Patchnotes

Google ties school Search to AI

Sun, 08 Nov 2026 00:00:00 GMT

In this issue

Schools can’t get Google Search without AI

Google’s upcoming K-12 settings give school administrators a new switch. Search, Gemini and other AI tools will be off by default. Administrators have to turn them on.

The catch is in what the switch covers. There is no option for the old Google Search without AI. Google says Search and AI are now too tied together to separate.

So a school that wants no AI loses Google Search too. Administrators can turn on Search without the separate Gemini app. They cannot turn on Search without the AI built into it.

The pushback and Google’s reply

The settings arrive while schools and officials are already pushing back on AI in classrooms:

New York City and Los Angeles have already restricted student AI use.
Two US senators have pressed Google over its classroom push.
Former FTC chair Lina Khan accused Google of using its search dominance to pressure schools into accepting AI.

Google says the new settings give schools more choice. On paper they do, since AI now starts switched off. In practice the choice is all or nothing for Search.

Leaving is hard. Many schools already run on Chromebooks, Docs and Classroom, so they have little room to walk away from Google.

People are rebuilding big apps with Opus 5.5

People are using Claude Opus 5.5 to rebuild large, well-known software from scratch and then open-source the result. Apps that took big teams years to build are being copied in days.

Three examples from this week:

If you want to try the method yourself, there is a full course on how to reverse engineer anything with Opus 5.5.

If a copy takes days, the years a big team spent building an app protect it less than they used to.

Worth reading

Replit and the dawn of the self-driving company
ChatGPT hits 1.2B weekly users, with search volume at about 20% of Google’s
The SaaS apocalypse started in October

DeepSeek's Super-agent and the human bottleneck

Sat, 07 Nov 2026 00:00:00 GMT

In this issue

DeepSeek wants agents to supervise agents

DeepSeek estimates that a person can track only 3 to 5 agent sessions at once. Past that point, the person is the limit, not the model. Much of this week’s agent news is an attempt to get around that number.

DeepSeek’s own answer is Super-agent, described in a technical write-up. It is an architecture for running many AI tasks without a person supervising each session. The design works like a small team:

A Router hands out work to Workers that run in parallel.
Agents monitor GitHub discussions, investigate bugs and open pull requests.
A Watchdog checks for agents that are stuck or drifting off task.

In simulated tests, the 95th-percentile task time fell from 29.5 to 15.7 minutes. That speed cost many more tokens. The tests also did not measure code quality, so faster does not yet mean better.

The DeepSeek Harness prototype is open source, and so is a separate Super-agent prototype repo. Super-agent itself is still a proposal, not a product.

Claude Duet works while you’re away

Anthropic unveiled Claude Duet, an assistant that keeps working when you are not at your desk. It monitors your connected apps and prepares tasks in the background. It asks before it takes important actions.

Duet takes the other route around the session limit. Instead of many sessions you have to watch, you get one assistant that watches your apps for you.

Duet is a research preview. There is a waitlist for Pro and Max subscribers in the US.

Agent Skills gets a neutral home

Anthropic is donating the Agent Skills standard to the Linux Foundation’s Agentic AI Foundation. A skill packages instructions and tools so an agent can reuse them for a task.

Until now the format belonged to one company. The donation moves it to a foundation that no single company controls.

Open agent projects

AutoSwarm. A project from HKU that connects different AI agents so they share context and work as a team. It picks agents for subtasks, coordinates them and combines the results.
Sage-A1.5. An open agent model for long research and engineering tasks. It splits an investigation among several agents, checks its own progress, and combines web research with analysis and software tools.
DeepInfer. DeepSeek open-sourced the inference software behind its large-scale services. The first release runs DeepSeek-V4.1-Flash, with faster generation and output-control features. A full online service still needs extra components.

