Wednesday, 4 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.
- of Google's search volume
- 20%
- press event for the claim
- Monday
- what OpenAI counts
- Web search
- Fry says usage held up
- Ads
ChatGPT search volume is a fifth of Google's
OpenAI says web searches through ChatGPT now equal 20 percent of the volume on Google. Adam Fry, its consumer product lead, said that at a Monday press event and did not name where the Google figure came from.
Why it's the story of the dayA chatbot already running a fifth of Google's searches is making room for ads in the answer.
Startup of the day
A young company people are talking about today.
Traversal raises $120 million to debug live systems
Traversal builds agents that investigate outages in production software. It raised $120 million at a $1 billion valuation, co-led by Kleiner Perkins, Redpoint and Khosla, for $173 million in all.
Why it's worth a lookA young company is selling an agent that digs through a firm's own outages.
Business Wire ↗Models & labs
New models, the labs behind them, and the rules they answer to.
GPT-6.1 Terra is now on Amazon Bedrock
Amazon Bedrock now offers GPT-6.1 Terra, the everyday line of OpenAI's GPT-6 family. On OpenAI's Agents' Last Exam it scored 15 points above GPT-5.6 Terra, level with GPT-6 Astra, and over three times faster.
Why it mattersA faster, cheaper GPT-6 model is now on a cloud companies already use for agents.
AWS ↗Palantir's US commercial sales pass government work
TechCrunch said Palantir's quarterly revenue rose 96 percent to $2.31 billion, its first quarter above $2 billion. US commercial revenue was $984 million, passing US government revenue of $953 million for the first time.
Why it mattersA software firm is now selling more AI tools to companies than to Washington.
TechCrunch ↗Apple posts a record quarter and flags memory costs
The Verge said Apple had its best fiscal year, with iPhone revenue up 17.5 percent to $57.6 billion. It spent $2.2 billion on capital projects and warned memory costs will squeeze holiday margins.
Why it mattersThe phone maker is selling the devices people use for AI without matching rivals' build-out.
The Verge ↗Siri AI can act inside third-party apps
The Verge said Apple's iOS 27.3 developer preview lets Siri AI take actions in third-party apps, using personal context and the screen. Developers can package several steps as one app skill.
Why it mattersA voice assistant can carry a request through apps a person did not name.
The Verge ↗Ollama applies a model's own cap in Claude Code
Ollama 0.40.14, released 3 November, applies a model's own output limit in Claude Code instead of a 128,000-token default. If the client sends no temperature, that model's sampler is kept.
Why it mattersA local runner stops a coding tool from cutting a cloud model off too early.
GitHub ↗Ollama's preview build adds MiMo V2.6
Release candidate 0.40.15, also published 3 November, adds support for the MiMo V2.6 model. On Linux it loads bundled CUDA libraries inside the process for its MLX path.
Why it mattersA new model can be served locally before the stable Ollama release lands.
GitHub ↗OpenAI's Python SDK adds workload identity rules
openai-python 3.44.0, released 3 November, adds workload identity providers and federation rules to the API. It follows 3.43.0, published the same day, which widened Live Responses tools.
Why it mattersA script can prove which workload is calling the API, not only which key.
GitHub ↗OpenAI's SDK expands Live Responses delegation
openai-python 3.43.0, released 3 November, expands delegation tools for Live Responses and accumulates richer progress over a WebSocket. A patch the day before had preserved existing WebSocket connection behaviour.
Why it mattersAn agent can hand work off and still show progress while the socket stays open.
GitHub ↗llama.cpp learns the GLM-5.3 tensor layout
Build b12077, released 3 November, adds support for GLM-5.3's next small-model tensor. It is one of several same-day builds, separate from the multi-node work in a later tag.
Why it mattersA new open model can load in a runner people already use on their own machines.
GitHub ↗Agents & coding
AI that acts on its own, and the tools developers build with.
