● TabBOOOne page. More perspective.✎ Customize
Your world, one good read at a time

TabLOID

Show your work. Then let someone else check it.

No. 28 · Miami BeachTuesday, October 27, 2026Read at your own pace
Conceptual illustration of a robot unfurling a huge manuscript while researchers inspect one dot.
Catch up quick
  1. OpenAI has released major proof claims for scrutiny. Formal verification and scientific understanding remain distinct steps.
  2. Anthropic reported a profitable June quarter at the operating level, alongside a large net loss.
  3. Cloudflare says a patched GLM-5.3 cache failure mixed customers' data in 155 responses.
  4. Microsoft has offered eligible staff a voluntary exit. Its announcement gives no total headcount.
  5. Meta's financing plan includes approximately $35 billion of offerings and a separate proposed $60 billion program.

Open what interests you

01THE PROOF DESKA very big proof meets a very big magnifying glassOpenAI's claims are now in the hands of researchersConceptual satire of AI proof claims under expert inspection.

OpenAI's October 26 release claims a refutation of the Exponential Time Hypothesis, a longstanding assumption used to derive conditional limits on computation. It says the result was formalized and checked in Lean. An external group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will examine and explain it.

The company also claims an all-dimensions local-smoothing proof and a rigorous existence result for quantum chromodynamics. Public materials differ: the mathematics and computer-science outputs are available, while the physics group led by David Gross is withholding its artifacts as it develops a write-up.

These are company-reported research results under expert examination. They do not establish practical speedups or settle P versus NP. Our read: publishing the work gives researchers something concrete to test; it does not make the testing redundant.

OpenAI's October 26 announcement and proof links

02THE BOOKS DESKOne quarter in the black. A different line in deep red.Anthropic's operating result and net loss tell different storiesA small folded note casts an enormous red shadow beside an accountant, a metaphor for accounting distinctions.

Anthropic's S-1 reports $11.555 billion in revenue and $974 million in operating income for the quarter ended June 30. Its net loss that quarter was $66.309 billion, including a $59.597 billion expense from changes in the fair value of convertible notes, derivative liabilities and SAFEs.

The balance sheet lists $131.651 billion in cash, cash equivalents and short-term investments at June 30. That is a dated liquidity measure, not a statement that nearly every dollar ever raised remains untouched. The filing also reports approximately 900% net revenue retention at September 30.

A fair-value charge is different from operating cash spending. It still belongs in the accounts. The useful question is whether operating performance persists as infrastructure commitments grow. One quarter cannot settle the economics of an industry.

S-1 filed October 26, including quarterly statements and balance sheet

03THE PRIVACY DESKThe robot handed over somebody else's coatCloudflare's AI cache failure crossed customer boundariesA robot cloakroom attendant gives a guest the wrong coat, a metaphor for mixed customer data.

Cloudflare says 155 GLM-5.3 responses contained data from other customers' recent requests between August 27 and September 9. All involved prompts longer than 681,280 tokens. Its October 26 disclosure says no other Workers AI model was affected and it found no evidence of deliberate exploitation.

The failure combined a memory pool smaller than the advertised context window with an SGLang error-handling bug. A rejected request continued through processing and read stale cache pages instead of stopping. Cloudflare says it corrected the configuration, patched the engine and restored the model on September 9.

This was a boundary failure in inference infrastructure, not evidence that every hosted model leaked. The uncomfortable lesson is ordinary: a successful HTTP response can still contain the wrong customer's information.

Cloudflare's incident report and remediation

04THE WORK DESKMicrosoft opens a voluntary exit doorAn offer to eligible teams, with the scale still unspecifiedAn office worker peers through an envelope-shaped door, a conceptual illustration of an exit offer.

Microsoft announced a Voluntary Exit Offer on October 26 for eligible employees at level 70 and below in participating organizations in Canada, India, the United Kingdom and the United States.

The company says the time-limited program includes financial support and career-planning resources. Eligibility varies by organization. Its public announcement does not say how many people qualify, how many will accept or how many roles will ultimately disappear.

An offer is not a completed departure count. We have left out the circulating 60,000-person figure because the primary announcement does not substantiate it.

Microsoft's announcement

One more worth opening
Meta's $95 billion headline comes in pieces

Meta announced approximately $35 billion in concurrent offerings, plus a proposed $60 billion at-the-market program. The latter is not assured. Its release also says chipmakers' interest in up to $5 billion of common shares is nonbinding. These are proposed financing steps, not $95 billion already received. Meta's October 26 financing announcement

Today's good news

A better forecast, with room for uncertainty

Researchers reported a machine-learning correction that improves probabilistic weather forecasts two to six weeks ahead. MIT's October 14 account describes gains in forecasting skill and extreme-event prediction. This could help planning; it does not promise an exact forecast for a particular day six weeks away.

Read the source ↗
Close ×Enlarged comic