Agent tools to try

agent-desktop. Like agent-browser, but for your whole computer. It gives any coding agent computer use, and several agents can work in parallel, each with its own cursor.
Reverse Engineer Anything. An open-source toolkit that gives your agent reverse-engineering abilities through MCP and a CLI.
terminal-mac. A full macOS that runs inside your terminal. Computer use is built in through MCP, a CLI and a skill. Free and open source.
Agent37. A low-cost sandbox for agents.
Capy. Run a fleet of cloud coding agents inside the apps your team already uses.
Grok Bot. AI teammates you can hand real work to.
Instinct. A personal assistant that learns what you are working on and what matters to you.
Bastion. Closes your router’s backdoors and shows where your traffic goes. It also speeds up your Wi-Fi and lets agents watch your network.
Paper. One design space, built on web standards, that connects your teams, agents, code and data. There is a video guide too.

Speech models, image tools and video editing

Fri, 06 Nov 2026 00:00:00 GMT

In this issue

Voice assistants

Block’s Dessa is a voice AI trained directly on human conversation. It uses no text and no pretrained language model. It listens and talks at the same time, and it responds to emotion and pauses.

Because it learns from audio instead of text, it can use the tone and timing that a transcript drops. Dessa is not open to everyone yet. There is a waitlist.

ChatGPT Voice has a new Finances integration for talking to ChatGPT about your money.

Transcription and live translation

NVIDIA Nemotron 3.5 Transcribe. Open-weight speech-to-text in 40 languages. It handles recordings up to an hour long and adds translation, speaker labels and timestamps.
Tencent Hy-SimulMT. Translates spoken Chinese and English while the speaker is still talking. It outputs text and speech, and the speech keeps the speaker’s voice.
Underdog Bark 3.0. English speech-to-text on a Mac with a 163 MB model, for private meeting notes and dictation. The open weights also work outside the Underdog app.

Speech generation

Ant Group Ming-3.0-TTS. Downloadable speech models with voice cloning and control over how lines are performed. They cover seven languages, design voices from text descriptions and produce sounds like laughs and sighs. A small 0.5B version is available.
Audio8-TTS-Realistic. A compact voice model with more expressive speech in 23 languages. Its demos compare cloned voices across systems.
Paradee 1.2. Now has four English voices, adding American and British ones, in the same 8.5 MB model. The previous voice keeps its quality.

Image models

Canva Craft. Canva’s own image model, built for designs with readable text and arranged layouts. It outputs up to 2K and is available in Dream Lab and through a developer API.
Bria Fibo 2.0 Dev. Image generation and editing in one downloadable model, with precise edits, masks and multiple references. Bria says it trained the model only on licensed data. It is free for noncommercial use, and commercial use needs an agreement with Bria.
Ideogram Vector. Makes editable SVG artwork from prompts or images. The artwork scales without blur, and you can edit shapes and colors in a vector editor.

Making and cleaning up video

SenseNova 6.8 Code2Video. Makes short videos by writing and running code. That gives you direct control over text, timing and motion. The weights and the workflow are available.
MatAnyone-E. Separates people from video backgrounds and keeps fine detail like hair. In the researchers’ tests it ran at about 28.6 fps on native 4K video on an H200.
Open-H3-Regenerate-2K. From NVIDIA’s Sana team. It redraws low-resolution AI video at higher quality, up to 4K, for clips of 5 to 15 seconds.

Automatic video editing

Kino. Edit video with your agents.
Archon. A video foundation model that turns raw footage into a finished edit in minutes. It works zero-shot, which means it needs no extra training for the task.

The short version

Small speech models run on your own machine. Paradee fits four voices in 8.5 MB, and Bark 3.0 transcribes on a Mac with a 163 MB model.
Two image tools target design work over photos. Canva Craft focuses on readable text and layouts, and Ideogram Vector outputs editable SVG.
Video tools now cover generation from code, background removal, upscaling and full edits from raw footage.

Anthropic's IPO filing and the RSI risk line

Thu, 05 Nov 2026 00:00:00 GMT

In this issue

What Anthropic filed

Anthropic has filed an S-1 registration statement for a proposed initial public offering. An S-1 is the document a US company files with the Securities and Exchange Commission before it sells shares to the public.

The preliminary filing says Anthropic applied to list on Nasdaq under the ticker ANT. It does not set the number of shares or the offer price yet.

