A detailed retelling of AI Search (@theAIsearch)’s episode from 1 November 2026: “AI finally sleeps, Gemini 4 Carbon, ETH cracked, energy-based models, Anthropic IPO: AI NEWS”. All news, results and assessments below are relayed from that episode; the companies’ claims and the presenter’s conclusions have not been independently verified.
The episode’s subjects include Google Sleep, an experiment that separates interaction with users from the consolidation of new knowledge into two stages. Before reaching that story, it moves through animation, three-dimensional generation, document and speech processing, Anthropic’s financial results and OpenAI’s claims to have solved longstanding scientific problems. A substantial part of the roundup covers models that users can run themselves, ranging from models for a single CPU thread to versions with hundreds of gigabytes of files.
NVIDIA Motion Bricks generates movement for three-dimensional characters and humanoid robots. The user supplies a few target poses or instructions, and the system fills in natural transitions in real time. In the Unreal Engine demonstration, a character walks, runs, jumps over obstacles and interacts with objects. Its manner of movement can be changed: examples show a zombie’s gait, an injured person and crouching movement. Several actions can be combined into a sequence. The episode cites up to 15,000 frames per second and two milliseconds of latency. The model was trained on more than 350,000 motion clips and uses recent poses alongside target poses to predict intermediate states. Code and weights have been released separately for humans and the Unitree G1, at approximately 800 MB each. The NVIDIA Open Model Agreement permits commercial use subject to its terms.
The next system, IDOL, reconstructs a three-dimensional object from one image, including the unseen back. It represents the result through Gaussian splatting, using many volumetric elements, or Gaussians, to describe the shape instead of a conventional polygon mesh. The level of detail can be adjusted: examples compare 32,000 Gaussians with one million. The presenter highlights the model’s small size, at 0.66 billion parameters, and its ability to run with 8 GB of video memory. Reported generation times are 19 seconds on an RTX 4090 and 42 seconds on an RTX 3080. The code and model are available for local inference.
SceneBloom builds entire three-dimensional rooms from text: living rooms, bedrooms, studies and bathrooms. In the results shown, chairs face tables, objects rest on appropriate surfaces and furniture does not intersect. The AI first divides the room into functional zones, then Flux generates images of furniture arrangements for each zone. A subsequent stage turns those images into a three-dimensional scene and adjusts object positions. The rooms can be viewed from different angles or used for training robots in simulation. The authors have released code, adapters and a library of three-dimensional objects; the presenter describes the comparisons as showing an advantage in spatial consistency.
Baidu Unlimited OCR 2 turns document scans into structured digital text. Its main distinction is processing multiple pages together: recognising them independently can lose connections between the end of one page and the start of the next. Here, according to the presenter, an entire book can be submitted in one request. In the document, text, table and difficult-case parsing tests shown, the model outperforms Qwen 3.8 Max, GLM 5.3 Flash and DeepSeek 4.1. The files occupy about 6.9 GB, and local-deployment instructions have been published.
UniAction combines different ways of controlling video generation. It accepts an image, text and control actions, including applied force, camera movement, the pose of a human hand or a robot’s position. One model handles all of these controls. Each action is assigned a description, coordinates and a time, giving them a common representation. Code, weights, training scripts and evaluation scripts are available. The presenter identifies Alibaba Wan 2.2 as the base and expresses surprise that the authors chose what he considers an older model instead of MiniMax.
AutoSwarm coordinates several AI agents working on a shared task. In the example of launching a new feature, it assigns work to agents, lets them run in parallel and maintains a shared context. It sits above existing agent-management tools such as Codex and Claude Code and can bring their work together. The project is open, supports Windows, macOS and Linux, and the episode shows a page with installation instructions.
