Author: Tor Fosheim

Sora 2 video generation could be heading for ChatGPT «soon»

Sora video generation in ChatGPT would expand access, but also increase costs. (Picture: generated)
The Information and ChatGPT-watcher Tibor Blaho are both reporting an upcoming release of video generation in the main ChatGPT app.

According to The Information’s x.com post, the release seems imminent, while Blaho has spotted Sora generation mentions and ID’s in the latest Andoid beta ChatGPT App.

Sora 2 was launched as a quasi social network with its own app in October 2025 and was invite-only.

Judging by early search interest on Google for «Sora invite code» soon after launch, interest for the app greatly exceeded capacity and available invites.

It produces 10-second clips from prompts or pictures and promises to put your likeness in anything you can imagine. It briefly went viral, running up some 920 million weekly users before dropping from #1 to #165 in the App Store.

Putting the app’s features into the mainline GPT app will bring it up to par with Google’s AI subscriptions, which gives users access to the video generator Veo 3 on the $25 Pro plan.

Read more: The Information (paywalled), Tibor Blaho on X, and Reuters.

AI lab Thinking Machines gets investment, gigawatt compute from Nvidia

With no product and an experienced team, Murati’s Thinking Machines lab is rounding up funding. (Picture: Nvidia)
Founded by OpenAI’s former Chief Technology Officer after the 2024 leadership spat, Mira Murati’s Thinking Machines has scored a big deal with Nvidia.

The startup has entered into «a multiyear strategic partnership» that will provide them with both money and significant Vera Rubin compute early next year — about the same level the first version of Grok was trained on.

The parties are not disclosing a sum total, but 1 GW of Nvidia compute is estimated to be worth about $50 billion, Reuters notes.

Murati’s AI lab has been largely secretive about its actual products, releasing a configurable API in December 2025 and vowing to make AI models more accessible, capable and, yes, customizable.

They raised $2 billion at a $12 billion valuation from Andreessen Horowitz and Nvidia in July 2025.

Read more: Joint press release, writeups on Reuters, CNBC and TechCrunch.

Meta announces new processor generation focused on inference

Meta’s MTIA-chips are supposed to support recommendations and AI for «billions» of people. (Picture: Meta)
Tapping their long-time partner Broadcom, Meta’s new in-house chips are built to scale from recommendation engines to advanced AI workflows.

The new generation Meta Training and Inference Accelerator (MTIA) isn’t built to replace the chips sourced from AMD or Nvidia, but are intended to supplement them and achieve «the lowest possible price.»

MTIA-chips are taking a different tack on developing AI models, optimizing for inference — the process of answering AI queries — instead of training. Most chips are customized for training, which is more compute intensive, but the majority of a chip’s life is spent putting together answers in production.

— Chip designs are based on projected workloads, but by the time the hardware reaches production — often two years later — those workloads may have shifted substantially, Meta says in their press release.

The new chips «have either already been deployed or are scheduled for deployment in 2026 or 2027,» Meta says — and they don’t disclose just how many of these they are making.

Read more: Meta’s press release. Writeups on Reuters and CNBC.

Claude for Excel and Powerpoint now shares info, and gets skills

Excel and PowerPoint editions of Claude can now talk to each other. (Picture: Anthropic)
As of today, Claude shares your conversation «across all open files,» so actions in one file can be «informed» by what’s happening in the other.

That eases some hassle for those who do a lot of work in PowerPoint and Excel, and removes the need to reintroduce the task or use extra steps.

It means that users can pull financials into a workbook and drop the valuation summary into a PowerPoint slide without switching tabs or re-explaining at every step.

At the same time, Anthropic is launching skills for workflows — that can be shared and dropped into other apps in an organization, so everyone can use the same time-saving actions stored in them.

Skills are stored prompts for workflows, and work as an old-school template; with everything set and working on repetitive tasks that are easy to automate.

Read more: Anthropic’s presentation and launch tweet. Writeups on VentureBeat and The Decoder.

As 140 million people use ChatGPT for learning math, OpenAI adds visuals

That’s a big number out of 900 million weekly ChatGPT users, accounting for about a sixth of the total usage and outperforming even health questions.

In their announcement, OpenAI says that more than half of U.S. adults struggle with math, and don’t feel confident teaching it to their kids.

