Saturday, 14 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.
- US Pro subscribers at launch
- 18+
- or a fingerprint, on every purchase
- Face ID
- stands in for the card number
- Token
- US rollout while real checkouts are watched
- Limited
ChatGPT Wallet lets a ‘dot’ agent check out after you approve
OpenAI launched ChatGPT Wallet on Friday so one of its agents, which it calls dots, can help choose a purchase and then pay. It is starting with a limited set of US ChatGPT Pro subscribers, 18 and older, using a US-issued Visa, Mastercard or American Express. The person reviews the items, the total and the card, and approves with Face ID or a fingerprint. The charge uses a payment token, so the card number is not shared with the dot or the merchant.
Why it's the story of the dayA purchase only happens after the person approves that exact total.
Startup of the day
A young company people are talking about today.
LangChain raises $300 million at a $3 billion valuation
LangChain said it raised $300 million at a $3 billion valuation. Existing investor CapitalG led, with GIC and DFJ Growth joining and IVP and Sapphire Ventures participating. The company said more than 7,000 organizations use LangSmith, including Harvey, Rippling, Listen Labs, Nvidia, Toyota and Salesforce. Clay uses it to monitor hundreds of millions of agent runs a month. Workday uses it to run agents across the company, and said the time to deploy a new agent fell from five months to under a day.
Why it's worth a lookThe round is for software that runs and checks a company's own agents, without locking those agents to one model.
LangChain ↗Models & labs
New models, the labs behind them, and the rules they answer to.
Thinking Machines posts Inkling 2 Small weights
Thinking Machines said on 12 November that Inkling 2 Small's weights are on Hugging Face, and that the model is integrated with Unsloth, SGLang, vLLM and TokenSpeed, with an API at Nvidia, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. The same post describes a mixture-of-experts model with 594 billion parameters, 17 billion active, image and text inputs, and a score of 31 on the Artificial Analysis Intelligence Index. The page had said the weights would ship under Apache 2.0 after safety testing.
Why it mattersA US lab's open weights are now something a local stack can actually load.
Thinking Machines ↗Perplexity posts Decider v1.3 for routing and extraction
The model card for pplx-decider-v1.3 describes a multimodal classifier on Qwen3.8-27B, under Apache 2.0, with a new Locate step that returns verbatim spans, character offsets and confidence. It says a CUDA GPU needs room for about 49 GiB of BF16 weights, and that the repository is private, so a download needs a Hugging Face token with access.
Why it mattersA small decision model can label a request and point at the words it used.
Hugging Face ↗PEFT 0.21.3 gathers a sharded adapter only when it must
PEFT 0.21.3, tagged 13 November, gathers adapter weights only when saving a DTensor tensor-parallel model. The note says this avoids the penalty in 0.21.2, which gathered the whole adapter when Trainer called save_pretrained on every rank.
Why it mattersA sharded fine-tune can checkpoint without every rank pulling the full adapter.
GitHub ↗LangChain's OpenAI package adds a decision classifier
langchain-openai 1.8.0, tagged 13 November, adds OpenAIClassifier for the generally available Decisions API. The package now requires Python 3.11 and langchain-core 1.8.0. Installs still on Python 3.10 keep resolving 1.7.x.
Why it mattersA pipeline can ask for a scored choice instead of a paragraph that has to be parsed.
GitHub ↗OpenAI's Python SDK widens agent, image and Live support
openai-python 3.51.0, published 13 November, expands agent, image, Live and external-key support on the API. The release note lists that as the feature change from 3.50.0.
Why it mattersCallers can use the newer agent and Live surfaces from the same SDK.
GitHub ↗The Z.ai provider suggests model IDs from its catalog
The AI SDK's @ai-sdk/zai 3.0.55, tagged 13 November, builds known model-ID suggestions from each provider's catalog declarations. Catalog updates then show up in autocomplete, while existing suggestions, endpoint-specific lists and custom model IDs stay.
Why it mattersAn editor can offer the models a provider actually lists, without a hand-maintained enum.
GitHub ↗ArchSpace keeps a public log of architecture experiments
ArchSpace is a public place to record an architecture proposal with the experiments and the reason, have it reviewed, and let later work cite it. It uses scaling laws to estimate loss at a larger run from smaller ones, and it ships Qwen3.5 and OLMo3 baselines, fitted scaling laws and a reusable training recipe.
Why it mattersA lab can see which architecture ideas were dropped, not only the ones that shipped.
arXiv ↗A paper separates three choices bundled in a vision encoder
The authors train every variant from scratch on the same data and compare visual question answering and language scores at matched compute. A jointly trained encoder matches a pretrained one, so the upstream visual prior is not required. An encoder-free model still trails when the backbone is enlarged by the encoder's parameter count, and vision-only modules inside early layers can match a pre-fusion encoder.
Why it mattersWhat a multimodal model needs from vision is dedicated capacity, not a frozen pretrained eye.
arXiv ↗LangChain.js turns its typesafe classifier into a decision model
@langchain/typesafe 0.1.0, tagged 11 November, implements TypeSafeClassifier as a decision model. It now extends BaseDecisionModel rather than returning a free-form label.
Why it mattersA TypeScript agent can score a choice the same way the Python decision clients do.
GitHub ↗$9B
Cisco raised its expectation for AI infrastructure revenue from hyperscalers to $9 billion in fiscal 2027.
The company said that is more than double the fiscal 2026 total. First-quarter revenue was $18.5 billion, up 24 percent. Cisco ↗
Agents & coding
AI that acts on its own, and the tools developers build with.
