["Hybrid Reasoning Agents: LLM agents that think in latent space, act through text", "If you don't run your own LLM proxy, how do you know whats going on?", "I built a programming language that LLMs can't use", "Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling", "How Far Will AI Assistants Go for a Refund They Can't Get?", "You could have invented transformers from memory", "How employment is changing at Anthropic", "Hyper Optimizing Kimi K3 in vLLM", "MegaTrain: Full Precision Training of 100B+ Parameter LLMs on a Single GPU", "I've been waiting over a month for Anthropic to respond to my billing issue"] ["Hybrid Reasoning Agents: LLM agents that think in latent space, act through text", "If you don't run your own LLM proxy, how do you know whats going on?", "I built a programming language that LLMs can't use", "Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling", "How Far Will AI Assistants Go for a Refund They Can't Get?", "You could have invented transformers from memory", "How employment is changing at Anthropic", "Hyper Optimizing Kimi K3 in vLLM", "MegaTrain: Full Precision Training of 100B+ Parameter LLMs on a Single GPU", "I've been waiting over a month for Anthropic to respond to my billing issue"]

Deep Reviews

Deep Reviews

Every AI Agent Leaderboard Is a Lie. Berkeley Has the Receipts.

UC Berkeley scored near-perfect on eight of the most-cited AI agent benchmarks without solving a single task. SWE-bench, WebArena, OSWorld, GAIA, Terminal-Bench, FieldWorkArena, CAR-bench, SWE-bench Pro — all gameable. Here's how and what to do.

Apr 13, 2026 7 min read