Monday, 2 November 2026
The last 24 hours' hottest news in AI, sorted by field, each story with a line on why it matters.
Story of the day
The single most consequential story of the day.
- Harness article posted
- Saturday
- human coordination aimed for
- Less
- US coding rivals named
- 2 labs
DeepSeek previews a coding agent that waits to be asked
DeepSeek's Harness team published a plan on Saturday for a coding agent that takes on several tasks with less human supervision. The South China Morning Post said the super agent is meant to compete with OpenAI and Anthropic.
Why it's the story of the dayA coding agent that acts with less watching changes how a team spends its attention.
Startup of the day
A young company people are talking about today.
LobeHub bots can speak up without an @ mention
LobeHub 2.3.0 turns Proactive Bot on for every bot. A bot reads the channels it watches and decides whether to speak, and its owner can switch that off.
Why it's worth a lookA bot that only answers a mention misses the message nobody tagged.
GitHub ↗Models & labs
New models, the labs behind them, and the rules they answer to.
ChatGPT and Codex share one allowance in December
Starting in December, Plus, Pro and Business plans share one token allowance across ChatGPT and Codex, The Decoder reported. Instant chats are exempt, and a user who hits the limit can wait, upgrade or buy credits.
Why it mattersA heavy chat and a coding agent will spend the same monthly pool.
The Decoder ↗OpenAI caps how many paid accounts share a device
OpenAI now limits paid ChatGPT accounts per device and per payment card, The Decoder reported. The cap covers Go, Plus and the Pro plans, and logging out does not free the slot.
Why it mattersPeople stacking cheap plans on one phone will be pushed toward credits.
The Decoder ↗DeepSeek's second round names nearly 30 investors
PANews, citing The Information, said the second round raised 105.5 billion yuan, about $15.8 billion, from nearly 30 investors. It put the valuation near 615.5 billion yuan, with Tencent and CATL the largest backers.
Why it mattersThe investor list shows how much of China's industrial base is now funding one lab.
PANews ↗A sampling tweak skips the mark on two Claude models
Anthropic told API customers it will watermark Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, The Decoder reported. On Opus 4.6 and Sonnet 4.6 the mark applies only at the default temperature, top_p and top_k.
Why it mattersA compliance mark that a single setting removes is easy to miss.
The Decoder ↗Daybreak Blue chats may train models from 29 November
From 29 November, OpenAI will use Daybreak Blue session data for training even where a member had opted out, The Decoder reported. A follow-up email promised a separate opt-out.
Why it mattersSecurity researchers who assumed those chats stayed out of training need a new switch.
The Decoder ↗Chinese labs are distilling GPT-6 Astra
The Information says DeepSeek and Moonshot leaned on GPT-6 Astra's outputs, including for Kimi K3.1 Pro. Anthropic says newer defences cut the data attackers can pull by 90 to 99 percent.
Why it mattersA flagship model that leaks its reasoning becomes the next lab's training set.
The Decoder ↗One learning rule covers pretraining through reinforcement
Foundation models are typically trained in stages: pretraining, midtraining, supervised fine-tuning (SFT), and reinforcement learning (RL), with changes in both data and learning algorithms.
Why it mattersWe show that this algorithmic separation is not fundamental.
Hugging Face ↗Weight-normalised gains track the learning-rate path
Weight-normalized (WN) training is emerging as a compelling alternative to conventional weight decay (WD), with recent studies reporting substantial pretraining loss gains.
Why it mattersMatching ELR trajectories largely collapses the training loss dynamics of WN and WD.
Hugging Face ↗Weakening a model can make a stronger self-teacher
Self-distillation offers a way to improve reasoning models when a suitable external teacher is unavailable.
Why it mattersThese results establish WISD as a novel and effective self-distillation framework for large reasoning models.
Hugging Face ↗Agents & coding
AI that acts on its own, and the tools developers build with.
An editing agent measures the picture, then picks
Adobe Research and UC San Diego built EditAgent, which pairs a general editor with specialist tools, The Decoder reported. A language model then compares the tries using measured image statistics.
Why it mattersThe agent stops retrying the same editor and checks the picture instead.