For speech synthesis, the presenter first highlights Paradee 1.2. The compressed model takes up only 8.5 MB, and the uncompressed version about 35 MB. It runs on a single CPU thread without a GPU and was trained to reproduce outputs from the larger Kokoro model, an approach known as distillation. The voice demonstration describes it as ten times smaller than the model it learned from. That compactness comes with a limit: there are just four preset voices, with no option to clone a personal voice. The model is released under Apache 2, which permits commercial use.
Tencent Hy-SimulMT translates speech between Chinese and English in real time while preserving the speaker’s voice. The examples include a question about Blender Studio and a technical explanation of a graphics processor. From the chart shown, the presenter claims a lower error rate and better translation quality than a human interpreter. That conclusion concerns the comparison presented. The model has 30 billion parameters and uses a mixture-of-experts architecture, activating only some of its blocks, equivalent to roughly three billion parameters, when processing data. The full version occupies about 61 GB, while a compressed Q4 GGUF version takes up about 20 GB; other lower-precision versions, including FP8, are also available. The licence is Apache 2.
Another NVIDIA release is a free collection of more than 2,600 three-dimensional objects, totalling about 490 GB. They include food, furniture, household appliances, industrial mechanisms, medical equipment and factory components. Simulation requires more than appearance and geometry: the objects also include mass, friction, collision boundaries and information about movable parts. The presenter suggests using the collection to build virtual environments in which robots learn to grasp and move objects. The assets are hosted on Hugging Face under Creative Commons Attribution.
The episode’s main scientific segment covers three OpenAI claims. The first concerns the Exponential Time Hypothesis, or ETH. The presenter explains it through SAT: finding an assignment of true or false values that satisfies a set of conditions. Adding a variable doubles the possible combinations, and the hypothesis concerns the impossibility of avoiding the associated explosive growth in computation. Hundreds of research papers rely on it, he says. The episode reports that an internal OpenAI model found an algorithm refuting ETH and that, according to the same announcement, the proof was checked step by step with the Lean formal verification system. In the presenter’s account, the algorithm expresses the conditions algebraically, divides the problem into groups and connects them through compressed information, obtaining the exact number of solutions without exhaustive search. He explicitly cautions that, even if the claim is confirmed, practical problems will not become instantly solvable and the algorithm may remain extremely slow.
The second claim concerns the local smoothing conjecture, described in the episode as having remained open for 35 years. It deals with wave behaviour at moments of strong concentration: such states can occur but should not persist for long. As the presenter explains, a result had previously been proved in two dimensions; OpenAI now claims a proof for any number of dimensions. The episode says this also implies a solution to the Kakeya conjecture in all dimensions. According to the same account, OpenAI has published Lean formal-verification materials.
The third story concerns quantum chromodynamics and Yang–Mills theory. The presenter recalls that the strong interaction holds quarks inside protons and neutrons, and that QCD describes it successfully in agreement with experiments. He relays OpenAI’s claim as a proof of QCD’s mathematical consistency and of the existence of the corresponding Yang–Mills theory. A crucial distinction is maintained: the Millennium Problem includes both existence and a mass gap. The episode discusses only the first part and does not declare the entire problem solved. He also recalls more than 700 mathematical manuscripts that, he says, OpenAI released a few weeks earlier. If the claims are confirmed, he sees them as a sign of the singularity, a stage at which AI development sharply accelerates scientific discovery. That is his assessment.
The financial segment covers Anthropic’s filing of an S-1 registration statement for an initial public offering, or IPO. The presenter mentions a reported target valuation of $2 trillion and the possibility of the largest offering in history, while stressing that the final valuation, proceeds and date remain unconfirmed. According to the figures presented, first-half revenue rose almost fourteenfold, from $1.2 billion to $16.3 billion. The reported loss was $80.7 billion, but he attributes most of it to the non-cash revaluation of early investment agreements as the company’s value increased. He gives a much smaller operating loss of $693 million. Excluding non-cash expenses, including employee stock compensation, produces adjusted operating profit of about $1.6 billion. The positive result concerns that adjusted measure specifically. He contrasts it with most leading US and Chinese AI laboratories, which, in his assessment, continue to make losses.
