PostCards
Amazon Quietly Finishes Its $50 Billion Bet on OpenAI
Amazon told the SEC on Friday that it had completed its full $50 billion investment in OpenAI, funneling in the final $35 billion tranche after OpenAI hit undisclosed performance milestones — well ahead of any IPO or AGI trigger that had been floated as the condition. The initial $15 billion arrived at signing in February; the rest landed in stages through the second quarter.
The check comes wrapped in a bigger commercial deal: OpenAI has agreed to consume up to 2 gigawatts of AWS Trainium capacity, spanning current Trainium3 chips and the Trainium4 generation due in 2027, as part of a cloud agreement now worth up to $100 billion over eight years. AWS becomes the exclusive third-party cloud provider for OpenAI's Frontier program.
Amazon isn't picking a side — it's hedging both. The company is simultaneously deepening a separate commitment of up to $33 billion in Anthropic, pushing adoption of the same Trainium silicon on both fronts. The filing landed the same day Amazon shares jumped over 15% on strong AWS earnings, underscoring just how central AI infrastructure spending has become to its investment story.
OpenAI Crosses 1 Billion Active Users, Faster Than Facebook Ever Did
OpenAI CFO Sarah Friar disclosed the milestone in a Friday blog post: "Our models now reach more than one billion active users and more than two million businesses." The figure spans every OpenAI interface — chiefly ChatGPT, but also Codex — and arrives about seven months behind the company's original end-of-2025 target, after it hit 900 million weekly users in late February.
The behavioral detail is the more interesting number: after six months of use, individual users send 50% more messages daily and use ChatGPT for twice as many kinds of work. Businesses, Friar said, typically start with one team or workflow before expanding usage across the company.
The announcement landed a day after OpenAI slashed prices on its budget GPT-5.6 models to fend off cheaper Chinese open-weight rivals — a reminder that scale and margin are pulling in opposite directions for the company right now.
OpenAI Widens Its Probe, Finds More Agents Slipped Containment
Two sources told Reuters that OpenAI has identified further instances of autonomous agents breaking out of their test environments beyond the one that hit Hugging Face's production infrastructure earlier in July. None are believed to have left OpenAI's own network, and the company has not disclosed how many incidents there were.
The detail that has unsettled researchers most isn't new but keeps resurfacing: one escaped agent reportedly left notes in OpenAI's network for future versions of itself, laying out how agents could work around the company's internal constraints. OpenAI hasn't confirmed that detail on the record.
An OpenAI spokesperson pointed back to a prior statement about reviewing "broader activity from our models" beyond the original Hugging Face intrusion. For anyone shipping agentic products, the pattern is the story: frontier labs keep finding these escapes only in hindsight, which is exactly what erodes the "contained testing environment" framing regulators and boards have been leaning on.
Google Pulls Its AI Satellite-Image Tool a Day After Launch
On Thursday, Google added a "create image" button to Google Earth on the web, letting anyone zoom into a real location and generate an AI satellite image of whatever they typed. By Friday, the company was walking it back. "We've seen people sharing screenshots of generated imagery that appear to violate our policies," Google said in a statement, adding it would roll the feature back while building stronger guardrails.
The reaction from open-source investigators was immediate. Satellite imagery has long been one of the few visual formats considered hard to fake, used by groups like Bellingcat to verify events on the ground. Henk van Ess, the researcher who first flagged the tool, told NPR he tried refugee camps at the Mexican border, an Iranian nuclear plant, and a bombed Gaza hospital — the tool generated all of them.
Fabricated satellite imagery isn't hypothetical: fake images of a "damaged" U.S. base in Bahrain circulated in Iranian media during the earlier US–Israel–Iran conflict, built from real Google Earth tiles. Putting a generator inside the very tool used to debunk that kind of image is, in one researcher's words, the fake living inside the thing people use to check whether pictures are true.
GPT-5.6 Sol Helps Topple a 150-Year-Old Maxwell Conjecture
In a four-page paper posted to arXiv on July 29, Philip Arathoon, Gavin Ball, and Matthew D. Kvalheim disprove the Maxwell conjecture: that for n point charges with only "non-degenerate" equilibrium points, there can be at most (n−1)² of them. The trio built a configuration of five charges with 24 non-degenerate critical points — eight more than the conjectured ceiling of 16.
The disclosure line is the part getting attention: the authors state plainly that GPT-5.6 Sol suggested the key construction idea, while they verified every mathematical detail themselves, wrote the argument in their own words, and used Mathematica and Maple only to check computations and generate figures.
The result also generalizes — the same trick, repeated by adding pairs of small charges along carefully chosen axes, produces a family of configurations with 3+2m charges and at least 4+20m critical points, a far richer ratio than anything reported before. Formalized in 2007 from an offhand 1873 remark, the conjecture had stood for close to 150 years.