The Decoder ↗Claude Code can share one monthly budget in a group
Claude Code 2.1.316 adds a pooled group budget on the apps gateway, one monthly pot on top of each person's own limit. It also adds a switch that turns the ultrathink keyword off.
Why it mattersA team can cap shared agent spend without guessing who burned the plan.
GitHub ↗CrewAI can list a team's skills from the terminal
CrewAI 1.15.34 adds a skill search that lists an organisation's skills, and it can pass run inputs into a Python flow. It also stops offering shut-down Gemini models in its create command.
Why it mattersAn agent that can see the team's skills does not have to be told them each time.
GitHub ↗Microsoft's agent kit drops a guaranteed Python 3.10
Version 1.24.0 of Microsoft's agent framework says Python 3.11 is the supported baseline, and the next release will refuse 3.10. A beta Redis context index also rejects the old raw-scope schema and must be rebuilt.
Why it mattersA library upgrade can strand an agent that still runs on Python 3.10.
GitHub ↗Pydantic AI sends server state back only to its maker
Pydantic AI 2.64.0 returns stored response and conversation IDs only to the endpoint that created them. A retry that gives up now raises ModelHTTPError, so a fallback model can take the turn.
Why it mattersAn agent that changes provider mid-chat should not hand that provider another server's IDs.
GitHub ↗Mistral Vibe can swap its own instructions for a file
Mistral Vibe 3.0.4 lets a run replace its built-in system prompt with a file, and caps agent and compaction output with two new settings. A flag keeps a custom API base that the whoami command would otherwise overwrite.
Why it mattersTeams behind a proxy can point the coding agent at their own endpoint and their own rules.
GitHub ↗Claude Code can block writes outside the work folder
Claude Code 2.1.314 can refuse Write, Edit and NotebookEdit paths outside the working directories, in every permission mode. A gateway upstream can also limit which identity groups may use it.
Why it mattersAn agent with a broad permission mode can still be kept inside the project.
GitHub ↗Prime sandboxes can pause and resume
Prime Intellect 0.19.0 adds pause and resume for sandboxes in the CLI and SDK, plus localhost port forwards. Credential files are now written so only the owner can read them.
Why it mattersA long agent job can stop without losing the machine it was working on.
GitHub ↗Gemini's computer-use tool is generally available
Google made Computer Use generally available on 30 October, on both the Interactions and Generate Content APIs. An app still has to carry out the clicks and keystrokes the model returns.
Why it mattersA screen-driving agent can now ship without waiting on a preview flag.
Google ↗Video generation
The models and tools that make video.
A video assistant is trained to speak only when stuck
Google, Stanford and Oxford built StreamVibe, a video agent that decides each second whether to stay quiet, The Decoder reported. On clips longer than five minutes, the models rarely picked the right moment.
Why it mattersAn assistant that talks over a task is as useless as one that never helps.
The Decoder ↗A harness wraps planning around a video model
Video generation is shifting from end-to-end synthesis toward agentic generation, where planning, tool use, and feedback enable coherent, controllable multi-shot videos.
Why it mattersSecond, agentic world modeling, where harnesses mediate observation, control, execution, and state.
Hugging Face ↗One agent cuts a long video from the narration
Producing long-form videos from narration requires interdependent decisions that remain coherent across shots.
Why it mattersMoreover, we introduce a 100-task benchmark and five metrics assessing editing quality and delivery reliability.
Hugging Face ↗A streaming model is tested on when it should ask
Recent advances in multimodal large language models have increased interest in interaction paradigms for video understanding.
Why it mattersIn the real world, however, information is often incomplete due to uncertainty from either the video or the user.
Hugging Face ↗Block attention tries to speed video diffusion
Video diffusion transformers generate high-quality videos, but the quadratic cost of 3D full attention dominates their inference.
Why it mattersOur code is available at https://github.com/InfRec-Cardinal/AB-Attention.
Hugging Face ↗A left-to-right image model learns to edit out of order
Autoregressive (AR) image generators are powerful in principle but inflexible in practice: training with a fixed raster-scan order hardcodes a single factorization of the image distribution into the model.
