Microsoft staff face a green card block; OpenAI's revenue is lower than we heard

Good morning.

Today’s claims face a check against the record.

Microsoft’s 6,000 layoffs last year became JD Vance’s case for a green card suspension on October 8. Microsoft’s H-1B application count measures something different from Vance’s visas.

OpenAI put its annualized revenue near $50 billion at September’s end. The figure investors were using deserves a second look before anyone starts comparing companies.

On October 6, OpenAI posted 372 families of mathematical results. Dana Moshkovitz’s 9-year-old son heard that “a robot solved the math problem you worked on for your whole career.” Three manuscripts were withdrawn the next day over a sign error.

Stay curious,

Marcus Schuler

The Labor Department said on October 8 it would stop accepting or processing permanent labor certification applications from Microsoft, Adobe and six IT outsourcing firms.

Labor Department certification is the first step most employer-sponsored green cards need. On October 8, Keith Sonderling said Microsoft and Adobe were suspended “due to multiple active federal investigations.”

Microsoft said 80% of roughly 6,000 H-1B applications last fiscal year were to extend or change the status of existing employees. Vance counted visas obtained, whereas Microsoft counted applications filed. His one-and-a-half ratio requires adding 6,300 visas to nearly 3,000 green cards against 6,000 layoffs last year.

Why This Matters:

• Microsoft employees seeking green cards face a blocked certification step toward permanent residency, which would free them from dependence on their employer.
• The Labor Department’s suspension leaves employers facing a duration Vance described as “as long as it needs to.”

OpenAI told investors annualized revenue approached $50 billion at September’s end, the Financial Times reported on October 8.

OpenAI excludes the cloud-partner sales Anthropic includes, explaining the roughly $20 billion gap from nearly $70 billion circulated in late September. Investors had adjusted OpenAI’s August figure to an estimated $40 billion. Without audited OpenAI figures, comparisons rest on run rates often calculated by multiplying monthly revenue by 12.

Stack Overflow's October 7 piece explains how to confine an AI agent's decisions inside ordinary, testable code.

Varun Jindal gives a separate, tested component sole authority to change business state after validating the model's proposal. "Structured output removes the parsing failure mode, not the judgment failure mode," Jindal writes.

• GlobalFoundries signed a five-year, $2 billion agreement with TSMC to make AI-chip packaging interposers in New York, with volume production expected to ramp in the first half of 2028.
• Arena raised a $200 million Series B at a $3.1 billion valuation on October 8 and released an Alignment Index checking whether models exceed user instructions.
• Manus raised more than $500 million in its first round since Chinese authorities ordered its $2 billion sale to Meta unwound in April.
• Google launched Playground on October 7 for US users aged 18 and older to build playable games from text prompts.
• OpenAI brought GPT-6 to Free and Go users October 8 after Luna's under-18 sexual-content safety score fell to 0.904 from August's GPT-5.6 Luna score of 0.984.
• Claude Haiku 5.5 matches GPT-6 Luna's $0.10 input price per million tokens on prompts up to 100,000 tokens, but used about three times Luna's output tokens per task at maximum effort.

A reference call often produces warm impressions instead of evidence. Use the candidate's own claims to verify the scope and result of their work.

Your raw input:

Paste the job description, the candidate's main claims, your unresolved concern, and the referee's role and working relationship with the candidate.

The prompt:

Why this works: The script anchors the call to claims already made and asks for observable examples. The answer notes help the interviewer spot ambiguity without steering the referee.

What to use: Any capable general model can handle this. For sensitive hiring data, use an enterprise tool with retention controls and remove protected personal details.

Credit: Midjourney

Prompt: Modern tarot card illustration, The Chariot. A woman drives a light purple vintage car along a winding coastal mountain road, viewed from behind at a distance.

OpenAI posted 372 families of mathematical results on October 6. It withdrew three manuscripts the next day over a sign error.

Our take: OpenAI had computer-checked proofs for 300 of its 719 top-line results as of October 7. Its own mathematicians do not yet understand many of the results. Publishing hundreds of proofs in that condition is an ambitious way to assign homework. Three manuscripts have already departed over a sign error.

Dana Moshkovitz needed an AI model to help read the AI's claimed proof of the Unique Games Conjecture. "It feels like something written by someone who's on psychedelics," she wrote. Her 9-year-old son had already heard that "a robot solved the math problem you worked on for your whole career."

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai