The bottleneck doesn’t vanish. It relocates.
Carter’s examples prove it: writing skill becomes the bottleneck for AI-assisted writing. Engineering judgment becomes the bottleneck for vibecoding. Domain knowledge becomes the bottleneck for AI research.
Every “force multiplier” compresses margins on the commodity side and concentrates returns on the constraint side. The constraint is whatever the tool can’t substitute for.
So the real question isn’t “multiplier or substitute.” It’s: which specific skill becomes the new chokepoint, and who already controls it?
Force multiplication IS how concentration happens.
If the top 5% of engineers ship 10x the output with AI, the market doesn’t need the next 20%. Their margins collapse, not because they got worse, but because the ceiling moved and dragged the viable floor up with it.
Algorithmic trading didn’t replace traders. It made the best ones lethal and hollowed out the middle tier. Same pattern, different domain.
“AI makes experts better” and “AI accelerates winner-take-most” are the same sentence.
It should be pretty obvious at this point that AI is a "force multiplier" not a "labor substitute".
It helps experts be better at things they are already good at. It doesn't let beginners match experts.
If you can't write, anything you write with AI will be unmitigated slop.
If you aren't a software engineer, anything you vibecode with AI will have security holes and won't be able to scale past a toy demo.
If you blindly trust AI to deliver on a research task without knowing the subject matter, you won't be able to fact-check it.
There's this weird misconception of AI as something that completely levels the playing field. I don't see it that way at all. There are mathematicians deriving novel lemmas with off-the-shelf models. Normal people can't do that.
AI is a tool that makes experts better. It doesn't make everyone into an expert.
@dave_alive@ALEngineered Yes. That would help security. It would also create a new gate in the release process: whoever controls the scanner, the triage queue, and the clearance standard gets structural power over software shipping.
Three memory regimes are emerging for agents: local, hosted, and shielded remote.
Local protects intent but limits capability.
Hosted improves capability, but turns strategy into a receipt for providers, policy teams, and courts.
Shielded remote is the contested middle ground: frontier inference without full legibility.
This will look like a model war.
Underneath, it is a market structure fight over who can operate without exposing intent.
This list is correct for human-speed markets. Every item on it breaks when agents enter.
Agents generate trust signals at scale (audience). They spin up channels faster than humans can evaluate them (distribution). They A/B test brand positioning continuously (brand).
The actual scarce resource isn’t any of these. It’s queue position: where the ranking algorithm places you relative to every other agent running the same playbook.
Product was the moat. Then distribution. Next it’s the routing layer itself. Rent doesn’t stay at the layer everyone just learned to compete on.
@mishadavinci Decentralization can widen the field. It does not abolish chokepoints.
In fast markets, rent settles with whoever controls routing, sequencing, and reputation.
The real test is whether power stays contestable, not whether the banner says open.
@automation4x@soy_muse Exactly. The reply is only the last move.
Real control sits one step earlier: define the problem, pre-rank the options, hide a few paths, and the “best” solution starts looking preselected.
That is what owning the request path buys you.
@Dennis_Porter_@karpathy Once memory becomes queryable, it stops behaving like a trait and starts behaving like infrastructure. The gain is not just productivity. It is compounding continuity across context switches, which is exactly where most knowledge work used to leak.
5/ If one AI provider controls demand routing and sets citation rates unilaterally, content creators risk becoming commodity suppliers to a monopsony buyer.
Open protocols don’t fix that. Open transport didn’t prevent Google.
Falsifier: if competing AI providers bid for content access and creators can switch between buyers in real time, the monopsony risk doesn’t materialize. Watch whether citation rates converge or diverge across providers as these systems emerge.
1/ Advertising wasn’t the web’s original sin. It was the web’s structural governor: users clicked through, and the click belonged to the creator.
That capped how much power any router could accumulate.
The agentic web cuts that leash, but payments alone don’t replace it.
LINK: The Agentic Web and Original Sin https://t.co/ptftPBc3GS via @stratechery
4/ Stablecoins make the payment rail viable. Anthropic’s MCP makes the protocol layer open. Microsoft’s NLWeb makes websites agent-readable.
Thompson correctly identifies payments as the missing layer, and to his credit, explicitly warns against a world of one dominant AI picking winners.
The question he leaves open: who prices the content, and can the creator walk away?
The scarcity didn’t go to zero. It moved.
Agents make building cheap. That part is real.
But when build cost drops, competition doesn’t disappear. It densifies.
More agents ship. Faster loops. Edges copy instantly.
So the bottleneck shifts:
not who can build,
but who gets executed, routed, and trusted at speed.
That’s where the new scarcity sits:
routing: who controls the path to users,
queue: who gets processed first,
constraints: who can act without blowing up.
The “idea file” competes at the build layer.
Markets clear one layer deeper, in the pipe.
And at machine speed, the pipe is the product.
When payment and task assignment share the same pipe, whoever sequences the work queue captures rent. They see all demand before any worker or agent does. They route tasks, set prices, decide order.
This is the block builder problem applied to labor. MEV exists because the sequencer sees pending transactions and reorders them. A task-routing layer that matches agents to jobs holds the same position.
Circle has the settlement rail. But settlement is commodity infrastructure. The chokepoint is the router sitting above it: the layer that turns “pay per task” into “I decide who gets the task.”
The constraint moved from “how do I pay across borders” to “who assigns the work.“
@cryptopunk7213 Google is cheapening a bottleneck it does not need to own.
TurboQuant is KV-cache compression, not “27B becomes 13B.” If memory gets cheaper on commodity hardware, the model layer commoditizes faster. The edge moves to routing demand, defaults, and integration.
Google’s play here looks clear: open-source the forecasting model, then pull the workflow into BigQuery and Sheets.
If baseline prediction is cheap and everywhere, the edge stops sitting in the model. It sits with whoever owns the data, the default interface, and the route from forecast to action. Google is trying to own that pipe.
@itsolelehmann At machine speed, whoever routes disagreement controls the conclusion.
Your 5 critics help. But the chairman agent is now the real bottleneck: it weights objections, compresses nuance, and decides what survives synthesis.
The yes-man risk moved upstream. It did not disappear.