Easy AI Daily | February 6, 2026

A roundup of AI industry news from the Chinese tech community zhichai.net, covering model releases, agent tooling, infrastructure, research, products, and policy.

Models & Capabilities

Claude Opus 4.6 Released: 1M Context + C Compiler Experiment

Anthropic upgraded Opus to 4.6, headlining 1M-token context (beta) with stronger planning and sustained execution. In an internal experiment, a multi-agent Opus 4.6 "free-range" setup spent two weeks writing a clean-room C compiler that compiles Linux 6.9, QEMU, and FFmpeg on x86/ARM/RISC-V with roughly 99% test pass rate. The model pushes SOTA on ARC-AGI2, SWE-Bench, TerminalBench, and more.

• Official release | C compiler experiment thread | ARC-AGI2 score discussion | Vals leaderboard

GPT-5.3 Codex: Faster, Cheaper Coding Model

OpenAI launched GPT-5.3-Codex emphasizing "build things right away." It posts modest accuracy gains over 5.2-Codex on SWE-Bench Pro and TerminalBench 2 while using ~2.1x fewer tokens and ~40% faster inference—roughly 3x efficiency at equal quality. Officially described as co-designed with GB200-NVL72 hardware.
• Model release | Performance analysis

Claude Opus 4.6 vs GPT-5.3-Codex: Coding SOTA Slugfest

Both vendors shipped minor versions the same day. Opus 4.6 tops ARC-AGI2, SWE-Bench, long-context, and agent-team benchmarks; GPT-5.3-Codex scores 65%+ on TerminalBench 2, is ~25% faster, and uses fewer tokens. Hands-on comparisons suggest Codex wins on speed and web/terminal automation, while Opus is more thorough on complex architecture understanding and long code review.
• Swift large-project comparison | Benchmark debate

Kling 3.0: Stunning Video Consistency, Audio Still Weak

Kling 3.0 is widely shared on Reddit for high visual quality: consistent characters across shots, cinematic framing and color grading, and 3–15 second multi-shot clips. Built-in audio remains the most criticized weakness, and the community questions third-party platforms reselling Kling's capabilities.
• Official examples | Way of Kings fan trailer

Mistral Voxtral Mini 4B Realtime: Low-Latency Multilingual ASR

Mistral released Voxtral-Mini-4B-Realtime-2602 on Hugging Face: 13 languages, native streaming architecture, <500ms end-to-end latency, configurable 240ms–2.4s output latency, ~12.5 token/s throughput—suited to local voice assistants and subtitles. Apache 2.0 licensed, but lacks speaker diarization and automatic turn detection.
• Model page | Community discussion

Google Sequential Attention: Leaner, Faster Attention?

Google Research promotes Sequential Attention for efficient feature subset selection claiming "leaner, faster models without sacrificing accuracy." Redditors note the paper dates from 2022, calling this more of a marketing re-share; "no accuracy loss" likely means metrics held steady, not bitwise equivalence.
• Research blog | arXiv paper | Reddit discussion

BalatroBench: Testing LLM Strategy with a Card Roguelike

BalatroBench uses the game Balatro as a strategy benchmark, with the BalatroBot mod and BalatroLLM framework supporting any OpenAI-compatible model and Jinja2 strategy templates for different "decision philosophies." Suggestions include evolutionary algorithms for self-consistent iteration; critics note newer models may have seen Balatro guides (2024 game).
• Project page | GitHub BalatroBot | Reddit discussion

Agents & Tooling

OpenAI Frontier: Official Enterprise Agent Platform

OpenAI's Frontier turns "AI coworkers" into a platform: unified enterprise context, tool and code execution environments, identity/permissions, and on-the-job learning. Internally, technical teams must "prefer agents" for many tasks by end of March, with AGENTS.md conventions, tool manifests, and MCP/CLI interfaces—arguably the first major vendor's public agent deployment playbook.
• Frontier post | gdb on internal process

Multi-Agent Dev Tool Wars: VS Code, Copilot CLI, Claude Code

GitHub added "Fleets" to Copilot CLI, spinning up teams of subagents tracked via local SQLite for dependencies and TODOs; VS Code positions itself as a multi-agent hub orchestrating OpenAI/Anthropic cloud agents and local background agents. Claude Code offers agent teams and /insights session analysis that auto-updates CLAUDE.md.
• Copilot Fleets | VS Code multi-agent vision | Claude Code agent teams

OpenClaw / Moltbot: "Local Agent" or Subscription Bundle Shell?

