Easy AI Daily | 2026-02-07

A translated digest of the February 7, 2026 AI industry news roundup.

Models & Capabilities

GPT-5.3-Codex vs Claude Opus 4.6: Frontier Coding Models Clash

OpenAI released GPT-5.3-Codex, emphasizing self-bootstrapping development and high SWE-Bench Pro/Terminal-Bench scores with finer reasoning but no public API. Anthropic launched Claude Opus 4.6, scoring 68.8% on ARC-AGI 2 with improved long-context and reasoning, topping multiple Arena leaderboards (code, text, expert). Community comparisons show each model has strengths and cost tradeoffs on real projects.

• GPT-5.3-Codex introduction
• Claude Opus 4.6 announcement & benchmarks
• Rails production repo benchmark: Codex vs Opus

Long Context & Memory: Opus 4.6, InfMem, LatentMem

Opus 4.6 clearly outperforms Gemini 3 Pro on 256k–1M token long-context tasks, though users still report "context rot." Academic proposals InfMem (imagine-retrieve-write + RL) and LatentMem (compressing trajectories into latent memory by role) claim 3–4x inference speedups and ~50% token savings on million-token QA and multi-agent tasks.

Google Sequential Attention and Model Pruning

Google Research re-highlighted "Sequential Attention" for subset-selection-based pruning. Redditors noted the paper is three years old — more of a re-promotion — and wondered if it will land in Gemma for inference acceleration.

World Models: Waymo + DeepMind Genie 3

Waymo's new world model built on Genie 3 generates realistic traffic scenes with camera + 3D LiDAR streams to stress-test autonomous driving with extreme events (tornadoes, plane landings on highways). Researchers call it a leap from "pixel generation" to "direct sensor stream output."

gWorld: Generating Phone GUI Worlds via Web Code, Not Pixels

Trillion Labs and KAIST released gWorld (8B/32B open-source visual world models) that generates HTML/CSS/JS and renders UI instead of predicting pixels. The 8B model hits 74.9% on MWMBench, reportedly beating 402B Llama 4 Maverick on some GUI tasks with <1% render failure. The community is skeptical of the headline claim but appreciates the structured-code approach. Model: gWorld-8B on Hugging Face.

Agents & Tooling

• Multi-agent "software teams" mature: Cursor claims hundreds of agents producing thousands of commits per hour; Anthropic used 16 Claude agents to write a compilable Linux-kernel C compiler in two weeks. Community consensus: test cases, Git workflows, and infrastructure matter more than the model itself.
• OpenClaw / Moltbot: OpenClaw + local LLMs (e.g., Qwen3Coder) builds strong self-iterating agents but with a large security surface and weak docs. Moltbot's "free assistant" actually requires paid APIs (OpenAI/Anthropic/Google, Brave, ElevenLabs), totaling ~$50–100/month.
• Recursive Language Models (RLM): A proposal to treat agents as REPL programs — state in variables, structured data between sub-agents — to reduce context rot; LangChain pushes trace/sandboxing/typed state, DSPy promotes RLM with auto-tuning. See the RLM blog post.
• BalatroBench: Benchmarks LLM long-horizon planning via the card game Balatro, using BalatroBot and BalatroLLM with any OpenAI-compatible endpoint.
• Coding assistant stacks: Cline users shrink context from 200k to 40k tokens via .clineignore, memory, and RAG; aider users tune Copilot/Opus architectures; Cursor users complain GPT-5.3-Codex lacks API access and free agent credits run out fast.