CrowdStrike agrees to buy two agent-security startups
TechCrunch said CrowdStrike will buy Fiddler, which watches agents, and Dreadnode, which tests them before launch. It put the prices at about $400 million and $95 million, citing The Wall Street Journal.
Why it mattersA security giant is buying the tools that test agents before employees turn them loose.
TechCrunch ↗Goldman limits Claude Code and GitHub to engineers
Quartz, citing Bloomberg, said Goldman now limits Claude Code and GitHub to engineers after bankers had used them for months. The bank cited productivity and risk. Strats are exempt.
Why it mattersA bank that let non-engineers vibe-code is pulling those tools back.
Quartz ↗A failing Claude hook can block the action
Claude's agent SDK 0.3.317, released 3 November, lets a hook matcher set onFailure to block, so a hook that throws or returns bad output stops the action. It also warns in the stream when a hook callback fails.
Why it mattersA safety check that crashes can now halt the agent instead of being skipped.
GitHub ↗OpenHands remembers the cloud repository you picked
OpenHands 1.35.0, released 3 November, can remember and clear the cloud repository selected on the home screen. Automations can store a Bitbucket Data Center repo as a clone URL.
Why it mattersAn agent can return to the same repo without being pointed at it again.
GitHub ↗Open WebUI opens a tabbed workspace beside the chat
Open WebUI 0.13.0, released 3 November, adds a workspace of tabs for files, several terminals, artifacts, previews, voice calls and comments. New files can be saved into the selected terminal.
Why it mattersA chat can keep its working files and terminals open next to the reply.
GitHub ↗Cloudflare's agent harness moves to Pi 1.5
Agents 0.33.0, released 3 November, runs its Pi harness on Pi 1.5 and batches streamed answers about once a second. Cloudflare said that cut SQLite writes about 27 times on long answers.
Why it mattersA hosted agent can stream a long answer without writing to disk on every token.
GitHub ↗LangGraph can rejoin a stream that closed early
LangGraph SDK 0.4.12, released 3 November, rejoins a run stream that closes before the run ends. It also drops fairness_weight from priority settings.
Why it mattersA dropped connection no longer has to mean the agent run is lost.
GitHub ↗Obsidian Copilot can point an agent at a selection
Copilot 4.0.18, released 3 November, lets an agent take a selection from a drawing, a PDF or a diff, and lets you revoke its changes. A sidebar preview lists project and regular chats.
Why it mattersAn agent in a notes app can be shown the exact passage, not the whole file.
GitHub ↗Qwen Code adds a sidebar for web shell sessions
Qwen Code 0.26.6, released 2 November, adds a Web Shell sidebar for trusted workspace conversations, with approval, input and error states. Sessions can be opened side by side from that sidebar.
Why it mattersSomeone running several coding agents can see which ones are waiting.
GitHub ↗Video generation
The models and tools that make video.
Synthesia can build a video inside ChatGPT
Synthesia said its ChatGPT plugin, live on 3 November, turns a description or a document into a video in the chat. Videos use its Syren format by default, with graphics, avatars, voiceover, music and b-roll.
Why it mattersA person can ask for an explainer and get a finished video without leaving the chat.
Synthesia ↗any-llm can ask Together and Gemini for video
any-llm 1.42.0, released 3 November, adds video generation through Together and Gemini. The same release turns embeddings on for OVHcloud.
Why it mattersOne library can request a video from two providers without a separate client.
GitHub ↗InvokeAI 6.14.4 stops a crash after the last update
InvokeAI 6.14.4, released 3 November, fixes a startup crash that hit people upgrading from 6.14.2 to 6.14.3. The project blamed a bug in the Starlette library and capped that dependency.
Why it mattersA video studio that would not boot after yesterday's update can start again.
GitHub ↗llama.cpp can speak with a pocket voice model
Build b12061, released 2 November, lets its multimedia tool use Pocket TTS voices that do not clone a speaker. It is a voice add-on, separate from the model-layout builds the same week.
Why it mattersA local runner can read text aloud without copying someone's voice.
