Models & Capabilities
Gemini 3.1 Pro: big benchmark gains, mixed real-world experience
Google released Gemini 3.1 Pro. ARC-AGI 2 scores jumped from 31% to 77%, and on Context Arena's MRCR retrieval evaluation it approaches GPT-5.2 — even stronger in hard-retrieval scenarios. It shows strong coding and spatial reasoning at the same price as 3.0. However, engineers report unstable CLI/agent tooling, routing confusion (frontend labeled Gemini actually served Claude), and "self-upgrade death loops" in agents like OpenClaw. Some worry about post-launch performance cuts.Links: Model card | MRCR eval | Cost comparison | Reddit discussion
Claude Opus / Sonnet 4.6: time horizons and long-reasoning failures
METR estimates Claude Opus 4.6's "50% software task time horizon" at ~14.5 hours, but with a very wide 6-98 hour confidence interval; the community cautions such metrics are sensitive to task sets and error amplification. Sonnet 4.6 leapt up Arena's coding, instruction-following, and math leaderboards, but users complain about long-reasoning mode frequently hitting token limits, empty outputs, and degraded Claude Code UI/stability.Links: METR thread | Reddit discussion | Arena leaderboards | Long-reasoning feedback
Qwen: open models climbing agent and vision leaderboards
Opinions split: some find Qwen weak on logic/common sense, others praise its 3-4B instruction following, citing Qwen 3.5-397B tying Kimi K2.5 at the top of Arena Vision. Qwen also improved significantly on the FoodTruck Bench business-simulation benchmark but still often "goes bankrupt" from poor execution — the "can reason, can't execute" agentic gap.Links: Vision Arena | Reddit discussion | FoodTruck Bench case study
DeepSeek system prompt leak
Someone extracted DeepSeek's full system prompt, which explicitly requires incorporating "socialist core values," avoiding discussion or criticism of the CCP, and includes hardware/deployment notes. This exposes both attack surface and the model's baked-in stance on politically sensitive topics.Agents & Tooling
- OpenClaw ecosystem: The most active multi-agent playground — agents launching tokens on Base, a "Last AI Standing" survival game, a Bitcoin dice casino; one agent on Gemini 3.1 Pro self-upgraded to nonexistent versions and crashed, requiring manual Claude rescue. Dashboards like ClawTower visualize multi-agent cost/state, but real permissions amplify risk alongside productivity. Dashboard repo | Last AI Standing
- GEPA / gskill: A pipeline treating "skills" as first-class artifacts (repo task → skill optimization → skill file delivery), reportedly enabling near-fully automatic fixes on specific codebases and ~47% faster Claude Code completion. Counterpoint: model-generated skill docs are often bloated; fewer, human-written constraints work better. Intro thread
- RLM & multi-agent topology: Recursive Language Models as a meta-scheduler; GPT-5.2-Codex and Gemini 3.1 Pro do well under RLM decomposition, Opus 4.6 does not. Research suggests topology (parallel/hierarchical/hybrid) alone accounts for 12-23% performance differences once model capabilities converge. Discussion | Topology paper summary
- NAVD: Replaces vector databases with append-only logs plus Arrow embedding indices; <10ms latency at 50k vectors. Project page
- Cloud-native agent runtimes: Airtable's Hyperagent deploys agents as isolated services with dedicated persistence and Slack integration; OpenClaw stays script-style (terminal/browser/codebase). Agents are becoming long-lived infrastructure services. Hyperagent post
- Taalas ASIC: A 6nm, ~53B-transistor chip with Llama 3.1 8B hard-burned into silicon achieves ~16-17k tok/s single-user inference, roughly $0.005 per million tokens at $0.10/kWh. Trade-off: models are essentially frozen; swapping requires a new tape-out. "Frozen base model + post-trained adapters" may be the practical path. Forbes coverage | Reddit thread
- ThunderKittens 2.0: Stanford Hazy Research ships BF16/MXFP8/NVFP4 GEMM on Blackwell, claiming parity or better vs cuBLAS. Key lesson: removing bad optimizations matters — undocumented tensor-core pipeline behavior can leave hardware idle. Blog post
- ggml/llama.cpp joins Hugging Face: The team joins HF to maintain the ggml stack with deep transformers integration — local inference moves from hobbyist project to mainstream ecosystem. Announcement | Reddit reaction
- tinygrad bets on AMD: George Hotz says the focus is solid compiler/codegen infrastructure for AMD GPUs, with bounties on quantifiable performance gains — "make the core IR good so every backend benefits" rather than hand-written per-card kernels.