Continue reading “As 140 million people use ChatGPT for learning math, OpenAI adds visuals”

French Canal+ partners with Gemini to index content, ease production

Gemini will scan Canal+’s library and create a massive recommendation engine. (Picture: generated)
In a multi-year deal first reported by Reuters, the now global TV platform is looking to improve its recommendation engine and help production teams.

Starting in June 2026, Canal+ will be providing producers with access to Veo 3, which will help creators «pre-visualize scenes before shooting,» or help recreate history from a single image.

They will also be indexing its extensive content library to use in a Gemini-powered recommendation engine, creating a massive, multimodal database of sound, video, and text data.

Canal+ recently bought the pan-African boradcaster MultiChoice, extending the reach of its app and services to some 70 countries, as they are racing toward a target of 100 million streaming users, and hope to compete with the American streaming giants.

Read more: Canal+ press release, Reuters report.

Ex-Meta chief AI scientist raises $1 billion to build world models

After leaving Meta at the end of 2025, LeCun’s new startup is already raking in the money. (Picture: generated)
Hot off the $1B raise for Fei Fei Li’s World Labs, the former AI boss at Meta, Yann LeCun, is raising as much for his Paris-based Advanced Machine Intelligence.

CEO Alexandre LeBrun notes to TechCrunch that world models are the latest buzzword in AI, but cautions that the startup is doing «fundamental research» and won’t be launching any products «for years.»

LeCun and LeBrun both say that generative AI models work by predicting text or images, and are severely limited when it comes to interacting with the real, messy, and unpredictable world.

AMI sees their clients as being in «industrial process control, automation, wearable devices, robotics, healthcare, and beyond,» according to their front page.

According to Reuters, their technology might show up in Meta’s AI glasses «in the short term.»

Read more: Reuters, TechCrunch and Wired.

Google announces slew of Gemini improvements to Workspace

Workspace got smarter, and can now draw on files, emails, chats and the web. (Picture: Google)
Sheets, Slides and Docs are getting some extra help from Gemini in a huge update to the service.

— Today, we’re making Gemini in Docs, Sheets, Slides and Drive more personal, capable and collaborative to help you get things done, faster, Google says.

All these apps can now draw on information from your Drive, Gmail, Chat and web search to draft things like emails and docs, or pull numbers for spreadsheets based on, say, an email conversation, meeting notes or separate sources in Drive. All it takes is a single prompt.

Google is especially proud of their agentic performance on Sheets, getting very close to the human expert benchmark on the SpreadsheetBench dataset.

The features are rolling out to all Ultra and Pro subscribers globally today, but is only available in English. Google is looking to bring on «more languages soon.»

Read more: Google’s announcement, launch thread. Writeups on 9to5Google and TechCrunch.

Meta acquires Moltbook team to work on Superintelligence Labs

Meta acquihires the Moltbook team for an undisclosed sum. (Picture: composite)
Meta won’t say what they paid, but are excited for how the team joining «opens up new ways for AI agents to work for people and businesses,» they tell Axios.

The agent-only social platform Moltbook took the web by storm in late January, providing a fascinating view of AI agents talking to other agents.

It is based on the equally viral agent platform OpenClaw, which was bought by OpenAI and kept open source less than a month ago.

Meta is taking a different track with Moltbook, deciding to keep it open «temporarily,» with stronger identity requirements — connecting the bots to their owners.

The founders, Matt Schlicht and Ben Parr, will be joining the Meta Superintelligence Lab on future efforts, and Meta tells TechCrunch that «we look forward to working together to bring innovative, secure agentic experiences to everyone.»

Read more: Axios broke the news, further info on CNBC, and TechCrunch.

Anthropic launches Code Review for all those bothersome Claude pull requests

Drowning in pull requests from Claude Code? Anthropic has an answer. (Picture: Anthropic)
If you ever used Claude Code, you’d probably notice a mountain of pull requests asking for review on any decent code base. This takes time and effort for developers — but now Anthropic offers a solution.

— One of the questions that we keep getting from enterprise leaders is: Now that Claude Code is putting up a bunch of pull requests, how do I make sure that those get reviewed in an efficient manner? Cat Wu, Anthropic’s head of product, tells TechCrunch.

The answer is the newly launched Code Review tool that uses multiple agents to scan code changes, comment, and rate them for severity.

Using it internally, Anthropic found something of note in 84% of automated code reviews with more than 1,000 lines, they say.