ChatGPT Wallet lets a ‘dot’ agent check out after you approve
OpenAI launched ChatGPT Wallet on Friday for a limited set of US ChatGPT Pro subscribers, 18 and older, with a US-issued Visa, Mastercard or American Express. One of OpenAI's agents, which it calls dots, can help choose a purchase and then check out. The user reviews the items, the total and the card, and approves with Face ID or a fingerprint. The charge uses a payment token, so the card number is not shared with the dot or the merchant, and card rewards and dispute rights stay in place.
Why it mattersAn agent can pay only after a person has signed the specific purchase.
TechCrunch ↗Companies can host Claude's agent service themselves
The Next Web reported that Anthropic now lets customers run Claude Managed Agents in their own cloud or data centre. Agents, session histories and credentials stay on those systems, and tool calls can be logged. Each model call still goes out to a provider the customer chooses, such as Anthropic, Amazon Bedrock, Google Cloud or Microsoft Foundry. Access is requested through an account team, and webhooks and scheduled runs are among the features that do not come over.
Why it mattersThe harness can sit inside a company network even though the model call does not.
The Next Web ↗Expo runs phone simulators in the cloud for coding agents
Expo said on 12 November that cloud iOS simulators and Android emulators are available on paid plans, so an agent can reproduce a bug, test a fix and attach a screen recording, screenshots and logs to a pull request. Sessions work with Cursor, Claude Code and Codex, including on Linux and on Windows through WSL, with no Mac required. Pricing is $0.10 a simulator minute, and Starter, Production and Enterprise plans include 60 minutes a month.
Why it mattersAn agent can show the bug is gone without a developer plugging in a phone.
PR Newswire ↗OpenCode runs subagents in the background
OpenCode 2.3.0, tagged 13 November, returns from a subagent immediately, lets it keep working, and reports back when it is idle, so several can run while the parent session continues. The terminal now collapses each run of tool calls into one activity row. Verbosity can be set back to medium or high. In the desktop app, Jump to message searches earlier prompts in the session.
Why it mattersA coding session can keep moving while other agents finish their own tasks.
GitHub ↗browser-use drops old screenshots in one batch
browser-use 1.0.4, tagged 13 November, waits until 20 tool results carry a screenshot and then drops the old ones together. Pruning on every new shot rewrote earlier messages and broke the prompt cache. On one 705-step task in 1.0.3, the run read 27.8 million input tokens and only 1.6 million from cache. Against 1.0.3, BU Bench v2.1 scored 76.1 versus 76.0, at $0.208 of model cost per task versus $0.434.
Why it mattersA browser agent can keep its cache, and the bill, if it stops rewriting the past on every click.
GitHub ↗n8n lets an agent turn off prompt caching
n8n 2.48.3, tagged 13 November, lets agents turn off prompt caching for endpoints that reject the cache fields. The same patch keeps trusted-source cache keys in one Redis cluster slot, and reads legacy-format data with the instance key when the stored key differs.
Why it mattersA workflow can call a model that refuses cache fields without the run failing on the request.
GitHub ↗Qwen Code 1.0 preview installs without Node
Qwen Code 1.0 preview, tagged 11 November, combines Qoder and Qwen Code for reading, changing, debugging and testing a repository. It supports Qwen and Qoder accounts, several model providers and custom compatible endpoints, and it ships native packages for macOS, Linux and Windows so a separate Node.js install is not required. Sessions can resume, and the tool can run non-interactively. The note says interfaces may still change during the preview.
Why it mattersA Qwen coding agent can be installed as an app, not only as a Node package.
GitHub ↗Claude's agent SDK can archive a session
Claude Agent SDK 0.3.325, tagged 13 November, adds archiveSession and unarchiveSession, and an archived field on the sessions returned by listSessions and getSessionInfo. reloadPlugins gains refuseIfBusy, so a host can have a plugin reload refused while the session is still working.
Why it mattersA host can put a finished session away without losing the record of it.
GitHub ↗A paper asks if a weak decision model can still teach
Decision models return a typed answer, such as a choice and a probability, so the same model can pick an agent's action, score an action already taken, or label actions for reinforcement learning. The paper compares three decision models and three general language models in those roles, and asks whether one that acts poorly can still be a good teacher.
Why it mattersThe model that should not drive the agent may still be the one that grades it.
arXiv ↗Tool of the day
Something you can try in the next five minutes.
OpenCode 2.3
Subagents now return immediately, keep working in the background, and report back when they are idle. The terminal collapses each run of tool calls into one row unless verbosity is raised.
Video generation
The models and tools that make video.
World Labs opens the Atlas world model to everyone
Radiance Fields reported on 13 November that World Labs opened Atlas beyond its earlier partners, with Viewfinder, Chisel, Splats to Mesh, Relight and a developer API that returns SPZ Gaussian splats. A refined reconstruction takes 2 to 512 frames for 5,000 credits, which the piece prices at $5. A draft takes 2 to 32 frames for 500 credits, about $0.50. Runs usually take 15 to 25 minutes. In the documented example, eight frames of a potter's studio became 2 million splats in 18 minutes.
Why it mattersA world model that used to be a private preview now has a public price for a splat.
Radiance Fields ↗Gradio's workflow canvas gets a 3D model node
Gradio's @gradio/model3d 0.20.0, tagged 12 November, adds 3D model node support on the workflow canvas. The package ships with the same release train as Gradio 6.32.
Why it mattersA Gradio workflow can pass a 3D model through a node instead of leaving it as a file on the side.
GitHub ↗PixVid lifts a latent video model into pixel frames
PixVid keeps a pretrained diffusion transformer and replaces the VAE decoder with a convolutional head that draws each frame in pixels. The authors say predicting clean video, rather than velocity, plus a larger noise shift, is what holds the quality. Finetuned from Wan2.1-14B, it matches the latent baseline within about 1.3 million training clips, and it generates at twice the resolution with the same per-frame token count.
