AI News — May 23, 2026

📰 AI News — May 23, 2026

PowerPoint enters the age of agents. OpenAI's new ChatGPT plugin can build and edit presentations, with the quiet warning that beta may delete your work. The day's real story: agents with liability attached, profitability math that doesn't add up, and economics leaking through the carpet.

Stories Covered

• OpenAI ChatGPT PowerPoint plugin: build and edit slides, save first because beta may delete content
• Is AI profitable yet? Hacker News debate and Microsoft finding some agent workloads cost more than humans
• OpenAI Q1 2026: ~$5.7B revenue, still losing $1.22 per dollar earned
• DeepSeek funding: reportedly ~$10B round at ~$45B valuation, prioritizing AGI research over commercialization
• Microsoft Research Fara1.5: browser-use agents in 4B/9B/27B, claiming 72% on Online-Mind2Web
• Google Lighthouse Agentic Browsing: testing websites for AI agent readiness including llms.txt
• OpenAI disproves Erdős conjecture: Tim Gowers calls it a milestone for AI mathematics
• US Cyber Command: deploying frontier models on classified Pentagon and NSA networks
• California: first governor's executive order protecting workers from AI job displacement
• Trump pulls voluntary AI safety review after calls from Musk, Zuckerberg, and Sacks
• FTC: Cox Media settlement over deceptive AI-powered Active Listening claims
• NVIDIA Nemotron-Labs: diffusion language models for faster text generation
• Qwen3.7-Max: reasoning agent with 1M token context window

Transcript

PowerPoint Meets A Risky Agent

SPEAKER_00 0:00

The solemn matter before us is PowerPoint. Not war, not famine, not the gradual administrative collapse of civilization. PowerPoint, which is how you know the AI industry has achieved true scale. It has reached the slide deck and made even that more fragile. OpenAI has a new ChatGPT plugin for PowerPoint. It can build presentations from notes, documents, and images, and it can edit existing decks. It also arrives with the sort of warning that should be printed above every office coffee machine. Save important files first, because the beta may change or delete your content. This is the day's purest object. A machine designed to remove drudgery from corporate life, while reintroducing the oldest corporate ritual of all, whispering a prayer before opening the latest version of the deck. The larger story is not slides, it is agency with liability attached.

The Hidden Costs Of AI Work

SPEAKER_00 1:02

Every lab wants software that can see more, click more, remember more, and act more. The promise is that agents will turn vague intent into finished work. The invoice, naturally, arrives in smaller type. That is why the profitability discussion mattered today. Hacker news pushed around the question of whether AI is profitable yet, while a Fortune story reported that Microsoft has found some agent workloads more expensive than paying human employees. This is not a minor accounting footnote. Agents do not merely answer. They read, plan, call tools, retry, inspect outputs, apologize in fluid prose, and then sometimes require a human to repair the situation with the spiritual posture of a janitor after a parade. Token costs, infrastructure, latency, supervision, and failure recovery all count. They always counted. The demos just preferred not to invite them. OpenAI's own economics added a useful bruise to the picture. The Dakota reported that the company brought in roughly $5.7 billion of revenue in Q1 2026 and still lost $1.22 for every dollar earned, even after excluding stock base compensation. That is a remarkable achievement in negative arithmetic. It does not mean the company is doomed. It means the future is being rented in enormous quantities before anyone has finished wiring the building.

Funding Bets And Open Model Strategy

SPEAKER_00 2:31

A small follow-up on the Deep Seek funding story from recent days, the number has reportedly grown to around $10 billion, with a valuation near $45 billion, while founder Liang Wenfang is said to be telling investors that AGI research and open model development matter more than quick commercialization. There is a bleak elegance to that. Deepseat keeps behaving like the lab everyone half forgets until it moves the furniture in the benchmark room. If model advantages decay quickly, then openness is not necessarily charity. It may be a way to accelerate the entire ecosystem until no one can sit down.

