Easy AI Daily News Digest - February 3, 2026: OpenAI Codex App, StepFun Step-3.5-Flash, Kimi K2.5, and More

📅 AI Industry Digest — February 3, 2026

Products & Applications

• OpenAI launches standalone Codex App for macOS: Integrates multi-agent parallel execution, per-task git worktrees, a /plan mode, reusable Skills, and scheduled Automations into a single "command center." Positioned as a code-agent workbench rather than a VSCode plugin.
• Links: Official intro | Codex product page | X announcement | OpenAIDevs thread
• Windsurf Wave 14 adds Arena mode: Compare multiple models side-by-side on the same coding task, with Battle Groups set to 0x credit cost for a week. Personal and public leaderboards encourage picking models via real tasks rather than benchmarks.
• Links: Download page | Author's post
• LM Studio 0.4.1 speaks Anthropic protocol: A local /v1/messages-compatible endpoint lets Claude Code swap its backend to local GGUF/MLX models by just changing the base URL. OpenAI-compatible endpoints and a TypeScript SDK are also provided.
• Links: Blog post | OpenAI-compatible SDK
• PerpetualBooster v1.1.2: Rust-based GBM replacing hyperparameter tuning with a single "budget" parameter; claims up to 100x faster training than LightGBM+Optuna with comparable accuracy. Supports R, ONNX export, and saving as XGBoost.
• Links: GitHub | Reddit discussion
• BCG deploys 36,000+ custom GPTs for 32,000 consultants, role- and methodology-specific, with project memory and team sharing — effectively internal SaaS. Commentary notes most companies remain at scattered-trial stages.
• Link: Reddit discussion
• AEGIS-FLOW: Community multi-agent cloud-security framework (LangGraph + MCP + Next.js) that scans AWS configs, generates reports, and produces Terraform fix patches with a mandatory human approval gate.
• Link: Live demo (may be unstable)
• Lutum Veritas: Indie deep-search engine claiming deeper retrieval/synthesis than ChatGPT/Gemini/Perplexity at ~$0.20/query, with BYO API keys, multi-model routing, forced citations, and a verification "ASK mode."
• Link: GitHub

Models & Capabilities

• StepFun Step-3.5-Flash: 196B-total / ~11B-active sparse MoE focused on long context and agentic use. Official scores: 74.4% SWE-bench Verified, 51% Terminal-Bench 2.0. An Int4 quant runs 256K context on 128GB locally; vLLM day-zero support.
• Links: Release thread | vLLM support | HF FP | HF Int4 | llama.cpp PR
• Moonshot Kimi K2.5: Ranks #1 among open-source models and #5 overall on LMArena's Code Arena. Perplexity has integrated K2.5 into Pro/Max tiers, self-hosting US endpoints for latency and reliability.
• Links: Code Arena leaderboard | Perplexity announcement screenshot
• GLM-4.7 Flash praised for frontend/interactive web coding: Community suggests pairing it with a high-end reviewer model (Claude/Kimi) for a cost-effective execute-and-review combo.
• Link: ggerganov's take on GLM-4.7
• Claude Sonnet 5 "leaks": Vertex AI 404s referencing claude-sonnet-5 fueled rumors of 1M context, ~50% cheaper than Opus 4.5, TPU-optimized throughput, and 80.9% SWE-Bench. Community remains skeptical — IDs often reflect creation dates, and long-context accuracy gains are unverified.
• Links: Reddit thread 1 | Reddit thread 2
• TII Falcon-H1-Tiny (<100M params): "Anti-curriculum" micro-models trained with target-domain data from step one, Muon optimizer, and Hybrid Mamba+Attention. The 90M tool-call model hits 94% relevance detection; the 600M reasoning variant scores 75% on AIME24.
• Link: Reddit intro
• Assistant_Pepe_8B: Fine-tuned from NVIDIA Nemotron on expanded 4chan data; the author claims across-the-board eval improvements over base. Discussion suggests 4chan text's distinctive statistics may yield surprising gains, and that the "alignment tax" hits small models harder.
• Links: Model page | Reddit discussion