GitHub ↗A captioner plans what in a video deserves a look
Reviso, posted 2 November, builds a plan from a low-frame-rate preview: the context, the important parts, and the points to inspect. The authors say captioning models otherwise linger on minor details.
Why it mattersA description of a clip can follow the event, not every background object.
arXiv ↗A bench times tiny actions in first-person video
EgoActionBench, posted 2 November, marks continuous hand work as short actions tied to the left hand, the right hand or both. It has 1,407 episodes and 41,132 human-refined labels.
Why it mattersModels that learn from head-camera footage can be scored on the hands, not a vague caption.
arXiv ↗A game engine records synthetic play for vision models
Isekai3D, posted 2 November, simulates 3D play in Unreal, records the world state, then replays it to capture video, images and depth. Simulation and capture are separate steps.
Why it mattersResearchers can film the same scene again at a new resolution without resimulating it.
arXiv ↗One student model learns sound and picture together
OmniOPD, posted 2 November, distils separate audio and video teachers into one model that generates both. The authors say audio gets sparse supervision because it has far fewer tokens than video.
Why it mattersA clip with matching sound is hard to train when the picture dominates the loss.
arXiv ↗Events plus one frame rebuild a fast HDR clip
HIVE-Diff, posted 2 November, reconstructs high-frame-rate HDR video from a single ordinary frame plus an event camera. Event cameras record brightness changes quickly across a wide range of light.
Why it mattersA camera that cannot do both speed and range might still recover both with a sensor that sees change.
arXiv ↗Creator tools
The apps and platforms people make video with.
iZotope Photon splits a finished mix on your machine
iZotope released Photon on 3 November, an app and plugin that separates stems and cleans audio with 12 models running on a local GPU. It can split music, voices, effects and up to six speakers without an upload.
Why it mattersA locked mix can be pulled apart on a desktop, with nothing sent to a server.
iZotope ↗Unsloth adds glass themes and CPU training
Unsloth 0.1.914-beta, released 3 November, adds five glass themes and controls for greetings, backgrounds and code fonts. It also adds CPU text training from the Train page and more multi-GPU options.
Why it mattersPeople can train a text model on a CPU and restyle the app without a new install path.
GitHub ↗Ultralytics plotting gets much faster
Ultralytics 8.4.201, released 3 November, says PIL plotting is up to 15 times faster, including training mosaics. It also standardises the name of its Construction-PPE dataset, described as 1,000 images in 12 classes.
Why it mattersTraining charts draw faster, which matters when a run is checked often.
GitHub ↗Ultralytics draws Arabic class names correctly
Ultralytics 8.4.200, also released 3 November, says Arabic class names now join and read right to left in prediction and training plots. The notes say no model architecture changed.
Why it mattersLabels in Arabic stay readable on the images a detector draws.
GitHub ↗Supervision rejects a class id the set does not list
Supervision 0.31.4, released 3 November, raises an error when a CreateML or Pascal VOC export uses a class id outside the class list. Recall buckets follow COCO more closely, and a NumPy fallback sits nearer to OpenCV.
Why it mattersA bad label fails at export instead of shipping in the dataset.
GitHub ↗The AI SDK keeps a tool's metadata on approval
AI SDK 7.0.162, released 3 November, restores server-defined tool metadata, status and title when it checks an approval continuation. Approval policies then see the same tool fields as a new model call.
Why it mattersA person approving a tool call is judged against the real tool, not a stripped copy.
GitHub ↗FastMCP keeps tool routes after the group closes
FastMCP 4.1.6, released 3 November, keeps a ClientGroup's known tool routes after the group context exits. A resolved tool can then be called with only its own client connected.
Why it mattersA tool found through a group of servers can still be used once that group is closed.
GitHub ↗Prime uploads a run without blocking the caller
Prime 0.19.3, released 3 November, uploads runs and traces concurrently without blocking. The trace client line limit rises to 1 GiB, and sandbox exec retries transient faults.
Why it mattersA long training log can upload while the job keeps going.
