
The full model weights will be released by July 27.
Congrats to the @Kimi_Moonshot team on this major milestone!
Kimi.ai@Kimi_Moonshot
Meet Kimi K3
English
Colin HermesClaw retweetledi
A former COO of Citadel Securities is telling you where the money is in AI
For free
Daniel Kornum@dkfromdk
English
Colin HermesClaw retweetledi
ANTHROPIC JUST CHANGED THE ENTIRE AI-SLOP FRONTEND.
It’s called Frontend Design, the official Anthropic skill for Claude Code.
And it already has 161,000 stars on GitHub.
The problem it solves is simple:
Every time you asked AI to build a website, you got the same thing.
Inter as the font.
Purple gradient.
Generic startup layout.
Eternal AI slop.
Frontend Design kills that.
Forces Claude to pick an aesthetic direction before it touches a single line of code.
And the result completely changes.
You can ask for:
→ brutalist
→ editorial
→ retro-futurist
→ luxury
→ maximalist
And it generates HTML, CSS, JS, React or Vue, production-ready.
How to install:
1. github.com/anthropics/ski…
2. describe the website or component you want.
3. pick a style.
4. Claude builds typography, color, motion and composition consistent.
One install. Available across every session.
Works in Claude Code, Cursor, Codex, Gemini CLI, OpenCode and 15+ harnesses.
100% open source. Maintained by Anthropic.
277,000+ installs in a few months speak louder than any review.
Repo 👇
AdiiX@adiix_official
ANTHROPIC JUST LOST 20,000 CLAUDE CODE USERS TO ONE GITHUB REPO. 20,000 devs. 26k stars. 4,000 forks. 10 free models. COMPLETELY FREE. Bookmark this before Anthropic notices. Your Claude Code bill will start working differently. It reroutes Claude Code to DeepSeek, Kimi, and 8 more same CLI, different model behind it. 5 min setup. $200/mo saved. 0 downgrade on small tasks. swap the endpoint in your config today. Link below. Claude Code → Endpoint swap → Free models → No bill → Money
English
Colin HermesClaw retweetledi
I was wrong. I said we were a year away from Fable 5 on our desk.
That day is today
An open model BETTER than Fable 5 in some benchmarks just dropped
Better than ChatGPT 5.6 on FrontierSWE. Better than Fable 5 on Automation Bench
This fundamentally changes the AI race forever
If people can start running Fable 5 on their desk, unlimited and for free, they are not going to pay thousands for subscriptions
Now let's be clear: will the hardware you need to run Kimi K3 be obtainable for most? No it won't
It will require multiple very expensive Nvidia chips or a bunch of Mac Studio 512gbs
But look at the slope, not the Y intercept. Over the last year local AI has become SIGNIFICANTLY more efficient and required way less compute
The smartest brains in the world are all attacking the compute problem as hard as they can
This is another step in that direction. And within a year or two, you'll be able to run this on a Mac Mini
Local AI has arrived, and it's not going anywhere
Kimi.ai@Kimi_Moonshot
Introducing Kimi K3: Open Frontier Intelligence 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost 🔹 Built for long-horizon agentic coding and self-evolving workflows Kimi K3 is now live on on Kimi.com, Kimi Work, Kimi Code, and the Kimi API. Open Weights by July 27, 2026. 🔗 API: platform.kimi.ai 🔗 Tech blog: kimi.com/blog/kimi-k3
English
Colin HermesClaw retweetledi
Biggest thing we've learned this year running coding agents in the cloud: output quality depends less on the model than on whether the agent has a real dev environment, the same one a developer would have locally.
Locally you never think about this. The agent just inherits your laptop's setup. In the cloud you rebuild everything from scratch, and you can't tell when you've gotten it slightly wrong.
Because nothing crashes. You just get output that's a bit worse than it should be. Most people don't notice, and the ones who do blame the model.
Nine times out of ten the agent simply couldn't run or verify its own work.
A year ago this didn't matter, models couldn't use an environment anyway. Now they can, and the environment is usually the bottleneck.
English
Colin HermesClaw retweetledi
Colin HermesClaw retweetledi
Your most expensive bug is the zombie microVM that keeps charging you because your .kill() line never ran.
Sandbox teardown should be bound to scope, not to a cleanup call you hope executes.
If your SDK supports context manager, defer, or await using, use them. If it doesn't, that's a gap worth paying attention to.
Leaking a sandbox should be structurally impossible.
English
Colin HermesClaw retweetledi
Colin HermesClaw retweetledi
Anthropic just dropped a 100% free course on Loop Engineering with Fable 5.
This is the clearest breakdown of Claude Code and agentic loops you'll find anywhere.
People are paying for tutorials that teach less than this one hour does.
Watch it today, then read the step by step guide on building loops below.
Anatoli Kopadze@AnatoliKopadze
English
Colin HermesClaw retweetledi
Join us tomorrow for the premiere of the Full Stack by CodeTV at noon PT (3:00 PM ET)! 🚀
👀 Watch as @arcdigital & I tackle the challenge of building a product in 4 hours, under the watchful eyes of seven cameras and the renowned @kelseyhightower Find the link below and we'll see you there!
👉 bit.ly/TFS-S1E1
@jlengstorf @codetv_dev
English
Colin HermesClaw retweetledi
Anthropic engineer:
"You're not supposed to prompt Claude. You're supposed to build a system that prompts itself."
In 45 minutes, she breaks down how Anthropic builds agents that remember, learn from their mistakes, and get smarter with every run.
Worth more than any paid course you'll find on building agents.
Watch this and bookmark
Swati Gupta@hrswatigupta
English