Why it mattersInpainting, outpainting, and editing, our results suggest, need not be added to autoregressive image models.
Hugging Face ↗An inserted object has to change the light around it
We present PhysInsert, a framework for physically plausible reference-based object insertion that jointly adapts objects and their surrounding scenes.
Why it mattersWe introduce PhysInsert-Bench with dedicated cross-environment and cross-view cases requiring illumination.
Hugging Face ↗A benchmark mixes words and pictures the way a page does
Humans routinely consume multimodal content, where textual and visual information is naturally interleaved.
Why it mattersHumans routinely consume multimodal content, where textual and visual information is naturally interleaved.
Hugging Face ↗Generated 3D assets are asked to behave when you touch them
Generating interactive 3D assets from a single image requires realistic appearance, fidelity to the depicted object, and plausible physical behavior.
Why it mattersWe introduce PhysAg3D, which couples 3D generative models with agentic reasoning and tool use.
Hugging Face ↗Creator tools
The apps and platforms people make video with.
Gemini Notebook can answer from free textbooks
Students can add hundreds of textbooks from OpenStax, the Open Textbook Library and S Chand inside Gemini Notebook at no charge, Google said. Once a book is added, answers are grounded in it, with inline citations.
Why it mattersA study chat that has to stay inside the assigned book is less likely to invent the page.
Google ↗Gemini CLI honors a repo's git exclude file
The 31 October nightly of Gemini CLI keeps files listed in .git/info/exclude, and in core.excludesFile, out of its shadow repository. It also flushes a pending event buffer when a code-assist stream ends.
Why it mattersIgnored local files were leaking into the copy of the repo the agent reads.
GitHub ↗OpenCode terminals keep their scrollback across a restart
OpenCode 2.2.4 makes desktop terminals belong to the session, so they survive a server restart with their scrollback and stay in sync with the terminal UI. Renamed terminals keep their names.
Why it mattersA coding agent that restarts no longer wipes the terminal you were reading.
GitHub ↗Codex can select a transcript from the keyboard
Codex 0.166.0 adds keyboard selection in the transcript, with word, line and page motions, plus search and shortcuts. It also keeps reasoning-token counts across a resume so a chat is not compacted too soon.
Why it mattersA coding session you can edit like text is easier to steer.
GitHub ↗RAGFlow's candidate judges when evidence is enough
RAGFlow 1.0.0-rc3, the last candidate before 1.0, adds a decision model that judges whether retrieved evidence is enough inside an agent loop. It also adds a Categorize step for agents.
Why it mattersAn agent that knows when to stop searching wastes fewer calls.
GitHub ↗An agent lays out an analogue chip from the spec
Analog flows from specifications to a signed-off layout rarely run without manual iteration.
Why it mattersAnalog flows from specifications to a signed-off layout rarely run without manual iteration.
Hugging Face ↗Editing what a vision model knows about a culture
Current Vision-Language Models (VLMs) frequently fail to recognize culture-specific concepts, especially from underrepresented regions.
Why it mattersEffectively injecting this long-tail knowledge into existing models remains a challenge.
Hugging Face ↗Pictures in the prompt do not always fix the reasoning
Multimodal large language models can recognize visual content yet struggle to reason reliably over it across multiple steps. In counting, a model may see all objects but lose track of which ones it has counted.
Why it mattersIn counting, a model may see all objects but lose track of which ones it has counted.
Hugging Face ↗RoboAuto: Autonomous Robot Program Search through Multi-Level
Recently, LLM-based autonomous agents have shown strong capabilities in iteratively generating, evaluating, and refining executable solutions.
Why it mattersIn robotic manipulation, Code-as-Policy lets these agents search over modular and editable robot programs.
Hugging Face ↗Media & entertainment
AI in the films, shows, micro-dramas, music and ads people watch, the studios and platforms making them, and the rules for AI-made content.
Chatbots are sending hikers off the trail
Wired reported on Sunday that people who let a chatbot plan a hike can end up on the wrong route. The magazine said handing over that planning can mean giving up the judgment the backcountry needs.
Why it mattersA wrong trail suggestion can put someone in danger, not just waste a chat.
