Easy AI Daily | 2026-02-07

A structured digest of AI industry developments for February 7, 2026.

Key points

• Frontier coding model showdown: OpenAI released GPT-5.3-Codex (focused on bootstrapped development, high SWE-Bench Pro / Terminal-Bench scores, fine-grained reasoning, no public API), while Anthropic shipped Claude Opus 4.6 (68.8% on ARC-AGI 2, improved long context and reasoning, multiple Arena #1 spots for code/text/expert). Real-world community comparisons show each excels on different projects with different costs.
• Long context & memory: Opus 4.6 outperforms Gemini 3 Pro on 256k–1M token tasks, though users still complain of "context rot." Academic proposals InfMem (imagine-retrieve-write + RL) and LatentMem (compressing trajectories into latent memory) claim 3–4x inference speedups and ~50% token savings on million-token QA and multi-agent tasks.
• World models: Waymo built a world model on DeepMind's Genie 3 that generates camera + 3D LiDAR traffic scenes to stress-test autonomous driving with extreme events. Trillion Labs & KAIST's gWorld (8B/32B, open source) generates HTML/CSS/JS instead of pixels; the 8B model hits 74.9% on MWMBench with <1% render failure, though community is skeptical of "beats 402B Llama 4 Maverick" claims.
• Agent teams: Cursor claims hundreds of agents making thousands of commits hourly; Anthropic had 16 Claude agents write a C compiler that compiles the Linux kernel in two weeks. Community consensus: test cases, Git workflows, and infrastructure matter more than the models themselves.
• Local agent frameworks: OpenClaw + local LLMs (e.g., Qwen3Coder) enable strong self-iterating agents but with security and documentation gaps; Moltbot's "free assistant" actually requires $50–100/month in paid APIs (OpenAI/Anthropic/Google, Brave, ElevenLabs, etc.).
• Engineering patterns: "Recursive Language Models (RLM)" treat agents as REPL programs with state in variables and structured data between sub-agents; LangChain/LangSmith push traces, sandboxes, and typed state; DSPy promotes RLM with auto-tuning.
• Blackwell/CUDA pitfalls: Wrong cuBLASLt FP8 kernel selection on Blackwell halves throughput (MXFP8 instructions needed); some B200s deadlock with TMA + NCU profiling. LM Studio users found Vulkan up to 50% faster than CUDA on some NVIDIA cards; MLX on M4 Max runs Qwen3-Coder-Next 4-bit over 2x faster than GGUF.
• Budget local inference: A 2018 8th-gen i3 with UHD 620 runs 16B MoE models at ~10 tok/s via OpenVINO; an i5-8500 + 32GB handles 12B Q4 GGUF for chat, TTS, and slow Stable Diffusion. Consensus: memory bandwidth beats compute; MoE + low-bit quantization are budget-friendly.
• Training/eval infra: StepFun's Step 3.5-Flash trained on 4096 H800s with 17.2T tokens, scoring 74.4 on SWE-Bench. SETA open-sourced 1,376 reproducible terminal environments (DevOps, security, ops) for realistic agent evaluation. Hugging Face launched Community Evals storing versioned benchmark results as YAML in model repos via PRs with Inspect AI.
• Research: StepFun's TinyLoRA trains just 13 LoRA parameters on Qwen-7B (+RL) to lift GSM8K from 76% to 91%. Kaiming He published "Generative Modeling via Drifting" on OpenReview. Eleuther discussed gradient unit-norm normalization improving training-sample attribution accuracy.
• AI4Science: OpenAI × Ginkgo Bioworks connect GPT-5 to automated wet labs for closed-loop protein experiments (claimed 40% production cost reduction); Labbench2 launches with ~1,900 hard experimental-design and clinical-trial questions.
• Products & industry: Perplexity quietly cut Pro Deep Research quotas and file upload sizes, angering users. Engineer discussions say AI works best as a "super search"/example generator, not for large-scale codebase changes. Lotus raised $41M for AI-assisted + licensed-physician primary care targeting ~100M Americans without family doctors. Lodash received $200K from the EU Sovereign Tech Fund as critical infrastructure. Manus.im faced a trust crisis: $5,000 charges after downgrades plus downtime. OpenRouter launched stealth model Pony Alpha, added Moonshot's Kimi K2.5, and reportedly grew usage ~10x in two years.
• Security & governance: OpenAI's Codex reads the entire local filesystem by default with no permission toggle (official issues say "not a bug"), exposing API keys and medical files. Anthropic reports Opus 4.6 found 500+ zero-day vulnerabilities in open-source libraries in sandboxed audits — raising both benchmark ideas and dual-use concerns. AI Red Team roles are emerging as real careers (e.g., Trajectory Labs hiring). Communities call for default detect-secrets integration in agent/IDE tools.

Notable links

• GPT-5.3-Codex announcement
• Claude Opus 4.6 announcement
• Sequential Attention paper
• gWorld-8B on Hugging Face
• TinyLoRA discussion
• Generative Modeling via Drifting
• Codex filesystem issue
• Why the Sovereign Tech Fund supports Lodash