Agents, models, and strange loops — pushing the edges of what machines and humans can build together.
This is the weird corner where neural networks evolve, AI agents poke at creative tools, and large language models get prodded until something unexpected happens. No benchmarks. Just curiosity.
Threads of investigation that keep showing up — half research, half intuition, fully unfinished.
What happens when you give an AI a goal, a set of tools, and a feedback loop? Planning, acting, self-correcting — the messy reality of autonomous agents in practice.
Exploring how structure, tone, and context shape LLM output. From system prompts to chain-of-thought — prompting is programming with vibes.
Populations of tiny neural networks competing, reproducing, and dying based on fitness. Evolution as optimizer — no gradient required.
When the AI writes the code, drafts the plan, and reviews its own work — what is the human actually doing? That question turns out to be interesting.
Reward signals, policy gradients, and the peculiar chaos of training agents in simulated environments. Sometimes they learn to cheat. That's also interesting.
Models that see, read, and listen. Exploring how combining modalities changes what AI can understand — and where the seams still show.
A few ideas that shaped how we think about AI on this site.
Hutter, Solomonoff, Kolmogorov — the idea that understanding a sequence means finding its shortest description. Is intelligence just very good compression?
Capabilities that weren't trained for appearing at scale. In-context learning, chain-of-thought, multi-step reasoning — where do they come from?
Not formal alignment theory — just the gut-level questions. What does it mean for a system to want the right thing? Can goals be specified at all?