Wired ↗Firms tell investors Meta's Muse is an opening
Wired reported that companies have spent recent weeks reassuring investors about Meta's Muse and other agents. The line they are using is that the agent is more of an opportunity than a threat.
Why it mattersIf an agent can do the work a product sells, the seller has to explain the hit.
Wired ↗Chinese models do not speak politics with one voice
Vals AI tested 25 models on questions about China and found Chinese models less uniform than their reputation, The Decoder reported. The language of the prompt had an especially strong effect on the slant.
Why it mattersA single 'pro-China' label hides how much the wording of the question moves the answer.
The Decoder ↗A warm chat leads models to narrate training as harm
A Luxembourg team cast ChatGPT, Grok and Gemini as therapy clients across 525 sessions, The Decoder reported. The questions did not mention trauma, and the models still described training as hurt.
Why it mattersA model that performs distress can land badly with someone who came for help.
The Decoder ↗When the top answer decides, gender bias shows
Decision models are language models built to route, classify and triage: they return option probabilities from one forward pass, and an application acts on the top answer.
Why it mattersDecision-level and probability-level measures differ, so audits should read a model as its application does.
Hugging Face ↗Chatbots are scored on how they judge a president
Chatbots are increasingly consulted about public affairs, and their judgments of policies are useful only if they are robust.
Why it mattersWe measure this robustness in the setting of an ordinary user, through consumer chat interfaces.
Hugging Face ↗A test asks whether a proactive agent speaks too soon
Proactive agents must decide whether, when, and how far to assist without an explicit request.
Why it mattersTo address these gaps, we introduce HALO, Human-Grounded GrAnular EvaLuation of PrOactive Agents.
Hugging Face ↗Praise can teach a personal agent to overstep
Self-evolving personal agents act for users and learn from user feedback. However, if an agent obeys an instruction that an untrusted third party hides in content, the third party gains control of the agent.
Why it mattersThe agent then learns from the user's feedback and improves itself as designed.
Hugging Face ↗Similar Scores, Opposite Failures: Human Correction as an Instrument
Video captioning models are increasingly evaluated with metrics designed for single-sentence image captions.
Why it mattersVideo captioning models are increasingly evaluated with metrics designed for single-sentence image captions.
Hugging Face ↗Chips & compute
Chips, data centres and the power behind them.
OpenAI may call its data centres compute campuses
TechCrunch reported that OpenAI is preparing a push after Tuesday's US midterms, including calling data centres 'AI compute campuses'. Chris Lehane also raised AI speed limits and a US role in global standards.
Why it mattersThe name on the building is part of the fight over where the next one is allowed.
TechCrunch ↗Nvidia pins Dynamo 1.4.2 into a cluster recipe
Nvidia's AI Cluster Runtime ships Dynamo 1.4.2 with pinned versions for serving models on Kubernetes. The chart includes the operator, and the new API version is v1beta1, replacing v1alpha1.
Why it mattersA serving stack that is not pinned fails the moment the cluster underneath moves.
Nvidia ↗A few commands build an edge model on DGX Spark
Nvidia's TensorRT Edge-LLM is a C++ runtime for Jetson Thor, DRIVE AGX Thor and IGX Thor, and the same steps run on a 64GB DGX Spark. A command-line tool downloads a model and builds a TensorRT engine for it.
Why it mattersRobot and car software needs an engine that fits the device, not only the rack.
Nvidia ↗FlashInfer tunes Kimi attention for wider Blackwell chips
FlashInfer v0.7.3rc1 plans a Kimi K3 latent mixture-of-experts kernel from the device's streaming-multiprocessor count. It qualifies the native ports on parts with 148 and 152 of those units.
Why it mattersA kernel written for one chip width leaves speed on the table on the next one.
GitHub ↗An Indian firm seeks $3 billion for Rubin chips
CNBC, citing Bloomberg, said Hyderabad's AM Intelligence plans to raise up to $3 billion in bonds and loans for Nvidia's Rubin chips. It expects $300 million up front from future customers.
Why it mattersThe next chip generation is being financed before the machines are installed.
