Show Your Work: Astra's Machine-Checked Proofs Land the Same Week as a Deepfaked President

Capability & Integration

• OpenAI previewed its next model family, Astra, by publishing ten machine-checkable Lean proofs of decade-old open problems in math and theoretical computer science — including a non-sofic groups construction and new sphere-packing bounds — for a reported $2,000 in compute, per The Decoder. Fields medalist Timothy Gowers said he’d recommend one proof to a top journal without hesitation — proofs anyone can verify independently, unlike most model claims.

  Review

  “$2,000 in compute” — traces to a single OpenAI self-announcement relayed via one outlet; an org’s own number about itself counts as single-sourced — add [!unverified] or cite a second outlet confirming the figure.

• Microsoft’s “Project Perception,” an agentic platform that hunts, triages, and patches security vulnerabilities, enters public preview today inside Microsoft Defender, paired with a new in-house model reserving GPT-5.4 for roughly the hardest 10% of tasks, per GeekWire.

• The EU’s AI Act transparency rules took effect August 2: chatbots must disclose they’re AI, and deepfakes must carry machine-readable labels, per the European Commission (existing systems have until December 2 to add the labeling).

Robotics

• BYD confirmed it will unveil a functional humanoid at its Di Space showroom in Zhengzhou in early August, per South China Morning Post. BYD’s own roadmap targets “application verification and routine deployment” only by year-end — a showroom reveal, not a deployed fleet.
• Figure AI manufactured its 1,000th Figure 03 unit on July 23, sustaining roughly one robot per hour at its BotQ facility, per eWeek. Its target is 12,000 units a year scaling toward 100,000 — a production claim, distinct from a confirmed count of robots doing paid work.

Hardware & Supply Chain

• TSMC posted record Q2 profit, up 77.4% year-over-year, with HPC and AI chips 66% of revenue; CoWoS packaging is sold out with 52–78 week lead times, per CNBC. TSMC raised 2026 capex guidance to $60–64 billion (from $52–56 billion) and added $100 billion to its Arizona buildout.

  Review

  “52–78 week lead times” — independent coverage of the same earnings call describes CoWoS as sold out through 2026/into 2027 but doesn’t corroborate this specific week range — verify the exact figure against the CNBC source.

  Review

  This bullet cites a July 16 CNBC article with no date context, sitting beside an explicitly dated July 30 “Update:” bullet — add a parenthetical date so readers don’t read both as equally fresh.

• Update: Amazon’s Q2 earnings, flagged in Friday’s brief as a capex-raise candidate, delivered: AWS revenue hit $42.2 billion, up 37% — its fastest growth since 2021 — on a $496 billion backlog, per CNBC. Amazon raised 2026 capex guidance to roughly $220 billion, citing AI demand and rising memory costs, and expects capacity constraints into 2027–28.

Environmental & Cultural Impact

• Global data-center electricity demand rose 17% in 2025, with AI-focused facility consumption up 50%; combined capex from the five largest operators (AWS, Google, Meta, Microsoft, Equinix) topped $400 billion in 2025 and is projected to rise a further 75% in 2026, per the IEA. Total data-center power use is projected to roughly double, to 950 TWh, by 2030.

  Review

  “five largest operators (AWS, Google, Meta, Microsoft, Equinix)” — Equinix’s capex is an order of magnitude smaller than the four hyperscalers; the IEA source page returned a fetch error during review — verify this grouping directly against the IEA report before next citation.

• Intel is preparing fresh layoffs within its Data Center and AI Group — a unit that posted 22% revenue growth last quarter — with no headcount figure disclosed, per Tom’s Hardware. Intel’s global headcount has fallen from ~132,000 in 2022 to ~81,000 today.

  Review

  “~132,000 in 2022 to ~81,000 today” — single-sourced dramatic (~40%) workforce stat with no [!unverified] flag — add one or cite a second outlet corroborating the range.

AI in the Wild

A July 24 political video circulating ahead of a Taipei rally used what Taiwan’s Criminal Investigation Bureau assessed as AI-cloned audio of President William Lai criticizing his own government’s handling of a cooking-oil scandal; the DPP called it deceptive, while the creator — a former KMT youth-department director — called the probe an attack on free speech, per Vision Times and Taipei Times. The paper ran an editorial today asking “is it time for a deepfake law” — the story isn’t really about the clip, but a democracy that hasn’t decided whether faking a president’s voice is fraud or satire.

Takeaway