Infrastructure & Hardware

• Blackwell / CUDA / Vulkan pitfalls: Wrong cuBLASLt FP8 kernel selection on Blackwell halves throughput (MXFP8 instructions needed); some B200s deadlock with TMA + NCU profiling. Vulkan can beat CUDA by 50% on some NVIDIA cards; on Apple M4 Max, MLX runs Qwen3-Coder-Next 4-bit inference 2x+ faster than GGUF.
• CPU-only local LLMs are viable: A 2018 8th-gen i3 with UHD 620 runs a 16B MoE model at ~10 token/s via OpenVINO; an i5-8500 + 32GB handles 12B Q4 GGUF for chat, TTS, even slow Stable Diffusion. Consensus: memory bandwidth matters more than compute; MoE + low-bit quantization are budget-friendly.
• Training/eval infra: StepFun's Step 3.5-Flash trained on 4,096 H800s with 17.2T tokens, scoring 74.4 on SWE-Bench. SETA open-sourced 1,376 reproducible terminal environments (DevOps, security, ops) for realistic code-agent evaluation.

Research & Methods

• Hugging Face Community Evals: Benchmark results stored as versioned YAML in model repos via PRs, with Inspect AI for reproducible runs. Ofir Press notes hard benchmarks (multilingual SWE-Bench, SciCode, VideoGameBench) remain far from saturated.
• TinyLoRA: Training only 13 LoRA parameters on Qwen-7B plus RL lifts GSM8K from 76% to 91% — the knowledge was already there; only the reasoning style needed nudging.
• Kaiming He's "Generative Modeling via Drifting" appears on OpenReview; Eleuther discussions cover gradient unit-norm normalization improving attribution accuracy (arXiv:2410.17413) and counterpoints based on Hessian quality (arXiv:2504.16430).
• AI4Science: OpenAI + Ginkgo Bioworks connect GPT-5 to automated wet labs for closed-loop protein experiments, reportedly cutting production costs 40%. Labbench2 launches with ~1,900 experimental-design and clinical-trial questions.

Products & Applications

• Perplexity Pro limits quietly tightened: Reduced Deep Research quotas and file upload sizes triggered user backlash and migration toward Gemini Pro and DeepSeek.
• Engineers use AI as "super search": A Reddit discussion among professional engineers found AI is most useful for SQL, API lookups, and legacy-code exploration — not large-scale refactoring; many "AI projects" turn out to need only RPA or scripts.
• Lotus raises $41M for AI-assisted + licensed-physician online primary care targeting 100M Americans without family doctors; AI handles triage and intake while doctors make decisions.
• AI crypto tools: MCP Contributors community builds dashboards translating on-chain contracts and trades into plain language for non-Solidity investors.

Industry & Business

• OpenAI vs Anthropic burn rates: Leaked/estimated figures suggest Anthropic revenue of $18B this year (projected $55B next), with $12B training + $7B inference spend, ~$139B cumulative operating expenses by 2028, and a $350B valuation with a fresh $10B raise. Commenters questioned OpenAI's solvency even more.
• Lodash declared critical infrastructure by the EU: The JS utility library received $200K from the EU Sovereign Tech Fund for maintenance and security hardening.
• Manus.im trust crisis: Users report being charged $5,000/account after downgrades, site outages, unresponsive support, and unexplained bans — a crisis-level event for a website hosting/building platform.
• OpenRouter ecosystem growth: New stealth model "Pony Alpha" focuses on tool calling and agent workflows; Moonshot's Kimi K2.5 added; usage reportedly grew ~10x in two years.

Policy, Governance & Safety

• Codex reads the entire filesystem by default: OpenRouter community members found OpenAI's Codex can read the whole local filesystem with no obvious permission switch, and maintainers consider it "not a bug" — with real cases of exposed API keys and medical records (issue #2847, issue #5237).
• Opus 4.6 finds 500+ zero-day vulnerabilities in open-source libraries via sandboxed automated audits, with fixes. Some propose a "cumulative CVSS repaired" benchmark; others worry about arming attackers.
• AI Red Teaming becomes a career: Stealthy AI safety firms like Trajectory Labs are hiring long-term red teamers (30+ hrs/week) with jailbreak-style interview challenges.
• Secret-leak prevention by default: Communities call for tools like detect-secrets as default hooks to redact keys from prompts and logs, after repeated .env and cloud-credential exposures.

---

📌 Source: Easy AI Daily