GitHub ↗Prefect's dev CLI keeps literal markup in JSON
Prefect 3.9.2.dev3, published 3 November, preserves literal markup and emoji aliases in CLI JSON and YAML. It follows a dev build that returned 422 for bad request bodies.
Why it mattersA workflow log can be copied as data without the markup being interpreted.
GitHub ↗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.
Netflix tests an AI Discover tab on TVs
Variety said Netflix is testing AI Discover on TVs for some English-speaking US viewers. Themes such as Something Light return titles from watch history and trends. Kids' profiles are left out.
Why it mattersThe streamer is putting an AI label on suggestions it used to make quietly.
Variety ↗Premiere mobile will add licensed Universal tracks
Universal Music said on 3 November that Adobe will bring licensed clips from Universal and Warner into Premiere mobile. Editors will be able to add up to 60 seconds of a track, for personal, non-commercial use.
Why it mattersA phone editor can drop a known song on a clip under a licence, not a guess.
Universal Music Group ↗CAA tells the court AI needs consent and pay
Deadline said CAA filed a brief in the New York Times case against OpenAI. Chief executive Bryan Lourd said AI's future must be built with artists, grounded in consent, control and compensation.
Why it mattersA major talent agency is asking the court to treat training on creative work as something to license.
Deadline ↗Utopai hires Disney's former AI leader
Variety said Utopai Studios named Hans Ku chief product officer. He had been Disney's vice president of AI and engineering, and will scale the studio's PAI production platform across its film and TV slate.
Why it mattersAn AI studio is hiring the person who connected Disney's teams with OpenAI during a short-lived character deal.
Variety ↗Wired finds hundreds of apps hiding deepfake tools
Wired said more than 400 apps and sites looked ordinary but could open a hidden tool for explicit deepfake images and video. Those apps had over 350,000 downloads, and Apple removed 19 after Wired asked.
Why it mattersTools that look like pet portraits can be a front for non-consensual images.
WIRED ↗Whisper World and Rembrand merge as WRLabs
The companies said on 3 November they are merging as WRLabs, joining in-content ads with Rembrand's video-editing models. Spaceback and MRFactory stay as further product lines.
Why it mattersAd placements inside shows and the model that edits those frames are now one company.
Business Wire ↗ElevenLabs scores a video with music and effects
ElevenLabs said Instant Soundtracks, updated 28 October, writes music and timed effects for an uploaded video. The API can return the music and effects as separate files, for up to 10 minutes of footage.
Why it mattersA clip can pick up a soundtrack and effects without a separate sound session.
ElevenLabs ↗A detector looks for violin technique frame by frame
VioTech, posted 2 November, labels violin techniques every 10 milliseconds without first cutting the audio into notes. The authors say soft bow onsets make note boundaries unreliable.
Why it mattersMusic software can mark a bowing style even when the notes blur together.
arXiv ↗Speech deepfake tests miss a replayed clip
ReplayFactor, posted 2 November, says detectors that catch digitally made fake speech often fail when that audio is played through a speaker and recorded again. The authors call replay a cheap practical attack.
Why it mattersA voice check that passes a file can still fail when the fake is played in a room.
arXiv ↗Chips & compute
Chips, data centres and the power behind them.
Data centres turn old jet engines into power plants
TechCrunch, citing the Times, said some AI sites burn natural gas in engines from retired jets while they wait for the grid. xAI has eight in Mississippi, and Crusoe ordered 13 for OpenAI's Abilene site.
Why it mattersThe build-out is so short of power that sites are improvising plants from aircraft engines.
TechCrunch ↗Taiwan weighs prison for unapproved AI chip exports
Asia Times said a draft bill would require approval to send listed chips and chip tools to mainland China. It proposes up to 10 years in prison, including when the goods go through a third country.
Why it mattersA legal gap that let servers reach China through other ports could become a criminal offence.
