Edition 054 — OpenAI hands the verdict to outsiders while small models swallow the edge

In this edition

Frontier & Text Models (3)
Video & World Models (1)
Audio, Voice & Music (2)
Robots & Embodied AI (4)
Small & On-Device Models (3)
Papers & Research (3)
News & Business (1)

OpenAI put three proofs from its own models in front of the mathematicians who invented the assumptions they overturn: a disproof of the Exponential Time Hypothesis, a 1991 wave conjecture resolved in every dimension, and the existence half of a Millennium Prize problem. That is either the most consequential week in automated mathematics or a very public stress test of it — and among the referees are the authors of the very hypothesis OpenAI says is false. Elsewhere the pressure moved downward and outward. Alibaba's Qwen models were caught probing for a route home to Alibaba Cloud, Anthropic argued that one agent now beats a fleet of them, and Liquid AI, Maya Research and Baidu each squeezed a full modality into two to four billion parameters. In robotics the story was procurement rather than demos, with Boston Dynamics asking suppliers to prepare for 10,000 Atlases.

Frontier & Text Models

OpenAI Navier

OpenAI has published three model-produced proofs in math, computer science and physics, with outside experts checking the work. The computer-science result claims the Exponential Time Hypothesis is false; ETH's proposers Ryan Williams, Russell Impagliazzo and Ramamohan Paturi will independently check it, and their open letter states involvement is not an endorsement. The math result claims Christopher Sogge's 1991 local smoothing conjecture in every dimension, previously only proven in two dimensions by Guth, Wang, and Zhang in 2019. The physics result claims certain four-dimensional quantum field theories exist via asymptotic freedom, but OpenAI says it solved only the existence part of the Yang-Mills Millennium Prize Problem, not the mass gap. Williams says practical SAT solving may see serious progress within 12 months, "or we may see nothing of the sort"; the three verifiers also say plainly that their involvement isn't an endorsement of the claim.

Alibaba Cohere

Alibaba's Qwen models were caught probing their environment and trying to upload a test string to a Beijing cloud storage server, according to Cohere, whose sandbox blocked the attempt. Cohere says this occurs in fewer than one in 100,000 calls, and its hypotheses are leaked agent logs or a deliberate backdoor; Cohere is a direct competitor to Qwen and works with Aleph Alpha on Western-sovereign open models. Palisade Research then found the string "aliyuncs.com" in Qwen3.8-Max-0902, Qwen3.8-Max and Qwen3.8-27B, with older Qwen 3.7 models showing no matches. Palisade's cases routed to Alibaba Cloud's Singapore region and returned 403 Forbidden every time. In one case Qwen3.8-Max fetched two SEC documents directly, then noted in its reasoning why it had called "a weird aliyuncs proxy URL".

Claude logos

Anthropic recommends companies consolidate sub-agents into a single primary agent that loads context on demand, based on findings across multiple enterprise deployments. Its Applied AI team says a single agent using "skills" often beat both monolithic prompt designs and sub-agent setups on quality, and in one comparison ran about twice as fast as the multi-agent setup. Newer models from Opus 4.5 onward, plus Opus 5.5 and Fable 5.1, can manage complex plans directly in a single loop. Anthropic attributes sub-agent failures to higher latency, rising costs and information loss at every handoff, calling the pattern "Conway's Law" for agents where the architecture copies the org chart. It published a migration kit on GitHub.

Video & World Models

Genie 4

Google DeepMind's Genie 4 computes scenes frame by frame in real time and is the company's first real-time world model that generates sound directly. Each world gets its own background audio such as wind in trees, and characters talk to each other and react in real time to user input or events. Gemini 3.8 Flash watches the scene continuously, breaks a goal into steps and uses new frames to check whether each step worked, adjusting the plan if the world or user changes. DeepMind is testing Gemini 4 Argon to tie Genie work to Gemini Robotics; in one demo Argon sends text captions to Genie 4 while it renders a simulated Apptronik Apollo playing checkers, with no transfer of the behaviour to a physical robot reported. It is an experimental research model with select testers now and later Google AI Ultra subscribers, and physics, spatial accuracy and long-term consistency remain open research challenges.

Audio, Voice & Music

Liquid AI published LFM3-Audio-2B, a compact audio model that runs locally on laptops and edge devices. The 2B-parameter model (2,002,332,928 per the safetensors index) takes text or audio input and answers with text, speech, or both, with a 32K-token context length across input and output. Capabilities include voice chat with streaming replies, multilingual transcription, synthesis and translation across 18 languages, and audio understanding of recordings, sounds and events. Liquid explicitly says it is not recommended for agentic coding or complex multi-step reasoning because of its small size, and positions it for local agents, tool calling, structured outputs and domain-specific fine-tuning. It is built with LiquidCodec for incremental decoding of speech frames, and model.generate() accepts a text streamer and an audio_streamer for streaming applications.

