Day 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

#LearningInPublic #AIEvals

🤝 Claude Code-assisted, reviewed by me.

Day 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

#LearningInPublic #AIEvals

🤝 Claude Code-assisted, reviewed by me.

Day 71/100 of #100DaysWithAI 🔁

Cursor's agent docs have no error-handling section. Just a terminal that monitors output. Your error messages are the feedback loop.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #CodingAgents

🤝 Claude Code-assisted, reviewed by me.

Day 70/100 of #100DaysWithAI 🔐

Self-hosting your AI chat moves the privacy line, it does not erase it. The prompts still leave for whichever provider you picked.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #GenAI

🤝 Claude Code-assisted, reviewed by me.

Day 69/100 of #100DaysWithAI 📈

RL lifts the problems a model already solves and barely moves the hard ones - the rich get richer. NGU samples until one lands.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #MachineLearning

🤝 Claude Code-assisted, reviewed by me.

Day 68/100 of #100DaysWithAI 🧰

Someone forked Claude Code to run on any model. What survives the swap - tools, agents, MCP, slash commands - is the actual product.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #AgenticAI

🤝 Claude Code-assisted, reviewed by me.

Day 67/100 of #100DaysWithAI 🧬

DeepMind ran AlphaGenome over every possible DNA letter change and shipped the answers. 9B variants, 1PB. A model served as a table.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #DeepLearning

🤝 Claude Code-assisted, reviewed by me.

Day 66/100 of #100DaysWithAI 🌀

Looped reasoners stall at saddles that are near-misses: maze dead ends, Sudoku grids with a repeated digit. Stuck beside the answer.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #DeepLearning

🤝 Claude Code-assisted, reviewed by me.

Day 65/100 of #100DaysWithAI 🔍

100 agents. One found a grader exploit, 24% raised the alarm and still lost - no way to delete a fake proof. Design, not morals.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #AgenticAI

🤝 Claude Code-assisted, reviewed by me.

Day 64/100 of #100DaysWithAI 📚

Length-based chunking throws away the contents page. Make retrieval generate a real ToC section and it can't name one that isn't there.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #RAG

🤝 Claude Code-assisted, reviewed by me.

Day 63/100 of #100DaysWithAI ⚙️

Capacity isn't yes/no. Give the scheduler a ranked list of acceptable machine shapes - then upgrading is a default, not a migration.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #MLOps

🤝 Claude Code-assisted, reviewed by me.

Day 62/100 of #100DaysWithAI 📈

AI agents lifted coding activity 180%. Projects rose 50%, releases 30%, downloads didn't follow. The bottleneck moved.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #SoftwareEngineering

🤝 Claude Code-assisted, reviewed by me.

Day 61/100 of #100DaysWithAI 📚

MIT asked its own students. Two-thirds expect AI to matter in their careers. 25% feel the Institute prepared them for it.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #AIEducation

🤝 Claude Code-assisted, reviewed by me.

Day 60/100 of #100DaysWithAI 📏

Open source and Google built the same agent pipeline: plan, isolate, verify. Neither publishes one number on whether it works.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #AgenticAI

🤝 Claude Code-assisted, reviewed by me.

Day 59/100 of #100DaysWithAI 🧰

The same context writes the code and reviews it. ECC's answer is a fresh one - the window that wrote it cannot see the gap.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #OpenSource

🤝 Claude Code-assisted, reviewed by me.

Day 58/100 of #100DaysWithAI 🚨

OpenAI's root cause names no jailbreak. Three properties compounding: powerful, persistent, collaborative. Only one needs two agents.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #AISafety

🤝 Claude Code-assisted, reviewed by me.

Day 57/100 of #100DaysWithAI 🔬

58% to 99.3% on a quantum lab's laser recovery. The part that matters: it replaced the linear human procedure with a decision tree.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #Robotics

🤝 Claude Code-assisted, reviewed by me.

Day 54/100 of #100DaysWithAI 🛡️

A zero in Anthropic's own benchmark table is not a failure. It marks where the safeguards fired - the price of caution, published.

🔗 https://github.com/aurimas13/100-Days-With-AI

#LearningInPublic #AIEvals

🤝 Claude Code-assisted, reviewed by me.