Why it mattersWorking in pixels lets the model answer the next frame instead of waiting on a compressed block.
arXiv ↗K2V lets a language model keyframe the physics
K2V moves the physical reasoning out of the diffusion sampler and onto a language-model step that proposes keyframes. The video model then interpolates between those frames instead of inventing the whole shot. The authors say this reaches state of the art on Physics-IQ Verified and PhyGenBench.
Why it mattersThe diffusion model only has to connect frames that were already supposed to be physically right.
arXiv ↗Test-time baking speeds up a view-synthesis decoder
Test-time baking trains the fast-weight MLP from a test-time-training layer to copy the input-output map of cross-view attention, instead of replacing attention outright. Context tokens are split so the baked MLP sees camera-only inputs and is trained against tokens that still contain the image. The authors report up to a 23 times rendering speedup with quality in line with the state of the art, and the work is a NeurIPS 2026 spotlight.
Why it mattersA view can be rendered from a small network that has already absorbed the scene.
arXiv ↗vid2vid-long copies correspondences into the new video
The paper says video-to-video models that keep a fixed causal window lose anything that falls outside it. vid2vid-long computes patch correspondences in the input video and uses them so regions that match in the source also match in the output.
Why it mattersA translated shot can stay consistent past the few frames the model can see at once.
arXiv ↗A paper matches only the end of a distilled diffusion run
Distribution-matching distillation trains a student to follow a teacher with the help of a fake score model. The authors say training that fake model on intermediate steps entangles the student's update with distributions the user never sees. eDMD instead matches the endpoint distribution, the frames that are actually delivered.
Why it mattersA few-step video student should be judged on the frame it shows, not on the noise in between.
arXiv ↗A neural proxy relights a CG shot while you watch
The paper splits rendering into path sampling and emission, and trains a neural render proxy so a static scene with a fixed camera and fixed materials can be relit interactively. The point is to avoid re-rendering the whole frame in an offline renderer for every lighting tweak.
Why it mattersA lighter can move a lamp without waiting on a full global-illumination pass.
arXiv ↗Moving traffic can drag a drone's stabilizer off the true frame
The paper defines consensus capture: on a busy intersection, vehicles that move together supply matching points that pull a robust homography toward their motion, and inlier counts do not flag it. A two-part benchmark measures the effect and the object masking meant to stop it, on 4K aerial frames with known homographies.
Why it mattersA hover that looks stable can be locked onto the traffic, not the ground.
arXiv ↗Korea's AI ad contest closes on 30 November
Maeil Business said entries for the third Korea AI Advertising Awards are due by midnight on 30 November, including generative-AI music videos for unreleased songs by HITGS and Genus.
The deadline is midnight on 30 November. Winners are announced on 16 December. Maeil Business ↗
Creator tools
The apps and platforms people make video with.
Vercel's UI SDK ships interface behaviors for agents
The UI SDK repository describes a beta library of interface behaviors, including keyboard navigation, focus, selection, dismissal, popups and drag and drop, with the ARIA those need, in @ui-sdk/core. Adapters cover React, Vue, Svelte, Solid and Web Components. The docs date the introduction to 12 November. A skill can be added so an agent writes the components.
Why it mattersAn agent can assemble a control from tested behaviors instead of inventing the focus handling.
GitHub ↗DSPy 3.5 graduates RLM, Flex and GEPA
DSPy 3.5, tagged 10 November, takes RLM, Flex and GEPA out of experimental. ReActV2 replaces ReAct, with a structured history and a typed submit tool, and the old name is deprecated until 3.6. Tools can return images and files on supported providers. The note says prompts, defaults and APIs change, and optimized programs should be evaluated again.
Why it mattersA program compiled on 3.4 can get different prompts after the upgrade, so the old score is not the new score.
GitHub ↗Sentry's Python SDK traces Bedrock Converse calls
sentry-python 2.75.0, tagged 12 November, adds support for Amazon Bedrock's Converse and ConverseStream APIs in the boto3 integration. The integration turns on automatically when boto3 is installed, and it can be passed explicitly to sentry_sdk.init.
Why it mattersA Bedrock call shows up as a span without a separate wrapper around the client.
GitHub ↗The AI SDK stops a replay from rewriting an approved tool call
AI SDK 7.0.173, tagged 13 November, stops in-place schema validation or input refinement from changing tool arguments that were already approved, when that approval is replayed. The same release logs the default step-limit warning when a WorkflowAgent stops at the default limit, and logs provider warnings from WorkflowAgent the way ToolLoopAgent does.
Why it mattersA second pass over an approved tool call keeps the arguments the person signed off.
GitHub ↗LangChain.js no longer crashes if a tool error is never read
langchain 1.6.6 for JavaScript, tagged 13 November, stops a failed tool call or subagent from crashing the process when nothing reads its output promise. A Command returned from wrapModelCall is also applied once when outer middleware passes it through.
Why it mattersAn agent can fail a tool call without taking down the process that forgot to await it.
GitHub ↗Inngest stops two runs from sharing a sleep id
Inngest 1.53.0, tagged 12 November, stops a newly scheduled sleep from colliding with another run that reuses an event key. Existing queued sleeps keep their ids. The same release stores lake JSON as text rather than a VARIANT, so Insights columns accept ->> and json functions.
Why it mattersA retry cannot wake the wrong run just because the event key was used again.
GitHub ↗WorkflowAgent treats a custom abort as an abort
The AI SDK package @ai-sdk/workflow 2.0.105, tagged 13 November, calls onAbort rather than onError when a WorkflowAgent call is aborted during a model step with a custom reason. generate still rejects with that reason. stream resolves with the steps that finished, as it does for an abort without a reason.
