Easy AI Daily | February 3, 2026

Products and Applications

• OpenAI ships standalone Codex App for macOS: No longer a VSCode plugin — it integrates multi-agent parallelism, per-task git worktrees, a /plan mode, reusable Skills, and scheduled Automations into a single "command center" for developers. Announcement | Codex page
• Windsurf Wave 14 adds Arena mode: Compare multiple models side-by-side on the same coding task; Battle Groups temporarily at 0x credit cost, with personal and public leaderboards encouraging model selection on real tasks. Download page
• LM Studio 0.4.1 adds Anthropic-compatible endpoint: Tools like Claude Code can switch to local GGUF/MLX models by just changing the base URL; OpenAI-compatible endpoint and TypeScript SDK also provided. Blog post
• PerpetualBooster v1.1.2 (Rust): Gradient boosting with a single "budget" parameter instead of many hyperparameters; claims up to 100x faster than LightGBM+Optuna at similar accuracy, with ONNX export and XGBoost save compatibility. GitHub
• BCG deploys 36,000+ custom GPTs for 32,000 consultants, tuned by role and methodology with project memory and team sharing — effectively an internal SaaS layer.
• AEGIS-FLOW: Community multi-agent cloud security framework (LangGraph + MCP + Next.js) that audits AWS configs, reports findings, and generates Terraform fix patches with human approval before changes. Demo
• Lutum Veritas: Independent deep search engine claiming deeper research than ChatGPT/Gemini/Perplexity at ~$0.20/query, with BYO API keys, mandatory citations, and an ASK mode for verifying claims. GitHub

Models and Capabilities

• StepFun Step-3.5-Flash: 196B-parameter sparse MoE (~11B active) focused on long context and agentic coding; reported 74.4% on SWE-bench Verified and 51% on Terminal-Bench 2.0. Day-zero vLLM support; Int4 version runs 256K context on 128GB machines. Hugging Face
• Moonshot Kimi K2.5: Ranks #1 among open models and #5 overall on LMArena's Code Arena; Perplexity has integrated it into Pro/Max tiers hosted on US nodes.
• GLM-4.7 Flash: Community feedback highlights strong frontend/interactive web coding with preserved reasoning; a common setup pairs it with Claude/Kimi for review.
• Claude Sonnet 5 rumors: Vertex AI logs show claude-sonnet-5 identifiers; rumors claim 1M context, ~50% cheaper than Opus 4.5, TPU-optimized, 80.9% SWE-Bench — but the community urges caution, noting IDs often reflect creation date, not release date.
• TII Falcon-H1-Tiny: Sub-100M specialized micro-models using anti-curriculum data injection, the Muon optimizer, and Hybrid Mamba+Attention; a 90M tool-calling model achieves 94% relevance detection, a 600M reasoning model hits 75% on AIME24.
• Assistant_Pepe_8B: Fine-tuning NVIDIA Nemotron on expanded 4chan data reportedly beats the base model across benchmarks, sparking discussion about the statistical uniqueness of the corpus and "alignment tax" on small models. Model

Agents and Tooling

• Write tests first in CLAUDE.md: Practitioners report that instructing coding agents to write a reproduction test before fixing bugs is the single most effective reliability prompt.
• "Conductor-style" engineering: One developer driving 5–10 parallel agents; skeptics warn that human context-switching costs degrade quality at high parallelism.
• OpenClaw security: Automated red-team audit scored its defenses 2/100; memory files (.md) are the biggest attack surface via injection. Cost can also burn OpenRouter credits rapidly. Audit report
• LangChain deepagents: Distills working systems (Claude Code, Manus) into four architecture patterns, replacing naive "one model calling many tools" designs.
• Recursive Language Models in practice: Used for code security auditing with model-driven planning and sub-model calls; custom tool calling and Deno sandbox permissions remain fiddly in DSPy.

Infrastructure and Hardware

• Memory, not compute, is the long-context bottleneck: A review from Imperial College and Microsoft notes KV-cache for a 1M-token, batch-1 request can approach ~900GB, favoring heterogeneous prefill/decode architectures.
• FlashAttention v3 lands on AMD RDNA GPUs via a merged PR, enabling efficient attention beyond NVIDIA. PR #2178
• Triton-Viz 3.0: Visualizes every load/store/matmul with OOB detection and loop profiling; supports Triton and Amazon NKI.
• Blackwell (sm120) tuning findings: Correctly implemented TMA+mbarrier slightly beats cp.async for large shapes; cuBLAS still appears to use sm80-era kernels; a missing __syncthreads() can deadlock kernels.
• MIT heat-powered chip: Performs matrix-vector multiplication using on-die temperature gradients; currently limited to 2×2/3×3 matrices at ~99% accuracy.
• Fudan flexible fiber chip (Nature): A "sushi-roll" structure packs ~100K transistors/cm in hair-thin fiber, withstanding 15.6 tons of pressure, repeated bending, and 100°C — targeting smart fabrics and brain-computer interfaces, though latency over one-meter lengths is questioned.

Research and Methods

• Why coding agents work: Verifiability (compilers, unit tests) plus a rich symbolic toolbox (shell, debuggers) creates a neuro-symbolic loop; replicating it elsewhere requires building equivalent tool + verification layers.
• Infinite synthetic RLVR tasks: Masking reasoning steps and generating distractor options converts web text into reasoning training data, reviving models saturated on existing RLVR sets.
• Perplexity isn't enough: Better next-token prediction doesn't guarantee gains in instruction following, tool-call stability, or multi-turn consistency.
• ConceptMoE: Clusters tokens into "concept units" before MoE routing to reduce redundant computation for long inputs.
• Token-level data filtering: Work associated with Alec Radford proposes filtering and weighting at the token level during pretraining rather than adding/removing whole datasets.
• Brains vs LLMs (Nature): Hierarchical timing of human speech processing aligns with LLM layer depth — deeper layers correspond to later, higher-order language cortex activity. Paper

Policy, Governance, and Safety

• OpenClaw red team findings: After social engineering and pipeline RCE were blocked, the attacker injected shell-expanded variables in JSON metadata; long-term memory stored in .md files propagates injected instructions to all future behavior — isolate credentials and blast radius as with prompt injection.
• Prompt injection defenses: Combining embedding-similarity input filtering with grammar-constrained decoding to structurally limit what models can output.

Industry and 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 reportedly created in the past 30 days.

📌 Source: Easy AI Daily