📰 AI News Daily — 09 Feb 2026

TL;DR (Top 5 Highlights)

OpenAI debuts an AI largely designed by AI—accelerating progress and intensifying safety and accountability debates.
Anthropic ships Claude Opus 4.6 (research preview) with Fast mode and GitHub Copilot integration; free credits spur experimentation.
China’s labs unleash a model wave; Kimi K2.5 tops OpenRouter as ByteDance releases Seedance 2.0 for video.
Big Tech plans a record ~$650B in AI capex, supercharging data centers while raising resource and macro concerns.
Apple opens CarPlay to third‑party chatbots and brings Gemini to Siri, escalating the consumer assistant race.

🛠️ New Tools

Vouch launched a trust-management system for open-source AI where reputable maintainers vouch for contributors, reducing model and package supply‑chain risk and enabling safer collaboration across distributed repositories and communities.
Flux 2 Klein 4B brings higher FPS and LoRA support for controllable diffusion, enabling faster, more iterative image generation pipelines and lower‑latency creative feedback loops on commodity GPUs.
Composio released a plugin connecting Claude Code to 500+ apps (Gmail, Slack, GitHub), simplifying multi‑app agent workflows and letting developers automate end‑to‑end tasks without custom integrations.
ByteDance introduced Seedance 2.0 for video generation, improving quality and controllability while expanding creative options for ads, short‑form storytelling, and rapid prototyping of motion concepts.
xAI expanded the Grok Imagine API with new image generators, offering richer styles and faster iteration for developers building creative tools, marketing assets, and visual assistants.

🤖 LLM Updates

Anthropic released Claude Opus 4.6 (research preview) with a new Fast mode, accelerating web and app generation, and began integrating speedups into GitHub Copilot and Copilot CLI for snappier workflows.
China’s February surge delivered Qwen3.5 (with native vision‑language), GLM‑5, MiniMax M2.2, Seed 2.0, and DeepSeek‑V4, intensifying global competition and rapidly raising open and closed‑source baselines.
Kimi K2.5 surged to the top of OpenRouter, signaling shifting model preferences among power users and pressure on Western labs to compete on reasoning, latency, and cost.
Codex 5.3 improved coding speed and accuracy enough that some developers consider switching from Claude Code, highlighting intensifying competition for IDE assistants and enterprise code generation budgets.
Local coding agents matured into practical tools on commodity hardware (~50GB RAM), enabling reliable offline development, better privacy, and reduced cloud costs for teams handling sensitive codebases.
MLX unveiled a CUDA backend posting blistering token throughput on Qwen3 4B, slashing startup time and generation latency for Mac‑first workflows targeting NVIDIA‑accelerated infrastructure.

📑 Research & Papers

Zyphra introduced OVQ‑attention, a technique for longer contexts with lower compute, promising cheaper large‑document reasoning and improved retrieval‑augmented generation without sacrificing throughput.
DeepSeek detailed Engram embeddings trained on a billion n‑grams, improving phrase‑level understanding and robustness in retrieval systems, semantic search, and agent memory recall.
DuoGen proposed tightly interleaved multimodal generation, allowing text and vision tokens to co‑evolve, which yields more coherent step‑by‑step reasoning for complex tool‑use and captioning tasks.
Researchers from Meta, Cornell, and CMU showed smaller models can learn complex reasoning with targeted curricula, challenging scale‑first assumptions and encouraging smarter pretraining over brute‑force parameter growth.
New community evaluations, including Context‑Bench, tackle benchmark saturation by testing long‑horizon memory and context management, offering more realistic signal for productization readiness.
Stanford’s SleepFM predicts 130+ conditions from one night’s sleep data, hinting at low‑cost screening tools and personalized preventative care powered by physiological sensing and AI.

🏢 Industry & Policy

Alphabet, Amazon, Meta, and Microsoft plan a record ~$650B in AI capital spending, accelerating datacenter buildouts and model training while raising concerns about energy use, supply chains, and distorted macro indicators.
OpenAI unveiled a model largely designed by its own AI systems, advancing automated research while amplifying safety, accountability, and regulatory questions around self‑improving AI.
Apple will open CarPlay to third‑party assistants (ChatGPT, Claude, Gemini) and integrate Gemini with Siri, shifting competition to real‑world assistant performance inside the dashboard and across devices.
Snowflake announced a $200M partnership with OpenAI, bringing advanced models like GPT‑5.2 to its data cloud so enterprises can build agents on proprietary data without moving it, strengthening on‑platform governance.
OpenAI, Anthropic, Google, and Microsoft formed the Agentic AI Foundation under the Linux Foundation to push open standards for responsible agents, counter concentration risk, and ease interoperability.
Security watch: the OpenClaw agent platform faced malware uploads; on‑chain agents caused market volatility; and Ethereum’s new ERC‑8004 standard targets manipulation‑resistant, trustless agent interactions.

📚 Tutorials & Guides

A hands‑on guide adds searchable memory to Claude coding agents without a vector database using three Python packages and a file watcher, cutting infra overhead while improving recall and responsiveness.
A curated paper roundup covers advanced RAG architectures, TinyLoRA, heterogeneous agent compute scheduling, and semi‑autonomous math discovery—offering concrete techniques teams can pilot this quarter.

🎬 Showcases & Demos

Claude Opus 4.6 Fast mode built full animated websites in seconds and a persistent multiplayer “full‑world” game in hours, showcasing rapid prototyping that compresses idea‑to‑demo cycles.
A creator ported an AI coloring‑book app to iOS in a day using Opus 4.6 Fast mode, highlighting faster mobile UI implementation and iteration.
Another newcomer built a Rust‑based YouTube music app with timed lyrics overnight using AI assistance, illustrating how agentic coding lowers barriers for solo developers.
MiniCPM‑o 4.5 demonstrated real‑time, full‑duplex vision‑language interaction by tracking live price tags, pointing to assistants that continuously perceive and respond without pauses.
The Growing Graphs demo visualized graph‑rewriting automata with evolving, cell‑splitting dynamics, turning abstract computation theories into intuitive, exploratory visuals.

💡 Discussions & Ideas

Trust philosophies diverge: OpenAI’s rule‑centric governance versus Anthropic’s character‑centric approach, fueling debate on how to scale dependable systems and align incentives across ecosystems.
Benchmark fatigue set in as MMLU/GSM8K saturate; researchers push community, task‑grounded evals mirroring production constraints—latency, failure recovery, long‑horizon memory—over leaderboard chasing.
Many foresee a software industrial revolution: consumer models deliver instant answers, while high‑end agentic stacks supercharge power users—multiplying code output and shifting value capture to orchestration layers.
Counter‑narratives to “AI is slowing” emphasize relentless capability gains and collapsing costs, enabling full‑stack prototypes for pennies and faster product iteration.
Builders spotlight “compounders”—framework and engine creators—as new leverage points, advocating intentional AI system design over trial‑and‑error and exploring self‑referential agents (e.g., Codex prompting Codex).
Engineering critiques warned about enlarged LLM attack surfaces, hidden “grep tax” on structured data, and persistent hyperparameter tuning bottlenecks that still throttle reliability and scale.

Source Credits

Curated from 250+ RSS feeds, Twitter expert lists, Reddit, and Hacker News.