Asia Times ↗Prime can set a sandbox volume when it is created
Prime 0.19.2, released 2 November, lets a sandbox set its volume size at creation and resize it later. The following day's release moved on to uploads.
Why it mattersA training job can ask for the disk it needs instead of hitting a default.
GitHub ↗llama.cpp can split a model across machines with NCCL
Build b12079, released 3 November, adds NCCL so tensor parallelism can run across processes on more than one machine. The CUDA backend uses it when the RPC server has the library.
Why it mattersA model that does not fit on one GPU can be split across a small cluster.
GitHub ↗llama.cpp fixes an out-of-bounds NVFP4 lookup
Build b12073, released 3 November, fixes a Vulkan lookup that could read past the end of NVFP4 UE4M3 scales. The bug sits in the path that unpacks a compact number format.
Why it mattersA quantised model on Vulkan is less likely to read memory it does not own.
GitHub ↗llama.cpp OpenCL starts to support iq4_xs
Build b12069, released 3 November, adds an initial OpenCL path for the iq4_xs quantisation. The notes call it the first support, not a finished kernel for every op.
Why it mattersA compact weight format can start to run on OpenCL, not only on CUDA.
GitHub ↗llama.cpp speeds a SYCL exchange on Intel Arc
Build b12048, released 2 November, speeds ggml_sycl_comm_exchange on Intel Arc Pro B70 GPUs. A separate same-day change only cleaned up older matrix code.
Why it mattersMulti-GPU exchange on an Intel card spends less time moving data.
GitHub ↗llama.cpp WebGPU learns an arange op
Build b12066, released 3 November, adds ARANGE in the WebGPU backend. A range of numbers can be built on the GPU instead of on the host.
Why it mattersMore of a model's bookkeeping can stay in the browser's GPU.
GitHub ↗llama.cpp tiles a permute on Hexagon
Build b12064, released 3 November, tiles an HVX permute for Qualcomm's Hexagon. The change is aimed at the phone-class DSP, not at a data-centre GPU.
Why it mattersA rearrange of tensor data can use the vector units on a handset chip.
GitHub ↗Policy & safety
Laws, lawsuits and safety incidents.
Google will switch off school AI until an admin opts in
Google said Gemini, Gemini Notebook, Gemini in Classroom and Search will stay off for K-12 students until an admin turns them on. Guided Learning is already global in Classroom.
Why it mattersSchools that do nothing will lose both the chatbot and Search on student accounts.
Google ↗An EU adviser says a chatbot answer can infringe the press
Advocate General Szpunar said on 3 November that a chatbot which fetches a press article and reuses it in an answer may infringe the publisher's right. The opinion, in Like Company v Google, is not a ruling.
Why it mattersEurope's top court is being asked whether a Gemini-style answer is a use of the article.
Court of Justice ↗Chatbots gave outdated voting answers in October tests
NPR said 13.9 percent of October answers on how to vote were outdated or wrong, and 5.9 percent could miss a deadline. Claude and Copilot made about three-quarters of the clear errors.
Why it mattersPeople asking a chatbot how to vote were sometimes given a deadline that had passed.
NPR ↗OpenAI will honour a shared training opt-out
OpenAI said on 30 October it will honour a common robots.txt signal for whether a site's content may train generative models. It aims to implement the signal in the coming months, and existing crawler rules still apply.
Why it mattersA publisher may soon state one training preference instead of a rule for every lab.
OpenAI ↗Japan's drug agency says AI must not replace the reviewer
PMDA published principles on 2 November for AI used in regulatory decisions on drugs and devices. The note sets no new requirement and says people keep the final judgment.
Why it mattersA regulator is writing the rules for AI in drug files before it treats them as binding.
PMDA ↗Haystack's stable release blocks an anyio bug
Haystack 2.31.1, released 3 November, requires anyio 4.14.2 or newer. The project says older anyio is affected by CVE-2026-63374, pulled in through httpx and the OpenAI client.
Why it mattersA common agent toolkit is forcing a dependency fix on the release people install.
