Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A debate between the author and Zack Korman on whether Anthropic CEO Dario Amodei is good or bad for the world highlighted deep disagreements over AI risk and governance. The author argues Korman underestimates AI capabi…
A developer trained a large language model from scratch using plain PyTorch, implementing the full post-training pipeline including SFT, reward modeling, DPO, PPO, and GRPO on public datasets, all runnable on a single GP…
A developer building with Claude Code discovered that AI task execution can be significantly sped up by splitting independent modules and running multiple agents in parallel. The approach requires a loosely coupled, high…
A developer wired an LLM as a first-class Yjs peer using y-prosemirror and the standard awareness protocol, demonstrating that AI agents break three silent assumptions in collaboration stacks: throughput, undo ownership,…
A developer building AgentLeash argues that as AI coding agents like Claude Code, Cursor, and Codex become more autonomous, the bottleneck shifts from code generation to supervision and control. They propose a local-firs…
Researchers released LOCUS, the largest open database of U.S. local laws, containing codes from 9,239 cities and counties. The corpus aims to fill a gap in machine-readable legal text for AI research, covering zoning, ho…
An engineer at Anthropic shares a setup guide for Claude Cowork, emphasizing that the tool works best with a delegation mindset rather than chat-style prompting. Key recommendations include creating dedicated project fol…
A developer built ccglass, an open-source local proxy and dashboard that lets engineers inspect the full request flow of AI coding agents, including system prompts, tool call schemas, tool results, per-request token usag…
A developer built a benchmark to test whether frontier language models can be trusted to interpret clinical genetic variants. Testing Claude Opus 4.8 against expert consensus from ClinVar, the model scored 60% accuracy, …
A developer built perso, a WebAssembly policy engine that enforces access rules for MCP (Model Context Protocol) tool calls. The engine compiles JSON-based policies into a portable .wasm binary, enabling fast O(1) evalua…
An analysis of frontier AI model release cadence shows Anthropic and OpenAI accelerating their release rates since 2023, while Google, Meta, and DeepSeek have not. The pattern is consistent with a form of offline recursi…
A Washington D.C. subscriber filed a class action lawsuit against Anthropic on June 14, alleging that its Claude Max 20x plan delivers only six to eight times Pro usage instead of the advertised twenty times. The complai…
A developer warns that connecting an MCP server to an AI coding agent introduces two distinct security threats: the agent executing destructive commands and the agent being manipulated by untrusted data returned from the…
A PingCAP VP stated that while coding is being solved by LLMs, shipping complex enterprise software remains difficult due to permissions, memory, collaboration, and security. The company has shifted to an agent-led model…
A new open-source tool called cc-fleet enables Claude Code to use third-party large language models as workers, allowing users to run models from providers like DeepSeek, GLM, Kimi, and Qwen within Claude Code's multi-ag…
Zalando researchers published a framework using multimodal LLMs to automate product retrieval evaluation, achieving human-level accuracy at up to 1,000 times lower cost and reducing evaluation time from weeks to 20 minut…
A developer has built 31 MCP servers providing AI agents with structured access to East African institutional data, including M-PESA, county government, drought classifications, land titles, and health records. The open-…
A developer improved a multilingual fact-checker's F1 score from 0.655 to 0.813 by fixing a fundamental input error: the model was trained on claims alone instead of claim-evidence pairs. The XLM-RoBERTa model fine-tuned…
Z.ai's GLM-5.2 model was claimed to rank No. 3 on Agent Arena and beat Google's Gemini 3.5 Flash, but the claim lacks context as Agent Arena is a live, multi-signal leaderboard where rankings vary by signal and timeframe…
A developer created AINAScan, a deterministic AST scanner that detects 15 structural bug patterns commonly generated by AI coding assistants like GPT-4, Claude, Gemini, and Copilot. The scanner catches issues such as fak…