Why it mattersStopping a run on purpose is no longer reported as if the model had failed.
GitHub ↗CrewAI documents the error when an app is not connected
CrewAI 1.15.39, tagged 9 November, documents the apps_not_connected error returned by the kickoff and status endpoints.
Why it mattersA failed run can say the app was never connected, instead of failing as a generic kickoff error.
GitHub ↗WorkflowAgent can take abort and error callbacks up front
@ai-sdk/workflow 2.0.104, tagged 13 November, lets WorkflowAgent take onError and onAbort in the constructor, as ToolLoopAgent does. Those run before callbacks passed to generate or stream. An error thrown inside the callback no longer replaces the run's own error or result.
Why it mattersA host can log a cancelled workflow without the logger becoming the failure the caller sees.
GitHub ↗In Q1, we delivered the strongest revenue and non-GAAP earnings per share growth rates for Cisco in over 15 years.
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.
Google opens pre-orders for Gemini glasses
Google said on 12 November that pre-orders are open for intelligent eyewear made with Warby Parker, Gentle Monster and Samsung. Warby Parker frames start at $395 and Gentle Monster frames at $449. The glasses go on sale in stores and online on 10 December, for Google accounts in the United States and South Korea, and they require an 18+ account. Gemini is the hands-free assistant.
Why it mattersA pair of glasses can take the assistant off the phone, for buyers in two countries first.
Google ↗Viant connects campaign controls to Claude, ChatGPT and Cursor
Viant's third-quarter release, dated 9 November, said revenue rose 29 percent year over year and that connected-TV advertiser spend rose nearly 40 percent, to more than half of spend on the platform. It said direct integrations with Claude, ChatGPT and Cursor let advertisers manage campaigns in natural language, outside Viant's own interface.
Why it mattersA buyer can change a TV campaign from the chatbot they already have open.
Viant ↗Dior's director says the holiday film used interpolation, not generative AI
afaqs reported on 12 November that Sylvain Chomet, director of Christian Dior's holiday short Night in Color, denied that generative AI made the artwork. Producer Aton Soumache told Variety the film is traditional 2D animation and that interpolation, software that draws the frames between two drawings, was a very small part of the process. He said 20 artists, including six animators, worked for seven months. Dior's creative director had told WWD the film was drawn entirely by hand.
Why it mattersThe argument is over what the studio disclosed, not only over which software filled a frame.
afaqs ↗A Korean contest will make idol videos with generative AI
Maeil Business said on 11 November that the third Korea AI Advertising Awards, hosted by MindSpark, takes entries until midnight on 30 November. A new category asks people to make music videos with generative AI for unreleased songs by the groups HITGS and Genus, with 1 million won for the winning video of each song. The overall grand prize is 5 million won. Fourteen brands have set advertising challenges. Winners are announced on 16 December, and the ceremony is on 22 December in Seoul.
Why it mattersRookie idols are handing unreleased songs to outside creators who will finish the video with generative tools.
Maeil Business ↗ElevenAgents gets a Unity SDK
An ElevenLabs changelog entry for 9 November says the ElevenAgents Unity SDK is available as the package io.elevenlabs.agents, from v0.1.0. It supports the Unity editor, desktop players on macOS, Windows and Linux, and WebGL, with iOS, Android and XR in beta, and it requires Unity 6.3 LTS or later. A scene connects through one Conversation API. The agent's voice plays through an AudioSource that can be placed for spatial audio.
Why it mattersA game can put a voiced agent in a scene without a separate voice client.
ElevenLabs ↗Citations on generated answers often are not the source of the claim
The paper replaces a fact in a retrieved passage with a false variant and asks the model again. Across five sizes of one model family and a second family, 1.2 percent to 15.7 percent of citable claims still came from the model's own parameters rather than the passage it cited.
Why it mattersA footnote can point at a document the answer did not actually use.
arXiv ↗Viant buys Waymark to generate the TV spot
AdExchanger reported on 12 November that Viant bought Waymark, a generative video-ad tool, for $20 million. Media companies including Comcast, Fox, NBCUniversal, Spectrum Reach and Gray Television have used it to make more than a million local ads in the past year. Waymark turns brand inputs into a finished spot with a script, visuals and a voice-over. Chief executive Tim Vanderhook said that ten years ago no one would let a vendor touch the creative, and that the bet is that this is changing.
Why it mattersThe company buying the TV advertising also wants to be the one that makes the ad.
AdExchanger ↗Gradio's image editor says when the camera is blocked
@gradio/imageeditor 0.21.3, tagged 12 November, shows an error when camera permission is denied for the webcam, instead of failing quietly.
Why it mattersA visitor who blocked the camera sees why the preview is empty.
GitHub ↗ElevenLabs opens an early-access app for meetings and dictation
TechCrunch reported that ElevenLabs opened early access to ElevenAI, an experimental app that takes meeting notes, dictates into other apps, and can hand follow-ups to an agent, such as drafting an email from the notes. It notifies other participants that the meeting is being recorded, can answer questions across past meetings, and dictates in more than 90 languages, including industry terms. The company showed it at its 11/11 summit in New York. Access is a signup, not a public download.
Why it mattersA voice lab is selling the meeting notes and the dictation as one team product.
TechCrunch ↗Ten years ago, no one would even let you touch the creative. Our bet is that's changing.
Chips & compute
Chips, data centres and the power behind them.
Cisco lifts its hyperscaler AI target to $9 billion
Cisco said on 12 November that fiscal first-quarter revenue was $18.5 billion, up 24 percent year over year, with GAAP earnings of $1.14 a share, up 58 percent, and non-GAAP earnings of $1.40 a share, up 40 percent. It raised its expectation for AI infrastructure revenue from hyperscalers to $9 billion in fiscal 2027, more than double the fiscal 2026 total. The quarter ended 24 October.