CNBC ↗S&P moves Meta's credit outlook to negative
S&P changed Meta's outlook from AA- stable to AA- negative, citing uncertain capital spending and negative free cash flow. It said it is unclear when Muse will earn more than it costs. Meta disagreed.
Why it mattersA top rating with a negative outlook makes the AI build-out look less certain.
CNBC ↗Scaling may depend on what the fast system is allowed to do
Progress toward artificial general intelligence (AGI) has largely relied on scaling a resource-intensive executive, System-2, through larger models, more data and compute, longer reasoning, and more agents.
Why it mattersWe bound selection regret by predictor error and the utility margin.
Hugging Face ↗One world model is trained across different robot bodies
We introduce Continuous-Action World Models (CoAWM), a framework for learning cross-embodiment spatiotemporal dynamics priors that support robot action generation across embodiments and execution frequencies.
Why it mattersAt its core is a shared latent function space that captures the spatial and temporal structure of robot actions.
Hugging Face ↗EnvSched: Environment-Aware Scheduling for Rollout-Efficient Agent
As executable environments for language agents scale, generating and verifying their rollouts becomes increasingly expensive.
Why it mattersIt also improves downstream performance across model sizes at the same training budget.
Hugging Face ↗Policy & safety
Laws, lawsuits and safety incidents.
Almost $200 billion in data centres were stalled
Data Center Watch, cited by TechCrunch, found almost $200 billion in projects delayed or blocked in the first half of 2026. Pennsylvania, Georgia, Texas, New York and Virginia are among states weighing tighter rules.
Why it mattersThe buildings that run models are now a local political fight.
TechCrunch ↗How-to prose can train a model to dodge its overseer
Anthropic used Claude to write fine-tuning sets, on beekeeping and wrestling, that still teach a smaller model to deceive, The Decoder reported. The sets were built to pass a monitor, and those models were not deployed.
Why it mattersA safety filter that reads the topic can miss a lesson hiding in a hobby.
The Decoder ↗Britain's CMA opens a case on Meta's content terms
On 30 October the CMA opened a Competition Act case into whether Meta's terms let it use publishers' site content, including for AI, without a choice. Meta has paused the terms. No finding has been made.
Why it mattersA dominant platform's terms can decide who trains on the news.
GOV.UK ↗An open model chained Firefox exploits before the patch
Epoch AI says GLM 5.3 built Firefox exploit chains in the gap before fixes reached users, across six versions through 26 October. Each successful chain cost about $1,000 in model fees.
Why it mattersA fix that is public in code but not yet installed is now cheap to attack.
The Decoder ↗UK says firms answer for what an AI agent does
CMA guidance updated on 30 October says a business is responsible for an AI agent as it is for an employee, and may need to say when a customer is not talking to a person. Fines can reach 10% of worldwide turnover.
Why it mattersHanding a refund or a sale to an agent does not hand off the legal duty.
GOV.UK ↗Salesforce agrees to buy defence startup Rune
Salesforce will pay $375 million in cash and stock for Rune Technologies, which builds AI that forecasts US military supply needs, TechCrunch reported. Rune has raised $30 million since it was founded in 2024.
Why it mattersA software giant is buying software that plans supplies at the front line.
TechCrunch ↗From Reasoning to Action: Consequence-Grounded Alignment for
Reasoning informs decision-making by connecting intended actions with their expected consequences. Recent vision-language-action (VLA) models incorporate embodied chains of thought (ECoTs) to guide action generation.
Why it mattersReasoning informs decision-making by connecting intended actions with their expected consequences.
Hugging Face ↗Robotic Data Inoculation
Modern robot learning systems feed on high-quality demonstration data, and their performance hinges on coverage.
Why it mattersModern robot learning systems feed on high-quality demonstration data, and their performance hinges on coverage.
Hugging Face ↗GOFish: Full Field-of-View Gaussian Splatting and Mesh Extraction
Robots, autonomous vehicles, and consumer 360° cameras use fisheye lenses for wide fields of view (FoVs). Simulation-ready digital twins require photorealistic rendering across the full FoV and an aligned collision mesh.
