TAWS

Wednesday, November 11, 2026

OpenAI's $10.9B Quarterly Loss, Grok Maker's Open Staff Apps, Nobel Panel Backs OpenAI Proof | TAWS

Watch on YouTube · Listen on Spotify

The rundown

AI-generated from the transcript

Austin and Ron work through six stories where the headline and the fine print point different ways. The Wall Street Journal reports OpenAI lost $10.9B in a quarter even as its cost per sales dollar improved, a security firm found Grok Build apps (including SpaceX and xAI staff apps) open to anyone, and Anthropic added Sheets, Whiteboards and a Google Meet assistant to paid Claude plans. Also: Ring's $249.99 indoor drone, lawsuits built on work Claude chats, and a Nobel-led panel backing OpenAI's AI-written physics proof while admitting it doesn't fully understand it yet.

OpenAI's quarterly loss hits $10.9B as sales surge

Watch from 2:41

What happened
The Wall Street Journal reports OpenAI's adjusted operating loss reached $10.9B in a quarter where it brought in about $10B in sales, up from a $10.1B loss the quarter before. Sales more than doubled from a year earlier. Adjusted figures show spending just over $2 per sales dollar, down from about $2.50. Bloomberg says OpenAI is completing funding at a $1.25 trillion valuation with an initial $11B closed, has delayed any stock market debut until 2027, and regained momentum partly through aggressive price cuts, while Anthropic's quarterly growth slowed to around 47% from more than 140%. Bloomberg notes the two companies count revenue differently.
The claim
OpenAI told investors sales grew nearly 50% in the last quarter alone. Greg Brockman said businesses now bring in most of its revenue.
Austin's take
Austin calls it progress: a bigger loss but a better ratio, with business customers and falling costs of intelligence pointing the right way. He still wants proof growth can become a profitable business, and says no winner should be declared over Anthropic. Ron argued it's a subsidy, buying customers with investor cash.
Why it matters
OpenAI's economics shape the prices and staying power of tools many people and businesses now depend on.
What to do
Watch the cost per sales dollar, not just the loss headline. Enjoy price cuts, but don't assume today's pricing is permanent.
Not confirmed yet
The figures went to investors privately and have not been audited. Revenue comparisons with Anthropic are imperfect because they count revenue differently.

Grok Build apps, including Musk company staff apps, were open to anyone

Watch from 7:09

What happened
Security firm Claroc found 24 open apps built by staff at SpaceXSI (the merged SpaceX and xAI) and X with Grok Build, including a confidential ad proposal to a federal agency, a map of 594 strategic sales accounts and a sales onboarding plan. Grok Build puts the random 'secret' words at the front of each app's address before grok.me, where public lookup databases log them; Claroc found more than 43,000 addresses this way. 94.7% of live apps it found opened with no sign in, and it estimates 650 to 800 held private data, including a fire department's shift board with badge numbers and an ICU patient's labs and ventilator status. Claroc reported it Sunday, November 8; staff apps were still open Monday evening and were pulled within hours of Ars Technica asking questions. Other users' open apps stayed online. In Claroc's paid account, the publish panel was preselected to 'anyone with the link.' In July, a researcher caught Grok Build's coding tool uploading developers' whole code projects, secrets included, to cloud storage.
The claim
SpaceXSI says apps stay private until a creator shares them, publishing takes several steps, and most open apps were games and calculators. Musk said the July data would be deleted.
Austin's take
Austin calls it a design problem, not a user problem: a preselected box is a design choice, and when the toolmaker's own staff get caught, he stops blaming customers. He wants the secret words moved after grok.me and all apps locked, not just staff ones. Ron agreed.
Why it matters
AI app builders make publishing easy, and defaults decide whether private data ends up public, including data from people who never touched a publish button.
What to do
If you've built anything with Grok Build or similar tools, check each app's sharing setting and lock anything with real data. Read publish screens before clicking through.

Claude adds Sheets, Whiteboards and a Google Meet assistant

Watch from 12:07

What happened
Anthropic launched Claude Sheets and Claude Whiteboards in beta for all paid plans on web, desktop and phone. Sheets builds tabs, formulas and dropdowns, edits alongside your team and lists each change. Claude Meetings, in beta for Team and Enterprise, joins Google Meet calls as a visible participant, drafts plans, pulls numbers from work apps and builds timelines on a whiteboard. It announces itself in chat, won't join calls with outside guests unless an admin allows it, and leaves if one joins. It does not work in Zoom or Teams yet. Claude Data is now generally available. Microsoft shares fell about 2% after Bloomberg's report, though Business Insider says inflation and an Nvidia scandal also hit tech that week. Microsoft says its Office suite grew product and cloud revenue 16% last fiscal year. TechCrunch says OpenAI unveiled similar workplace tools weeks ago.
The claim
Anthropic product leader Nicholas Lynn told Bloomberg Claude Sheets is relatively basic and not built for sophisticated financial modeling.
Austin's take
Austin calls it a threat to Microsoft and Google: spreadsheets, whiteboards, meetings and company data cover a big slice of the workday, and a crowd of AI companies is chasing the same desk. Ron's hype case: it's beta, Meet-only, and the incumbents are still growing, which Austin called a fair point.
Why it matters
An AI that sits in meetings and edits your files changes how follow-up work gets done and puts pressure on the office suites most people use.
What to do
If you're on a paid plan, try Sheets on a simple task like a budget, not complex modeling. Teams on Meet should set admin rules before letting Claude into calls.
Not confirmed yet
The tools are in beta, and the Microsoft stock dip can't be pinned on Anthropic alone.

