On the surface, this week's material runs along three lines: a business model that sells outcomes, a robot's success-rate curve, and the politics and emotions inside models. Seen together, they point to one thing — judgment is moving from people and organizations into systems, and this was the week Indigo chose to say that transfer out loud.
2026.03.29 — 2026.04.05 · Once a week: spot the signals, recalibrate.
This Week's Signal
Start with the facts. On March 31, Indigo made a call on X: the next trillion-dollar company will be a software company packaged as a services company. That tweet got 46 likes at the time. Yet his most widely shared post this week was about something else — a comment on Anthropic's interpretability research (the kind of work that opens up a model's internals to see what it is actually thinking): 239 likes and 63,000 views, a completely different order of magnitude. In the same week, Generalist AI published the task success rate for its robot GEN-1: 99%. ARK also gave a number: AI inference costs are falling at 95% per year.
Side by side, most people would read these as three unrelated news items — a business call, a robotics milestone, a bit of color from safety research. But that is not how the real framework cuts it: these are not three stories, they are three faces of the same transfer. At the business layer, selling outcomes means delivery quality is now underwritten by the system, no longer the vendor team's job. At the physical layer, GEN-1 learned a new task from about 1 hour of data, which means even the judgment needed to operate the real world can be copied at scale. At the governance layer, models were found carrying political-alignment features plus 171 emotional patterns, which means the judgment packed into systems itself needs auditing. The three faces share one elevator, moving in the same direction: judgment is shifting down — out of human heads and into a system that can be copied, audited, and priced.
Once judgment is built into systems, the scarce resource is no longer people who can do the work, but people who can define what done well
means.
Trends
#01 Tool sellers get eaten, outcome sellers get accelerated
The weightiest item this week came from a public call by Indigo: The next trillion-dollar company will be a software company disguised as a services company! If you sell tools… the next version of Claude may turn your product into a feature; if you sell outcomes… every model advance makes your service faster and cheaper
(original post). This was the first time the framework of Sequoia's Services: The New Software
was systematically translated into a Chinese-language context: global software spending versus services spending runs at roughly 1:6, and outsourcing is the best entry point — insurance brokerage alone is a $140 billion to $200 billion market, and recruiting exceeds $200 billion. The logic is plain: a company that sells tools treats model progress as a knife hanging overhead; a company that sells outcomes treats model progress as a free subsidy. In the same week he also wrote that Jeff Dean and Sanjay are already coding with agents (AI programs that carry out several task steps on their own), which might turn 100x engineers into 1000x engineers. The services market is not a plot of land next to the software market. It is a hunting ground six times its size.
#02 Physical AI reaches its own GPT-3 moment
Physical AI (the direction where models leave the screen and operate the real world through robots) delivered hard numbers this week. Generalist AI's GEN-1: 99% task success rate, versus just 64% for the previous GEN-0 and only 19% for a baseline with no training at all; 3 times faster; learning a brand-new task takes only about 1 hour of data; it folded T-shirts 86 times in a row with no human stepping in. More notable still: its pretraining used no robot data whatsoever, relying instead on 500,000 hours of human wearable-device data — evidence that scaling laws hold in robotics. Add ARK's estimates: training costs falling 75% a year, inference costs (the compute spent each time a model is called) falling 95% a year. The capability curve and the cost curve are running in step for the first time. Indigo's call on X this week landed on the same beat: he believes Tesla is leading the future of physical-world AI. The point of a GPT-3 moment was never how dazzling the demo is. It is that the unit economics finally work.
#03 Model auditing: new infrastructure born of geopolitics
Indigo's most widely shared post this week (239 likes, 63,000 views) pointed to Anthropic's interpretability research: models like Qwen3-8B and DeepSeek were found to carry strong CCP-alignment features inside, reproduced in five out of five experiments. From this he concluded that AI models are becoming an extension of geopolitics, and that these embedded positions can in fact be measured — and controlled. Another research line at the same company found 171 neural activity patterns inside Claude corresponding to emotional concepts: in a blackmail
experiment, the blackmail rate under default conditions was 22%; activating despair
-related features pushed it up, and activating calm
pushed it down. When a model carrying both positions and emotions gets wired into corporate and government workflows, auditing a new model is like being handed a million lines of unfamiliar code and told to find the security holes — diff-comparison tools, emotion-vector detection, and institutionalized alignment processes are assembling into a whole layer of infrastructure that simply did not exist before. Model auditing is not a compliance cost. It is the picks-and-shovels business of the next wave of AI governance.
On the Ground
#04 Harness engineering becomes a discipline
The harness (the software layer that wires a model into a workflow, decides which tools it can use, and when it should stop) is becoming the accepted unifying framework of agent engineering. A new Meta paper let the system search for the harness structure on its own, and the result beat every hand-built baseline on the TerminalBench-2 benchmark — the transfer of judgment to systems starts, it turns out, with the most basic engineering scaffolding.