Why it mattersThe networking vendor is telling investors the AI build is still in its early orders, not a one-quarter spike.
Cisco ↗llama.cpp tunes GCN matrix configs for HIP
Build b12370, tagged 13 November, tunes GCN MMQ configurations on the HIP path. The change is the CUDA backend's HIP-specific kernel config, not a new model.
Why it mattersAMD GPUs that go through HIP get matrix tiles chosen for GCN rather than a generic setting.
GitHub ↗llama.cpp adds SM103 machine code to non-native CUDA builds
Build b12367, tagged 13 November, adds SM103 SASS to CUDA builds that are not compiled natively for that architecture.
Why it mattersA binary built on an older GPU can still carry code for Nvidia's SM103 parts.
GitHub ↗llama.cpp runs RMSNorm in BF16 on CPU and CUDA
Build b12366, tagged 13 November, adds BF16 RMSNorm on CPU and CUDA, and uses one helper to launch the F32 and BF16 CUDA kernels.
Why it mattersA BF16 model can normalize without a detour through a wider type.
GitHub ↗Vulkan in llama.cpp gets a top-k mask
Build b12365, tagged 13 November, adds TOP_K_MASK to the Vulkan backend, fixes the barrier for that op, and declares the kernel in the common header.
Why it mattersA Vulkan device can mask to the top-k locally instead of pulling that step back to the CPU.
GitHub ↗Speculative decoding warns when a probability floor drops good drafts
Build b12364, tagged 13 November, warns when --spec-draft-p-min is greater than zero and draft sampling is probabilistic. Under rejection sampling, the draft's top probability does not say whether the target would accept the token, so the cutoff was dropping drafts the target would have kept. The note says p_min now behaves as before, and the log suggests setting it to zero in that mode.
Why it mattersA speed setting can throw away draft tokens the main model would have accepted.
GitHub ↗Ollama's release candidate maps Vulkan integrated GPUs with mmap
Ollama v0.40.17-rc0, tagged 10 November, uses mmap for Vulkan integrated GPUs. The same candidate uses thinking metadata when it writes OpenCode settings, and it changes how the ChatGPT CLI restores and restarts.
Why it mattersAn integrated GPU can load a model through the mapping the discrete path already used.
GitHub ↗llama.cpp moves the mixture-of-experts cache out of the graph file
Build b12361, tagged 13 November, moves the mixture-of-experts cache code out of llama-graph.
Why it mattersThe expert cache can change without every graph edit having to carry it.
GitHub ↗LiteLLM's long-term image line moves to 1.108.8
LiteLLM 1.108.8, tagged 13 November, is the current patch on the long-term support line. The release note says the lts image tag always points at the latest LTS release.
Why it mattersA deployment that tracks the lts tag picks up this patch without moving to the newer dev line.
GitHub ↗We are in the very early stages of a multi-year networking supercycle, and we're confident in our ability to continue capturing the multi-billion-dollar opportunity it represents for Cisco.
Policy & safety
Laws, lawsuits and safety incidents.
Kindo raises $30 million to govern agents and what they spend
Kindo said it raised $30 million co-led by Crosspoint Capital and Forgepoint Capital, with TFX Capital and CircleRock Capital joining existing investors, bringing the total to $57.6 million. The company describes one place for security, IT, compliance and finance to discover agents, including ones nobody approved, watch their access and cost, and apply identity controls to people and agents.
Why it mattersThe budget for an agent is becoming something a security team can set in one console.
Kindo ↗Claude Code 2.1.325 fixes decoding of zstd responses
Claude Code 2.1.325, tagged 13 November, fixes a security issue in how zstd-compressed responses are decoded. The same release adds a one-time tip in VS Code and similar terminals about re-running a notebook cell after Claude has read or edited it, and a line under a policy-block error that says to run /rewind and change the message, or /clear.
Why it mattersA compressed reply is a place a client can be attacked, not only a way to save bandwidth.
GitHub ↗n8n stops redacting runs that use a global credential
n8n 2.48.4, tagged 13 November, stops redacting executions that use global credentials. The change is a bug fix on the 2.48 line.
Why it mattersAn operator can read a run that used a shared credential without the log being wiped.
GitHub ↗The E2B CLI stops using a stored API key
E2B CLI 2.24.0, tagged 13 November, authenticates every request with the login's access token instead of a stored API key. e2b auth login opens the CLI auth page, tokens refresh before they expire, including during long commands, and refreshes are serialized so concurrent commands do not sign the user out. The project API key in the config file is no longer used or written.
Why it mattersA long sandbox log can keep its login without a static key sitting on disk.
GitHub ↗Sentry's next Python SDK replaces the PII switch
sentry-python 3.0.0b4, a pre-release tagged 9 November, makes span streaming the default tracing model and says the fine-grained data_collection setting supersedes the flat send_default_pii flag.
Why it mattersWhat a SDK is allowed to send is becoming a list of categories, not one on-off switch.
GitHub ↗Claude's agent SDK reports a machine policy that did not load
Claude Agent SDK 0.3.323, tagged 12 November, adds machinePolicy.didNotLoad on the result of resolveSettings, in alpha, so a host can see when a machine policy source may exist and did not load. matchDeniedModels, also alpha, reports which models a deniedModels list blocks.
Why it mattersA host can tell a blocked model from a policy file that never loaded.
GitHub ↗The AI SDK will not fetch a remote file unless you say how
AI SDK 7.0.172, tagged 13 November, requires an explicit experimental_download hook before experimental_extract or experimental_decide will fetch a remote file URL. Callers can pass createDownload with a DNS-protected transport for untrusted URLs. Inline file data still works without the hook.
