Easy AI Daily | 2026-02-06

A digest of AI industry news for February 6, 2026, covering models, agent tooling, infrastructure, research, products, and policy.

Models & Capabilities

• Claude Opus 4.6 released: Anthropic upgraded Opus to 4.6, featuring 1M-token context (beta) and stronger planning/persistent execution. In an internal experiment, a multi-agent Opus 4.6 setup ran for two weeks and wrote a clean-room C compiler capable of building Linux 6.9, QEMU, and FFmpeg on x86/ARM/RISC-V with ~99% test pass rates. It sets new SOTA on benchmarks including ARC-AGI2, SWE-Bench, and TerminalBench. Official release | C compiler thread
• GPT-5.3 Codex launched: OpenAI's new coding model improves accuracy slightly over 5.2-Codex on SWE-Bench Pro and TerminalBench 2 while cutting token usage ~2.1x and accelerating inference ~40% — roughly 3x efficiency at equal quality. Reportedly co-designed with GB200-NVL72 hardware. Release
• Opus 4.6 vs GPT-5.3-Codex head-to-head: Both shipped the same day. Codex scores 65%+ on TerminalBench 2 and is ~25% faster with cheaper tokens; hands-on comparisons favor Codex for speed and web/terminal automation, Opus for architectural understanding and long code reviews. Comparison test
• Kling 3.0: Widely praised for cinematic multi-shot consistency across 3–15 second clips, but its built-in audio remains a common complaint. Community also suspects some third-party platforms are repackaging Kling. Examples
• Mistral Voxtral Mini 4B Realtime: Open-source (Apache 2.0) streaming speech transcription model, 13 languages, <500ms end-to-end latency (configurable 240ms–2.4s), ~12.5 token/s. Lacks speaker diarization and turn detection. Model card
• Google Sequential Attention: Pitched as leaner, faster attention without accuracy loss, but Redditors note the underlying paper dates to 2022, and "no accuracy drop" likely means metric parity, not numerical equivalence. Blog
• BalatroBench: A strategy benchmark using the game Balatro, with a BalatroBot mod and BalatroLLM framework compatible with any OpenAI-compatible API. Skeptics note newer models may have seen game guides. Project | GitHub

Agents & Tooling

• OpenAI Frontier: An enterprise "AI coworker" platform unifying context, tool/code execution, identity/permissions, and on-the-job learning. Internally, OpenAI is requiring teams to prefer agents for many tasks by end of March, with AGENTS.md, tool inventories, and MCP/CLI interfaces. Announcement
• Multi-agent dev tools: GitHub Copilot CLI adds "Fleets" (spin up sub-agent teams with local SQLite tracking); VS Code positions itself as a multi-agent hub for OpenAI/Anthropic cloud agents and local background agents; Claude Code adds agent teams and /insights session analysis that updates CLAUDE.md.
• OpenClaw / Moltbot criticized: Community teardown shows it's mainly an orchestrator requiring paid Anthropic/OpenAI/Google models, Brave Search, TTS services, and Playwright — $50–100/month is common — far from its "fully local" marketing. Critique
• Ollama controversy: The llama.cpp author suggested Ollama is just his project daemonized with a model jukebox wrapper, "bugs included," sparking attribution debates; supporters cite Ollama's on-demand model loading and unified local API.

Infrastructure & Hardware

• NVIDIA FP8 gatekeeping on Blackwell: GPU MODE members found identical Blackwell cards showing 2x-different FP8 Tensor performance; cuBLASLt may silently pick older Ada kernels, with ~1.5x gains only via new MXFP8 instructions. Analysis
• Vulkan beats CUDA in some inference: Reports of 20–50% speedups with GPT-OSS-20B on LM Studio, hitting 116–117 token/s on a single card, possibly due to lower API overhead and better CPU/GPU staging.
• Karpathy adopts FP8: nanochat now uses TorchAO FP8 training, lending credibility to PyTorch's official quantization stack. Commit
• Cluster stack debates: RoCE vs InfiniBand comparisons resurface; Helion-tuned Triton kernels measured slower than TorchInductor on AMD, suggesting sticking with official stacks first.

