The Long Game: Meta Launches Muse While Everyone Else Ships Targets

Capability & Integration

• Meta launched Muse on September 8: a personal AI agent meant to run continuously in its own “Secure VM,” while a separate system called Sentinel — not the model itself — decides whether each proposed action is auto-allowed, blocked, or sent back for approval, per Meta and Axios. It’s free up to 100M tokens/week, rolling out first on iOS, Android, and web in the US.

• Update: Artificial Analysis revised its Intelligence Index twice this week; v4.3 now scores GPT-6 Astra and Claude Fable 5.1 tied at 53 points apiece, reversing the gap reported here yesterday that had Astra behind, per The Decoder and OfficeChai. Astra still costs far less per task — $3.26 versus Fable 5.1’s $7.63 on average — per Artificial Analysis.

  Review

  The $3.26/$7.63 cost figures could not be independently confirmed (the x.com source link is unreachable); other reporting on the same AA comparison shows a similar ratio but different absolute dollar figures (~$1.67 vs $3.76 in one snapshot), consistent with AA’s live index having moved since this snapshot. Suggested action: verify directly against Artificial Analysis’s current page, or soften to relative terms (“roughly 2x cheaper”) if exact figures can’t be reconfirmed.

Robotics

• BMW moved Figure AI’s newer Figure 03 into logistics-sequencing work at its Spartanburg plant, following a completed 10-month pilot in which the prior Figure 02 helped build more than 30,000 X3 vehicles, per BMW Group and The Robot Report. It’s a continuation of one existing automaker relationship, not a broader commercial launch.

  Review

  Independent coverage (The Robot Report, Repairer Driven News, Interesting Engineering) describes the Figure 02 deployment as an 11-month pilot, not 10 months. Suggested action: fix “10-month” to “11-month.”

  Review

  The “30,000 X3 vehicles” figure traces to BMW’s own press release about its own program — single-sourced, no independent corroboration. Suggested action: add > [!unverified].

• Tesla is converting part of its Fremont line into a first-generation Optimus line (target: 1 million units/year) and breaking ground on a much larger Giga Texas factory (target: 10 million/year, long-term), per The Robot Report and Teslarati. Tesla’s own timeline puts first Giga Texas production in 2027 — today that’s a construction site and a stated capacity target, not a robot coming off a line.

Hardware & Supply Chain

• Qualcomm and Amazon announced a multi-generation partnership on September 8 to co-design custom AI inference silicon and 1.6-terabit optical interconnects for AWS data centers; Qualcomm issued Amazon a warrant for up to 25 million shares (roughly $4 billion), vesting through 2036 and contingent on Amazon purchasing up to $60 billion of Qualcomm products, per SiliconANGLE and Time.

  Review

  The warrant/vesting/purchase-commitment figures trace to the two companies’ own joint disclosure of the deal terms — single-sourced by organizations reporting on themselves, despite pickup by CNBC/TipRanks. Suggested action: add > [!unverified].

• A September 7 KB Securities note reports Samsung and SK hynix’s combined finished-memory inventory has fallen below 10 days’ supply, as HBM4 — which consumes roughly three times the wafer capacity of standard DRAM — continues to displace conventional production, per Seoul Economic Daily.

  The brokerage’s forecast that 2027 will be “the most severe shortage in history” is one firm’s projection, not an industry consensus.

• China’s Ministry of Industry and Information Technology published a five-year plan targeting 9,800 exaflops of AI computing capacity by 2030, up from 1,590 exaflops in 2025, backed by roughly $532 billion in planned infrastructure investment, per South China Morning Post. It’s a capacity target running to 2030, not a statement about chips available today.

  Review

  The exaflop/dollar figures trace to the ministry’s own plan for its own targets — single-sourced despite SCMP’s reputable pickup (an organization’s own announcement of its own numbers counts as single-sourced). Suggested action: add > [!unverified].

Environmental & Cultural Impact

• A Bloomberg analysis found 70% of U.S. locations with the steepest electricity-rate increases sit within 50 miles of significant data-center activity — part of the reasoning behind New York’s first-in-the-nation statewide moratorium on new data centers using 50 megawatts or more, in effect for up to a year, per Bloomberg. Household electricity rates rose 13% nationally in 2025.
• More than a dozen other states have active data-center moratorium or restriction bills moving through 2026 sessions, per NCSL and MultiState trackers — but Maine’s governor vetoed a similar statewide pause in April, so New York’s approach remains the exception, not an emerging norm.

AI in the Wild

Valentino’s fully AI-generated ad campaign for its “DeVain” handbag, released late last year, resurfaced this week in a Yahoo Finance retrospective on brands pulling AI-generated campaigns after backlash — its AI origin was never in dispute, only whether audiences would accept it, per Yahoo Finance and eWeek. Nine months on, one failed handbag ad is still doing more to shape brand AI strategy than any success story has.

Takeaway

Review Notes