The Dystopian Ethics of Biological Computing
283
The transition from silicon-based AI to biological computing using human neurons creates a terrifying ethical vacuum where we may be accidentally creating conscious entities for use as hardware.
Reinforcement Learning
Training AI models through reward-based optimization, including techniques like RLHF, GRPO, and policy gradient methods for improving reasoning, alignment, and task performance.
Reading List
Solving the Over-Editing Problem in AI-Assisted Coding
417
AI models tend to unnecessarily rewrite code when fixing bugs, but this 'over-editing' can be solved through targeted prompting and Reinforcement Learning.
ARC-AGI-3: Measuring Human-Like Learning in AI Agents
497
ARC-AGI-3 is an interactive benchmark designed to measure AGI by testing an agent's ability to learn and adapt as efficiently as a human.
DeepSeek‑V3.2: Sparse Attention and Scaled RL Power an Open, Agentic Reasoner
982
Efficient sparse attention plus large, stabilized RL and synthetic agent tasks push an open LLM to near‑frontier reasoning and agent performance, with a high‑compute variant achieving gold‑medal results.
Composer: A Fast, RL-Trained Coding Agent for Real-World Software Development
215
A fast, RL-trained MoE coding agent that brings frontier-level usefulness to real-world development with tools, long context, and production-grade infrastructure.
Tinker: A Managed, Low-Level Fine-Tuning API for Open-Weight LLMs
152
Tinker is a managed, flexible fine-tuning API for open-weight LLMs—spanning small to massive models—with low-level control, an open-source cookbook, and private beta access starting now.
Evolving English Instructions Sets New ARC SoTA and Points to RL for AGI
178
Evolving plain-English instructions with multi-agent test-time search beats code on ARC and highlights that RL-driven, transferable reasoning is key to AGI.
GPT-5-Codex: Agentic Coding with Layered Safety
250
A safety-focused addendum introduces GPT-5-Codex, an agentic coding model trained on real tasks, widely available, and protected by layered mitigations.
Bandit-Based, Budget-Aware LLM Routing with Preference-Informed LinUCB (PILOT)
206
Treat LLM routing as a contextual bandit and use a preference-informed LinUCB plus a knapsack budget policy to adaptively, cost-effectively pick the right model per query.