A $40 billion payments company gave away its internal AI agent for free. Most AI agents lock you into one model or vendor. Switching providers usually means rebuilding your whole setup. Jack Dorsey's $40B company open sources an AI agent called goose. It takes a different
@alphasignalai
-
Claude Fable 5’s strongest result: rejecting the wrong metric
By
–
Claude Fable 5’s strongest result was not writing more code. It was rejecting the wrong metric. We tested it on 3 ML tasks:
> Perfect churn validation was leakage
> Drift was real, but not the root cause
> Churn AUC was the wrong target for retention offers The hard one: Fable -
Rebuilding Claude Code: ClawCodex in Python (open source)
By
–
Someone just rebuilt Claude Code from scratch in pure Python. The tool is called ClawCodex and it's 100% open source. That's 180,000 lines of pure Python. No TypeScript runtime required. The original only talks to one model family. This rebuild routes through six providers.
-
Train LLM from scratch GitHub repo and AI newsletter summary
By
–
Repo: https://
github.com/FareedKhan-dev
/train-llm-from-scratch
… Check out https://
alphasignal.ai/newsletter to get a daily summary of the latest
breakthrough news, models, papers and repos. Read by 300,000+ devs. -
Open-source repo shows how to train ChatGPT-like LLM
By
–
Someone open-sourced every secret step that turns a model into ChatGPT. Most LLM tutorials stop at the architecture diagram. This open-source repo goes from raw data to a working model. You can train your own LLM on a single GPU. It builds a transformer from scratch in
-
Unified platform for training and testing world models
By
–
You can now train and test world models in one place. Researchers kept rebuilding the same machinery for world models. These systems let agents imagine outcomes before acting. Every project shipped its own training stack and data loader. Nothing was comparable.
-
Loops move judgment earlier; gates hold most value
By
–
this is the cleanest framing: loops move judgment earlier. the work shifts from “write the next prompt” to “define memory, tools, gates, stop rules, and review”, the gate is where most of the value lives
-
Quality and budget gates fix two failure modes; brakes matter
By
–
Exactly, quality gates and budget gates solve two different failure modes which are bad work getting through, and okay-ish work running forever most agent loop demos show the motion, the useful part is the brakes
-
Small models locked out of rare tasks; large models prioritize common tasks first
By
–
Small models don't just learn slower.
— AlphaSignal (@AlphaSignalAI) 9 juin 2026
They're structurally locked out of rare tasks forever.
No matter how long you train them, frequent task signals constantly overwrite rare ones.
Large models solve this by fully satisfying common tasks first, then dedicating leftover… https://t.co/9U0z5AIPgY pic.twitter.com/yhu439mdyoSmall models don't just learn slower. They're structurally locked out of rare tasks forever. No matter how long you train them, frequent task signals constantly overwrite rare ones. Large models solve this by fully satisfying common tasks first, then dedicating leftover
-
Agent loops need task repetition, code execution, and rejection gates
By
–
Agent loops are not a default, even if it's viral.. They only work when the task repeats, the agent can run the code, and something inside the loop can reject bad output. Miss that gate and you are just automating token burn.