Key Takeaways
- OpenRouter annualized token volume passed 10 quadrillion, roughly doubling every 10.5 weeks, with the past nine weeks exceeding all prior volume combined.
- OpenBMB released the 1B-parameter MiniCPM-OCR document-parsing model, with a Hugging Face Space comparing its output against Baidu Unlimited OCR v2.
- A Saturate-based re-OCR of 14,483 pages of fossil-fuel industry documents ran at 8.9 pages/sec on a single GPU for $1.47, with concurrency auto-set to 512.
- Mistral Large 4 wrote a quantum-chemistry SCF solver from a short prompt, self-verified energies against PySCF to 1e-13, and chose the STO-3G basis set on its own.
- Ethan Mollick framed January 2026 as the plausible start of real agentic work, while flagging that GDPx lacks a human baseline and that GPT-6 Pro found errors in scoring.
- Yann LeCun disputed responsibility for Llama-4's disappointing results, pointing to rushed product pressure and decisions moved to Menlo Park.
1. AI Models and Infrastructure
- OpenRouter has surpassed 10 quadrillion annualized tokens processed, with token volume roughly doubling every 10.5 weeks; the past nine weeks alone exceeded all prior volume combined 1. The figure is a platform-level usage signal, not a measure of model quality.
- OpenBMB released MiniCPM-OCR, a 1B-parameter document-parsing model, and a Hugging Face Space now compares its output against Baidu Unlimited OCR v2 1. Separately, a Saturate-driven re-OCR of 14,483 pages of fossil-fuel industry documents cost $1.47 on a single GPU, ran at 8.9 pages/sec, and auto-set concurrency to 512; Saturate batch-calls any OpenAI-compatible endpoint or inference engine 2.
- Mistral Large 4 was given a short prompt to write a quantum-chemistry SCF solver from scratch, self-validated its output, and built a 3D visualization, deciding on its own how to break bonds and which orbitals to display 1. A Mistral employee said the model chose the STO-3G basis set for speed and qualitatively correct orbital shapes, and checked energies against PySCF across geometries to 1e-13 precision 2.
2. Agents, Evaluation, and Research Direction
- Ethan Mollick estimated GPT-4 could independently complete about four minutes of work at a 50% success rate, while current models can handle weeks of work; he called January 2026 a reasonable dividing line for agentic work actually beginning, with significant productivity impact so far mostly limited to programming 1. He endorsed GDPx for evaluating the value of AI work but said it needs a human baseline like GDPval, and noted GPT-6 Pro found errors in scoring where a 0% score masked economically meaningful differences in success rates; the cited post said GDPx covers high-value domains and the best model scored only 12% (1/7) 2.
- Mollick argued labs treat software with practices like mutation and poor communication because of a core belief that these products are temporary vehicles to raise money in the one or two years before "real" AGI makes all products obsolete 1. A forwarded item said Anthropic stated in a revised S-1 that, based on internal estimates of future models, it may "soon" meet its definition of AI driving significantly accelerated progress—under its RSP, progress at twice the rate expected without significant AI contribution, compressing two years into one (e.g., a 9x yearly gain becoming an 81x threshold) 2.
3. Industry Positioning and Dispute
- Yann LeCun clarified his Meta roles: he created FAIR in 2013 and its Paris and Montreal branches in 2015 and 2017, stepped down as head in 2018 to become Chief AI Scientist, and FAIR was co-led by Antoine Bordes and Joëlle Pineau from 2018 to 2023 1. He said he never worked on or developed LLMs, citing only defending Galactica in October 2022, pushing Llama internally, helping persuade Zuckerberg to form a GenAI product group around Llama, and persuading management to open-source Llama-2 2. On Llama-4's disappointing performance, he blamed pressure to ship too fast, which made the GenAI team overly conservative and less innovative; he said the team should have spent more time integrating FAIR research and let Paris's original Llama creators keep more technical decision rights instead of moving decisions to Menlo Park 3.
- A forwarded item said LeCun's work with Randall Balestriero shows scale is not the key to a model understanding the world—what matters is whether two views shown to the model are perspectives on the same thing; paired image-text training reached a comprehension level with 21.6 million samples that next-token prediction needed 900 million samples to reach 1. The same forwarded item said a 109M-parameter text model came closer to an image model's worldview than a 72B-parameter language model, and stated that LeCun left Meta in November and founded his own lab in Paris 2. These forwarded claims are disputed/unverified and are not independently confirmed in the supplied material.
- Hamel Husain grouped AI benchmark companies with Gartner vendor rankings, mortgage-backed-securities credit rating agencies, and US college rankings, saying they are in very similar businesses and that enough people have not yet recognized the fourth category but eventually will 1. A forwarded item said a popular benchmark's tasks were taken down from Hugging Face and called on @HelloSurgeAI to fix it 2.
4. Applications and Demos
- Aravind Srinivas shared an unofficial GTA VI 3D map project pairing Leonida landmarks with fan-identified Florida locations, allowing side-by-side map comparison and approximate 3D exploration of Leonida 1. He also shared an interactive globe of Earth's 4.5-billion-year evolution, built on 14 open datasets with a custom engine driving 258 plates across 109 paleo-elevation maps, letting users drag a timeline or enter a city to see its position 400 million years ago 2.
- AI Engineer conference talks have recently begun uploading to YouTube; the forwarding account called it the most important AI conference globally, said companies present on introducing and scaling AI at SoTA level, and listed 10 of the most relevant talks in a long thread while calling the talks high quality 1.
