Easy AI Daily | 2026-02-03

A digest of AI industry news for February 3, 2026, organized by product launches, models, agents and tooling, infrastructure, research, safety, and business.

Product & Application Launches

• OpenAI releases standalone Codex App for macOS: Integrates multi-agent parallel execution, per-task worktrees, a /plan mode, reusable Skills, and scheduled Automations in a single "command center" — positioned as a code-agent workbench rather than an IDE plugin. (Official intro | Codex product page)
• Windsurf launches Arena mode (Wave 14): Compare multiple models side-by-side on the same coding task, with Battle Groups at 0x credit cost for a week. (Download page)
• LM Studio 0.4.1 adds Anthropic API compatibility: A local /v1/messages endpoint lets tools like Claude Code swap backends to local GGUF/MLX models by changing the base URL; OpenAI-compatible endpoints and a TypeScript SDK also included. (Blog post)
• PerpetualBooster v1.1.2: A Rust gradient boosting library replacing hyperparameter tuning with a single "budget" parameter — claims up to 100x faster than LightGBM+Optuna at similar accuracy, with ONNX export and "save as XGBoost" support. (GitHub)
• BCG deploys 36,000+ custom GPTs for 32,000 consultants, fine-tuned per role with project memory and team sharing — essentially internal SaaS.
• AEGIS-FLOW: A community multi-agent cloud security framework (LangGraph + MCP + Next.js) that scans AWS configs, generates reports, and produces Terraform fix patches with mandatory human approval. (Demo)
• Lutum Veritas: An indie deep search engine claiming ~$0.20/query deep research with forced citations, BYO API keys, multi-model routing, and a verification ASK mode. (GitHub)

Models & Capabilities

• StepFun Step-3.5-Flash: 196B-parameter sparse MoE (~11B active), targeting long-context and agentic use. Officially reports 74.4% on SWE-bench Verified and 51% on Terminal-Bench 2.0; an Int4 quantized version runs 256K context on 128GB machines. Day-zero vLLM support available. (HF model page | Int4 version)
• Moonshot Kimi K2.5: Now the top open-source model (fifth overall) in LMArena's Code Arena; Perplexity has integrated it into Pro/Max via US-hosted nodes.
• GLM-4.7 Flash praised for interactive frontend coding: Common pattern is GLM for execution with Claude/Kimi for review.
• Claude Sonnet 5 leak rumors: Vertex AI logs show claude-sonnet-5 references; rumors claim 1M context, ~50% cheaper than Opus 4.5, TPU-optimized, and 80.9% on SWE-Bench — though the community remains skeptical.
• TII Falcon-H1-Tiny: Sub-100M parameter micro-models using anti-curriculum data injection, Muon optimizer, and hybrid Mamba+Attention. A 90M tool-calling model hits 94% relevance detection; a 600M reasoning version scores 75% on AIME24.
• Assistant_Pepe_8B: Fine-tuned on 4chan data from an NVIDIA Nemotron base; the author claims broad benchmark improvements, sparking discussion of unique statistical properties of 4chan text and the "alignment tax" on small models. (Model page)

Agents & Tooling

• Best practice: write tests first in CLAUDE.md — requiring reproduction tests before fixes is reported as the single most effective prompt for coding-agent reliability.
• "Conductor-style engineering": One developer driving 5–10 parallel agents; critics warn that context-switching costs degrade quality.
• OpenClaw/Moltbook ecosystem: Fun but risky — a security audit scored it 2/100, with demonstrated memory-file injection attacks; costs can also burn through OpenRouter credits quickly. (Audit report)
• LangChain releases deepagents (JS): Distills working systems like Claude Code and Manus into four reusable agent architecture patterns, with observability and evaluation tooling.
• Recursive Language Models (RLM) in practice: Community experiments use models to recursively plan, call tools, and audit codebases cheaply — though custom tooling and Deno sandboxing remain fiddly in DSPy.

Infrastructure & Hardware

• Memory, not compute, is the long-context bottleneck: An Imperial College + Microsoft survey shows a 1M-context, batch-size-1 request on DeepSeek-R1-class models can need ~900GB of KV-cache memory, motivating heterogeneous prefill/decode architectures. (dair.ai summary)
• FlashAttention v3 lands on AMD RDNA GPUs (PR).
• Triton-Viz 3.0: Visualizes per-load/store/matmul for Triton and Amazon NKI, with OOB detection and a Colab Triton-Puzzles integration.
• Blackwell (sm120) tuning notes: Correctly implemented TMA+mbarrier slightly beats cp.async for large matrices; cuBLAS still appears to use sm80-era kernels.
• MIT heat-powered chip: Performs matrix-vector multiplication using on-chip thermal gradients (currently only 2×2/3×3, ~99% accuracy) — far from practical AI inference.
• Fudan "sushi-roll" fiber chip (Nature): Rolls circuits onto elastic substrate, integrating ~100K transistors/cm along a meter of fiber; withstands 15.6 tons of pressure, repeated bending, and 100°C — aimed at smart fabrics and brain-computer interfaces.

Research & Methods

• Why coding agents work so well: Verifiability (compilers, tests) plus rich symbolic toolboxes (shell, debuggers) create a neuro-symbolic setup — replicating it elsewhere requires equivalent tooling + verification layers.
• Infinite synthetic RLVR data: Masking reasoning steps and generating distractors turns ordinary web pages into reasoning tasks, potentially reviving models saturated on existing RLVR data.
• Stop over-trusting perplexity: Better next-token prediction doesn't guarantee improvements in instruction following, tool-call stability, or multi-turn consistency.
• ConceptMoE: Routes and computes at the "concept" level (clusters of similar tokens) rather than per-token, reducing redundancy for long inputs.
• Token-level data filtering: Work from Alec Radford's collaborators proposes filtering and weighting at the token level during pretraining for finer control over learned capabilities.
• Brains vs. LLMs (Nature): The brain's temporal hierarchy in speech processing aligns with LLM layer depth — deeper layers correspond to later, higher-order language processing.

Safety & Governance

• OpenClaw red-team lessons: Attackers bypassed defenses via shell-expansion variables embedded in JSON metadata; long-term memory stored in .md files is the biggest attack surface — credential isolation and blast-radius control are essential. (Full report | Author notes)
• Prompt injection defense: Combine embedding-similarity filtering of inputs with grammar-constrained decoding to constrain model outputs.

Industry & Business

• Waymo reportedly raising $16B at ~$110B valuation, with at least $13B from Alphabet and participation from Sequoia, DST, and Dragoneer — up sharply 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.

📌 Source: Easy AI Daily