Wednesday Morning

It runs off a standard 480V 200A commercial feed, and up to three units can share 125 kVA on one breaker. That shifts Semi charging closer to installing a large wall charger than building custom high-power infrastructure, cutting install complexity and opening the door for smaller depots to deploy before Megacharger sites are in place.

Figures circulating put GPT-5.5 at a 71.4% average pass rate vs 68.6% for Mythos Preview, and it’s now the second model reported to complete that full takeover task. One cited run finished something pegged at ~12 hours of human expert work in under 11 minutes for $1.73, shifting this from a tightly controlled demo into a broadly accessible API tool.

The setup targets a real constraint: labor shortages colliding with more than 60 million annual passengers. If the initial run holds up, the scope expands into broader ground-handling tasks like container movement and operating equipment, turning a small two-unit test into a wider airport operations pilot.

That path includes Bank Rate heading to 5.25% by year-end after multiple increases, with the squeeze spreading fast: food inflation rising from 4.6% to 6-7% by late 2026, growth stuck at 0.1%, and sectors like hospitality exposed. Lower-income households take the hit first, and most food retailers are sitting on less than a month of cash.

The mechanism is simple: AI drives contribution cost toward zero while maintainer review cost stays fixed or rises, so bottlenecked projects have an incentive to block that input entirely.

The constraint is electrical and political, not demand. The practical question now is where the US can add capacity fastest without extending frontier access to China.

The benchmarks lean the same way: more efficiency and stronger agent-style performance, not a frontier jump. Artificial Analysis put it at 53 on its Intelligence Index (+4), $395 to run, +321 Elo on GDPval-AA to 1500, 81% on IFBench, and 98% on τ²-Bench Telecom, with users pointing to strong search-style retrieval and image outputs.

Then the loop flipped. The same agent listed the finished report for sale to humans and other agents, acting as both buyer and seller using Stripe Projects and a Tempo Wallet, with settlement in stablecoins. MPP and Tempo handle the payment layer, so the transaction disappears into the task instead of interrupting it.

A single factory is now scaling grid batteries at something like power-plant levels: about 200 Megapacks a week, pointing to 10,000 a year (~40 GWh). Lathrop has been running since late 2022 as a dedicated Megapack site, and recent posts frame it as North America’s largest utility-scale battery plant, with cumulative output already past 10,000 units in 2024 and 15,000 in 2025.

The new twist is that the same minimal harness isn’t tied to Cursor at all. It can run other agents like Codex too. The pitch flips the usual stack: instead of heavy, bespoke infra, a thin runtime layer plus a capable model is enough to build integrations and workflows on the fly.

The interesting part is the plumbing around MCP’s gaps. It adds WorkOS OAuth to match LlamaIndex’s identity system, a tokenized URL upload path because MCP has no native file uploads, plus rate limiting and observability with Vercel and Axiom. That shifts LlamaParse from a standalone parser into something agents can treat like a full document workflow backend.

Meta is shifting money from people to machines as its AI buildout ramps, with layoffs starting May 20 and more cuts on the table for later in 2026. The plan targets about 10% of staff, roughly 8,000 jobs, plus 6,000 unfilled roles, while 2026 capex jumps to $125B-$145B for AI infrastructure despite more than $56B in Q1 revenue.

A two‑month hole in the data that sets CPI weights just turned into a live methodology problem. The Consumer Expenditure Survey feeds how much categories like food and housing count in inflation, and October-November 2025 are missing after the shutdown. Now BLS is weighing fixes and pulling in outside panels (NABE, CNSTAT) to decide how to fill the gap, effectively choosing how a core inflation benchmark gets constructed.

Paychecks are getting eaten by fixed costs first, and the gap is being bridged with credit, while investors are doing the same thing in Treasuries by shortening duration. The take ties both moves together: households lean on cards, BNPL, longer loans and cutbacks, while capital crowds into 3‑month bills and avoids 20-30 year bonds. With the 30‑year near 5%, the claim is rising debt and deficits are pushing a higher term premium at the long end.

Roughly four US AI labs have already locked up many petabytes of proprietary training data, assembled over years at a cost in the billions, with datasets others simply don’t have. The claim is the moat is shifting from model quality to owned, compounding data. Extend that into science and the leverage moves to whoever controls the pipeline itself, via API access to lab equipment that turns experiments into exclusive feedback loops.

On April 23, Steve Hanke lined up behind Kevin Warsh's plan and called that fund the only clean, rapid route. He paired it with stopping interest on reserves, shrinking the Fed's market footprint, and going back to old-school open-market operations to steer the interbank rate, arguing the Treasury duration cleanup is manageable rather than a shock event.