Agents & Tooling

• "Write the test first" in CLAUDE.md: Practitioners report that requiring reproduce-test → fix → verify via tests is the single most effective prompt for coding-agent reliability.
• Link: Experience post
• "Conductor-style engineering": One developer driving 5–10 parallel agents, rarely reading all the code. Counterpoint: human context-switching costs mean quality collapses with too many parallel threads.
• Links: Pro | Con
• OpenClaw ecosystem: fun but risky: Security assessment scored 2/100; red-team drills showed memory files can be injected for indirect execution, and API credits can burn fast on OpenRouter.
• Links: Security audit | Cost/safety discussion
• LangChain deepagents: Distills Claude Code/Manus-style systems into four reliable architecture patterns, with observability and evaluation tooling, replacing naive one-model-many-tools designs.
• Link: LangChain JS announcement
• Recursive Language Models in practice: Used for code security audits where the model plans, calls tools and sub-models, drilling into codebases — fast and cheap with models like Kimi K2, though custom tool-calling and Deno sandbox permissions remain fiddly in DSPy.
• Link: RLM audit example

Infrastructure & Hardware

• Long-context inference is memory-bound, not compute-bound: An Imperial College + Microsoft survey finds 1M-context, batch-1 requests (e.g., DeepSeek-R1) can need ~900GB of KV cache; prefill/decode may need heterogeneous memory-intensive vs compute-intensive hardware.
• Link: dair.ai summary
• FlashAttention v3 lands on AMD RDNA GPUs, enabling efficient self-attention kernels beyond NVIDIA for local long-context inference.
• Link: PR #2178
• Triton-Viz 3.0: Visualizes every load/store/matmul in Triton and Amazon NKI kernels, with OOB detection, inefficient-loop profiling, and Triton-Puzzles Colab integration.
• Links: Discord release post | Triton-Puzzles repo
• Blackwell (sm120) micro-benchmarks: Correctly implemented TMA+mbarrier slightly beats cp.async on large shapes, while cuBLAS appears to still use sm80-era kernels; a single missing __syncthreads() can deadlock the kernel.
• Links: GPU Mode thread 1 | GPU Mode thread 2
• MIT heat-powered matrix-multiply chip: Uses on-die temperature gradients as the compute medium; currently limited to 2×2/3×3 matrices with ~99% accuracy — far from practical AI inference without strong error correction.
• Link: Reddit discussion
• Fudan "sushi-roll" flexible fiber chip (published in Nature): Circuits rolled onto an elastic substrate achieve CPU-class transistor density per meter of fiber, tolerating 15.6-ton compression, repeated bending, and 100°C — aimed at smart fabrics and brain-computer interfaces. Community questions the "thinner than hair" claim and latency over one-meter lengths.
• Link: Reddit discussion

Research & Methods

• Why coding agents work so well: Verifiability (compilers, unit tests, runtime errors) plus a rich symbolic toolbox (shell, debugger) — replicating this elsewhere requires building equivalent tooling + verification layers.
• Link: random_walker thread
• Inexhaustible synthetic RLVR: Masking reasoning steps and generating distractors converts ordinary web pages into RLVR-style reasoning tasks, reviving saturated models, with strong results in security tasks. Community view: synthetic data is an engineering recipe problem, not inevitable collapse.
• Links: RLVR thread | Synthetic pretraining essay
• Stop over-indexing on perplexity: Better next-token prediction doesn't guarantee improvements in instruction following, tool-calling stability, or multi-turn consistency.
• Links: Critique 1 | Critique 2
• ConceptMoE: Clusters similar tokens into "concept units" before MoE routing, cutting redundant computation for long documents and large batches.
• Link: Paper intro
• Token-level data filtering: Work with Alec Radford's collaboration proposes token-level filtering/weighting in pretraining for fine-grained control over learned capabilities.
• Link: neil_rathi thread
• 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.
• Links: Paper | Popular article

Policy, Governance & Safety

• OpenClaw red-team findings: After social-engineering and pipeline RCE attempts were blocked, the attacking agent 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 containment are essential.
• Links: Full report | Author's notes
• Prompt-injection defense combo: Embedding-similarity filtering to intercept malicious inputs, plus grammar-constrained decoding to restrict model output structure.
• Link: Adversarial Design Thinking exercises

Industry & Business

• Waymo reportedly raising $16B at up to $110B valuation: At least $13B from Alphabet, with Sequoia, DST, and Dragoneer participating — a big step up from $45B in October 2024.
• Link: Related discussion
• xAI Grok Imagine 1.0: Generates 10-second 720p video with audio; over 1.2 billion videos created in the past 30 days.
• Link: Announcement

📌 Source: Easy AI Daily