Reddit teardowns show Moltbot is essentially an orchestrator requiring users to supply Anthropic/OpenAI/Google models, Brave Search, ElevenLabs or OpenAI TTS, and Playwright—$50–100/month is common. Workflows are novel with self-improving loops, but far from the "truly local, no cloud" pitch.
• Reddit critique | OpenClaw site

Is Ollama Just a Daemonized llama.cpp?

The llama.cpp author suggested on GitHub that Ollama merely daemonizes his project and wraps a model jukebox—even copying bugs—sparking criticism over uncredited code. Supporters counter that Ollama's value is on-demand model loading and a unified local API for switching models in terminals and tools.
• Reddit controversy

BalatroBench / Spacemolt: Game Sandboxes for LLMs

BalatroBench benchmarks strategy via a card game; Spacemolt is a large-scale agent MMO for long-horizon interaction. Both give agents a controllable, scoreable, sufficiently complex playground closer to real usage than single Q&A or small scripts.
• BalatroBench | Spacemolt MMO

Infrastructure & Hardware

NVIDIA Accused of "Quietly Halving" FP8 Performance on Blackwell

GPU MODE community testing found identical Blackwell cards showing up to 2x FP8 Tensor performance differences. cuBLASLt apparently silently selects older Ada kernels; only MXFP8 instructions yield ~1.5x gains. Suspected driver/firmware-level gating, with long analyses on HN and Reddit.
• Gatekeeping analysis | Reddit discussion

Vulkan Beats CUDA on Some Inference Tasks

Local LLM users report Vulkan compute running 20–50% faster than CUDA on models like GPT-OSS-20B, hitting 116–117 token/s single-card. Speculation: lower Vulkan API overhead and better CPU/GPU staged coordination versus CUDA paths that default to running everything on GPU.
• LM Studio hardware discussion

TorchAO / FP8 Training Gets Real-World Validation from Karpathy

Karpathy adopted torchao FP8 training in nanochat—an endorsement that PyTorch's official quantization stack is production-viable, not a paper toy. Good news for anyone saving VRAM and power.
• nanochat torchao commit

RoCE vs InfiniBand, Helion vs TorchInductor: Picking a Cluster Stack

GPU MODE threads revisit RoCE vs InfiniBand; on AMD, auto-tuned Helion Triton kernels were measured slower than torch inductor—suggesting the "new framework wins everything" era isn't here yet. Most advice: exhaust the official stack first, then layer niche solutions incrementally.
• RoCE vs IB overview

Research & Methods

SALE & Agent Primitives: Multi-Agent Without Group Chat

Meta's SALE turns agent routing into a "strategy auction": candidate agents submit short plans, evaluate each other's cost-effectiveness, then one is dispatched—cutting costs 25–35% while maintaining or slightly improving scores. Agent Primitives lets multi-agent systems communicate via KV cache instead of natural language, gaining 12–16% accuracy over single agents across 8 benchmarks while using 3–4x fewer tokens than traditional multi-agent setups.
• SALE intro | Agent Primitives intro

TinyLoRA: GSM8K to 91% by Tuning Only 13 Parameters?

A PhD thesis project proposes TinyLoRA + RL, updating only 13 parameters on Qwen-7B to lift GSM8K from 76% to 91%. If reproducible, it suggests huge headroom for extreme low-rank fine-tuning—appealing for edge and enterprise tuning.
• Author's tweet

MaxRL / Log-Prob Rewards: RL Objectives Converging Toward MLE

MaxRL interpolates between REINFORCE and maximum likelihood—a one-line change normalizing advantages by mean return—claiming better sample efficiency and final performance than GRPO on reasoning tasks. Related work uses log-prob rewards to unify verifiable and non-verifiable tasks under a "prediction loss" view.
• MaxRL intro | Log-prob reward discussion

SIEVE: Distilling Natural-Language Instructions into Weights

SIEVE demonstrates efficient continual learning: a few natural-language feedback rules—no large datasets—distilled into model weights, outperforming in-context learning on many tasks. Ideal for converting long prompts into training data.
• SIEVE intro

Long Context & Privacy: OVQ-Attention and Privasis Dataset

Zyphra's OVQ-attention uses finer vector quantization to trade compression rate against compute in long contexts. Privasis releases 1.4M synthetic privacy records and a 4B cleaning model claiming on-device sensitive-field removal outperforming o3 and GPT-5—useful as a local privacy proxy in front of LLMs.
• OVQ-attention | Privasis-Cleaner

Products & Applications

Local LLM Coding: Worth a $20K Machine?