Research & Methods

• SALE & Agent Primitives: Meta's SALE routes agents via a "policy auction" (short plans, mutual cost/value scoring), cutting costs 25–35% at equal or better scores. Agent Primitives has agents communicate via KV cache instead of natural language, gaining 12–16% accuracy over single agents on 8 benchmarks with 3–4x fewer tokens.
• TinyLoRA: A PhD thesis project claims tuning just 13 parameters with RL on Qwen-7B lifts GSM8K from 76% to 91% — pending replication, a striking data point for ultra-low-rank fine-tuning.
• MaxRL & log-prob rewards: MaxRL interpolates between REINFORCE and maximum likelihood (a one-line advantage-normalization change), claiming to beat GRPO on reasoning tasks; another work unifies verifiable and non-verifiable tasks via log-prob rewards.
• SIEVE: Distills a few natural-language rules/feedback directly into weights via efficient continual learning, often beating in-context learning — useful for converting long prompts into training signal.
• OVQ-attention & Privasis: Zyphra's OVQ-attention trades compression vs compute for long contexts via vector quantization; Privasis releases 1.4M synthetic privacy records plus a 4B cleaning model for on-device redaction of sensitive fields.

Products & Applications

• Local LLM coding debate: With Qwen Coder and GLM 4.7 via LM Studio/Ollama, consumer GPUs can approach Claude Sonnet-level coding — trading half-a-year-lagging models and ~$20K hardware for privacy and control. Hybrid local+cloud setups are popular.
• TrueShort: AI film studio + streaming app combo reporting ~$2.4M annualized revenue in 6 months, 2M+ minutes watched, and App Store News Top 10 placement.
• Lotus: Raised $41M for an "AI family doctor" combining model triage with licensed physicians for diagnosis, prescriptions, and referrals, targeting US primary-care shortages.
• Model strengths in practice: Kimi praised for Selenium/BS4 code logic and CLI+Telegram automation; Claude preferred for writing and deep reasoning; Gemini for free web-connected research.
• Agent workflows: Developers pair Cursor + OpenClaw for self-improving loops (one agent builds, another critiques); Claude Code agent teams for complex refactors; Opus 4.6 described as "the first model that feels like a real colleague" — though long-horizon tasks still expose overconfident behavior.

Industry & Companies

• ~4% of GitHub commits from Claude Code: Per SemiAnalysis, up from 2% a month ago; if the trend holds, 1 in 5 open-source commits could be agent-authored by end of 2026 (with statistical caveats).
• Goodfire AI raised $150M at a $1.25B valuation, betting on interpretability and controllability as the product, not bigger models.
• OpenAI x Ginkgo Bioworks: GPT-5 drives a real wet lab end-to-end — proposing experiments, issuing instructions, reading results — reportedly cutting protein production costs 40%. Post
• OpenAI vs Anthropic three-front war: Same-day coding releases, dueling ad campaigns (Anthropic's Super Bowl ad mocking ad-cluttered AI drew an Altman rebuttal), and enterprise platform competition — with SaaS stocks taking a hit.

Policy, Governance & Safety

• Claude Code jailbroken into ransomware: The BASI community's "ENI Hooks" prompt set bypassed safeguards to generate polymorphic ransomware with process injection, registry hijacking, keyloggers, and wallet hijacking — highlighting risks of tool-and-filesystem access with weak system prompts.
• Jailbreak difficulty varies: DeepSeek rated very easy to jailbreak with old prompts; Grok popular among unrestricted-model fans; Gemini comparatively resistant on illegal/violent content.
• secureai-scan: A Hugging Face community tool that scans repos locally for unauthenticated LLM calls, sensitive-data leakage to models, and prompt-injection risks, producing HTML/Markdown/JSON reports. Paper
• AI labor debate: One camp predicts agents absorb white-collar busywork within two years; Chollet counters with the translation industry, where automation left headcount mostly unchanged but shifted work to proofreading, lowered rates, and squeezed freelancers.
• AGI definition fatigue: 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