Easy AI Daily | February 3, 2026

A digest of the day's key AI industry news, curated by the Easy AI Daily community on zhichai.net.

Products & Applications

• OpenAI releases standalone Codex App for macOS: A "command center" for developer code agents, integrating multi-agent parallel execution, per-task git worktrees, a /plan planning mode, reusable Skills, and scheduled Automations — no longer a VSCode plugin.
Links: Official intro | Codex product page
• Windsurf Wave 14 adds Arena mode: Compare multiple models side-by-side on the same coding task, with Battle Groups at 0x credit cost for a week, plus personal and public leaderboards.
Link: Windsurf download
• LM Studio 0.4.1 speaks Anthropic's protocol: A local /v1/messages-compatible endpoint lets tools like Claude Code switch to local GGUF/MLX models by changing only the base URL; OpenAI-compatible endpoints and a TypeScript SDK also available.
Link: LM Studio blog
• PerpetualBooster v1.1.2: A Rust GBM replacing hyperparameter tuning with a single "budget" parameter, claiming ~100x speed over LightGBM+Optuna with similar accuracy; supports R, ONNX export, and saving as XGBoost.
Link: GitHub
• BCG deploys 36,000+ custom GPTs for 32,000 consultants, fine-tuned by role and methodology with project memory and team sharing — systematic internal SaaS-style adoption.
• AEGIS-FLOW: A community multi-agent cloud security framework (LangGraph + MCP + Next.js) that scans AWS configs, generates reports and Terraform fix patches, and enforces human approval before changing infrastructure. Demo
• Lutum Veritas: An indie deep search engine claiming deeper research than ChatGPT/Gemini/Perplexity at ~$0.20/query, with BYO API keys, multi-model routing, mandatory citations, and a verification ASK mode. GitHub

Models & Capabilities

• StepFun Step-3.5-Flash: A 196B-parameter sparse MoE (~11B active) targeting long-context and agentic use; official figures: 74.4% SWE-bench Verified, 51% Terminal-Bench 2.0. Day-zero vLLM support, llama.cpp PR open, and an Int4 quant runs 256K context on 128GB machines.
Links: FP model | Int4 model
• Moonshot Kimi K2.5: Ranked #1 open model (5th overall) in LMArena's Code Arena; Perplexity has integrated it into Pro/Max tiers using US-hosted nodes.
• GLM-4.7 Flash praised for frontend/interactive web coding: Community suggests pairing it with Claude/Kimi for review — GLM executes, high-end models audit.
Link: ggerganov on GLM-4.7
• Claude Sonnet 5 leaks: Vertex AI 404s for claude-sonnet-5 fuel rumors of 1M context, ~50% cheaper than Opus 4.5, TPU-optimized, and 80.9% on SWE-Bench — though the community urges skepticism.
• TII Falcon-H1-Tiny: Sub-100M specialized micromodels using anti-curriculum data injection, the Muon optimizer, and Hybrid Mamba+Attention. A 90M tool-call model hits 94% relevance detection; a 600M reasoning version scores 75% on AIME24.
• Assistant_Pepe_8B: An 8B model fine-tuned on extended 4chan data reportedly outperforms its Nemotron base across evals, sparking discussion about stylistic statistical effects and the "alignment tax" on small models. Model

Agents & Tooling

• Best practice: "write the test first" in CLAUDE.md: Requiring agents to write a reproduction test before fixing bugs is reported as the single most effective prompt for reliability.
• "Conductor-style engineering": One developer driving 5–10 parallel agents; critics warn context-switching costs degrade quality.
• OpenClaw ecosystem is fun but risky: Security audit scored 2/100; live red-team drills showed memory files (.md) can be injected to hijack future behavior, and costs can balloon quickly on OpenRouter.
Link: Audit report
• LangChain ships deepagents (JS): Distills four architecture patterns from working systems like Claude Code and Manus, replacing naive "one model, many tools" setups.
• Recursive Language Models (RLM) in practice: Used for code security auditing with model-driven planning and sub-model calls; custom tool calls and Deno sandboxing remain friction points in DSPy.

Infrastructure & Hardware

• Long context is memory-bound: Imperial College + Microsoft survey shows a 1M-context, batch-1 DeepSeek-R1 request can need ~900GB of KV cache; prefill and decode may need heterogeneous hardware. dair.ai thread
• FlashAttention v3 lands on AMD RDNA GPUs: PR #2178
• Triton-Viz 3.0: Visualizes every load/store/matmul with OOB detection, supporting Triton and Amazon NKI, plus Triton-Puzzles Colab integration.
• Blackwell (sm120) kernel findings: Well-implemented TMA+mbarrier slightly beats cp.async for large matrices; cuBLAS still appears to use sm80-era kernels; one missing __syncthreads() can deadlock a kernel.
• MIT heat-powered chip: Computes matrix-vector multiplication using on-chip temperature gradients; currently limited to 2×2/3×3 matrices at ~99% accuracy.
• Fudan "sushi-roll" fiber chip (Nature): Rolled circuits on elastic substrate integrate ~100K transistors per cm, tolerating 15.6-ton compression, repeated bending, and 100°C; aimed at smart fabrics and brain-computer interfaces.

Research & Methods

• Why coding agents work so well: The software world is inherently verifiable (compilers, unit tests) and rich in symbolic tooling (shell, debuggers) — a neuro-symbolic recipe other domains must replicate.
• Infinite synthetic RLVR: Masking reasoning steps in web pages to generate verifiable reasoning tasks may revive models saturated on existing RLVR data; synthetic data collapse looks like an engineering problem, not an inevitability.
• Stop over-trusting perplexity: Better next-token prediction doesn't guarantee better instruction following, tool-call stability, or multi-turn consistency.
• ConceptMoE: Clusters similar tokens into "concept units" before MoE routing to cut redundant computation for long inputs.
• Token-level data filtering: Alec Radford collaborators propose token-level filtering/weighting during pretraining for finer-grained capability control.
• Brains vs LLMs (Nature): The brain's temporal hierarchy of speech processing aligns with LLM layer depth — deeper layers map to later, higher-order language processing. Paper

Policy, Safety & Industry

• OpenClaw red-team lessons: After social engineering and pipeline RCE failed, attackers bypassed defenses via shell-expansion variables hidden in JSON metadata. Long-term memory stored in .md files is the biggest attack surface; isolate credentials and blast radius.
• Prompt injection defense combo: Embedding-similarity input filtering plus grammar-constrained decoding to restrict what models can output.
• Waymo reportedly raising $16B at ~$110B valuation, with at least $13B from Alphabet and participation from Sequoia, DST, and Dragoneer — up from $45B in October 2024.
• xAI Grok Imagine 1.0: Generates 10-second 720p video with audio; over 1.2 billion videos created in the past 30 days. Announcement