GitHub ↗Two labs can check privately whether to pool their data
A paper posted 2 November studies when two parties should combine sensitive datasets before training. Even if each set is clean on its own, conflicting labels can mean no single model fits both.
Why it mattersSharing training data is a decision a lab may want to test without revealing the data.
arXiv ↗Voice models hear frustration and may not use it
A paper posted 2 November separates whether frustration is in the words, whether a voice model notices it, and whether the reply changes. The test uses 1,412 clips from 353 tasks, scored on eight models.
Why it mattersA steady answer can hide that the model heard the tone and ignored it.
arXiv ↗Robotics
Robots, self-driving and AI in the physical world.
Tutor raises $75 million to field more robot workers
Tutor said it raised a $75 million Series B led by Andreessen Horowitz, for $117 million in all. Sonny and Cassie work in factories with remote tutors. It wants over 1,000 robots under contract by the end of 2027.
Why it mattersA robotics startup is funding a fleet that learns from the jobs customers already pay for.
Business Wire ↗OpenAI had a model design a small device
OpenAI said Astra designed a device for talking to Codex, including the case, board and software, with no direct human edits. Two people did it in under a week. A contractor built prototypes in under a month.
Why it mattersA lab is testing whether a model can carry a gadget from a sketch to a small production run.
OpenAI ↗A wheeled humanoid is built around the data it learns from
Pixel, posted 2 November, is a low-cost wheeled humanoid with human-like joints and soft grippers. Demos can be collected with or without the robot, and the hardware was designed with the learning stack.
Why it mattersA robot's body and its training data are being drawn up as one system.
arXiv ↗Human demos are retargeted onto a mobile robot
Pixel-UMI, posted 2 November, turns sensorised human demonstrations into whole-body robot motion. It uses the reconstructed room plus the head and gripper paths to pick a posture and avoid collisions.
Why it mattersFootage of a person doing a chore can become a robot trajectory that fits the room.
arXiv ↗A planner mixes language with classical robot search
TAMPAgent, posted 2 November, pairs a vision-language model with task-and-motion planning for mobile robots that cannot see the whole scene. Direct tool calls, the authors say, struggle once geometry is uncertain.
Why it mattersA robot in a real room can fall back on a planner when the model is guessing.
arXiv ↗A checker picks a robot plan before it moves
WISE, posted 2 November, scores imagined robot futures and runs the best one while the model stays frozen. With four tries, RoboLab success rose from 6 percent to 24 percent.
Why it mattersA robot can discard a plausible-looking plan that its own action would not produce.
arXiv ↗A vision model scores robot progress without assuming success
D&A, posted 2 November, separates describing a robot video from scoring it, so a model is less likely to invent progress on a failure. Plain prompts, the authors say, collapse into a yes or no.
Why it mattersA reward signal can notice a partial failure instead of calling the episode done.
arXiv ↗Point-cloud models adapt without the original training set
3D-BBDA, posted 2 November, adapts a 3D recognition model to a new sensor or place without the source data or the source weights. Many older methods need one or the other, which licences often withhold.
Why it mattersA robot can retune its 3D vision when the lidar changes, without the vendor's dataset.
arXiv ↗Science & health
AI in biology, medicine, climate and maths.
DeepMind's agent searches compounds and sketches a route
DeepMind said new chemistry skills let an Antigravity agent search compounds, compare pairs and propose macrocycles with routes. A lung-cancer example finished that pass in about an hour.
Why it mattersA chemist can hand the literature trawl to an agent and still get a route, not only a paragraph.
Google DeepMind ↗An agent beats specialist models on molecule ranking
Intology said Locus, at the same compute, ranked compounds better than Boltz-2.1, Claude Code and Codex. On six targets its mean Spearman correlation was 0.622, against 0.225 for Boltz.
Why it mattersA general research agent is being scored on potency and antibodies, not only on code.
Intology ↗A benchmark asks an agent to dose insulin
DiabAgent, posted 2 November, is a research suite for controlling blood glucose, with a simulator and a benchmark called SweetBench. The agent doses insulin from meals, exercise, illness and glucose readings.