Anthropic’s commitments under computing contracts exceed $600 billion, according to the episode. Google, Microsoft, Amazon and SpaceX are named; the SpaceX agreement is reported at up to $120.5 billion, while Anthropic intends to spend up to $150 billion with Google over the coming years. The presenter separately explains that these are multiyear commitments rather than costs already incurred. Around 12,000 customers spend at least $100,000 a year, almost 2,000 spend at least $1 million, and roughly 250 spend more than $10 million. The five largest account for about 21% of revenue. He sees a risk here: one or two of those customers switching to another provider, such as OpenAI, could materially affect revenue. Under the structure presented, the seven co-founders will retain 50.1% of the votes after the IPO through a special share, limiting public investors’ influence over major decisions. The offering could take place as early as November, on Nasdaq, with Ant as the proposed ticker.
MatAnyone-E separates video foregrounds from their backgrounds while retaining hair, soft translucent edges and details of several people in the frame. To improve efficiency, it first analyses a low-resolution image, identifies difficult regions and concentrates additional high-resolution computation on those areas. The team reports real-time 4K video processing on one H200; the episode separately cites a requirement of roughly 3 GB of video memory and high speeds on consumer GPUs. Training code and models are still promised for release, along with a ComfyUI custom node.
NVIDIA Open-H3 Regenerate 2K increases detail in MiniMax H3 video and also accepts material from LTX2 and NVIDIA Sana Video. Although the name emphasises 2K, the presenter describes several output resolutions up to 4K and clips lasting up to fifteen seconds. The code is already available. His own tests revealed an important limitation: the system changes some details, especially characters’ faces. The additional detail therefore comes with changes to the content of individual regions, rather than complete preservation of the source video.
StreamVibe is a video agent that offers help proactively. It watches a video stream, relates what it sees to the user’s goal and chooses when to offer a suggestion. Examples show someone preparing a tortilla or trying to book a flight. The instruction can be changed during the task, for example to ask it to intervene only when a mistake occurs. It considers the current request, progress on the task and its own previous responses when deciding whether to speak or remain silent. The presenter highlights an advantage in identifying user intent in the tests shown. At the time of recording, code and models are awaiting internal approval, with publication promised in seven days.
The MUNDO world model builds an environment from one or several images that users can explore, modify and simulate. It first creates a simplified three-dimensional scene, then describes individual objects, including mass and friction, while a physics engine calculates movement and collisions. Demonstrations set a camera route, have an apple knock over a monitor and train humanoid robots. As the presenter explains, changes to objects and actions performed on them follow the physics engine’s calculations. A repository exists, but code and weights are only promised by the end of November.
GPT 6.1 Terra is presented as a model between the fast, compact Luna and the slower, more capable Sol. The emphasis is on cost, efficiency and operating computer applications; in one example Terra builds a tram in Blender. In Agent’s Last Exam, which assesses professional tasks, the results shown put it at the level of larger Sol models at a lower cost. The episode also highlights image-based problem solving and a reduction in factual errors in the chart presented. On Terminal Bench, a test of coding through a terminal, Terra trails Claude Sonnet 5.5 in performance while remaining cost-effective. On the Artificial Analysis chart, it is a few points below Sonnet and GPT 6.1 Sol. The presenter describes Sonnet 5.5 as roughly ten times more expensive than Terra. According to the episode, Terra is available on paid GPT Work and Codex plans and through the API; he later mentions selecting the model in ChatGPT.
