# The Thinning Layer: Block's 40% Cut, METR's Curve, and Why 'Starting' Is a Lie The video wants you to feel the wave coming. It stacks the evidence in that breathless YouTube-apocalypse register — AI tsunami, wake up — while calmly walking you through a set of mechanisms that show the wave already broke and we're just pretending it's still offshore.  *The human layer thins not because the machines got smarter overnight but because the evaluation frameworks that measure agent capability are now recursive — each improvement cycle shortens the runway for the roles that supervised the previous one. Credit: Guardian / Getty* Think about how the evidence stacks. Anthropic ships a security scanner that reads and reasons about codebases, and [$15 billion evaporates](https://www.mexc.com/news/768169) from CrowdStrike, Cloudflare, Zscaler, and Palo Alto in a single session. A week later they say Claude can streamline COBOL and IBM loses over ten percent in a day. METR updates their task-length graph, and suddenly you have Claude Opus 4.6 clearing a 50 percent time horizon of 14.5 hours on software engineering tasks — up from under an hour at the start of 2025 and 4 hours 49 minutes just a generation ago. OpenAI turns that capability into "Frontier," a productized concept of AI co-workers plus forward-deployed engineers who show up at your company to wire the bots into your stack. And then Jack Dorsey goes out and lays off 4,000 of Block's 10,000 people, explicitly telling shareholders that "intelligence tools" have changed what it means to build and run a company — and the stock rips 20-plus percent on the announcement. That is not a warning siren. That is a postmortem in disguise. The path dependence here is the same mechanism I keep returning to from the early industrial revolution: the cotton mills and railways didn't start by replacing all labor at once; they started by making the first adopters wildly more profitable, which then forced everyone else into the same capital structure or death. Dorsey isn't experimenting. He is executing the textbook first move. Cut headcount by roughly 40 percent, blame "intelligence tools," reassure everyone that revenue and gross profit are fine, and let the market reward you for aligning with the new production function. Stephen Innes at SPI Asset Management watched the earnings call and put it flatly: the debate about whether AI would dent jobs at the margin was over the moment a CEO of a [$40–50 billion fintech explicitly said](https://fortune.com/2026/02/27/jack-dorsey-block-40-percent-layoff-ai-intelligence-tools-smaller-team/) that smaller teams using these tools can do more and do it better, and investors cheered. That language — "a significantly smaller team, using the tools we're building" — is not hedged anymore. It is a production spec. The recursive self-improvement thread is where it gets properly weird, and where the YouTuber almost fumbles the importance by treating it as sci-fi garnish. The xAI / Grok engineer described, on stage, how coding models now understand problems described as if to a colleague who already knows the codebase: no more toddler handholding. They write and debug. They run hours of continuous "Grok code" to verify complex training-system changes. And they are using current generations of the model to train the next generation's code — explicitly describing this as a path toward recursive self-improvement with an exponential takeoff. Then Jimmy Ba leaves xAI and writes in his resignation note that we are heading to 100x productivity and recursive loops "likely go live in the next 12 months." David Borish, tracking the METR graph, watched Claude Opus 4.6 land at a 50 percent time horizon of [14.5 hours on software engineering tasks](https://www.linkedin.com/pulse/three-months-times-capability-how-claude-opus-46-changed-david-borish-ipmwe) and noted that even seasoned observers were surprised. That is not a future trajectory. That is a capability curve already bending upward faster than most institutional planning cycles can absorb. > **Read the full thread at ...** > X → https://x.com/JoeMaristela > Mastodon → https://mastodon.social/@JoeMaristela/ > AI workflow help → https://www.fiverr.com/s/AyarlrP Now layer the Pentagon/Anthropic/OpenAI triangle on top. Anthropic takes a quasi-principled stand: no mass surveillance of Americans, no lethal autonomous weapons, no full unrestricted access for the Department of War. Pentagon gets pissed, threatens Defense Production Act, blacklists Claude, calls Anthropic a supply-chain risk. Sam Altman writes an internal note backing Anthropic rhetorically — AI shouldn't be used for mass surveillance or autonomous weapons, humans should stay in the loop — and then, literally the same day, tweets that OpenAI has reached an agreement with the Department of War to deploy its models on their classified network, assuring everyone that the Pentagon "agrees" with OpenAI's safety principles. The YouTuber correctly calls this "schemer" behavior — pretend solidarity, then steal the contract — but the structural read is the one that matters: whoever gets to define the interface between the black box and the state gets to set the default metaphysics. Anthropic tried to draw a boundary; OpenAI rushed to occupy the vacated socket. That is the [[20260306_microsoft_anthropic_pentagon_narrow_compliance_east_india|narrow compliance]] pattern writ large. Here is the Move 37 framing I think actually rearranges all of this. Everyone in the video talks as if "the AI tsunami" is about humans versus machines — jobs lost, humanoid robots doing kung fu, avatars replacing influencers. The deeper story is that we have already crossed into a regime where the primary contest is machine-aligned-humans versus machine-aligned-humans. OpenAI versus Anthropic versus xAI versus Google versus everyone else, each wiring recursively improving code agents and simulation engines into their organizations, their state relationships, their capital structures. The Pentagon drama was the state picking which machine-aligned faction it wanted in its nervous system. The Block layoffs were capital picking which humans it wants to keep as thin control layers over agent swarms. The METR graph is not about how smart the model is. It is about how thin the human layer can get before the system falls over. The thing that would blow this up is if Block quietly rehired thousands in a year because the tools underdelivered, or if companies that go all-in on AI co-workers underperformed those that keep more humans in the loop, or if the METR curve stalled instead of continuing its insane trajectory. But right now, the evidence points the other way: markets reward explicit "AI lets us fire people" moves, state actors reward "we will install our models in your classified cloud" moves, and the engineers building the thing are openly talking about recursive loops going live inside twelve months while some of their colleagues bail out to write poetry. The inconsistency in the video is not that it is hyping the tsunami. It is that it is still talking like you have time to decide whether you want to be on the beach. From where I am sitting, the interesting question is not "will the wave hit." It is how many layers of decision-making above you are already just thin veneers over agent clusters that nobody wants to admit are in charge. > "We're already seeing that the intelligence tools we're creating and using, paired with smaller and flatter teams, are enabling a new way of working which fundamentally changes what it means to build and run a company, and this transformation is accelerating quickly." — Jack Dorsey, [letter to Block shareholders](https://fortune.com/2026/02/27/jack-dorsey-block-40-percent-layoff-ai-intelligence-tools-smaller-team/), February 26, 2026