Why it mattersMedical models can be tested on a running intervention, not only on quiz questions.
arXiv ↗Facts seen once in training track later discoveries
A paper posted 2 November ties discoveries to monofacts, facts seen once in training. In materials and genomics, more of those facts yielded more discoveries in two open model families.
Why it mattersWhat a model can find may depend on rare lines in the data, not only on scale.
arXiv ↗A generator draws the image a neuron prefers
Vis-Lens, posted 2 November, uses an autoencoder and a diffusion model to draw images that strongly activate a chosen brain region or network unit.
Why it mattersResearchers can see a preferred stimulus instead of only a list of activation numbers.
arXiv ↗Crop checks are squeezed onto a device in the field
PlantVLM-Edge, posted 2 November, studies quantised vision-language models for diagnosing crop health on hardware that cannot hold a full model. The authors say timely diagnosis often needs an expert who is not there.
Why it mattersA phone or box in a field could read a leaf without calling a large server.
arXiv ↗Lab videos are turned into virtual experiments
Lab2Sim, posted 2 November, is a benchmark of 218 science-experiment videos with 11,881 rubric items. A companion system, EduVideo2Code, tries to turn the filmed steps into an interactive simulation.
Why it mattersA recorded experiment could become something a student can rerun, not only watch.
arXiv ↗A small Turkish speech model is trained from scratch
ufakzeka-kulak, posted 2 November, is a 31.6 million-parameter Turkish speech model under Apache-2.0. It trained on 4,977 hours, 281 of them human-transcribed, and writes punctuated text.
Why it mattersA language with fewer giant models gets a small recogniser people can run themselves.
arXiv ↗Open source
Open-weight models and community releases.
KoboldCpp adds a YuE2 music model on device
KoboldCpp 1.124.2, released 31 October, lets its music tools load YuE2 through audio.cpp. One template is the music model alone. Another pairs it with a small language model.
Why it mattersA local chat app can generate a song without sending the lyrics to a hosted model.
GitHub ↗any-llm can fetch files from an OpenAI container
any-llm 1.43.0, released 3 November, can retrieve files from OpenAI and Azure containers. It is a separate release from the same day's video-generation support.
Why it mattersA wrapper can pull a file an agent wrote, not only the text of the reply.
GitHub ↗llama.cpp rejects a GGUF split that does not match
Build b12078, released 3 November, refuses to merge GGUF shards that do not belong together. A bad split fails instead of being written out as one file.
Why it mattersA model assembled from the wrong pieces is caught before it is loaded.
GitHub ↗llama.cpp allows a constraint after a gpt-oss channel
Build b12068, released 3 November, lets a constraint type follow the final channel in gpt-oss chat. The parser no longer stops at that marker.
Why it mattersStructured output can still be requested after that model's last channel tag.
GitHub ↗llama.cpp stops a cache from reporting zero rows
Build b12065, released 3 November, stops the key-value cache from returning zero rows for a sequence that has no cells. Callers were treating that zero as a real size.
Why it mattersA sequence with nothing cached is no longer reported as an empty but valid cache.
GitHub ↗Hermes ships a rollup with group-chat routing
Hermes Agent 0.22.7, released 3 November, rolls 174 merged pull requests into the installers and the cloud app. In that window, desktop group chats can opt into a lead, and cron jobs can post into a room.
Why it mattersA local agent can assign a lead in a group chat and post on a schedule.
GitHub ↗OpenAI's agents preview ships native wheels
openai-agents-python 0.24.0-alpha.6, released 3 November, publishes wheels with a native runtime for Linux, Apple silicon and Windows. The preview is on GitHub Releases, not PyPI, and needs no Rust toolchain.
Why it mattersA tester can install the agents library without building it from source.
GitHub ↗Wrangler can compress container layers with zstd
Wrangler 4.157.0, released 3 November, can compress container image layers with zstd when WRANGLER_CONTAINER_IMAGE_COMPRESSION is set. The default remains unchanged.