Why it mattersA model feature that used to pull a URL on its own now waits for the application to choose the client.
GitHub ↗LlamaIndex raises the Python floor on its integrations
LlamaIndex 0.14.26, tagged 13 November, relocks integration dependencies for security fixes and raises requires-python across those packages. One change in the notes touches lockfiles only. The requires-python bump is marked breaking.
Why it mattersA security bump can drop older Python, so a pinned integration may not install until the environment moves.
GitHub ↗Robotics
Robots, self-driving and AI in the physical world.
Agile Robots raises $500 million at an $8.5 billion valuation
Agile Robots said it closed a $500 million Series F at an $8.5 billion valuation, taking total invested capital to $2 billion. SoftBank led, with Nvidia, Hitachi, Hitachi Ventures and T.Capital among the participants. The Munich company said it has more than 20,000 robotic systems installed, and that its Fürstenfeldbruck site can build up to 10,000 legged and wheeled humanoids a year. The money is for US expansion and more production, including in Germany.
Why it mattersA European humanoid maker is raising at a valuation that assumes factory volume, not a lab prototype.
Agile Robots ↗UFORCE raises $700 million at a $6 billion valuation
UFORCE said it raised a $700 million Series A at a $6 billion valuation, co-led by Valor Equity Partners and the UK National Wealth Fund, with the US-Ukraine Reconstruction Investment Fund participating. The London company, founded in Ukraine, builds uncrewed sea, land and air systems and the software that commands them. General Christopher Cavoli, retired, joins the board. The release names the MAGURA, NEMESIS, LIUT and PREDATOR families among the products the round will build.
Why it mattersA defence company is pricing autonomous systems as something allies will manufacture, not only import.
UFORCE ↗A navigation model imagines the view before it moves the drone
ProtoVis-Nav predicts feature prototypes and combines them to estimate what the next local view should look like, then uses those prototypes as context for the waypoint. The paper says this is aimed at language-guided drone navigation, where a small error in 3D becomes a collision, and that it improves on existing baselines.
Why it mattersThe drone commits to a move only after it has pictured the place that move should reach.
arXiv ↗A suite scores short, language-guided drone maneuvers
UAV-Flow++ frames Flying-on-a-Word: a drone turns a short instruction, an egocentric view and its current motion into continuous flight actions. The task covers primitive moves and maneuvers aimed at an object, rather than only long-route planning.
Why it mattersThe hard part of a drone instruction is often the last few metres, not the route across the map.
arXiv ↗XBot-Omni adds body motion and a face to an omni model
XBot-Omni uses a Thinker, a Talker and an Actor. The Thinker reads text, images, audio and video. The system extends that to body motion and facial expression so a robot can show what it is about to do, not only say it.
Why it mattersA robot's answer can be a gesture and a look, not only a sentence.
arXiv ↗A clustering loop assigns mixed robot teams
The paper describes an iterative clustering method for giving tasks to teams of different robots, including cases where the team can do something no single robot can. The allocation has to account for those joint skills, not only each robot's own list.
Why it mattersA fleet planner that assigns one robot at a time misses jobs that only a pair can finish.
arXiv ↗A paper asks why a relaxed origami design snaps to a bad fold
The authors study discrete design choices, using origami as the example, where the usual method optimizes a continuous relaxation and then picks the nearest discrete design. They say that last step can land in a poor basin, and they want the failure to be interpretable.
Why it mattersA design tool that rounds at the end can commit to a fold the continuous search never really liked.
arXiv ↗llama.cpp stacks Hexagon mappings to avoid fragmenting IOVA
Build b12368, tagged 13 November, combines Hexagon queue and fence buffers, sizes the queue from the thread count, and stacks mmap regions on the host so IOVA space fragments less. It also retries a failed mmap.
Why it mattersA Hexagon DSP, the kind built into phones and small machines, can keep a larger model mapped.
GitHub ↗Vulkan can keep a device awake for a set number of seconds
Build b12363, tagged 13 November, adds an optional Vulkan device keepalive. The backend owns it, and the environment variable is now a number of seconds.
Why it mattersA device that sleeps its GPU between tokens can be told to stay awake for a short job.
GitHub ↗We believe Physical AI will transform industries. Agile Robots combines deep robotics expertise with the ability to execute at industrial scale, enabling it to bring Physical AI into everyday operations.
Science & health
AI in biology, medicine, climate and maths.
Orakl raises $34 million to predict a cancer trial before it starts
Orakl Oncology said it raised a $34 million Series A co-led by Norrsken VC and Singular, with InfraVia, Serena, Bpifrance's Patient Autonome Fund, teampact.ventures and Plug and Play among the new investors. The company said the round brings total funding to $50 million. It grows patient-derived tumour models and pairs them with clinical data. Orakl said its model predicted cancer-trial outcomes with 90 percent accuracy, including pancreatic-cancer results, and that 18 of the world's top 20 drug companies are in its commercial pipeline.
Why it mattersThe bet is that a lab-grown tumour, read by a model, can fail a drug before a Phase 2 trial does.
InfraVia ↗Q-CTRL raises $290 million for quantum control software
Q-CTRL said it raised US$290 million, A$415 million, in a Series C led by Salesforce Ventures, with Bullhound Capital and Affinity Partners as large participants. New investors include AMD Ventures, EDGE Group and Australia's National Reconstruction Fund Corporation. The company called the round the one that makes it a quantum-sector unicorn, and said the software is used both to run quantum computers in data centres and to navigate when GPS is not available.
Why it mattersThe cheque is for the control layer that makes a noisy quantum device usable, not for a promise of a bigger chip.
Q-CTRL ↗A hyperspectral index builds a graph of land cover
The paper proposes MALNISI, a measure of how similar two hyperspectral pixels are based on their neighbourhoods, and uses it as the weights of a graph. A graph attention network then learns an unsupervised representation for land use.
