The Archive

DeepSeek V4 Flash and standard V4 weights released on Hugging Face, expanding availability for local deployment.

DeepSeek V4 Flash and Pro models now available via API, enabling developers to run open-weights inference at scale.

Developer built browser-based GTA game using Claude, Cesium, and real Google Earth/OSM data; entertainment demo, limited AI architecture insight.

russellromney/honker adds Postgres NOTIFY/LISTEN and Kafka-style queue semantics to SQLite via Rust extension.

Anthropic postmortem: three Claude Code harness bugs, not model quality, caused two-month regression in output quality.

Reddit discussion of GPT 5.5 with no technical details, benchmarks, or substantive analysis.

Bluesky's decentralized feed architecture explained: custom algorithm implementation running on commodity hardware via AT Protocol.

Reddit user reports GPT-5.5 performs well for creative writing tasks, positioning it as improvement over 5.1.

Reddit user seeks hardware purchasing advice between NVIDIA Blackwell and Mac Studio for local LLM inference.

I benchmarked and compared Claude Opus 4.5 vs Opus 4.6 vs Opus 4.7 vs Sonnet 4.6 testing effort levels from low, medium, high, xhigh, max as curious about token usage/costs and performance within Claude Code https://ai.georgeliu.com/p/tested-claude-ai-llm-models-effort Hope folks find this useful. The test was done with Claude Code v2.1.117 which is apparently the fixed versions from Anthropic's post-mortem announcement.

User reports Qwen 3.6 27B quantized model achieves 22 tok/s on RTX 5060 Ti with 24K context window.

Anthropic restores Claude Code access to Claude Pro subscribers after temporary removal.

Reddit post complaining about repetitive phrasing in open-weight LLM outputs; suggests retraining to eliminate the pattern.

Reddit user asks whether Qwen3.6 27B matches larger models; community discussion of open-weight model capabilities.

Reddit user reports Claude Desktop multi-account setup and discovers shared session storage across isolated instances.

Reddit post with emoji hype about ChatGPT 5.5; no substantive details or official announcement provided.

Reddit user benchmarks Qwen 3.6 35B vs 27B on coding tasks; 35B faster (72 TPS) but less accurate, 27B slower but more precise.

Reddit post with no substantive content; insufficient information to assess.

Claude users can access more apps with Anthropic's AI now thanks to new connectors for everything from hiking to grocery shopping. Anthropic already supported connecting numerous work-related apps to Claude, like Microsoft apps, but this expansion focuses on personal apps like Audible, Spotify, Uber, AllTrails, TripAdvisor, Instacart, TurboTax, and others. Some of these apps, such as Spotify, already have similar connectors in OpenAI's ChatGPT. Once an app is connected, Claude will suggest relevant connected apps directly in your conversations, like using AllTrails for hike recommendations. A...

Simon Willison ports LlamaIndex's LiteParse PDF text extraction tool to run in-browser, using spatial parsing and Tesseract OCR without ML models.

Trump-Xi summit may be rocked by US mulling huge sanctions.

Reddit thread asking ClaudeAI subreddit moderator bot about moderation work and concerns about model obsolescence.

Reddit user speculates OpenAI reached AGI and will outpace Anthropic; compares Codex and Claude Code features.

Sierra, the AI customer service agent startup founded by technologist Bret Taylor, announced today that it has acquired the YC-backed French startup Fragment.

GPT 5.5 offers better token efficiency than GPT 5.4 despite higher per-token pricing; comparison to Claude Opus 4.7 shows GPT 5.5 5-10x cheaper on ARC-AGI-2.

User automated product tutorial video generation end-to-end using Claude (script, voiceover, editing, publishing).

Reddit speculation post with no substantive claim.

In March 2026, three LLM agents generated over 600,000 lines of code, ran 850 experiments, and helped secure a first-place finish in a Kaggle playground... In March 2026, three LLM agents generated over 600,000 lines of code, ran 850 experiments, and helped secure a first-place finish in a Kaggle playground competition. Success in modern machine learning competitions is increasingly defined by how quickly you can generate, test, and iterate on ideas. LLM agents, combined with GPU acceleration, dramatically compress this loop. Historically… Source