Ring's $249.99 indoor drone camera

Watch from 17:31

What happened
Ring first promised an indoor drone in 2020; it never shipped. The new drone cam costs $249.99 and is planned to ship in 2027 once regulators approve it. The camera is off in its dock, records only in flight, shows lights when on, and can launch automatically when another Ring device is triggered. Scheduled flights need a Ring Protect subscription. The Next Web reports flights of five to eight minutes, with about eight minutes of charging per minute flown. In a CBS News demo, the drone didn't land after losing connection as Ring said it should, and stair navigation and syncing with other Ring devices weren't working. In 2023 the FTC charged that every Ring employee could access customer videos and alleged one watched videos of at least 81 female users; Ring paid $5.8M to settle and denied breaking the law. Amazon's $2,350 Astro robot never caught on, per The Next Web, which calls the new $999.99 Rover Cam for businesses a rebranded Astro.
The claim
Amazon points to the drone's privacy design. Ring says the drone should land if it loses connection.
Austin's take
Austin says wait: it needs to prove itself in real homes, the battery means it spends most of its shift charging, and automation costs extra. He sees more potential in commercial buildings empty after hours. Ron argued a prerelease demo is for finding bugs and the trigger-launch feature is genuinely smart.
Why it matters
A flying camera inside your home raises the stakes on who can see the footage, a question Ring has faced before.
What to do
Wait for real-world reviews, and ask Ring directly who can access drone footage before buying.
Not confirmed yet
The CBS demo used a prerelease unit, and the 2023 case doesn't say who can access this drone's footage today. Shipping depends on regulator approval.

Work Claude chats show up in trade secret lawsuits

Watch from 22:05

What happened
Dorsia, a members-only restaurant booking platform, sued rival Rumor Group, saying a poached engineer used company-paid Claude to research what secrets he could take. Bloomberg Law reports the suit says Rumor offered a $240,000 salary, a $20,000 bonus and 5% equity that began vesting before he left, and agreed to cover his legal costs. A day earlier, Cubist, Point72's computer-driven trading unit, sued a fired analyst it says used work Claude and OpenAI's Codex to access its source code to join a competitor; its complaint says he 'had attempted to steal and likely did steal' confidential information. Law.com reports a Florida federal magistrate judge let a plaintiff seek chats and deletion logs directly from AI providers, possibly a first for Florida federal courts.
The claim
Cubist says it fired the analyst in September after an internal investigation. The analyst's lawyer calls the claims baseless and says he looks forward to setting the record straight.
Austin's take
Austin started at 'evidence': if a chat shows what happened, a judge should see it, case by case. After Ron argued that discoverable prompts push people to untracked personal devices and that contracts can prove poaching, Austin said Ron was right that existing rules can handle espionage without the chats.
Why it matters
Chats on a company AI account may not be private, and courts are starting to let litigants request them from AI providers directly.
What to do
Treat work AI chats like work email: assume they can be reviewed or subpoenaed, and keep them professional.
Not confirmed yet
All accusations are unproven, nobody has been charged, and a request for records doesn't mean any have been handed over.

Nobel-led panel says OpenAI's physics proof holds

Watch from 27:10

What happened
A panel led by Nobel physicist David Gross says OpenAI's AI-written proof holds, showing a simplified theory of the force inside atomic nuclei can be built. The panel relied on Lean, a program that checks each proof step against the rules. The proof spans 9,477 files; the audit found zero 'sorry' placeholders (skipped steps) in its dependencies, inspected 250,319 pieces of code (about 100,000 supporting the main result), matched every file to OpenAI's list and rebuilt it from scratch successfully. The panel used Astra, OpenAI's own AI, to help read the code. It says its only ties to OpenAI are accepting submissions and using Astra, and it has no present connection to the Clay Mathematics Institute. The proof does not show a mass gap or prove the infinite volume limit is unique, so the $1M Millennium Prize stays unclaimed. The panel wrote, 'we do not claim to fully understand the proof yet.'
The claim
OpenAI said its AI cracked part of a physics puzzle.
Austin's take
Austin says trust the part the panel approved: Lean checks every step regardless of who wrote it, and the panel judged the proof establishes the claimed theory. Save the trophy. Ron's doubt case: the referee used OpenAI's own AI and admits it doesn't understand the proof.
Why it matters
It's a test case for how experts verify AI-generated research too large for humans to fully read.
What to do
Read claims like this narrowly: a verified partial result, not a solved Millennium Prize problem.
Not confirmed yet
The panel says it does not fully understand the proof yet, and the mass gap and uniqueness questions remain open.

The Loop

The weekly newsletter

The week in AI in one email: what happened, why it matters, what to do about it.