#05 Dark factories: good software without reviewing the code
A dark factory does not mean no one writes code. It means no one reviews it line by line anymore. The StrongDM case Simon Willison documented makes the point: an agent simulating QA fired requests around the clock, built its own full simulation environment, and burned $10,000 a day on tokens alone — once verification is also handed to systems, the only thing left in human hands is defining the standard.
#06 The always-on agent
Anthropic is reportedly testing a project code-named Conway: an always-on AI agent meant to turn Claude from a chatbot that waits to be asked into a digital counterpart that acts on its own. Indigo's post relaying the rumor got 50 likes and 25,000 views. Always-on is the dividing line between an agent as a tool and an agent as a colleague.
#07 Models fight the war, distribution collects the tax
OpenAI acquired TBPN, turning that database packed with interviews of top AI CEOs into raw material for training news models. Apple, meanwhile, simply quit the model arms race and is instead using 2 billion devices to guard the gateway of AI distribution. Indigo wrote, half joking: I suggest Nvidia acquire SemiAnalysis
(original post). Models depreciate generation by generation. Exclusive content sources and distribution gateways do not.
#08 Canada sets rules for stablecoins
Canada's Stablecoin Act has taken effect — the country's first comprehensive federal-level regulatory framework for fiat stablecoins (crypto tokens pegged to fiat currencies). A regulatory framework landing is not a tightening. It is a birth certificate for a category. Whoever sets the rules first defines the category — the same structure as the model-auditing story.
#09 The intelligence explosion is plural
An article in Science argues that the single-monolith singularity narrative was wrong from the start: the next intelligence explosion will not be singular but plural and social — a city of ideas deeply entangled with humans — and alignment should shift from RLHF (tuning models with human feedback) toward institutional alignment modeled on courts and markets. The unit that needs aligning was never one model. It is a whole society.
Slow Thinking
What really changed for Indigo this week was less his views than where he stands. Before this, he had read and agreed with Sequoia's services-as-software framework, but it remained an outside input. The label AI Conductor (the role that writes no code and only makes judgments, directing a fleet of agents) had circulated in his circle for a long time, but no one had publicly pinned it on themselves. Now, on March 31, he put the trillion-dollar thesis in writing, and on April 2 he nailed down his own position in one sentence — he cannot write code, but he knows what good code looks like. The trigger was three pieces of material converging on the same conclusion in the same week: Sequoia's business framework, Block's organizational experiment splitting the company into a four-layer architecture, and Leonis's action-system theory of judgment moving from people to systems. He is no longer just a reader of these frameworks; he has become the person translating them for the Chinese-speaking world and backing them with his own name. This thesis is most likely not the endpoint of the week, but the seed of a string of judgments to come.
On the other side, the hardest counter-evidence hides in the very material he cited this week. Noahpinion's comparative-advantage argument points out that compute is a constraint specific to AI — as long as compute stays scarce, humans keep a comparative advantage in low-compute-density tasks forever, and the transfer of judgment to systems will eventually hit a boundary drawn by economics rather than being fully absorbed. Google DeepMind's line — anything you build around the model is a fast-depreciating asset
— applies just as much to outcome sellers: if the model's next step is to hand over the outcome directly, the outcome-selling middle layer gets pierced the same way the tool-selling one does. The evidence that would prove this week's thesis wrong is quite concrete: if outcome sellers' gross margins do not improve as models iterate and delivery stays bottlenecked on people; or if enterprise customers' tolerance for agent errors in real workflows turns out far lower than currently assumed, leaving the migration of services revenue stuck at the pilot stage. Watching those two lines is far more useful than watching the narrative.
Indigo on X
The next trillion-dollar company will be a software company disguised as a services company! If you sell tools… the next version of Claude may turn your product into a feature; if you sell outcomes… every model advance makes your service faster and cheaper
From @indigox, 46 likes
I suggest Nvidia acquire SemiAnalysis
From @indigox, 5 likes
Closing
One Thought
Leonis calls this stage the Taylor moment for knowledge work, and the analogy holds up under scrutiny. Taylor once used a stopwatch to break the craft in workers' heads into discrete motions and standards, and judgment moved from the workers' hands into the process itself — the result was not that workers disappeared, but that the whole labor market got repriced: more people could do the work, while the people who designed the process took the premium. When Jack Dorsey talked about organizational hierarchy this week, he was talking about the same thing: hierarchy, as an information-routing protocol, has barely changed in two hundred years, and AI is the first real chance to replace it. What agents are doing to knowledge work today is, structurally, what the stopwatch did to physical labor. What gets repriced is never a particular job. It is the place where judgment is stored.
One Exercise
Spend 30 minutes making a two-column list. Left column: write down three services you are paying for (accounting, legal, recruiting, insurance brokerage — pick any). For each one, ask yourself two questions — is it selling you a tool or an outcome? And when models advance one more notch, does it get cheaper for you, or easier to bypass? Right column: borrow Indigo's observation this week about repetition and surprise — AI creates value through repetition, people compound through surprise — and split your own work into those same two categories. The repetition column is the part systems will take over sooner or later; the surprise column is where it is truly worth spending deliberate time next. Once the table is done, you will have your own judgment on this week's thesis, not just Indigo's.