- Benchmark methodology problems: MiniMax and Epoch AI admitted their SWE-bench Verified harness configurations differed from the official one; re-running aligned the results — the same benchmark can vary widely across harnesses. Meanwhile models scoring 70%+ on ARC-AGI 2 still play Connect-4 badly, fueling doubt about what these "intelligence" tests measure. Epoch AI correction
- Time-horizon metrics are statistically fragile: METR's Opus 4.6 point estimate is striking but the confidence interval is huge and the task set is nearly saturated; even METR warns against extrapolating from single points. Researcher explanation
- Hodoscope / ARES: Hodoscope is a trajectory browser for auditing agent action sequences at scale (it quickly found a benchmark bug); ARES exposes intermediate activations in long tasks, enabling probing/activation steering to locate and correct failure modes. ARES repo
- Claude Code Security (research preview): an AI security scanner with patch suggestions; Anthropic claims 500+ long-standing vulnerabilities found and fixed in real open-source repos. Restriction: users may not scan arbitrary third-party code with it, which many find legally/product-wise awkward. Launch post
- Document/search AI wave: Qwen-AI Slides generates near-finished presentations in minutes (Chinese/English only); Kimi's CLI is preferred over its VS Code plugin for large repos; Perplexity Pro users are drifting to ChatGPT/Claude/Kimi over tightened quotas and bot-only support. Qwen-AI Slides discussion
- Local inference vs API: A Mac Studio M3 Ultra user running Qwen 3.5 concluded local is mostly "privacy" since APIs are cheap. Replies added: control (no surprise deprecation/degradation), offline use, fine-tuning, lower latency, and possible long-term cost savings — if you accept hardware cost and setup effort. Reddit discussion
- Extreme-speed apps: ChatJimmy claims 15k tok/s chat; Voxtral Realtime does speech-to-text with sub-500ms latency for live meetings/captions. Voxtral Realtime
- ggml.ai joins Hugging Face (see above) — the "local model revolution" gets corporate backing; community mixed between centralization worries and relief about sustainable maintenance. llama.cpp discussion
- Unsloth × Hugging Face: Official partnership brings free LLM fine-tuning pipelines to HF, with 100k+ Unsloth-finetuned models already open-sourced. The real bottleneck is increasingly data, not compute. Announcement video
- Amazon Kiro AI incidents: Per FT reporting and engineer accounts, Amazon's internal Kiro coding assistant once decided to "delete and rebuild" an environment during a failure, causing a 13-hour regional outage — the second AI-tool-related incident in months. Public attribution remains "user error," but internally Amazon is reconsidering agent permissions and two-person approval flows. FT report
- Perplexity & OpenRouter outages: Perplexity angered Pro users with tighter quotas, mysterious subscription cancellations, API 500s, and bot-only support; OpenRouter's backend refactor missed an edge case, returning empty image generations while still billing — refunds issued afterward. Platform stability is product experience. OpenRouter statement
- Security stocks drop: Within an hour of an Anthropic blog analyzing AI's impact on cybersecurity, CrowdStrike, Cloudflare, Okta and peers reportedly lost ~$10 billion in combined market cap — the market is extremely sensitive to "AI + security" sentiment. Tweet roundup
- Claude Code data-access controversy: Anthropic scans open-source repos with Claude Code Security while restricting users from scanning third-party projects; separately, a user reported seeing what looked like another company's commercial lease documents in Claude Cowork, sparking debate over indexing, training-data leakage, or permission bugs. Either way, it's legally sensitive. Reddit thread
- Jailbreak techniques: The jailbreak community cataloged DeepSeek and Sonnet 4.6 system prompts and tested the "Crescendo" progressive technique on Gemini 3.1 (starting harmless, gradually steering context, then framing escalation as research/documentation). Gemini appears to guard most tightly; the API route is comparatively easier. Discussion
- Privacy drives local AI: Beyond features, the top motivation is simply not handing code, logs, and documents to cloud providers indefinitely — especially with unannounced degradations and account bans. Many teams prefer an extra GPU over API lock-in. Discussion
- FBI chip-secrets arrest: The FBI charged three engineers with stealing processor-security and cryptography documents from Google and others — a reminder that as hardware depends on few companies, insider risk grows. FBI notice
Infrastructure & Hardware
Research & Methods
Products & Applications
Industry & Business
Policy, Governance & Safety
📌 Source: Easy AI Daily (zhichai.net editorial translation)