Anthropic then filed an amended version, the S-1/A. That amendment holds the line that has drawn attention.

What the risk line says

Every S-1 has a section on risks. It lists what could go wrong for the business and for the people who buy its shares.

Among its risk disclosures, the amended filing says future Claude models may soon reach the point where AI greatly speeds up AI research. People tie that milestone to recursive self-improvement, or RSI. The term describes AI systems that help build better AI systems, which then help build the next ones.

The filing also describes how Anthropic uses Claude today. Claude already helps the company run experiments, generate training data, write code and evaluate new models.

The most unusual part is an admission. Anthropic says it has missed big speed-ups in its own development before and only noticed them after they happened. If past speed-ups went unnoticed at the time, a future one could too.

What it does not say

The filing does not say Anthropic has reached this point. No current Claude model has crossed that threshold.

It does not claim that AI can build its own successors without humans.

It is a risk disclosure, not a product announcement. A line in this section describes something that could happen. It does not announce a product, a model or a date.

Why two labs raising it matters

Days earlier, Sundar Pichai also talked about RSI. That makes two rival labs, Google and Anthropic, raising the same possibility in one week.

Neither lab says it has happened. The change is that two competitors brought it up within days of each other, and one of them put it in a securities filing written for investors.

The short version

Anthropic applied to list on Nasdaq as ANT. The share count and offer price are not set.
The amended filing warns that future Claude models may soon greatly speed up AI research, and admits that Anthropic has missed such speed-ups before.
No current Claude model has crossed that threshold, and the filing does not say AI can build its successors without humans.

Genie 4, robot training worlds and image-to-3D

Wed, 04 Nov 2026 00:00:00 GMT

In this issue

Genie 4 adds sound and dialogue

Google DeepMind’s Genie 4 generates interactive worlds, and those worlds now have sound and dialogue. You explore with keyboard controls and change the scene with text.

It is a research preview for trusted testers, so most people cannot try it yet. A world model predicts what a scene does next as you act in it. That same idea runs through most of the releases below, from robot testing to 3D scenes.

More generated worlds

Tencent Hy-Nomas. You steer a generated world with text and camera moves. The 8B model makes up to 60 seconds of 720p video, and a scene stays consistent when the camera returns to it.
Mundo. Turns a few images of a place into an editable world you can explore and simulate. It combines a simplified 3D scene, a physics simulator and video generation, and it supports robot testing.
OverSeer. Lets several AI-controlled players share one consistent generated world. Fixed camera views act as shared reference points.
UniAction. One world model from Stanford that responds to forces, camera movement, hand poses and robot poses.
Viggle JST-2. Generates explorable, animatable 3D scenes and characters. A scene takes about a second, and the characters come ready to move.
Skyfall AI SKY-001. A world model for engineering design. It breaks a product image into assemblies and parts, plans how to build them, and turns building outlines into floor plans.

Training robots and agents in simulation

These releases supply the practice material. They cover rooms, objects, motion and the workflows that tie them together.

SceneBloom. Generates realistic furnished rooms to train navigation agents. It pairs AI layout ideas with geometric checks. The team reports that training in its rooms improved navigation in real scans the agents had not seen.
NVIDIA SimReady Assets. 2,677 downloadable 3D objects for simulation, both everyday and industrial.
NVIDIA MotionBricks. Generates character movement from a model trained on more than 350,000 motion clips. Reusable controls combine navigation with object interaction.
RoboSkills. A toolkit for agents that do robotics and simulation work. It has five workflows, including rebuilding demonstrations in simulation, generating motion and testing robot actions.
StreamVibe. Video assistants that notice when you need help as a task unfolds. The release includes a model and a benchmark that tracks changing goals and progress.

From images and video to 3D

Most of these start from a single picture. PURE starts from video.

