Easy AI Daily | 2026-02-07

A roundup of the day's AI industry news, covering models, agents, infrastructure, research, products, business, and safety.

Models & Capabilities

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

• OpenAI released GPT-5.3-Codex, emphasizing self-bootstrapping development and high SWE-Bench Pro / Terminal-Bench scores, though without a public API.
• Anthropic shipped Claude Opus 4.6: 68.8% on ARC-AGI 2, improved long-context and reasoning, topping multiple Arena leaderboards (code/text/expert).
• Community benchmarks show both models have distinct strengths and cost profiles on real projects.

Links: GPT-5.3-Codex | Opus 4.6 announcement | Rails production benchmark

Long Context & Memory

• Opus 4.6 performs notably better than Gemini 3 Pro on 256k–1M token tasks, though users still report "context rot" in practice.
• Academic proposals: InfMem (imagine-retrieve-write + RL) and LatentMem (role-based trajectory compression into latent memory), claiming 3–4x lower inference latency and ~50% token savings on million-token QA and multi-agent collaboration.

Links: InfMem discussion | LatentMem discussion

Google Sequential Attention

• Google Research highlighted "Sequential Attention" for subset selection-based pruning/feature selection. Reddit users noted the paper is three years old — possibly recycled promotion — and asked whether it will land in Gemma models.

World Models: Waymo + Genie 3, and gWorld

• Waymo built a world model on DeepMind's Genie 3, generating realistic camera + 3D LiDAR traffic scenes (tornadoes, plane landings on highways) to stress-test autonomous driving — a shift from "pixel generation" to "sensor-stream generation."
• Trillion Labs & KAIST released gWorld (open-source 8B/32B vision world models) that generates HTML/CSS/JS instead of pixels. The 8B model hits 74.9% on MWMBench, allegedly beating 402B Llama 4 Maverick on some GUI tasks with <1% render failure. Community remains skeptical of the "crushes 402B" framing.

Links: gWorld-8B on Hugging Face | Waymo announcement

Agents & Tooling

Multi-Agent "Software Teams"

• Cursor claims hundreds of agents producing thousands of commits per hour weekly; Anthropic used 16 Claude agents to write a C compiler that builds the Linux kernel in two weeks.
• Community consensus: test cases, Git workflows, and infrastructure configuration matter more than the model itself.

Links: Cursor experiment | Claude Code agent teams

OpenClaw / Moltbot: Powerful but Costly

• OpenClaw + local LLMs (e.g., Qwen3Coder) can build strong self-iterating agents, but with a large security surface and weak docs.
• Moltbot's "free assistant" requires paid APIs (OpenAI/Anthropic/Google, Brave, ElevenLabs), totaling $50–100/month.

Recursive Language Models (RLM)

• A Twitter proposal treats agents as REPL programs: state in variables, structured data between sub-agents instead of pasting logs into prompts. LangChain/LangSmith push trace/sandboxing/typed state; DSPy promotes RLM with auto-tuning. See RLM blog post.

BalatroBench

• A benchmark wrapping the roguelike Balatro via BalatroBot's state API + BalatroLLM prompts, testing long-horizon planning. Supports any OpenAI-compatible endpoint. Links: BalatroBot | BalatroBench

AI Coding Assistant Stacks

• Cline users squeeze context from 200k to 40k tokens using .clineignore, memory banks, and RAG; aider users tune Copilot/Opus configs; Cursor users frustrated that Codex 5.3 remains API-less and Agent mode stops when free credits run out.

Infrastructure & Hardware

Blackwell / CUDA / Vulkan Pitfalls

• Wrong FP8 cuBLASLt kernel selection on Blackwell costs 2x throughput; MXFP8 instructions are needed. Some B200s deadlock under TMA + NCU profiling.
• LM Studio users found Vulkan can beat CUDA by 50% on some NVIDIA cards; on Apple M4 Max, MLX 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, even slow Stable Diffusion.
• Community takeaway: bandwidth matters more than compute; MoE + low-bit quantization are the budget user's friends.

Training & Eval Infra

• StepFun's Step 3.5-Flash: trained on 4096 H800s with 17.2T tokens, SWE-Bench 74.4, with eval-first methodology.
• SETA open-sourced 1,376 reproducible real terminal environments (DevOps, security, ops) for code agents.

Research

Hugging Face Community Evals

• Benchmark results stored as versioned YAML in model repos via PRs, with Inspect AI for reproducibility. Researchers note hard benchmarks (multilingual SWE-Bench, SciCode, VideoGameBench) remain far from saturated.

TinyLoRA

• StepFun's TinyLoRA trains only 13 LoRA parameters on Qwen-7B + RL, lifting GSM8K from 76% to 91% — the knowledge is already in the model; you just redirect the reasoning style.

Kaiming He & Attribution Work

• Kaiming He posted "Generative Modeling via Drifting" on OpenReview. Eleuther community discussed gradient unit-norm normalization improving training-sample attribution accuracy vs. arguments that good Hessian estimates make normalization unnecessary (paper 1, paper 2).

AI4Science

• OpenAI × Ginkgo Bioworks connected GPT-5 to automated wet labs for closed-loop protein experiments, reportedly cutting production costs 40%. Labbench2 launched with ~1,900 experimental-design and clinical-trial questions.

Products & Business

• Perplexity Pro quota cuts: Users discovered reduced Deep Research and file upload limits without communication, prompting migration to Gemini Pro and DeepSeek.
• How engineers actually use AI: Reddit consensus — best as supercharged search/example generators (SQL, APIs, legacy code); poor at large refactors; many "AI projects" could be RPA or simple scripts.
• Lotus raised $41M for an AI-triage + licensed-doctor telehealth platform targeting 100M Americans without family doctors.
• Crypto AI tools: MCP Contributors community building dashboards + natural-language explanations of on-chain contracts for non-Solidity-reading investors.
• OpenAI vs Anthropic burn rates: Estimates put Anthropic at $18B revenue this year ($55B next), but $12B training + $7B inference spend and $139B cumulative opex before ~2028 breakeven, with a $350B valuation and $10B raise discussed.
• Lodash received $200K from the EU Sovereign Tech Fund as "critical software infrastructure."
• Manus.im trust crisis: Users charged $5,000 after downgrades, site outages, and unresponsive support.
• OpenRouter: Launched stealth model Pony Alpha (tool calling/agent workflows), added Moonshot's Kimi K2.5, and reportedly grew usage ~10x in two years.

Policy, Governance & Safety

• Codex full filesystem access: OpenAI's Codex reads the entire filesystem by default with no obvious permission toggle; issues #2847 and #5237 document exposure of API keys and medical files, with maintainers deeming it "not a bug."
• Opus 4.6 finds 500+ zero-day vulnerabilities in open-source libraries via sandboxed audits, with fixes — sparking debate over a new "cumulative CVSS repaired" benchmark vs. arming attackers.
• AI Red Team roles heating up: Trajectory Labs hiring long-term red teamers (30+ hrs/week) with gamified interviews.
• Secret leakage: Communities call for default detect-secrets hooks to redact credentials from prompts and logs.

📌 Source: Easy AI Daily