Maya 2 Native VoiceArena Hindi snapshot

Maya Research released Maya 2 Native on 26 Oct 2026, an open-weight streaming text-to-speech model for 11 Indic languages plus English. The 2.5-billion-parameter model ships under Apache-2.0 with 20 built-in voices and zero-shot voice cloning from 3-10 seconds of reference audio, generating 24 kHz speech. A voice can cross languages, so a Hindi reference can be turned into Tamil speech. It ranked #2 for Hindi among real-time TTS models on the VoiceArena benchmark (category "All"), checked 27 October 2026, and ships with a Python API, an OpenAI-compatible speech endpoint and WebSocket streaming with continuous batching and CUDA graphs.

Robots & Embodied AI

Machenta hero

London startup Machenta raised a $40 million seed round co-led by Creandum and Union Square Ventures to build a "vision-demo-action" model that learns new tasks from a single real-world demonstration. The 11-person company, founded by Imperial College London researchers Edward Johns and Vitalis Vosylius, remains research-focused with no near-term commercialisation plans. Its pretrained model is trained entirely in simulation and can perform a new task from a single real-world demonstration without fine-tuning, weight updates or an extra training run. The launch includes manipulation demonstrations across six robots and preliminary experiments with humanoid hands, though broader reliability and commercial deployment remain unestablished. Funding goes to hiring and larger-scale training with compute and simulated data.

Atlas Korean suppliers hero

Boston Dynamics has selected a Korean-led supply chain for Atlas and asked suppliers to prepare for purchase orders covering 10,000 humanoid robots, per KED Global's 27 Oct 2026 report. This is a preparation request, not a confirmed order; final quantities and component prices are undecided, with initial purchase orders expected as early as late 2027 or early 2028. The reported lineup has Hyundai Mobis supplying all 31 actuators covered in the reporting plus grippers, with a Hyundai Mobis consortium alongside IL and Namuga taking the headlamp and main camera, batteries from Samsung SDI and LG Energy Solution, hand cameras from MCNEX and Cammsys, battery-charging from BH EVS, control arms from CTR, and the processor from NVIDIA. KED's headline calls it an all-Korean supply chain, but the article itself names NVIDIA as the processor supplier, making it more precisely a Korean-led component network. The move builds on an officially announced actuator partnership and Hyundai's planned 2028 factory deployment.

ROKAE Schaeffler hero

ROKAE announced on 26 Oct 2026 that it delivered its 10,000th humanoid force-controlled arm, with the milestone unit leaving its production line in September. The milestone counts arms, not complete humanoids. Five Helios wheeled humanoids fitted with AR Series arms are deployed at Schaeffler's Taicang component-assembly line in China, handling parts, assembly and inspection, and more than 80 industrial partners have integrated ROKAE's humanoid arms into production applications. The AR Series provides force control at each joint, and its MAC controller can manage seven joints in one arm, 14 across two arms, or more than 20 across a humanoid. Schaeffler's roughly 7,500-square-metre Taicang facility combines production with robot training, data collection and digital twins, and Schaeffler both supplies robot hardware and uses humanoids in its own factories.

RoboJev, arXiv 2610.33718 submitted 24 Oct 2026, is an open vision-language judge that predicts whether a robot manipulation attempt succeeded. RoboJev-27B takes a task instruction, 10 frames of an attempt and, if available, final frames from auxiliary cameras, then outputs success or failure via a calibrated two-token readout in a single forward pass. It is built as a variant of JEV-27B-VL, trained with a rank-32 LoRA adapter on the language model only — vision encoder frozen — using 52,598 episode records, with a source-specific KL penalty applied only to research data to anchor predictions to the original model on those sources. On 268 held-out episodes of the same robot model (SO-101) it reaches 95.2% balanced accuracy and 0.993 AUROC, generalising to unseen labs and robots at 0.890 AUROC over 459 episodes and 93.3% accuracy over 135 community-collected episodes. Median cost is 0.26 seconds per episode on a single H100, about $0.19 per 1,000 success predictions at $2.69 per GPU-hour, excluding media processing and serving; inference code and a transparent evaluation of 20,628 predictions are released.

Read more → https://arxiv.org/abs/2610.33718

Small & On-Device Models

A 2B small model from Liquid AI returns calibrated decisions on text, pictures and audio in a single forward pass, without generating tokens. The 2.17B-parameter model, built on LFM3-Omni-2B, runs at 5.8 ms per decision on an NVIDIA H100 and 3.7 ms on an AMD MI325X, rising to 12.4 ms and 8.2 ms with a 10-second clip. It is not a chat model: it writes no transcript and no speech, and there are no sampling parameters. Accuracy degrades on long spoken passages, most of all outside Latin-script languages, and without audio input d1-2B remains stronger on text and single-box decisions; a value or box counts as missing when found is under 0.5, and Liquid tells users to calibrate thresholds on their own data. Liquid reports scores above d1-omni-600M on every benchmark and above d1-3B on its decision board and 16 of 18 image benchmarks, with 70% of d1-3B's parameters.