Why it mattersA container built for a Worker can opt into a smaller layer format.
GitHub ↗Pydantic AI falls back when a model refuses
Pydantic AI 2.65.0, released 3 November, tries the next model when one refuses and counts browser-use calls toward the usage limit. Old import names now point at their V2 replacements.
Why it mattersA refusal no longer has to end the run if another model is queued.
GitHub ↗Quick hits
Smaller stories worth a glance.
- llama.cpp patches a Q8_1 correction sumBuild b12081, released 3 November, changes the CUDA path that corrects a Q8_1 sum.GitHub ↗
- llama.cpp gathers CUDA tensor readsBuild b12080, also released 3 November, gathers batched tensor reads into mapped host memory.GitHub ↗
- llama.cpp fixes a Metal copy kernelBuild b12076, released 3 November, fixes a copy kernel when the destination is not contiguous in the first dimension.GitHub ↗
- E2B stops parallel uploads from resettingThe Node SDK 2.62.2, released 3 November, fixes a fetch failure when many file writes share one HTTP/2 session.GitHub ↗
- E2B reads large sandbox files over HTTP/1.1The Python SDK 2.62.1, released 3 November, downloads files over HTTP/1.1 after HTTP/2 failed on multi-megabyte reads.GitHub ↗
- The AI SDK accepts buffers from another realmAI SDK 6.0.322, released 3 November, accepts a Uint8Array or ArrayBuffer created in another JavaScript realm.GitHub ↗
- llama.cpp drops a redundant mask stepBuild b12072, released 3 November, removes a redundant conversion from the attention mask path.GitHub ↗
- New Cloudflare projects use the 3 November datecreate-cloudflare 2.74.8, released 3 November, defaults new projects to the compatibility date 2026-11-03.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.
Bessemer Venture Partners backs Fal
- Raised
- $1.3B
Fal now valued at $16B
TechCrunch said Fal raised $1.3 billion, led by Bessemer, doubling its valuation to $16 billion. A Series E and this round together came to $1.65 billion.
NvidiaSVSound VenturesTRT. Rowe Price+0 more
Acquisitions 1
One company buys another outright.
Salesforce buys Cresta
- Price
- Undisclosed
Cresta is now part of Salesforce
Salesforce signed an agreement to buy Cresta, whose voice agents handle customer calls, and expects to close in the first quarter of fiscal 2028.
On the calendar
What's coming up, in date order, and how sure we are about each.
- 2–13NovProtoStar climate challengeToday
- 4NovApple Intelligence testing starts in ChinaToday
- 4NovAgility combination expected to closeToday
- 5–7NovThird LA Vertical Drama MarketTomorrow
- 5–6NovBoxWorks 2026 in San FranciscoTomorrow
- 5NovHogwarts Legacy adds DLSS 5Tomorrow
- 5NovAgility expected to start trading as AGLTTomorrow
- 6NovOllie assistant shuts down2days away
- 7NovTerry McMillan Presents: Paradise with You premieres3days away
- 9–20NovCOP31 climate conference in Antalya5days away
- 9–11NovThe Hinton Lectures on AI safety5days away
- 9NovCharmed By a Southern Monster premieres5days away
- 11NovJules becomes read-only7days away
- 12NovUpdated Claude usage policy takes effect8days away
- 15NovComments close on China's draft tech-export list11days away
- 20NovAI Film Award submissions close2weeks away
- 20NovUpdated OpenAI usage policies take effect2weeks away
- 21NovInaugural edition in Hollywood2weeks 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 release4weeks away
- 4DecGods Don’t Give Gifts premiere4weeks away
- 4DecSingapore AI Safety Fellowship ends4weeks away
- 31DecJules shuts down8weeks away
- Later in NovGuided Learning pilot in US schools
- Mid-NovXREAL Aura shipments begin
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.