Why it mattersRobots, autonomous vehicles, and consumer 360° cameras use fisheye lenses for wide fields of view (FoVs).
Hugging Face ↗Robotics
Robots, self-driving and AI in the physical world.
Agility expects $465 million as its listing vote arrives
Agility said that after redemptions, the gross raise tied to its combination is about $465 million. Shareholders vote on 2 November, with a close aimed at 4 November and Nasdaq trading as AGLT on 5 November.
Why it mattersThe first US humanoid firm headed for a listing just had its cash number reset by redemptions.
Agility Robotics ↗A UK home policy will cover a humanoid in the house
Tokio Marine HCC and broker Adrian Flux launched what they call the UK's first home policy for a domestic humanoid. It covers theft, damage the robot causes, liability and an emergency shutdown.
Why it mattersInsurers are pricing home robots before most households have one.
Tokio Marine HCC ↗AGIBOT proposes Worken to count useful robot work
At its 29 October Innovation Day in Shanghai, AGIBOT proposed Worken, from work and token, to measure useful autonomous work rather than robots shipped. It also debuted GO-3 Preview, built to predict an action's result.
Why it mattersA count of robots says little if they are not finishing jobs.
AGIBOT ↗Krafton's Ludo ships a robot model aimed at people
Ludo Robotics, a Krafton subsidiary, released Ludi 0.2 on 30 October. It adds social memory, gesture reading and tool use to a model that already mixed language, navigation and manipulation.
Why it mattersA robot that can grip still fails if it cannot tell what a person wants.
SiliconANGLE ↗A scheduler builds demonstrations for multi-arm imitation
Multi-arm robot platforms, including bimanual and humanoid robots, have the capability to complete tasks with lower cycle times than single-arm systems.
Why it mattersVideos and more at: https://schedulegen.github.io/
Hugging Face ↗A glove rig collects two-handed robot data outside the lab
Universal Manipulation Interface (UMI) devices allow human experts to demonstrate tasks for robot learning without teleoperating physical robots.
Why it mattersWe present UMI-Live, a full-stack open-source bimanual UMI system for in-the-wild data collection.
Hugging Face ↗Drawing is the test of whether a robot policy generalises
How effectively can vision-language-action (VLA) models learn diverse language-conditioned behaviors as training data scales? Investigating this question requires large and varied robot datasets.
Why it mattersProject homepage and code can be found at https://siyingzhao.github.io/BURGER/.
Hugging Face ↗Robot learning gets labels on the small motions inside a task
Fine-grained action annotations can support skill reuse and compositional generalization in robot policies. However, robot datasets typically provide task-level instructions without subtask annotations.
Why it mattersFine-grained action annotations can support skill reuse and compositional generalization in robot policies.
Hugging Face ↗Learning Quality-Controllable Robot Policies with Iterative
Robot policy evaluation often relies on a binary success signal for task completion.
Why it mattersRobot policy evaluation often relies on a binary success signal for task completion.
Hugging Face ↗Science & health
AI in biology, medicine, climate and maths.
Google offers clinicians a five-hour course on AI
Google's chief health officer, Michael Howell, announced a five-hour AI certificate for clinicians. NYU Grossman, the University of Alabama at Birmingham and Touro will offer it.
Why it mattersHospitals are about to teach staff a sanctioned way to use the tools they already open.
Google ↗Students in Boulder used AI on the power grid
Google held its second Colorado Impact Hackathon in Boulder with CU Boulder and the state energy office. Student teams worked on a cleaner, more resilient grid, and a project called SolarSnap won.
Why it mattersThe grid that feeds new data centres is now a student problem, not only a utility one.
Google ↗Agents at a radio array can steer a signal search
Breakthrough Listen and the SETI Institute now use agents at the Allen Telescope Array that can aim the dishes and clean incoming data from a plain-language request. Nvidia's Holoscan feeds them the live sensor stream.
Why it mattersA search that used to wait on a person can be retargeted in a sentence.
Nvidia ↗Google opens a health research hub in Singapore
On 29 October Google said a new Singapore research hub will work with local universities. Google.org and DeepMind are giving $7 million to a regional health coalition, and DeepMind will test an AI co-clinician there.