Why it mattersSatellite bands can be clustered by who their neighbours look like, not only by a pixel's own spectrum.
arXiv ↗Agents take a federated-learning job from a plain request
Cortex-FL lets someone describe a training goal in natural language. Agents then look for collaborators, set up the task and run federated learning so each party's data stays where it is. The paper presents this as a way around the specialist setup that usually blocks an organisation from joining a shared model.
Why it mattersA hospital or a factory can ask for a joint model without first standing up a federation stack.
arXiv ↗Standard LoRA updates do not fill the rank they advertise
The paper says adapters, especially higher-rank ones, frequently fail to use the capacity their rank allows, and that this shows up as a tradeoff between learning and forgetting across fine-tunes. It treats unused rank as a cause, not only as a smaller matrix.
Why it mattersA fine-tune can forget the base model and still not have used the adapter it paid for.
arXiv ↗Diagonal state-space layers record the inputs and lose their order
State-space models often keep the parallel scan cheap by scaling each hidden channel with a single number. Those scales commute, so the state does not encode order. The paper uses representation theory to explain the resulting failures at tracking state, and to describe a representation that would not commute.
Why it mattersA memory that is linear in time can still be the wrong memory for a task that depends on sequence.
arXiv ↗Runtime still struggles to use the zeros in a pruned convolution
Pruned convolutional kernels on drones, phones and medical devices are often sparse enough that the arithmetic should shrink, the paper says, but kernels that skip those zeros are not yet the default. The work analyses sparse convolution as its own problem.
Why it mattersA model that is small on disk can still run as if every weight were present.
arXiv ↗Vision transformers separate objects in attention, not in the features
Across eight pretrained vision transformers, the paper finds that individual objects are distinguished more reliably by query-key routing than by value similarity or the residual representation.
Why it mattersInstance perception can be a property of where attention goes, which is a different place to intervene than the features.
arXiv ↗TRL 1.18 removes the experimental async GRPO imports
TRL 1.18.0, tagged 13 November, removes the submodules of trl.experimental.async_grpo. Imports of AsyncGRPOConfig from async_grpo_config, and imports from async_rollout_worker, packing, vllm_client or weight_transfer in that package, no longer resolve.
Why it mattersTraining code that reached into the experimental GRPO path has to move before this version will import.
GitHub ↗Open source
Open-weight models and community releases.
llama.cpp 1.0.0 freezes a public C API
llama.cpp v1.0.0, tagged 13 November, removes the deprecated public API and promotes the NextN speculative API into llama.h. The release adds text architectures for Kimi-K3.1-Pro, Ripple and xing4_0, drama-3 speech, a host-memory cache for mixture-of-experts experts, and an OpenAI-style /v1/audio/speech endpoint. It also updates ggml to v1.0.0. From here, a breaking change to the public C headers is a major version.
Why it mattersLibraries that bind llama.cpp can treat 1.0 as a line that will not move under them without a version bump.
GitHub ↗llama.cpp factors grammar alternatives before it matches
Build b12362, tagged 13 November, speeds up matching of large grammars by left-factoring them. The change keeps character-class boundaries intact and stops the factoring step from expanding without limit.
Why it mattersA constrained grammar can be large without the matcher walking every alternative from scratch.
GitHub ↗LangChain.js merges streamed tool arguments into the tool call
@langchain/core 1.4.1, tagged 13 November, merges streamed input_json_delta blocks into the preceding tool_use or server_tool_use block at the same index.
Why it mattersA streaming tool call arrives as one block instead of a trail of JSON fragments.
GitHub ↗E2B's template commands go through the SDK
E2B 2.64.0, tagged 13 November, makes e2b template list and e2b template delete use the SDK's Template.list and Template.delete. TemplateInfo gains maxCpuCount for resizable templates. JSON output from e2b template list now prints the SDK objects, with templateId and buildId, instead of the raw API fields.
Why it mattersA script that parsed the old JSON field names has to change with this version.
GitHub ↗Goose rewords a compaction prompt Opus 5.5 refused
Goose 1.58.1, tagged 10 November, rewords the compaction prompt that Opus 5.5 was refusing as distillation. It also refreshes Databricks OAuth tokens before they expire.
Why it mattersA summariser that a model treats as an attempt to copy it can be rewritten so the run continues.
GitHub ↗Goose can score which context is still relevant
Goose 1.58.0, tagged 9 November, adds an opt-in context-relevance operation and a context-window report and viewer. The release also adds declarative providers, including AIHubMix and RouterPlex.
Why it mattersA long session can show what is still inside the window, and drop what the check says is not.
GitHub ↗LangChain.js 1.6.5 can route a turn to another model
langchain 1.6.5 for JavaScript, tagged 11 November, adds modelRouterMiddleware, which picks the agent's model, and an auto mode. The note describes it as middleware for choosing the model on a turn.
Why it mattersOne agent definition can send an easy turn to a cheaper model and keep a larger one for the rest.
GitHub ↗The JavaScript Anthropic package reads tool calls from tool_calls
@langchain/anthropic 1.6.5, tagged 13 November, falls back to tool_calls in AnthropicToolsOutputParser so withStructuredOutput still works when streamed content has been normalized.
Why it mattersA structured-output parser keeps working after the stream has been turned into ordinary tool calls.
GitHub ↗Dockerignore classes stay ASCII in E2B's helper
E2B's @e2b/dockerfile-utils 0.1.2, tagged 13 November, keeps bracket whitespace classes in Dockerignore patterns ASCII-only, matching Moby. A pattern such as [\s] no longer matches a non-ASCII space such as U+00A0.