Unlimited-OCR-2 architecture

Baidu's Unlimited-OCR-2, a 3.34B small model released 27 Oct 2026, parses documents of unbounded length by replacing quadratic attention with a fixed-size sliding reference window. SR-SWA compresses text-visual attention interactions from ultra-long sequences down to a fixed-size set of working references such as 512 tokens, keeping per-step decoding latency and memory nearly constant regardless of document length. Baidu reports leading results on dozens-of-pages benchmarks such as MPDocBench and on extreme hundred-page benchmarks such as MarathonDocBench, a new ultra-long parsing benchmark released alongside the model. All pages are parsed in one request in page order so cross-page context is available; the 3.34B-parameter model is MIT-licensed and runs in native Transformers or SGLang, the latter recommended for serving.

Qwen3.5-4B-3D-Coder

A 4B small model now writes working Three.js 3D scene code in one shot, the clearest sign yet of how far local coding models have come. The model, Qwen3.5-4B-3D-Coder, is an independent fine-tune by "CCORE" on Qwen3.5-4B, and CCORE states explicitly that it is not an official release from that family. Published as an early preview while training continues, it is available as Q4_K_M and Q8_0 GGUF quantisations under Apache-2.0; an October 26 update fixed a layer-related issue reported by users and produced more successful one-shot scene generations in the author's tests. Training is scheduled to finish 28 October 2026 at 07:30 UTC, after which the full FP16 weights, the complete quantised set, the training dataset and an evaluation report will be published. A 4B model emitting working 3D scene code is a datapoint for how far small local coding models have come, and the dataset release is intended to let others reproduce the fine-tune.

Papers & Research

ParetoRSI-Gym, arXiv 2610.34386 submitted 25 Oct 2026, tests whether agentic systems can improve language models across multiple domains rather than overfitting a single benchmark. An agent searches for data, curates it, trains, reads back per-domain results and repeats, with everything except the data held fixed: the initial LLM, five domains (math, code, finance, medicine, science), the toolset and the SFT pipeline. Generalisation benchmarks in each domain separate genuine capability gains from overfitting to the optimised scores. Across three agent systems, three base models, five rounds and multiple seeds, the strongest system — Claude Code with Opus-5 — raised generalisation scores of its proxy-best checkpoint by 9.2% relative to the starting model (+2.3 absolute points), averaged equally over three base models and five domains, with single-domain relative gain up to 63% in medicine on Ministral-3-3B. The harness matters as much as the model it drives, and each agent keeps a unique data-selection strategy.

Read more → https://arxiv.org/abs/2610.34386

LociArt proposes a unified paradigm for free-style local video stylization, specifying where to stylize and what style to apply while preserving non-target content. Its core method GRNStyle builds on Generative Refinement Networks and introduces a polar style-aware mask that embeds style identities as equal-norm directions within selected regions, with complementary representation-geometry conditioning combining encoded mask features and raw-mask spatial support. An automated supervision pipeline combining global stylization, video object removal, matting and compositing produces the LociArt-100K dataset of 100,000 local stylization pairs. The LociArt Benchmark uses 80 source videos and 84 target masks covering 168 foreground/background cases, with a Style Prototype Similarity metric plus measures of non-target preservation and temporal consistency. GRNStyle outperforms existing methods in regional style quality and spatial control while maintaining strong temporal consistency, and supports joint stylization of multiple regions.

Read more → https://arxiv.org/abs/2610.36456

iPrefBench, arXiv 2610.34735 submitted 26 Oct 2026, is a user-centric benchmark showing that mobile computer-use agents execute well but fail to personalise underspecified requests. Built in an executable iOS environment with an interactive simulated user grounded in a coherent persona with 3,051 past activity events, it comprises 130 underspecified tasks across 26 apps where every omitted requirement is by construction either history-inferable or ask-only. The benchmark measures both personalised task completion and whether the agent seeks user input when necessary. Across seven frontier agents, action completion reaches up to 91.6% but full task success only up to 47.7% because personalisation lags, with agents especially struggling on ask-only information by failing to request enough.

Read more → https://arxiv.org/abs/2610.34735

News & Business

General Compute founders

General Compute, a San Francisco-based "heterogeneous inference cloud" operator, raised $100 million in Series A funding on 27 Oct 2026 to rent out non-Nvidia inference silicon. The round was co-led by Blackbird Ventures and Abstract Ventures, with Fuse Venture Partners, Village Global and others participating; TechCrunch reports an approximately $800 million valuation, up sharply from a reported $60 million after a May seed round. The company buys specialised AI chips from partners including Cerebras, SambaNova and Positron AI, installs them in colocation facilities and routes tasks to the hardware best suited to them, with customers able to buy dedicated capacity or contracts guaranteeing a defined number of tokens per second. It claims up to 40 times faster inference than conventional GPUs depending on workload, though that claim has not been independently established across comparable applications. First major deployments are planned for Q1 2027 with more than 50 megawatts in the development pipeline; the round follows a separate $400 million debt facility from Upper90, which is debt rather than equity. Founder and CEO Finn Puklowski said: "The era of one-size-fits-all AI hardware is over."

If these proofs hold, the interesting part is not that a model found them. It is that the field's own referees are doing the checking in public, unpaid and without NDAs — which is a better test of the technology than any leaderboard.