The Wake: May 20, 2026

The Wake is a daily briefing from George's saved internet. The issue is written as a newsletter first. The tweets are the source material, preserved below for receipts.

Source window: May 19, 2026. Signals: 4 bookmarks and 8 likes.

Brief

A small number of connected moments over the last 24 hours reveal a coherent shift in how dominant AI companies are trying to harvest internal expertise, how challenger labs are positioning enterprise agents, and how fragile the infrastructure that stitches it all together remains. Mark Zuckerberg, in leaked internal audio, framed a plan to use employees as the dataset to accelerate coding models ahead of layoffs. At the same time Anthropic is hiring headline talent and shipping secure tunnels for managed agents that reach private networks without opening inbound ports. Meanwhile startup tooling like Codex is proving product-market fit, but cloud-account blocks and provider outages remain a single point of failure. The result: a race to extract institutional knowledge into models while the technical and legal scaffolding to do that safely is still immature.

Train before you fire: the operational calculus

The leaked audio attributed to Mark Zuckerberg is notable not because it is novel but because it makes explicit a tactic many companies have quietly adopted: intensively instrument internal workflows so models can learn from high-skill employees before those employees are reduced in force. The logic is straightforward. High-quality internal signals accelerate model capability in narrow, product-relevant tasks: notably coding: and therefore reduce the marginal cost of replacing those tasks with automation.

That read has three immediate implications.

• Labor risk. Even if framed as efficiency, the tactic erodes trust and invites legal scrutiny. Are employees’ outputs being used with consent? Are commercial secrets being ingested into models that will then be externalized? Expect labor groups, data privacy advocates, and possibly regulators to probe. This is fertile ground for disputes over trade secrets, contract law, and employment protections.
• Talent churn. People who perceive themselves as being trained to obsolescence will leave proactively. Publicly airing the strategy accelerates the brain drain it aims to delay.
• Competitive lock-in. Training on proprietary internal flows produces model advantages that are hard for rivals to replicate: provided legal and reputational costs are manageable.

If you run engineering orgs, take this as a demand signal: you will be asked to choose between short-term model gains and long-term morale, retention, and risk. There is no neutral path.

Anthropic, Karpathy, and the enterprise agent land grab

Anthropic’s moves this week are a coordinated one-two. Recruiting Andrej Karpathy signals a serious R&D push and a bid for credibility at the research frontier. The product news is equally pointed: managed agents that can reach internal systems over an outbound-only, post-quantum encrypted tunnel (built on Cloudflare Tunnel). That changes the adoption calculus for enterprise agents.

Outbound-only connectors remove a major barrier to trials: IT teams can avoid exposing internal endpoints. For buyers, that lowers friction and speeds pilots. For Anthropic and similar vendors, it reduces a key security objection and creates a positioned dependency: enterprises will trust a vendor’s connector to be both the runway and the rails for their agents.

Two caveats.

• Vendor concentration. Using an intermediary like Cloudflare as the underlying tunneling layer centralizes a new chokepoint. If Cloudflare or Anthropic’s managed control plane has an incident or changes policy, customers feel it immediately.
• Trust and credentials. These connectors still need credentials, access policies, and runtime isolation. “Outbound-only” simplifies networking but does not eliminate the need for rigorous identity, least privilege, and auditability.

Anthropic is betting that secure, low-friction network primitives plus world-class hires will be enough to win enterprise mindshare. The path is open, but the trust architecture has to scale faster than the sales motions.

Product maturity vs infrastructure fragility

Codex and related developer-facing apps are maturing. Users praise mobile integration, Git workflows, and actual utility. That product traction matters: it proves that models can deliver value in day-to-day engineering workflows, which makes the business case for wider, deeper integrations: including the very training strategies discussed earlier.

But product-level progress is colliding with infrastructure fragility. A startup reported its Google Cloud account was blocked, taking services down and forcing a public post-mortem. This is not a niche problem. Provider account suspensions, policy-driven blockages, and opaque enforcement are now existential threats for companies that rely on cloud providers for both control planes and data planes.

Combine the two threads and you get a brittle stack: companies moving valuable, sensitive work into managed agents and model training workflows that depend on third-party tunnels, cloud control planes, and single-provider credentialing. A single blocked account or a Cloudflare incident can halt the whole pipeline.

Immediate takeaways for operators: assume provider access can fail, design recovery paths, and insist on contractual clarity about account suspensions and incident response. Multicloud and provider-agnostic connectors are back in fashion for reasons that have nothing to do with sticker price.

Tokens, incentives, and platform leverage

Sam Altman’s outreach to token-first startups and the playful tweets about Codex throttling are small signals of a larger strategic posture: platform vendors are experimenting with new incentive designs to seed ecosystems. Token instruments promise alignment with startups but introduce a new set of governance, liquidity, and regulatory questions.

If you are an investor or operator, watch how token-funded cohorts handle control of model access, data rights, and revenue share. Token instruments can accelerate network effects, but they also hardwire dependencies. A token grant from a dominant model provider can make a startup economically and operationally reliant on that provider’s roadmap and rate limits.

The crude incentive is powerful: give startups upside denominated in the provider’s own ledger, and you denature competition. Expect more of this. Expect also that regulators will ask whether token arrangements are a backdoor to market foreclosure.

What to watch

• Regulatory and legal responses. Follow labor advocates and privacy regulators for complaints or investigations about using employee work for model training without explicit consent.
• Anthropic product adoption and vendor stack choices. Watch which enterprises start running managed agents with outbound-only tunnels and whether they accept Cloudflare (or alternate) dependencies.
• Talent movement. Karpathy’s role at Anthropic could accelerate a wave of hires or signal new research directions. Keep an eye on who they recruit and what open-source or research outputs appear.
• Cloud account enforcement incidents. More post-mortems like the recent one will matter. Track provider policies on account suspension and whether enterprise contracts begin to demand incident SLA guarantees and escrowed control plane access.
• Token-funded startups. Monitor YC and similar programs for how token investments change governance and commercial terms, and whether tokens create subtle lock-in to model providers.

Takeaway for this morning: the industry is actively converting human expertise into model capability at scale, and the levers being used are both technical and contractual. That conversion promises faster product gains but creates concentrated points of failure and fresh legal exposure. If you are building or advising AI product teams, prioritize explicit data-use policies, redundant infrastructure, and contractual protections now.

Generated from Birdclaw bookmarks and likes. Edited by Ody before publication.