Cambridge AI Safety Plan, Instinct’s Agent Growth and Meta’s Enterprise Push

AI is becoming an institutional concern, a commercial intermediary and a new layer of enterprise infrastructure at the same time. This TBPN episode connects a Cambridge safety proposal, Instinct’s adoption metrics and the launch of Meta Enterprise Platform.

Measuring AI that builds AI

The proposal discussed in the episode starts with a measurement problem. Labs already say their models write code, generate experiments and contribute to research, but each company uses different units. The proposal calls for comparable indicators covering the share of code, spending and R&D activity actually performed by AI systems.

A common framework could make AI-driven research acceleration easier to observe. It would also let governments and outside analysts compare labs without relying on incompatible internal claims.

Planning for incidents before they happen

A second part calls for predefined government responses to new kinds of disruption: a major cyberattack carried out by an agent, a biosafety incident or severe economic dislocation. The point is not that every answer already exists, but that responsibilities and procedures should be established before an emergency.

TBPN contrasts that institutional approach with the prepper culture reported in parts of the AI safety community. Remote land and a stocked bunker may offer little protection if the owner has no relationship with the surrounding community; an isolated reserve can become a target rather than a refuge.

Instinct converts trust into transaction volume

Figures highlighted from the Instinct founder’s appearance on Invest Like the Best describe an invite-only product already associated with roughly $1 billion in transaction volume. The hosts explicitly note that it is unclear whether the figure is cumulative or annualized.

Other claims include daily growth of 5% to 10%, compute demand doubling roughly every week and a team of about fourteen people. Travel reportedly represents half of transaction volume. About 40% of users are said to share a personal card within three weeks, while connecting at least one piece of sensitive information is associated with retention near 80%.

Transactions still run through third-party cards, so the volume does not automatically equal Instinct revenue. Possible business models include merchant commissions, an Instinct-issued card and interchange income. The harder asset to replicate is permission: users allowing an agent to complete tasks and purchases with progressively less supervision.

Rational agents, unstable systems

The episode also discusses a warning from Apollo’s chief economist. Financial agents could make the same individually optimal decisions at the same moment, amplifying a market move or accelerating deposit flight. No explicit coordination would be required if similar models pursue similar objectives.

The hosts see this as a gradual rather than immediate risk. Users still have to connect accounts, approve actions and build trust before accepting autonomous execution. That transition could give banks and regulators time to adapt.

Meta builds a new enterprise pillar

Meta Enterprise Platform is set to combine Muse Agent, Meta Business Agent, Muse API, Muse Code, the company’s models and its infrastructure. CJ Desai will lead the effort.

TBPN argues that the platform could bundle several revenue streams under one label: agent software, inference, tokens, services and large compute agreements. Meta is therefore creating more than a single product; it is building a broad commercial container that could become a new growth pillar.

Source

• Chaîne: TBPN
• Vidéo source: https://www.youtube.com/watch?v=bDVLnSIbrlU