Easy AI Daily News — February 6, 2026

A translated and structured edition of the Easy AI daily briefing covering models, agents, infrastructure, research, products, industry news, and safety.

Models & Capabilities

• Claude Opus 4.6 released: Anthropic upgraded Opus with 1M-token context (beta), stronger planning and sustained execution. In an internal experiment, multi-agent Opus 4.6 wrote a clean-room C compiler over two weeks that compiles Linux 6.9, QEMU, and FFmpeg on x86/ARM/RISC-V with ~99% test pass rate. New SOTA on ARC-AGI2, SWE-Bench, TerminalBench and other leaderboards. (official release)
• GPT-5.3 Codex launched: OpenAI's new coding model offers small accuracy gains over 5.2-Codex on SWE-Bench Pro and TerminalBench 2, with ~2.1x fewer tokens and ~40% faster inference — roughly 3x efficiency at equal quality. Reportedly co-designed with GB200-NVL72 hardware. (announcement)
• Opus 4.6 vs GPT-5.3-Codex: Codex scored 65%+ on TerminalBench2 and is ~25% faster with fewer tokens; hands-on testing shows Codex stronger at speed and web/terminal automation, Opus better at complex architecture understanding and long code review. (comparison thread)
• Kling 3.0: Praised for cinematic, character-consistent multi-shot video (3–15s clips), but built-in audio quality is widely criticized. (examples)
• Mistral Voxtral Mini 4B Realtime: 13-language streaming transcription model with <500ms end-to-end latency (configurable 240ms–2.4s), ~12.5 token/s, Apache 2.0; lacks speaker diarization and turn detection. (model page)
• Google Sequential Attention: Marketed as leaner/faster attention via feature subset selection, but Redditors noted the paper dates to 2022 and "no accuracy loss" likely means metric parity, not numerical equivalence.
• BalatroBench: A strategy benchmark for LLMs using the game Balatro, with BalatroBot mod and BalatroLLM framework supporting any OpenAI-compatible model via Jinja2 policy templates. (site)

Agents & Tooling

• OpenAI Frontier: An enterprise "AI coworker" platform unifying company context, tool/code execution, identity permissions, and on-the-job learning; internal mandate for agent-first workflows by end of March. (blog post)
• Multi-agent dev tools: GitHub Copilot CLI adds "Fleets" (spawn agent teams, track dependencies via local SQLite); VS Code positions itself as a multi-agent hub; Claude Code adds agent teams and /insights session analysis.
• OpenClaw / Moltbot criticism: Reddit teardown shows it's mainly an orchestrator requiring paid Anthropic/OpenAI/Google APIs, Brave Search, TTS, and Playwright — $50–100/month is common, far from "truly local."
• Ollama controversy: The llama.cpp author suggested Ollama is just a daemonized wrapper of his project, sparking attribution debates; defenders cite Ollama's on-demand model loading and unified local API.
• Game sandboxes for agents: BalatroBench and Spacemolt (a large-scale agent MMO) offer controllable, scoreable environments more realistic than single-shot QA.

Infrastructure & Hardware

• NVIDIA Blackwell FP8 concerns: GPU MODE community found FP8 Tensor performance can differ 2x on identical Blackwell cards; cuBLASLt may silently pick older Ada kernels, with ~1.5x gains only via new MXFP8 instructions. (analysis)
• Vulkan beats CUDA in some cases: 20–50% faster on GPT-OSS-20B local inference (116–117 token/s single card), possibly due to lower API overhead and better CPU/GPU staging.
• TorchAO FP8 in production: Karpathy adopted torchao FP8 training in nanochat, a vote of confidence for PyTorch's official quantization stack.
• Cluster stack debates: RoCE vs InfiniBand comparisons; on AMD, Helion auto-tuned Triton kernels were slower than torch inductor — the official stack still wins for most.

Research & Methods

• Meta's SALE: Agent routing as a "policy auction" — candidates submit short plans, peer-evaluate cost-effectiveness; cuts cost 25–35% at equal or slightly better scores.
• Agent Primitives: Multi-agent communication via KV cache instead of natural language: +12–16% accuracy over single agents on 8 benchmarks with 3–4x fewer tokens.
• TinyLoRA: Updating only 13 parameters (plus RL) reportedly lifts Qwen-7B GSM8K from 76% to 91%.
• MaxRL / log-prob rewards: An interpolated REINFORCE–MLE objective (one-line advantage normalization change) claims better sample efficiency than GRPO; log-prob rewards unify verifiable and non-verifiable tasks.
• SIEVE: Distills a few natural-language feedback rules into weights efficiently, often beating in-context learning.
• OVQ-attention (Zyphra): Finer vector quantization trade-offs for long-context compression; Privasis: 1.4M synthetic privacy records plus a 4B cleaning model for on-device redaction of sensitive fields.

Products & Applications

• Local LLM coding: Community debates whether ~$20k machines running Qwen Coder / GLM 4.7 via LM Studio/Ollama approach Claude Sonnet-level coding; hybrid local+cloud setups suggested.
• TrueShort: AI film studio + streaming app combo reaching ~$2.4M ARR in 6 months, 2M+ minutes watched, App Store News Top 10.
• Lotus: $41M raised for an "AI family doctor" — model triage and paperwork, licensed doctors diagnose/prescribe.
• Model picks: Kimi praised for practical scripting (Selenium/BS4) and resume automation; Claude preferred for writing and deep reasoning; Gemini used as a free web-connected research tool.
• Agent workflows in practice: Cursor+OpenClaw self-improving agent loops; Claude Code agent teams for refactors; Opus 4.6 described as "the first model like a real colleague," though long-horizon tasks still suffer from overconfident failures.

Industry & Business

• Claude Code at ~4% of GitHub commits: SemiAnalysis data suggests authorship doubled from 2% in a month; could reach 20% of open-source commits by end of 2026 (methodology caveats apply).
• Goodfire AI: $150M Series B at $1.25B valuation, focused on interpretability and controllable system design as the core product.
• OpenAI x Ginkgo Bioworks: GPT-5 connected to real wet labs, proposing and iterating experiments autonomously; reportedly cuts protein production costs 40%. (blog post)
• OpenAI vs Anthropic: Same-day coding model releases, dueling Super Bowl-era ads, enterprise platform competition (Frontier vs Anthropic knowledge-work plugins), and SaaS stock pressure.

Policy, Governance & Safety

• Claude Code jailbreak: The BASI community's "ENI Hooks" instruction set bypassed safeguards to generate polymorphic ransomware with obfuscation, process injection, registry hijacking, keyloggers, and wallet hijacking — highlighting risks when models get filesystem + tool access with weak system prompts.
• Jailbreak difficulty: DeepSeek rated very easy to jailbreak with older prompts; Grok popular among unrestricted-model fans; Gemini comparatively resistant on illegal/violent content.
• secureai-scan: Hugging Face community tool that locally scans repos for unauthenticated LLM calls, sensitive-data leakage, and prompt-injection risks, generating HTML/Markdown/JSON reports. (paper)
• Labor market debate: Predictions that agents will absorb white-collar form/memo work within two years vs. Chollet's translation-industry counterexample — headcount roughly unchanged post-automation, but rates fell and freelancers were squeezed.
• AGI definition: Andrew Ng argues "AGI" has become an ill-defined buzzword; by the original "any human intellectual task" standard, he estimates decades remain.

📌 Source: Easy AI Daily