Easy AI Daily Digest | February 7, 2026: GPT-5.3-Codex vs Claude Opus 4.6, World Models, and Agent Teams

Key points

Models & Capabilities

• GPT-5.3-Codex vs Claude Opus 4.6: OpenAI's GPT-5.3-Codex emphasizes self-bootstrapping development and high SWE-Bench Pro/Terminal-Bench scores, but has no public API. Anthropic's Claude Opus 4.6 hits 68.8% on ARC-AGI 2 with improved long-context and reasoning, topping multiple Arena leaderboards. Community comparisons show each has strengths and different costs on real projects.
• Long context & memory: Opus 4.6 outperforms Gemini 3 Pro on 256k–1M token tasks, though users still report "context rot." Academic proposals InfMem (anticipate-retrieve-write + RL) and LatentMem (role-based trajectory compression into latent memory) claim 3–4x lower inference latency and ~50% token savings on million-token QA and multi-agent tasks.
• Google Sequential Attention: subset-selection pruning/feature-selection technique, criticized as a re-promotion of a three-year-old paper (arXiv:2209.14881).
• World models: Waymo built a world model on DeepMind's Genie 3 generating photorealistic camera + 3D LiDAR traffic scenarios to stress-test autonomous driving with extreme events. Trillion Labs & KAIST released gWorld (8B/32B open-source), generating HTML/CSS/JS instead of pixels; the 8B model scores 74.9% on MWMBench, with community skepticism about "beats 402B" claims.

Agents & Tooling

• Cursor reports hundreds of agents making thousands of commits per hour; Anthropic used 16 Claude agents to write a C compiler that compiles the Linux kernel in two weeks. Community consensus: test cases, Git workflows, and infrastructure matter more than the model itself.
• OpenClaw + local LLMs (e.g., Qwen3Coder) enable strong self-iterating agents but raise security concerns. Moltbot's "free" assistant requires paid APIs totaling $50–100/month.
• Recursive Language Models (RLM): treat agents as REPL programs with typed state and structured data between sub-agents; LangChain/LangSmith add trace and sandbox updates, DSPy promotes RLM with auto-tuning.
• BalatroBench: benchmarks LLM long-horizon planning via the card roguelike Balatro using BalatroBot APIs and BalatroLLM prompts.
• Cline users trim context from 200k to 40k tokens with .clineignore, memory banks, and RAG; Cursor users complain Codex's agent mode stops after free quota.

Infrastructure & Hardware

• On Blackwell, wrong cuBLASLt FP8 kernel selection halves throughput; some B200s deadlock with TMA + NCU profiling. Vulkan can be 50% faster than CUDA for local LLMs on some NVIDIA cards; MLX is 2x+ faster than GGUF for Qwen3-Coder-Next 4-bit on M4 Max.
• CPU-only inference: a 2018 8th-gen i3 runs a 16B MoE model at 10 token/s via OpenVINO; older i5-8500 + 32GB handles 12B Q4 GGUF. Community takeaway: bandwidth matters more than compute; MoE and low-bit quantization help.
• StepFun's Step 3.5-Flash: trained on 4,096 H800s with 17.2T tokens, SWE-Bench 74.4. SETA open-sourced 1,376 reproducible terminal environments for code agents.

Research & Methods

• Hugging Face Community Evals: benchmark results stored as versioned YAML in model repos via PRs, using Inspect AI; researchers note hard benchmarks remain far from saturated.
• TinyLoRA: training only 13 LoRA parameters plus RL lifts GSM8K from 76% to 91% on Qwen-7B.
• Kaiming He published "Generative Modeling via Drifting"; Eleuther discussions cover gradient-norm normalization for training sample attribution (arXiv:2410.17413, arXiv:2504.16430).
• AI4Science: OpenAI x Ginkgo Bioworks connect GPT-5 to automated wet labs, reportedly cutting production costs 40%; Labbench2 offers ~1,900 experimental-design and clinical-trial questions.

Products & Industry

• Perplexity quietly cut Pro Deep Research quotas, prompting user exodus toward Gemini Pro and DeepSeek.
• Engineers report AI is most useful as "super search" (SQL, API docs, legacy code) rather than large-scale code rewrites.
• Lotus raised $41M for AI-assisted telehealth with licensed doctors targeting 100M Americans lacking primary care.
• Anthropic revenue projected at $18B this year, $55B next, with ~$139B cumulative operating spend by 2028; investors weigh a $350B valuation.
• Lodash received $200k from the EU Sovereign Tech Fund as critical infrastructure; Manus.im faces a trust crisis over $5,000 charges after downgrades; OpenRouter launched stealth model Pony Alpha and Kimi K2.5, with usage up ~10x in two years.

Policy & Security

• OpenAI Codex reads the entire local filesystem by default with no clear permission controls; GitHub issues note exposed API keys and medical files.
• Claude Opus 4.6 reportedly found 500+ zero-day vulnerabilities in open-source libraries, raising both benchmarking interest and dual-use concerns.
• AI red team roles are heating up (e.g., Trajectory Labs hiring); communities call for default secret-scanning hooks like detect-secrets.