Key Takeaways
- Thinking Machines Lab released Inkling 2 Mini, a 26B-parameter looping transformer under Apache 2.0, with support in transformers, llama.cpp and MLX.
- Hugging Face disclosed a self-built CDN moving nearly 1 EB per month, built for large model downloads rather than small web files.
- NVIDIA detailed Alpamayo 2 Edge, a 2B-parameter vision-language-action model for in-car trajectory generation and explanations.
- Yann LeCun amplified claims about 800,000+ advanced AI chips approved for Chinese firms, alongside new JEPA world-model results.
- Perplexity pushed its agent orchestration story with Chrome task delegation and a Mac app supporting frontier and local models.
1. Open Models and Infrastructure
- Thinking Machines Lab released Inkling 2 Mini (A3B/26B) under Apache 2.0, described as reaching the pareto frontier for open agent models, with Hugging Face support in transformers, llama.cpp and MLX. It is a looping transformer that runs on a laptop, scores 20.1 on the AA Index, and Mira Murati said the looping mechanism lets the model "think more, write less"; she noted its reasoning process is still monitorable at this scale but may not always be, and the team released a variable-loop research model for further study 1 2 3 4.
- Hugging Face said it has run its own CDN since this summer, moving nearly 1 EB per month — about 420 GB/s and 80,000 requests/s — because AI downloads are dominated by a few very large files pulled by machines, mostly GPU clusters concentrated in a few cloud regions, unlike the many-small-files assumption behind traditional CDNs; caching must handle huge objects and release-day download surges 1.
- Hugging Face also highlighted Diffusers LoRA fine-tuning of NVIDIA's RFD4-Proteina for protein design, training a rank-32 adapter on the denoiser with a flow-matching loss applied to backbone coordinates and local latents rather than pixels, with training code broadly similar to standard image and video generation models 1.
2. Chips, Policy and World Models
- Yann LeCun shared a column arguing the debate around NVIDIA and Jensen Huang obscures a larger behind-the-scenes contest between open-system advocates and closed-system holdouts, reflecting a long-standing US tension between professed free-market support and a determination to control global flows of strategic technology. He also amplified a claim that disclosed classified information shows the Commerce Department under Howard Lutnick approved sales of more than 800,000 advanced AI chips to Chinese companies in the year through May 11, including several Chinese military firms; these are claims relayed in the posts, not independently verified here 1 2.
- LeCun shared Fast-JEPA, whose world-model planning is reported at 2.41x faster than INTACT with 96.50% success in 10.88 ms, and a second JEPA paper describing a content-geometry tradeoff: world models can learn smooth navigation maps while losing the locations of objects that require interaction. Slow SVReg reportedly raised planning success from 21.2% to 92.9% in hard first-person maze navigation and from 44.0% to 87.3% on PushT manipulation 1 2.
- LeCun argued energy-based models are often mistaken for a kind of probabilistic model when they are more general, with probabilistic models a special case requiring a computable, convergent partition function and a probabilistically interpretable loss; many EBMs need no normalization and can be trained with losses lacking a probabilistic interpretation. He said his NYU deep learning course has used EBM as a unifying framework for classification, regression, structured prediction, unsupervised and self-supervised learning, including JEPA and diffusion models, for about 20 years 1 2 3.
3. Agents and Developer Tooling
- NVIDIA introduced Alpamayo 2 Edge, a 2B-parameter reasoning vision-language-action model that generates driving trajectories and causal explanations in real time in the vehicle and reasons only when needed; it is built on the Nemotron backbone from NVIDIA Cosmos 3 Edge, supports fine-tuning, and is published on Hugging Face. NVIDIA also set out eight design principles for efficient sparse attention on its accelerated computing, noting that finding the right tokens itself carries cost even as sparse attention cuts computation 1 2.
- Perplexity's Aravind Srinivas said tasks can be delegated to Perplexity Computer from Chrome, with the extension running directly in Chromium browsers so the agent can operate across tabs and tools after a task is handed off from the sidebar. He said the Mac app has improved substantially with rising retention and is the only AI app supporting both frontier and local models, including hybrid and fully local setups, and he recommended Gemini Neon on Perplexity Computer while saying it may become the default based on retention data 1 2 3 4.
- Ben Tossell argued the binding constraint on personal AI agents is increasingly the connection layer and information access rather than model intelligence, since each lab's assistant layer must enter through a desktop, browser or phone app. He rated Grok Bot the best personal-agent product despite weaker models, said Claude Duet is phone-only and he stopped using it after testing, and argued Apple's hardware quality and iMessage network effects mean consumers will not run their own Mac Mini iMessage servers while Apple is unlikely to let personal agents operate phone apps 1.
4. Research Verification and Evaluation
- Deedy said the Lean community has confirmed Levent Alpöge's proof of the Riemann hypothesis as correct, citing a joint statement from Lean FRO and the Mathlib Initiative certifying that the proof meets all necessary verification standards. The cited post says Alpöge published the proof on November 20 as a 17-million-line Lean library and that the Lean and Mathlib communities completed verification within a day; this is a community claim relayed in the posts 1.
- Ethan Mollick argued OpenAI and Anthropic models are hard to simply distill, which may help explain the relative slowdown in open-weight model progress, though how central distillation is to rapid open-weight gains remains unclear and may soon be answered. He also said focusing only on top AI research labs' tasks likely understates AI's capability on economically valuable work, citing Epoch's MirrorCode results in which AI autonomously reimplemented software researchers estimated would take humans 2-17 weeks, and an OpenAI report in which a researcher estimated a month of team work would take three months without coding agents 1 2.
- Hamel Husain stressed that in evals good process matters more than tools, and a book launch for Evals for AI Engineers, co-authored with sh_reya, drew many agent builders at an event hosted by danolsen and LeanProdMeetup 1 2.
5. Industry Signals
- Elon Musk said SFT is more like "cutting hair with a chainsaw" than fine-tuning, arguing it should be built into main training and that system prompts are superfluous if default behavior is sensible; he said this is already done in Tesla AI, which he called the most efficient AI on Earth, with FSD V15 rolling out fully next month. These are his claims, not independently verified here 1.
- Clement Delangue, citing a new US House report, said US researchers, startups, nonprofits, government and students all benefit broadly from open-source AI regardless of origin, and argued the best response to Chinese open-source AI is not a ban but building more American open-source AI 1.
- NVIDIA's GTC Washington, D.C. session published an agenda focused on AI infrastructure, open models, trusted agents and physical AI, with Ian Buck, Justin Boitano, Kari Briski, Bryan Catanzaro and Rev Lebaredian participating and online viewing available December 1 1.