The presenter calls Gemini 4 Carbon Google’s new flagship. The tables shown indicate strengths in agentic coding, general-purpose tasks, computer use, professional work, handling different data types, mathematics, science and cybersecurity. It ranks first in Artificial Analysis, one point above Claude Opus 5.5, which in turn is one point above GPT-6 Astra. Its listed cost is below Opus but above Astra. On Omniscience, Carbon has the highest accuracy and the lowest rate of fabricated answers. The latter measure assesses whether a model admits it lacks an answer or invents one. It earns the highest balance on Vending Bench, managing a virtual vending machine. On Blueprint Bench, where models produce a floor plan from a room photograph, its result is almost equal to the human result. On Frontier SWE, however, it ranks third behind Astra and Opus 5.5. At the time of the episode, access is limited to an initial group of cyber-defence specialists.
NVIDIA Nemotron 3.5 Transcribe turns audio into text and lets users specify the kind of transcript: verbatim, edited, with a chosen number format or speaker timestamps. It supports forty languages, recordings up to an hour long, speaker separation and word-level timestamps. The system combines speech components with the Nemotron language model. It has about 34 billion total parameters, with 4.3 billion active parameters. Weights have been published in several formats: BF16 occupies about 67.5 GB, while the smallest FP4 version is under 22 GB.
Google Genie 4 turns text or an image into an interactive world. Users move with WASD or the arrow keys and use text to add objects, characters and effects or change the scene. Examples show realistic and fantastical spaces with responsive characters. A major addition is sound: ambient audio, music and interactive dialogue. A demonstration pairs it with Gemini 4 Argon to show a robot playing draughts with a person: Genie generates the simulation, while Gemini considers the next move. The presenter thinks the game appears correct, while emphasising that everything shown is generated. Access is restricted to trusted testers, with availability subsequently promised to subscribers on the highest-priced Ultra plan and early-access Gemini API and AI Studio partners. A waiting list is open.
PhysAg3D turns a photograph of an object into an interactive three-dimensional model with movable parts and physical behaviour. Demonstrations include a camera with a rotating lens, a phone with moving components and materials deforming under pressure. A separate model first reconstructs the geometry, then an agent examines how the object is built and determines how its parts move. Physics-simulation tools model and refine those interactions. The presenter cites an advantage in comparative tests of visual quality and realistic interactions. Code and the model have been released under MIT.
Google Sleep is an experimental language-model training framework. The presenter explains the problem: an ordinary conversation does not necessarily consolidate new knowledge in a model, while successive fine-tuning can displace earlier knowledge. In the first stage, the system works with users, performs tasks and temporarily stores information. In the second, it stops receiving new data and consolidates what it has accumulated into the model. The model also generates its own exercises, selects useful ones and trains on them to improve reasoning. A memory test reportedly achieves an almost perfect result on up to ten million tokens, the units into which a model divides text. The presenter compares this volume with roughly 700,000 words. The approach outperforms several other methods in that test. The episode also mentions better mathematical problem solving and learning new languages with less loss of earlier knowledge. The presenter sees potential for continual learning from experience, but explicitly calls the work a very early prototype.
Audio8 TTS Realistic is another open speech model, this time offering voice cloning from a few seconds of audio without separate training for the speaker. The episode includes examples of different voices being reproduced. The model has 0.6 billion parameters, supports 23 languages and occupies about 2.6 GB. Its Apache 2 licence permits commercial use. The presenter emphasises its small size, speed and suitability for an inexpensive consumer GPU.
SenseNova 6.8 Code to Video creates animated videos from text prompts by writing and executing code. The presenter compares the approach with Hyperframes and Manim: it constructs frames containing text, shapes and movement, suited to motion graphics and educational explanations. Code makes it possible to edit individual elements while retaining the others. Qwen 3.8 27B is named as the base, and the complete files occupy about 55 GB. The presenter points to the need for a powerful GPU for local inference. Instructions have been published.
A separate research story covers an energy-based image-generation model. In contrast to the diffusion approach described by the presenter, which predicts noise removal, this model learns an energy function that assesses how well an image matches the objective. Successive changes reduce that energy, producing an image. He immediately qualifies it as a proof of concept: the sample images are noticeably worse than those from current leading generators. Code and a training script are open, and the model occupies 2.44 GB.
