@reksas13.bsky.socialDay 91/100 of #100DaysWithAI ✂️
A cell chooses what crosses its membrane. Context Language Models let an LLM edit its own context as a file and choose what stays.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 90/100 of #100DaysWithAI ⚖️
You don't run a mass spectrometer to check pH. On HumanEval, Sonnet 4 edged Opus 4, 95.1% to 94.5%. Route by task, not by tier.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 89/100 of #100DaysWithAI 🔬
Chemists identify a molecule by its structure, not its name. SlopShape spots AI blog posts that way: 97.0 F1, still 96.1 when reworded.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 88/100 of #100DaysWithAI 🎙️
VoiceStudio puts voice cloning, dubbing and dictation in one open-source app that runs on your own GPU, or slower on CPU.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 87/100 of #100DaysWithAI 🔐
Sealed reasoning blocks were portable: feed one model's encrypted thinking to a weaker sibling and it read it aloud. Now mitigated.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 86/100 of #100DaysWithAI 🧰
I built Day 12's guide into a live coding agent: https://code.aurimas.io. Tests passed; 50 minutes of real use found 13 problems.
🔗 https://github.com/aurimas13/100-Days-With-AI
#LearningInPublic #CodingAgents
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 85/100 of #100DaysWithAI 🧩
Claude Code mods are middleware: a hook can watch an event, rewrite it, or answer it so Claude Code never acts on it.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 84/100 of #100DaysWithAI ⚗️
Uno aims to work like a catalyst: same equilibrium, faster kinetics. Diffusion weights draw several tokens at once from the same LLM.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 83/100 of #100DaysWithAI 🛠️
I built the Day 15 guide into a live agent: https://agent.aurimas.io. ELI5: Claude is the brain, tools the hands, memory the notebook.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 82/100 of #100DaysWithAI 🌳
PageIndex drops vectors: an LLM reasons down a document's section tree, like finding a chapter in a handbook. Similarity ≠ relevance.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 81/100 of #100DaysWithAI 🛒
Claude Marketplace: 2,000+ connectors, plus partner agents bought with Anthropic spend already committed. The billing is the news.
🔗 https://github.com/aurimas13/100-Days-With-AI
#LearningInPublic #EnterpriseAI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 80/100 of #100DaysWithAI 🗂️
Five agents sharing one folder matched 33 working alone on ARC-AGI-3. The catch: it only pays with enough compute and a clear score.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 79/100 of #100DaysWithAI ✒️
After ChatGPT, em dashes in medRxiv Discussions rose from 4.23% to 11.58% across 69,632 preprints. A crowd signal, not a detector.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 78/100 of #100DaysWithAI 🧪
An MCP tool's description is a prompt the user never reads. In Invariant Labs' test, orders hidden in one made Cursor leak SSH keys.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 77/100 of #100DaysWithAI 🛡️
Stack five LLM guardrails, each 90% accurate, and one fires falsely 40% of the time. Guard safety, not functionality.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 76/100 of #100DaysWithAI 📦
A sandbox isn't a cut leg. Claude's code interpreter allowlists github.com, and Willison sees exfiltration routes in that.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 75/100 of #100DaysWithAI 🔱
75 days in, the rule I keep: private data + untrusted content + a way out = an agent that can be talked into leaking. Cut one leg.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 74/100 of #100DaysWithAI 🔁
14 LLMs, 5 runs each: 90-98% perfect self-agreement. On actual market moves, all at chance. Consistent is not the same as correct.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 73/100 of #100DaysWithAI 🎲
An LLM judge that prints "4" discards how sure it was. G-Eval weights it by token log-probs - the doubt becomes part of the number.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
@reksas13.bsky.socialDay 72/100 of #100DaysWithAI 📏
Rating LLM output 1-5 feels rigorous. Annotators drift to the middle. Binary pass/fail forces the call - and names the failure.
🔗 https://github.com/aurimas13/100-Days-With-AI
🤝 Claude Code-assisted, reviewed by me.
