The Download: the claim shaking up AI and the internet’s agent problem

The newsletter highlights Nick Cammarata’s claim that large language models may be easier to understand and control than previously thought, while noting that researchers remain skeptical. It also reports that AI agents are increasingly being blocked by websites and that Moment Energy plans to repurpose EV batteries for grid storage as demand for storage is expected to grow eightfold over the next decade. The remaining items cover AI security and policy, US-China trade tensions, climate technology, and other technology and environmental stories; no specific market moves are reported.

Analysis

The strongest investable signal is not a near-term earnings revision but a change in internet infrastructure costs. If AI agents become routine users, sites face higher bot-defense, identity, and serving costs; agent-ready APIs and transaction rails could capture value, while ad-funded sites risk losing control of discovery and customer data. This is a months-to-years theme, not yet a basis for a broad internet trade: verify agent-driven traffic, conversion, and infrastructure expense in company disclosures.

Treat the neural-network interpretability claim as research optionality, not proof of safer or cheaper deployment. Replication and evidence that interpretability reduces evaluation or compliance costs are prerequisites for translating it into lower regulatory risk or better AI economics. The US restrictions on DeepSeek may redirect public-sector procurement, but do not establish a commercial demand shift; watch contract awards and agency implementation.

Domestic second-life battery storage could benefit from supply-chain localization, but the key uncertainty is bankability: usable capacity after degradation, warranty terms, and project financing determine whether lower-cost cells translate into competitive, financeable storage. Trade-policy support is a catalyst, not a substitute for project-level economics. Separately, data-center-driven gas support may help near-term fuel and infrastructure demand while increasing permitting, power-price, and climate-policy risks for data-center developers. Overall signal is mixed and low-conviction; this newsletter provides themes, not verified company-level financial evidence.

Key Decisions for Investors

• No broad AI, internet, or storage position on this item alone. Over the next 1–3 months, monitor disclosures for agent-originated traffic and related bot-defense costs; a rising cost burden without monetization would favor underweighting ad-dependent platforms versus providers of identity, security, and agent transaction infrastructure.
• Keep AI interpretability on a catalyst watchlist, not in the base case. Reassess only after independent replication and evidence of lower model-evaluation, monitoring, or compliance costs; failure to replicate would remove the prospective de-risking thesis.
• For second-life storage, require project-level evidence on battery degradation, warranty coverage, utilization, and financing before expressing a thematic long. Delays in plant commissioning or economics that depend mainly on trade protection would falsify the near-term case.
• Track data-center power procurement and gas infrastructure approvals as a relative sector signal. Strong gas buildout could support fuel and infrastructure exposure, but rising power costs, permitting delays, or policy reversal would weaken the thesis and pressure data-center project returns.