Jason Says
Opus 5 drops and Cognition buys Poke in the same news cycle — cheaper models are table stakes now; the real moat is whether your AI has a personality worth coming back to.
Anthropic released Opus 5, its new flagship model priced below Fable with fewer usage restrictions. This positions Opus 5 as the go-to choice for most developers and enterprises, lowering the cost of building on Anthropic's ecosystem significantly.
Black Forest Labs' FLUX 3 multimodal flow model outperforms Seedance 2.0, Gemini Omni, and Grok Imagine on generation benchmarks. It also ships FLUX-mimic, a video-action model for robotics — a major leap for open-source generative AI developers.
Cognition (maker of Devin) acquired Poke for a low nine-figure sum to bring its conversational personality to its coding agent. The deal signals a shift: AI differentiation is moving from raw capability to interaction style and user affinity.
ego-lite is a browser that lets AI agents run tasks in parallel Spaces while you keep your own tabs. Unlike browser-use or agent-browser frameworks, it shares login state natively and reduces token overhead — a natural fit for Claude Code and Cursor power users.
Kronos is the first open-source foundation model trained specifically on financial candlestick (K-line) data from 45+ global exchanges, accepted at AAAI 2026. For quant developers, it's the first serious LLM backbone purpose-built for structured price data — no more forcing GPT onto OHLCV.
SANA-Video 2.0 combines gated linear attention with periodic softmax anchors to generate 720p video on a single GPU at 5B and 14B scales, matching full-softmax quality. For indie developers, this dramatically lowers the hardware barrier to high-quality video generation.
As Washington debates responses to Chinese AI and alleged model distillation, Nvidia, Mistral and others are pushing back against broad restrictions on open-weight models. The outcome could reshape global developers' access to top-tier open-source AI.
This paper proposes distilling an agent's interaction history directly into model weights, so learned experience persists after the context window clears. For developers building long-running AI agents, this means agents can genuinely accumulate expertise across sessions — not just within one conversation.
专注科学与算力交叉领域的风险基金，第三期规模较上期扩大 60%，重点押注 AI 基础设施与科学计算。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free