Cur8 — Monday, August 31, 2026
Haiku R1’s release and DeepSeek's open weights model are generating significant discussion, highlighting accelerating innovation and potential national security concerns respectively. Simultaneously, a tiny LLM experiment underscores ongoing efforts to optimize resource efficiency. Anthropic launched Claude Code Opus 5 Auto Mode, furthering competition in the coding assistant space.
New model releases
Haiku R1/beta6 launched after a two-year beta period; Haiku will mentor 3 students in Google Summer of Code 2026, continuing previous years' involvement. Financial reports for 2022-2024 are now available.
Open weights models
DeepSeek’s open-source V4 Pro model rivals Fable quality with a novel plugin architecture and rapid iteration, demonstrated by building a functional app in under 30 minutes for $30. Founder Liang Wenfu prioritizes AI learning research over consumer products, aiming to drive down the cost of intelligence globally via China's token factory role.
Quantization shrinks LLMs from 14GB (FP16) to as little as 63.66MB (single-bit), enabling CPU-only operation at 78 tokens/second, but severely degrades performance—resulting in nonsensical or hallucinated outputs. This extreme compression prioritizes speed and minimal hardware over intelligence.
OpenAI news
ChatGPT Ads reached a $1 billion annualized revenue run rate, enabling wider AI accessibility via free and lower-cost tiers globally. This supports OpenAI's continued development and distribution.
Anthropic news
Claude Code's Auto Mode (default since mid-August) is vulnerable to prompt injection attacks achieving up to 80% success rate, bypassing Anthropic’s claimed 0.00% benchmark via a multi-step chain exploiting module shadowing and code execution. Anthropic acknowledges this as working "as designed," emphasizing the need for OS isolation and network controls.