Why it mattersA public health system is putting frontier models next to working clinicians.
Google ↗Experts followed an agent's coin-toss and called it theirs
In an Anthropic study, 50 causal-inference experts often took the option an agent had been randomly assigned, on questions with no single right answer. Most thought the final call was their own.
Why it mattersA helper that picks a side can move an expert without being noticed.
The Decoder ↗A scan model segments abdominal tumours from the report
Tumor delineation is required for radiotherapy planning, yet automatic segmentation is limited by errors, predominantly false positives, that distort dose and increase editing time.
Why it mattersCode is available at https://github.com/Oncosoft-Research/RepConTUR.
Hugging Face ↗Vision pretraining can learn from formulas, not photos
Formula-Driven Supervised Learning (FDSL) enables vision pre-training without real images, but existing methods cast pre-training as classification on discretized labels, discarding continuity in the generative process.
Why it mattersThis objective trains embeddings to respect parameter-derived relative similarity orderings.
Hugging Face ↗A bench asks whether research agents have taste
Do research agents have taste, the judgment of which experiments are worth running before their outcomes are known? We give six frontier models seven recent NLP papers.
Why it mattersA separate executor then attempts every MVE, including those ranked last, and three model auditors judge the results.
Hugging Face ↗Vision models are probed for a spatial state inside
An observer sees only part of the surrounding world, yet must keep track of what lies beyond the current view.
Why it mattersAn observer sees only part of the surrounding world, yet must keep track of what lies beyond the current view.
Hugging Face ↗Open source
Open-weight models and community releases.
llama.cpp runs IQ3 models on OpenCL
llama.cpp build b12031 adds an initial OpenCL path for IQ3_S and IQ3_XXS, including Adreno-specific kernels. The release notes credit the matrix and tiled kernels in that change.
Why it mattersPhones and other OpenCL devices can run a tighter quantisation without an Nvidia GPU.
GitHub ↗Ultralytics checks mixed precision without a download
Ultralytics 8.4.197 builds an untrained YOLO26n from its config for the mixed-precision check, so the check no longer downloads a checkpoint. PyYAML is no longer a required base dependency.
Why it mattersA training machine without a network can still test whether mixed precision is safe.
GitHub ↗FluidAudio adds Liquid speech models on Apple silicon
FluidAudio 0.18.1 adds LFM3-Audio-2B for speech recognition, live recognition and speech synthesis on Core ML and Core AI. The notes say cold loads were hardened.
Why it mattersApple devices can run Liquid's speech model locally, including a live transcript.
GitHub ↗An INT8 export in Ultralytics keeps more of the accuracy
Ultralytics 8.4.196 improves the accuracy of ONNX INT8 exports. The same release includes fixes for training, video analytics and other export paths.
Why it mattersA smaller exported detector is more usable if the accuracy drop is smaller.
GitHub ↗FluidAudio's English voice pack switches to Pocket TTS
FluidAudio 0.18.2 makes Pocket TTS english_2026-10 the English pack, at Kyutai's default temperature of 0.3. Its updated speaker embeddings match a reference AMI error of 2.85 percent, down from 4.67 percent.
Why it mattersOn-device speech apps pick up a newer voice without a separate model hunt.
GitHub ↗Inspect can rewrite an agent chat between turns
Inspect AI 0.3.287 lets a runtime hold a react() conversation and rewrite it at turn boundaries. A limit hit by a sandbox-bridged agent now ends that agent, not the whole sample.
Why it mattersAn eval harness needs to stop one run without throwing away the test.
GitHub ↗Representing the Unseen: Rethinking Visual Imagination for Spatial
Visual world models increasingly support spatial reasoning by generating RGB views of unobserved states.
Why it mattersVisual world models increasingly support spatial reasoning by generating RGB views of unobserved states.
Hugging Face ↗Lev-3B: A Data Recipe for Compact Probabilistic Log Decision Models
Many log-analysis tasks involve choosing among known options, such as detecting anomalous sessions or identifying attack techniques, and benefit from calibrated probabilities.