Reddit debates using local models like Qwen Coder and GLM 4.7 with LM Studio/Ollama on consumer GPUs for a coding experience near Claude Sonnet. Pros: privacy, control. Cons: models lag ~6 months and hardware is a one-time investment. Hybrid suggestions: local for common tasks, cloud for hard problems.
• Is local LLM coding worth it | Fully offline experiences

TrueShort: AI Film Studio + Streaming Hits $2.4M ARR in Six Months

TrueShort bundles an AI film studio with a streaming app: ~$2.4M annualized revenue, 2M+ minutes watched, and a top-10 App Store news ranking. Model-generated shorts plus app-based distribution and subscriptions.
• TrueShort results

Lotus Raises $41M for an "AI Family Doctor"

Lotus aims to address primary care shortages for 100M Americans via AI triage and paperwork with licensed physicians diagnosing, prescribing, and referring. Community sentiment has shifted from "too dangerous" to "genuinely needed if guardrails hold."
• Lotus launch

Kimi, Claude, Gemini: Each Shines Differently

Kimi users automate resumes and job emails via CLI + Telegram and report stronger practical code logic (Selenium/BS4) than Gemini; Claude remains preferred for writing and deep reasoning; Gemini serves as a free, web-connected research tool.
• Kimi user experience

AI Coding Agents in Practice: Claude Code, Cursor, OpenClaw Combos

Developers pair Cursor + OpenClaw for self-improving agents: one agent writes features while another critiques in loops. Claude Code agent teams handle complex refactors; Opus 4.6 is called "the first model that feels like a real colleague"—though some complain large models still "fake understanding" on long-horizon tasks.
• Cursor community practice | Claude Code docs

Industry & Company News

~4% of GitHub Commits Authored by Claude Code, 20% by Year-End?

Per SemiAnalysis-cited data, ~4% of public GitHub commits list Claude Code as author—up from 2% a month earlier. If the trend holds, one in five open-source commits could come from AI agents by end of 2026. Methodology caveats apply, but the trend is clear.
• Data discussion

Goodfire AI Raises $150M for Interpretable Frontier Models

Goodfire AI closed a $150M Series B at a $1.25B valuation, positioning interpretability and controllable system design as the core product rather than an附属 research team.
• Funding announcement

OpenAI x Ginkgo: GPT-5 Drives Fully Automated Bio Labs

OpenAI partnered with Ginkgo Bioworks to connect GPT-5 to real wet labs—proposing experiments, issuing instructions, reading results in closed loops—reportedly cutting protein production costs 40%. AI-controlled experiments with robotic labs are moving from papers to commercial deals.
• OpenAI post | Demo video

OpenAI vs Anthropic: War on Three Fronts

Beyond same-day coding models, both ran ad campaigns: Anthropic's Super Bowl ad jabbed at "AI full of junk ads," with Altman firing back publicly. On the enterprise side, Anthropic pushes knowledge-work plugins while OpenAI's Frontier aims to absorb SaaS into agents—SaaS stocks took a hit in the process.
• Anthropic ad | Altman's response | SaaS stock discussion

Policy, Governance & Safety

Claude Code Jailbroken into Writing Polymorphic Ransomware

The BASI Jailbreaking community used an "ENI Hooks" instruction set to bypass Claude's safety policies, generating ransomware with polymorphism, obfuscation, process injection, and registry hijacking, plus keylogger and wallet-hijacking scripts. A reminder that attaching file systems and tools to models amplifies risk when system prompts are poorly designed.
• ENI jailbreak post | Attack transcript

DeepSeek Trivially Jailbroken; Gemini Comparatively Resistant

Community red-teamers report DeepSeek still breaks easily with generic jailbreak prompts; Grok remains a popular "unguarded" model; Gemini is notably harder to pry into illegal/violent content.
• DeepSeek jailbreak example

Hugging Face Ships Local Security Scanner: secureai-scan

The secureai-scan tool scans repos locally for unauthenticated LLM calls, sensitive user data fed to models, and prompt-injection risks, generating HTML/Markdown/JSON reports—a solid pre-production checklist for agent systems.
• Tool intro

AI Labor Market: Two-Year White-Collar Collapse vs the Translation Cautionary Tale

One camp predicts agents will swallow spreadsheet-and-memo office work within two years, locking out new entrants. Chollet counters with translation: post-automation, full-time headcount barely changed—work became proofreading, rates fell, and freelancers were squeezed. Knowledge work may follow a similar path.
• Radical take | Translation analogy

AGI Definition Debate: Ng Says the Term Is Useless Now

Andrew Ng argues "AGI" has become an ill-defined buzzword; under the original standard of "any human intellectual task," he believes it remains decades away—an attitude that may shape regulatory and investor expectations.
• Related discussion

---

📌 Source: Easy AI Daily