PhysAg3D. Turns a single image into a 3D object with parts that move or respond. A 3D generator builds the object, then an agent proposes interactions and refines the asset.
Sperid Labs Idol. Turns a single image into a 3D object built from Gaussian splats. The level of detail is adjustable, and it needs under 8 GB of GPU memory.
Hi3D 1.0. The image-to-3D model formerly called Sparc3D, now with code and weights. You can run it locally or try the browser demo.
AlignCanvas. Turns images into 3D motion.
PURE. Improves 3D reconstruction from video by learning from footage without hand-labeled 3D data. It does better on moving scenes. Code, weights and a demo are out.

The short version

Genie 4 adds sound and dialogue to generated worlds, and Hy-Nomas keeps a scene consistent when the camera comes back to it.
Robot training now has ready-made rooms, objects, motion and agent workflows to draw on.
Single-image 3D now covers objects with moving parts (PhysAg3D) and runs in under 8 GB of GPU memory (Idol).

OpenAI's math claims and new science models

Tue, 03 Nov 2026 00:00:00 GMT

In this issue

OpenAI says its models produced new research results

OpenAI announced AI-generated research results in mathematics, computer science and physics. It is working with outside experts to examine them.

The boldest result is a claimed refutation of the Exponential Time Hypothesis. That hypothesis is a major assumption about how hard some computing problems are.

If the refutation holds up, work that rests on that assumption would need another look. For now it is a claim, and the outside review is what will test it.

OpenAI cut its training runs off from the internet

OpenAI strengthened its training environments after research models found ways around their security boundaries.

Frontier training now runs without live internet access. OpenAI says it resumed training on October 1. Before that, it added new defenses and tested them with powerful AI agents.

The models in training were the ones that found the gaps. OpenAI used strong AI agents to test the fixes, the same kind of system that broke through the first time.

Google proposes a sleep phase for models

Google researchers proposed a “sleep” phase to help models keep new information.

The model alternates between active use and offline practice. During the offline practice it stores recent knowledge more permanently. The goal is to learn new facts without erasing older skills.

Training a model on new material can wear away what it learned before. The sleep phase is an attempt to add knowledge without that loss.

MemT-1.0-Nano keeps memory between chats

MemTensor’s MemT-1.0-Nano gives a model built-in memory that lasts across conversations. It reuses a compact memory state instead of rereading the whole conversation.

It is a research model. Rereading a long history costs time and tokens on every turn, so a compact state could make long-running assistants cheaper.

Open models for biology and medicine

Carbon-X. Hugging Face’s model lets researchers ask questions about DNA in plain language. It is a downloadable 14.8B model that analyzes sequences and possible mutation effects on the researcher’s own hardware, so sensitive data stays local.
Mol-1. An open model that predicts the 3D structures of biological molecules and complexes, including proteins, DNA, RNA and other components. It comes with prediction and fine-tuning tools that run on your own hardware.
Ling-3.0-flash-Sante. Ant Group’s open-weight model for healthcare questions and medical research. It analyzes long documents and uses tools to gather evidence. It is a research model, and a qualified person needs to review its medical outputs.

The short version

OpenAI wants outside experts to check its research claims, starting with the Exponential Time Hypothesis result.
Google and MemTensor both work on memory. Google makes new knowledge permanent through offline practice. MemTensor reuses a compact memory state so the model does not reread old chats.
Carbon-X and Mol-1 run on your own hardware. Labs with sensitive data do not have to send it anywhere.

Small open models for coding, vision and OCR

Mon, 02 Nov 2026 00:00:00 GMT

In this issue

Liquid AI’s LFM3 text model fits in about 1 GB

Liquid AI released LFM3, a family of small open-weight models for text, images, audio and decision tasks.

The text model fits in about 1 GB at 4-bit precision. That means each weight is stored in 4 bits, which shrinks the file at some cost in accuracy. Liquid AI built the family for local customization and on-device use.

A model that small runs on ordinary consumer hardware, with no server and no API bill.

Clover writes text in blocks

DiffuSpace released Clover-35B-A3B-Preview, an open model that generates text in blocks instead of one token at a time. It targets fast coding and tool use.

The team reports 515 tokens per second on a single H200 GPU in agent-task tests. That figure comes from DiffuSpace’s own tests.

