Easy AI Daily — November 7, 2025
A curated digest of AI industry news from the Easy AI teaching project.
Model Updates
- Moonshot AI's Kimi K2 Thinking: An open-weight 1T-parameter INT4-quantized MoE reasoning model supporting a 256K context window and 200–300 consecutive tool calls. Achieves SOTA on HLE (44.9%) and BrowseComp (60.2%).
- Tech blog | Hugging Face | Kimi.com
- SANA-Video lands in Diffusers: The video generation model is merged into Hugging Face Diffusers (PR #12584), compatible with the Diffusers scheduler/pipeline ecosystem.
- Polaris Alpha rockets to Repo Bench top 3: An anonymous model climbs to #3 on Repo Bench, sparking speculation it is GPT-5.1 or Gemini. Some users found Claude 4.1 outperforming Claude 4.5 on certain tasks.
- GPT-5 passes Gemini 3 Pro on VoxelBench: Community screenshots show GPT-5 beating Gemini 3 Pro (Lithiumflow) at 3D model generation (results).
- GPT-5.1 source code leak (unconfirmed): A Reddit user shared snippets suggesting a "GPT-5.1 Thinking" model name in OpenAI source code; not officially confirmed. Reddit discussion.
- Google Ironwood AI chip: 4x faster than the previous generation; a single pod supports 9,000+ TPUs and can train 100-trillion-parameter models, aiming to challenge Nvidia. Google announcement
- Silicon and inference stack updates: TPU v7 (Ironwood) approaching GA (docs); llama.cpp adds Apple M-series Neural Accelerators support (PR); Together's ATLAS adaptive speculative decoding delivers 4x speedup (blog).
- GPU systems deep dive: NVIDIA Blackwell supports FP4→FP16 block conversion (PTX ISA 8.8); memory bandwidth tests reach 92% of spec; Triton dynamic compilation with C++ JIT for kernel optimization (docs).
- Agent frameworks, wallets, and managed RAG: LangChain releases Deep Agents for JS; Privy + LangChain agent wallets; Perplexity Comet adds multi-tab browsing; Google Deep Research integrates Gmail/Drive (link).
- CodeClash: humans still win: LLMs played 1,680 matches, but human experts won 37,500–0; Claude Sonnet 4.5 was the best model. Results
- fastWorkflow snags Tau Bench SOTA: A small model with context engineering matches larger models on retail and airline workflows; paper forthcoming. Repo | Tau Bench
- Tiger Data coding agent cookout (NYC): A coding agent meetup in Brooklyn on November 13. RSVP
- DroidRun AI discussion: Reddit users discuss this Android automation tool and the open-source status of Gemini 2.5 Computer Use. Thread
- Memorization vs. generalization; agent/data-science evals: GoodfireAI decomposes MLP weights into memorization and generalization components (blog); Google releases the DS-STAR data-science agent benchmark (pub); MIRA reveals visual reasoning flaws (arXiv).
- Equivalent linear mappings paper: EleutherAI research shows reasoning in Qwen 3 14B and Gemma 3 12B can be represented as linear mappings, with low-dimensional semantic structure found via SVD. OpenReview
- Anthropic postmortem on fp16 vs fp32 sampling bugs: Precision mismatch caused top-p/top-k sampling errors, underscoring the importance of verifying dtype pipelines. Postmortem
- Yannick Kilcher Discord slow mode debate: Discussion over 1/2/6-hour slow modes in the ML papers channel, balancing content quality with user experience; leaning toward moderate enforcement.
- Hugging Face regulation pause: HF paused some Spaces over potential new rules, sparking debate on whether this is a more responsible approach to avoid security gaps.
- Tinygrad remote reboot: tinybox devices now support BMC remote reboot, confirmed by George Hotz.
- XPeng IRON humanoid robot: Its gait mimics female pelvic sway, showcasing advanced biomechanics, though practical market use is questioned; users compare it with Tesla Optimus. Reddit
- Apple eyes Google's 1.2T model for new Siri: Reuters reports Apple is considering Google's 1.2-trillion-parameter model for Siri, weighing model choice against privacy. Reuters
- OpenAI mid-run prompt editing: Users can interrupt long queries and add new context without restarting, improving flexibility for GPT-5 Pro queries. Demo video
- Soumith Chintala leaves Meta/PyTorch: The PyTorch creator announced his departure, reflecting on PyTorch's growth and open-source culture. Tweet
- David Sacks: "There will be no federal bailout for AI": Sacks argued market competition suffices; Sam Altman clarified OpenAI is not seeking government guarantees and supports public AI infrastructure. Sacks | Altman
AI Hardware
Agents & Tooling
Research & Benchmarks
Community & Events
Company & Industry News
*Source: Easy AI teaching project.*