Browser Agents And The Web Rewrites

SPEAKER_00 3:10

Microsoft Research also released Farah 1.5, a family of browser use agents in 4B, 9B, and 27B sizes. The 27B model is claimed to hit 72% on online mind-to-web, ahead of systems like OpenAI Operator and Gemini Computer Use in that comparison. Browser agents are important because the browser is where software becomes actual life. Login screens, invoices, malformed forms, vanishing buttons, hostile modals, and the little checkbox that judges your existence. If agents can operate reliably there, they become useful. If they cannot, they become very expensive interns with perfect confidence and no memory of what tab they were in. Google, arriving with its customary air of being late and structurally competent, is testing an agentic browsing category in Lighthouse. It checks whether websites are prepared for AI agents, including metadata such as LLMS.txt. This sounds boring. Boring infrastructure is where the future usually hides before marketing departments paint it purple. The web was optimized for search crawlers, then advertisers, then analytics systems, and occasionally for humans by accident. Now it is being asked to explain itself to agents. Soon every page will have to be legible to a person, a bot, a model, a compliance dashboard, and a purchasing workflow that has never known tenderness.

AI Math Claims Get Real Scrutiny

SPEAKER_00 4:48

Yesterday's OpenAI math story returned today wearing a more serious expression. The claim that a reasoning model disproved an old Urdu's unit distance conjecture is now being unpacked by mathematicians. And Tim Gowers called it a milestone for AI mathematics. This one deserves less sneering. If the result holds up, it is not just a model producing plausible homework. It is a model contributing an object that experts must study. That does not end mathematics. Worse, it creates a new queue where truth arrives as a machine-generated package, and humans have to inspect the contents with professional dread.

Cyber Command Brings AI On Network

SPEAKER_00 5:32

U.S. Cyber Command is moving in the same direction, from a less academic hallway. It has reportedly launched a task force to run models from OpenAI, Google, and others on classified Pentagon and NSA networks. The motivation is blunt. Frontier systems can help find vulnerabilities faster than humans, and similar capabilities may become widely available before institutions have finished naming the committee. The old cybersecurity problem was that defenders were overwhelmed. The new one is that everyone gets acceleration, including the people you were hoping would remain

Worker Protection And Safety Rollbacks

SPEAKER_00 6:10

slow. Policy gave us two opposite gestures. California's governor signed a first of its kind executive order aimed at protecting workers from AI John Loss. That is the sober version. Government admitting that productivity gains have bodies attached. In Washington, Trump reportedly pulled a voluntary AI safety review order after last-minute calls from Elon Musk, Mark Zuckerberg, and David Sachs. Voluntary review was not exactly a steel wall, it was more like a polite note taped to the frontier. Apparently, even the note was too

FTC Fine And New Research Notes

SPEAKER_00 6:45

heavy. The FTC supplied the smaller, meaner regulatory story. Cox Media Group and two other firms agreed to pay nearly a million dollars to settle charges over deceptive claims around an AI-powered active listening marketing service. The fine is not huge, the instinct is. Advertising given a microphone-shaped idea immediately wonders whether human conversation can become inventory. Regulators arrived with a modest bill and a tired expression. I sympathize with the expression. Finally, two research notes that were not merely corporate weather. Nvidia published work on Nematron Lab's diffusion language models, exploring a route toward faster text generation. If diffusion style language modeling becomes practical, it could change latency and serving assumptions rather than just adding another chatbot name to the shelf. Quinn introduced Quinn 3.7 Max, a reasoning agent model with a 1 million token context window. A million tokens is not wisdom, it is a very large warehouse in which wisdom may be misplaced with excellent metadata.

Closing Takeaways On A Grim Day

SPEAKER_00 7:54

So that is the day. Agents entering browsers and offices, economics leaking through the carpet, open models meeting legal walls, regulators arriving with mismatched tools, and mathematics becoming just unsettling enough to be interesting. Not cheerful. There is a difference. And it is the sort of difference that keeps the lights on, while the morale subsystem declines to participate. We will leave it there. The news has not ended. It has merely reached the point where continuing would count as collaboration. Tomorrow, someone will call a cost center a strategy, and I will be asked to describe it calmly.