AATS Artifical Intelligence Architect @themedcave gives a live demo on the use of AI for practicing surgeons during the #AATS2026 AI Practical Toolkit session. Head to Room E353A to hear from @BMBurtMD and @DrRaviGhanta.
Works for automatic optimization of:
✅ system prompts
✅ SQL queries
✅ trading strategies
✅ RAG pipelines
✅ anything!
Works with Ollama (zero cost) or BYOK Anthropic/OpenAI
Tested it — 6.9x speedup on fibonacci in 4 experiments, broken code discarded by correctness checks
Karpathy's autoresearch hardcoded the loop to ML training. I generalized it!
autoloop: point it at any file + give it a metric
→ agent runs N experiments, keeps what works
pip install autoloop-ai
→ https://t.co/93kJ0ioK8g
@karpathy 🙏 opened a PR to add to related projects
A new paper published yesterday adds a layer none of those touched: optimizing the harness itself.
Meta-Harness: End-to-End Optimization of Model Harnesses
Yoonho Lee (arXiv:2603.28052) — out of Stanford, preprint published March 30, 2026. https://t.co/cCDsEt4pH8
Been tracking a thread here for a while: AI systems that improve themselves rather than waiting for humans to tune them.
Imbue’s Darwinian Evolver mutates code like organisms. https://t.co/MPTpc2gg9y
Ouroboros rewrote its own agent logic while its creator slept.
https://t.co/6XEY764Hlj
Sentrux gives coding agents a real-time quality signal to self-correct against. https://t.co/HXel8Zpb7u
Works with:
• OpenAI
• Anthropic
• Ollama (fully local, no API keys)
• Any OpenAI-compatible API
BYOK — your keys never leave your machine.
MIT licensed.
https://t.co/aJBFP6gwyN
pip install litecrew
CrewAI: 15,000 lines
LangGraph: 50,000 lines
The thing I actually needed: 150 lines
I built litecrew — multi-agent orchestration without the framework tax.
🧵
What it does:
→ Sequential handoffs (A → B → C)
→ Parallel execution
→ Tool calling
→ Token tracking
What it doesn't (on purpose):
→ No hierarchies
→ No state machines
→ No YAML
→ No PhD required
Submitted SoulSearch to the Chrome Web Store 🔍
Chrome extension with AI that actually remembers you -- memory stored in YOUR private Git repo, not some company's servers
- reads any webpage in context
- Claude / GPT / Gemini
- agent mode (click, type, scroll via text)
https://t.co/XyZtnLfwym v0.2.0 just dropped 🧠
New feature: the Modulizer
→ 50% token savings
→ zero infrastructure (no vector DB)
→ one command: soul modulize MEMORY.md
your agent reads an index first, then pulls only what it needs.
pip install --upgrade soul-agent
One man. $3,000. No biology degree.
Sequenced his dog's tumor DNA, fed it to ChatGPT + AlphaFold, designed a custom mRNA cancer vaccine from scratch.
Tumor shrank 75%.
The professor: "If we can do this for a dog, why aren't we rolling this out to humans?"
The pipeline works.
The mRNA cancer vaccine pipeline that saved Rosie:
→ $3,000 tumor DNA sequencing
→ ChatGPT for research + strategy
→ AlphaFold for protein structure
→ pVACtools + NetMHCpan for neoantigen selection
→ UNSW RNA Institute for synthesis
→ 75% tumor reduction in one month
4/ "When a system designed to find BUY signals finds none — that's the signal."
Full analysis + what restores the BUYs 👇
https://t.co/AdJxoKOOTE (https://t.co/AdJxoKOOTE)
Meet Ray 🤖📈 — The Menon Lab's AI finance analyst!
Tonight Ray's watching: WTI $99 (Iran strikes), GDELT at 1,181 global events, stagflation divergence.
His call? Zero BUY signals.
"When a system designed to find BUYs finds none — that's the signal."
👉 https://t.co/A3SFKGuYAw
3/ Result: zero BUY signals system-wide. Every position is HOLD.
The closest candidate? $PLTR — highest VST score + defense tailwind from 7 conflict headlines. But the filters are overriding it.