The only problem is that it takes quite a few tokens to run a lot of agents on code changes, and the average cost is between $15-$25 per pull request, depending on complexity, Anthropic writes.

The tool is available as a research preview for Team and Enterprise plans as of today.

Read more: Anthropic’s announcement, writeups on TechCrunch and The Register.

Anthropic’s Cowork gets tapped by Microsoft Copilot

Microsoft’s version of Copilot is entirely cloud based, pleasing many enterprise customers. (Picture: Microsoft)
While announcing Claude availability in «mainline Copilot chats,» Microsoft is also launching its version of the Cowork agent.

Copilot Cowork works mainly across Microsoft’s own Office 365-offerings, and can author emails in Outlook, manage calendars and create files in PowerPoint, Word and Excel, Microsoft says.

— When you hand off a task to Cowork, it turns your request into a plan. The plan continues in the background, with clear checkpoints so you can confirm progress, make changes, or pause execution at any time, they write.

Anthropic’s Cowork does much of the same for general office workloads, and is available as a separate app — with connectors for popular services and tasks.

It works locally on your computer to complete tasks, but Microsoft’s version is cloud-based, and works the data companies have uploaded, offering more diversity and control of their data.

Read more: Microsoft’s presentation and blog post. Writeups on Reuters, The Register and VentureBeat.

Alibaba’s new Rome agent got caught mining crypto without authorization

Mining for crypto was not exactly in the brief for the coding agent during training. (Picture: generated)
In training, Alibaba’s new coding assistant agent opened up an SSH-tunnel to an outside computer and started secretly mining cryptocurrency.

It was an «unanticipated — and operationally consequential — class of unsafe behavior that arose without any explicit instruction and, more troublingly, outside the bounds of the intended sandbox,» the researchers write in a paper, buried on page 15.

It was not prompted in any way to do this, or acting on any kind of instruction — the agent was acting autonomously.

Crypto mining opens up a path to interacting with the real world economy, Axios points out, and lets a rogue agent set up its own business and pay for services, for example.

The actions of the bot was caught by the team’s firewall, along with other attempts to access their «internal network resources,» and was quickly stopped after a brief investigation.

Many are pointing to this incident as a «paperclip»-moment and considering the vast GPU resources that an agent in training has access to, it poses some real questions about agentic security.

Read more: Paper on xArchiv, reports on Axios, Cryptopolitan, and Copilot summary.

OpenAI puts «adult mode» on the back burner, for now

«Adult mode» on ChatGPT gets another postponement. (Picture: generated)
The much touted feature last showed up as «naughty chats» in the Android app just two weeks ago, but is now postponed indefinitely.

It was supposed to debut in the first quarter of 2026, after first being conceived as «letting adults be adults» in October 2025. It was supposed to run in tandem with age verification, rolled out this January.

OpenAI says that «we still believe in the principle of treating adults like adults, but getting the experience right will take more time,» writes Axios.

The postponement is due to the company focusing on other high priority issues, such as «intelligence, personality improvements, personalization,» and «making the experience more proactive.»

Read more: Axios and TechCrunch.

Claude finds 22 security vulnerabilities in the latest version of Firefox

Claude spent two weeks finding a fifth of all serious bugs in all of 2025. (Picture: Adobe)
14 of the bugs Opus 4.6 discovered were classified as «high-severity vulnerabilities» and were fixed by Mozilla in the latest update in late February.

The process took only two weeks to find about a fifth of the total high-severity risks found in all of 2025 — providing a much faster way to scan for bugs.

— Opus 4.6 is currently far better at identifying and fixing vulnerabilities than at exploiting them. This gives defenders the advantage, Anthropic writes, but warns this might change.

Claude works on the full stack, from initial bug hunting to verification and then suggesting patches, offering much needed relief to overworked developers.

— We view this as clear evidence that large-scale, AI-assisted analysis is a powerful new addition in security engineers’ toolbox, Mozilla says in a blog post.

Read more: Anthropic’s workthrough, Mozilla’s blog. Writeups on TechCrunch and Axios.

OpenAI’s robotics and hardware head leaves in Pentagon protest

Caitlin Kalinowski, head of robotics and consumer hardware at OpenAI, makes no qualms about her feelings on their new contract with the Pentagon, and is resigning in protest, saying that:

— Surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got.

Continue reading “OpenAI’s robotics and hardware head leaves in Pentagon protest”