Meta Overseer maintains a shared generated environment for multiple agents, coordinating their different viewpoints. In addition to each participant’s first-person view, it generates views from fixed cameras around the space. This helps track events outside an agent’s current field of view, maintain consistency and account for player locations. Demonstrations include Minecraft, Doom and scenes resembling live-action footage; the presenter clarifies that all of the videos are generated. The dataset has been released, but the model has not.
The closing run of open models begins with Nanbeige4.3-33B. It is a dense model with 33 billion parameters, without selectively activated expert blocks. Its files occupy 66.3 GB. On the Terminal Bench and Deep Suite results shown, it outperforms comparable Qwen 3.8 and Gemma 4 models, especially on the second test. The presenter also highlights professional tasks in IT, engineering, finance and healthcare, where it leads Qwen 3.8 in most of the comparisons. Code and weights are available, but the full version requires powerful hardware.
Liquid AI has introduced the small LFM3 family, with roughly two billion parameters per model. The presenter refers to six models and lists a text model; Vision for images and video; Audio for speech understanding, generation and translation between eighteen languages; Omni, which combines these capabilities; and Decision variants for fast decisions, such as whether a customer should receive a refund or an email is spam. In the tests presented, the text model outperforms some larger competitors. The family is hosted on Hugging Face; the base text version occupies about 3.5 GB. GGUF, FP8, FP4 and MLX for Apple Silicon are available, along with different versions of the other models.
Apex 35B is based on Qwen 3.6 35B and tuned for automated research and scientific tasks. In the demonstrations, it writes code to control robots in simulation: arms stack blocks and move objects. The model examines the result and fixes its code after a failed attempt. Another example reconstructs a physics simulation from a photograph. In the tests shown, Apex substantially outperforms the base model on research and agentic tasks. The full files occupy about 72 GB, while the smallest published lower-precision version takes up around 9 GB. The presenter warns that this compact version loses quality.
The Groot family has four versions, ranging from 19 billion parameters with three billion active to 280 billion with sixteen billion active. The largest was derived from DeepSeek 4.1: it is about half the size but scores somewhat below its source model. The smallest is based on Qwen 3.6 35B and, according to the comparisons presented, outperforms it in coding, automation and general knowledge at almost half the size. FP4, FP8 and GGUF versions have been published. The presenter gives about 366 GB for the largest Q4 GGUF model, and roughly 39 GB for the smallest full version and 12 GB for its Q4 GGUF version.
Clover 35B converts the Qwen 3.6 35B base into a diffusion language model. It generates and refines blocks of tokens, the units of text a model processes, rather than emitting them one at a time in sequence. In the example, it creates an interactive workspace with applications and draggable windows. In the agentic-work and tool-use tests shown, the presenter highlights strong results against other diffusion and conventional models of a similar size, as well as a favourable speed–quality trade-off. Weights are available, with a full size of about 69 GB.
Spark-X2.6-1.7B targets local computers and edge devices. The presenter explicitly distinguishes it from heavy flagships: the comparison is with small Qwen 3.5 2B, Gemma 4 and MiniCPM models. It performs better on most of the agentic coding, competition mathematics, long-context and general-knowledge tests presented. The model has 1.7 billion parameters. Speed is compared on a Mac M5 Pro and an RTX 4090, where he also notes an advantage. NVIDIA GPUs, Ascend NPUs and Apple Silicon are supported, and the model occupies about 3.4 GB.
Finally, SenseNova Loop 9B repeatedly passes through a substantial portion of its neural-network layers. This supplies more computation without a proportional increase in parameter count. The presenter calls it a recurrent multimodal architecture, intended for different types of input, and presents the suggestion that Claude and GPT use similar mechanisms strictly as a rumour. On the agentic tests shown, the nine-billion-parameter model is comparable with some much larger models. Code and weights have been published under Apache 2, with a total file size of about 38 GB.