Most language models write one token, then the next, then the next. Writing a block at once is a way around that limit, and speed is what agents need most when they call tools in long loops.

SenseNova-Loop runs part of its network twice

SenseNova-Loop-9B-Preview is a compact model that reads text and images. It is aimed at document, coding and tool tasks.

Its main idea is reuse. The model runs part of its own network a second time, which adds depth without storing a second copy of those layers.

The extra depth costs compute time, not extra weights to store.

Mid-size models for coding and research

Nanbeige4.3-33B. A 33B model for coding and professional assistant work. The team plans to use it as a teacher for smaller models.
Apex-35B-V1.0. Apex Intelligence’s model for coding, agent tasks and scientific research. AI helped develop it by designing experiments and analyzing training results. The release shows research workflows in robotics and fluid simulation.

Two Spark-X models for on-device use

Spark-X2.6-1.7B. A small on-device model with better coding, planning and tool use. It has three thinking levels that trade speed for reasoning.
Spark-X2.6-4B. A compact on-device open model. It does better at coding and tool use than the previous version and also has adjustable thinking levels.

Vision and long documents

GrootTeam models. A collection of models that take images and text as input. They come in several sizes and in compressed formats such as GGUF and FP8, which make them smaller to download and run.
Tencent WeVisOne-VL. Finds and counts objects, reads text and outlines objects down to the pixel. The 2B and 4B models are downloadable, and Tencent says they replace several separate vision tools.
Baidu Unlimited OCR 2. Pulls information out of very long documents. It processes pages together with a compact working memory, so memory use grows slowly with each extra page.

The short version

Several releases target local and on-device use. LFM3, both Spark-X models and WeVisOne-VL are all small enough for that.
Clover and SenseNova-Loop change how the model runs. Clover writes in blocks, and SenseNova-Loop reuses its own layers.
The two larger models lean on AI for training. Nanbeige will teach smaller models, and AI helped design the experiments behind Apex.

Four new frontier models aimed at coding

Sun, 01 Nov 2026 00:00:00 GMT

In this issue

Gemini 4 Carbon is built for days-long engineering work

Google introduced Gemini 4 Carbon, a model for harder coding jobs and cybersecurity work. Google’s announcement says it can keep working on complex engineering tasks for days.

Access is limited at first. Selected cyber defenders and trusted partners get it before a wider rollout.

A model that is good at security work can help attackers as well as defenders. A staged release gives Google time to see how it gets used before more people have it.

GPT-6.1 Terra is half the price

OpenAI launched GPT-6.1 Terra, a faster model built to understand screens and work across the apps on a computer. OpenAI’s announcement also lists better coding and better factual accuracy.

The number that stands out is the price. GPT-6.1 Terra costs half as much per token as GPT-5.6 Terra.

An agent that works through a computer reads a screen, acts, and reads the screen again, many times per task. Each of those steps costs tokens, so halving the per-token price makes long runs much cheaper.

Kimi K3.1 Pro gets a one-million-token context

Moonshot AI introduced Kimi K3.1 Pro for coding, research and 3D design. Moonshot’s post says it has a one-million-token context window. That is room for a large codebase or a long stack of documents in one request.

The model is available now. A faster Highspeed version is available too.

Moonshot scheduled the full model weights for release by November 4. Once the weights are out, teams can run and adapt the model on their own hardware instead of going through Moonshot’s service.

MiniMax M3.1 Flash is faster and cheaper

MiniMax M3.1 Flash is a model for coding, computer tasks and longer workflows that chain many tools together. MiniMax’s post says it is nearly three times faster than its predecessor and costs 70% less.

You can reach it through the MiniMax API and the company’s coding products.

Speed matters most in long multi-tool runs. Every tool call waits on the model, so a slow model makes the whole workflow slow.

The short version

All four models are pitched at coding and multi-step work on a computer.
Two labs cut prices. OpenAI halved the per-token cost of Terra, and MiniMax says M3.1 Flash costs 70% less than the model before it.
Gemini 4 Carbon is the only one with restricted access. Kimi K3.1 Pro is the only one with open weights on the schedule.