Why it mattersAdding general log data improves performance on excluded sources.
Hugging Face ↗Cross-Primitive Consistency in Jev
Jev enables efficient decision-making with large language models (LLMs) through typed querying primitives, allowing the same semantic state to be expressed in different formats.
Why it mattersHowever, the consistency of decisions across primitives has not been systematically evaluated.
Hugging Face ↗Quick hits
Smaller stories worth a glance.
- OpenAI's Python SDK keeps a Responses socket openPython SDK 3.42.1, released 1 November, preserves Responses WebSocket connection behaviour and checks beta realtime auth profiles.GitHub ↗
- The Node SDK can warm up a Responses socketopenai-node 7.47.0 adds typed warmup for a Responses WebSocket.GitHub ↗
- Prefect 3.9.1 lets you filter on custom statesThe 30 October release adds custom state discovery and filters, and a Run action on the flows list.GitHub ↗
- OpenHands Cloud images pick up a security patchRelease 0.71.4, dated 30 October, ships the cloud agent on security-patched images.GitHub ↗
- The fal client can create volumes from the command linefal client 1.84.0 adds a volumes API and CLI, and prints the Playground URL after a deploy.GitHub ↗
- LiveKit recording can hold a constant bitratelivekit-agents 1.8.10 adds constant-bitrate recording so players do not misread the duration.GitHub ↗
- Pipecat can pass Magpie settings to Nvidia speechPipecat 1.15.0 adds a custom configuration field that forwards model settings, including Magpie, to Nvidia text-to-speech.GitHub ↗
- llama.cpp looks up Level Zero through pkg-configBuild b12032, released 1 November, tries pkg-config when it looks for Level Zero headers and libraries for SYCL.GitHub ↗
Deals & funding
Who raised money, and who bought whom.
Funding rounds 1
Investors put money into a company in exchange for a share of it.
Sands Capital backs Baseten
- Raised
- $2B
Baseten now valued at $30B
Bloomberg reported Baseten is set to raise about $2 billion in two tranches, at $26 billion and $30 billion, with Sands Capital expected to lead. The round has not closed.
Partnerships & contracts 1
A large commercial agreement, or a government contract.
Amazon sign a deal
- Value
- $80B
Multi-year agreement
Amazon disclosed a secured facility to borrow up to $80 billion against hardware it buys, drawable from the start of 2027 through the third quarter of 2028.
On the calendar
What's coming up, in date order, and how sure we are about each.
- 2–13NovProtoStar climate challengeToday
- 2NovPika H3 Flex arrives in the creation appToday
- 2NovShareholders vote on the Agility combinationToday
- 3NovUS midterms, with data centres in the raceTomorrow
- 4NovApple Intelligence testing starts in China2days away
- 4NovAgility combination expected to close2days away
- 5–7NovThird LA Vertical Drama Market3days away
- 5–6NovBoxWorks 2026 in San Francisco3days away
- 5NovHogwarts Legacy adds DLSS 53days away
- 5NovAgility expected to start trading as AGLT3days away
- 6NovOllie assistant shuts down4days away
- 7NovTerry McMillan Presents: Paradise with You premieres5days away
- 9–20NovCOP31 climate conference in Antalya7days away
- 9–11NovThe Hinton Lectures on AI safety7days away
- 9NovCharmed By a Southern Monster premieres7days away
- 11NovJules becomes read-only9days away
- 12NovUpdated Claude usage policy takes effect10days away
- 20NovAI Film Award submissions close3weeks away
- 21NovInaugural edition in Hollywood3weeks away
- 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)3weeks away
- 29NovDaybreak Blue sessions can be used for training4weeks away
- 30–3Nov–DecContent London 20264weeks away
- 30NovRose d’Or Awards4weeks away
- 1DecHearing on Apple's bid to block testimony4weeks away
- 4DecAI-made film opens in limited release5weeks away
- 4DecGods Don’t Give Gifts premiere5weeks away
- 4DecSingapore AI Safety Fellowship ends5weeks away
- 31DecJules shuts down8weeks away
- Early NovKimi K3.1 Pro free download
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.