Why it mattersA sandbox image ignores the same files Docker would, including when the pattern was copied from a document that used a special space.
GitHub ↗Cost saving tip
One change that makes the same work cheaper, from today's news.
Prune screenshots once, not on every click
browser-use 1.0.4 waits until 20 tool results carry a screenshot and then drops the old ones together. Pruning on every shot was rewriting earlier messages and missing the prompt cache. On one 705-step task the previous version read 27.8 million input tokens.
GitHub ↗Upgrade past 1.0.3 if a browser agent is rereading the whole transcript. The bench cost per task fell from $0.434 to $0.208 at the same score.
Quick hits
Smaller stories worth a glance.
- Langfuse sends ClickHouse parameters in the request bodyLangfuse 4.80.0, tagged 13 November, sends ClickHouse query parameters in the request body by default.GitHub ↗
- Langfuse tries OpenRouter as a gateway providerLangfuse 4.79.0, tagged 12 November, adds OpenRouter as a pre-release provider in its AI gateway.GitHub ↗
- Prefect's automations page can be searched and orderedPrefect 3.9.3, tagged 13 November, adds search and ordering on the automations page.GitHub ↗
- Tokenizers' release candidate restores the Python APIs other libraries needTokenizers 1.0.0-rc.4, tagged 9 November, brings back the Python APIs that libraries such as Transformers and vLLM use.GitHub ↗
- n8n 1.123.96 makes JSON highlighting visible on a light themeThe 13 November patch also updates the Model Context Protocol SDK.GitHub ↗
- Qwen Code Desktop retries an enrollment whose answer was lostDesktop 0.27.3, tagged 13 November, lets an Agent Host retry an enrollment when the answer never came back.GitHub ↗
- The Qwen Code TypeScript SDK bundles CLI 0.27.3sdk-typescript 0.1.33, tagged 13 November, bundles a CLI built from the same revision, version 0.27.3.GitHub ↗
- Qwen Code can post a finished task into DingTalkQwen Code 0.27.3, tagged 13 November, adds a send-only DingTalk webhook that posts notifications and successful task results as Markdown.GitHub ↗
Deals & funding
Who raised money, and who bought whom.
Funding rounds 2
Investors put money into a company in exchange for a share of it.
Fifth Down Capital backs Databento
- Raised
- $100M
Databento now valued at $1.1B
Databento said it raised $100 million at a $1.1 billion valuation. Fifth Down Capital led, and Fitch Ventures joined as a strategic investor. The company said revenue has more than doubled since its Series B and that it has enterprise contracts with five of the six largest frontier AI labs.
FVFitch VenturesNNEADVDRW Venture CapitalTCTribe CapitalRVRedpoint VenturesHHedosophia+0 more
Planven backs Nebuly
- Raised
- $21M
Nebuly valuation not disclosed
Nebuly said it raised a $21 million Series A led by Planven to measure the return companies get from AI agents. Customers named in the announcement include Oura, Iveco Group, Fastweb+Vodafone, CNH and Zeiss.
EVExor VenturesTCTenacity CapitalTCTGVG CapitalFFastweb+VodafoneCVCDP Venture Capital+0 more
· Nebuly ↗
On the calendar
What's coming up, in date order, and how sure we are about each.
- 9–20NovCOP31 climate conference in AntalyaToday
- 15NovComments close on China's draft tech-export listTomorrow
- 15NovWorkBuddy and CodeBuddy free period endsTomorrow
- 17NovThird-quarter results and earnings call3days away
- 18–19NovTrump and Xi meet in Shenzhen4days away
- 18NovWebinar on running a dock with Stretch4days away
- 18NovNvidia invited to brief Congress4days away
- 18Nov‘Dot’ agents become the default in workspaces4days away
- 19NovKinference demo livestream5days away
- 19NovClaude Data turns on for Enterprise5days away
- 20NovAI Film Award submissions close6days away
- 20NovConductor cloud service shuts down6days away
- 20NovUpdated OpenAI usage policies take effect6days away
- 21NovInaugural edition in Hollywood7days away
- 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)8days away
- 23NovIn-person registration closes for the FDA robot workshop9days away
- 25NovQwen4-Max preview ends11days away
- 26NovHuang's deadline to testify12days away
- 29NovDaybreak Blue sessions can be used for training2weeks away
- 30–3Nov–DecContent London 20262weeks away
- 30NovRose d’Or Awards2weeks away
- 30NovGemini 4 Carbon introductory price ends2weeks away
- 30NovKorea AI Advertising Awards entries close2weeks away
- 1DecHearing on Apple's bid to block testimony2weeks away
- 2–3DecWorkshop on autonomous surgical robots3weeks away
- 4DecAI-made film opens in limited release3weeks away
- 4DecGods Don’t Give Gifts premiere3weeks away
- 4DecSingapore AI Safety Fellowship ends3weeks away
- 10DecIntelligent eyewear goes on sale4weeks away
- 16DecKorea AI Advertising Awards winners5weeks away
- 22DecKorea AI Advertising Awards ceremony5weeks away
- 31DecHy4 half-price offer ends7weeks away
- 31DecJules shuts down7weeks away
- 1JanFrontier labs must register under the RAISE Act7weeks away
- Later in NovGuided Learning pilot in US schools
- Mid-NovXREAL Aura shipments begin
- DecemberNothing's first AI-native product
- 2027Two Nighthawk r3 processors linked
- H1 2027Digit 5 early customer access
- Nov–Dec 2026Fort Benning robotaxi pilot
- Later this yearAge checks reach Threads
- 2028R200 systems ready to deploy
- through Feb 2028Language Explorer in Real Talk
- early 2027Firefly always-on agents enter beta
- End of NovExpo to detail Agent Sessions
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.






