Announcement

Apr 20, 2026

Agents Don't See Ads. And Four Other Shifts Rewiring AI.

Five signals from the Stanford CS 153 lectures — the founders and CTOs building the next trillion-dollar AI stack, speaking on the record.

The Quiet Course Where the Real Thesis Lives

Stanford's CS 153: Frontier Systems is not a normal class. Anjney Midha and Michael Abbott run it as a closed-room lecture series where the people actually building the AI infrastructure of the next decade — Jensen Huang, Sam Altman, Lisa Su, Satya Nadella, Andrej Karpathy, the founders of Mistral, ElevenLabs, Palantir, Luma, Black Forest Labs, Reddit — walk in front of fewer than 100 students and tell the truth about what they are building. No press. Minimal marketing. The lectures go up on the CS 153 channel a few days later if the speaker signs off.

The signal-to-noise ratio is the highest you will find anywhere in AI right now.

We went through seven of these lectures and pulled out the five shifts that matter for anyone building, investing in, or positioning against AI. Not predictions. Things happening in the present tense, on the record, from the people making them happen.

If you run an operating business, these five shifts rewrite your next eighteen months. We are leading with the one that is the single slide that belongs in every operator's next board deck.

Shift 01 — Agents Don't See Ads

Steve Huffman co-founded Reddit in 2005 with $12,000 from Y Combinator, sold it eighteen months later, watched it nearly die down to a single employee, came back in 2015, rebuilt it, took it public. Reddit's corpus now produces the equivalent of English Wikipedia every two weeks, and Huffman confirmed on the record that every major LLM in existence was trained on it, "whether folks admit it or not."

The data-licensing story is important — Google, OpenAI, the first-ever public content policy distinct from a privacy policy — but it is not the main event.

The main event is this: the agent economy is coming, and agents do not see ads.

Right now, AI agents interact with the internet through browsers. They load HTML, interpret it, extract information. Huffman sees the near future clearly — platforms will publish APIs designed specifically for agents. Not scraping. Dedicated agent endpoints. And the trillion-dollar question follows: if agents use Reddit instead of humans, and agents do not see ads, what is the fair exchange of value?

This is not a Reddit problem. This is the platform economics question of the next decade. It applies to every internet business. When agents book flights for you, they ignore the banner ads for hotel upgrades. When they research products, they skip sponsored results. When they browse, they do not scroll past native advertising. The entire $700 billion digital advertising economy was built on the assumption that humans see ads. Agents break that assumption.

Huffman split Reddit users into two archetypes. Scrollers come for community — they want human connection and cannot be replaced by agents. Seekers come for information — and agent summarization is already eating that behavior. Reddit Answers, their RAG product citing back to source posts, is their bet on owning the seeker experience before an agent owns it for them.

Operator read. If your business cannot be read, understood, and transacted with by a machine, it is becoming invisible. Structured data, clean APIs, and machine-interpretable product information are no longer nice-to-haves. They are existential. Agent-readable is the new mobile-responsive.

Shift 02 — Post-Training Is the New Frontier

Guillaume Lample, co-founder of Mistral, said it on the record: "These last couple of years we were focused on pre-training. Now we are shifting, putting more effort on post-training. The companies that will be the most innovative will not necessarily be the ones with the most compute, but the ones with the most flexibility."

Pre-training is buying ingredients. Post-training is knowing how to cook. You can buy wagyu beef and truffles and still serve a mediocre dinner. A small team with basic ingredients who knows how to season, fine-tune, and plate will beat you every time. That is the current dynamic playing out between billion-dollar incumbents and lean operator teams.

A few hundred well-structured examples from a specific use case now outperforms the most expensive general-purpose models on the planet. That is not an exaggeration. That is the current benchmark reality. Open-source checkpoints are raw material — the value lives in the last mile: fine-tuning, synthetic data generation, deployment infrastructure, evaluation loops. Ninety percent of companies cannot even deploy an open-source checkpoint competently. That is the arbitrage.

The real lesson from DeepSeek R1 is not the model itself. It is that reasoning-via-reinforcement-learning is the next scaling axis, and it runs on a fundamentally different infrastructure profile than pre-training. A new capability surface just opened up under the incumbents.

Operator read. If you hold deep domain knowledge in any real-world vertical — healthcare, law, finance, real estate, manufacturing, insurance — you have a window to build something a big company cannot easily copy. The window closes as the tooling standardizes. You are in the hoarding phase, not the golden age. Move.

Shift 03 — Voice Is Becoming Infrastructure

Mati Staniszewski, founder and CEO of ElevenLabs, walked into that Stanford lecture hall and dropped numbers that should change how every operator thinks about customer interaction. $430 million in revenue in 36 months. Over $100 million in additional ARR in a single quarter. 450 employees. That is roughly one dollar of ARR per dollar of payroll, putting ElevenLabs in the same metabolic tier as Anthropic.

The number is not the point. The category is the point.

ElevenLabs is not building a voice product. They are building the audio cloud — the infrastructure layer through which every business will interact with its customers via voice inside of two to three years. Transcription, reasoning, generation, emotional control, and real-time turn-taking, stitched into a single platform.

The defining technical debate in voice AI right now is cascaded versus fused. Cascaded pipelines — separate speech-to-text, reasoning, and text-to-speech models — win on reliability and auditability. You can swap components, enforce guardrails, call tools, authenticate accounts, process payments. Fused end-to-end models win on latency and emotional coherence. They respond in 300 milliseconds and maintain emotional state across a conversation, but sacrifice observability. Mati's on-record prediction: the future is hybrid. Low-stakes interactions run fused. The moment the agent needs to authenticate, transact, or execute a tool, the system switches to cascaded.

The real breakthrough was not quality or latency. It was controllability. Until roughly six months ago, you could not direct a voice model like a director on set. Now you can say "slower, more reassuring" or "more dramatic, hold that line" and the model delivers. ElevenLabs invested heavily in annotating emotional delivery — peppy, stressed, reassuring, flat — and that single capability unlocked studio and enterprise adoption simultaneously.

Mati sees the endgame as three to five conversational platforms — analogous to three to four cloud providers — mediating all business-to-customer voice. Support. Sales. Marketing. Education. Internal training. The whole stack.

Operator read. If you are building any consumer or enterprise product, design for voice as a primary input and output channel. The controllability breakthrough means 2025–2026 is the adoption window. Waiting until 2027 puts you on the wrong side of a platform shift.

Shift 04 — The Software-Industrial Complex Is Broken

Shyam Sankar has been at Palantir for nineteen years. Employee number thirteen. CTO and EVP. His lecture was the most contrarian and arguably the most important of the entire course.

His diagnosis: "There is a legitimation crisis. Our institutions — government and commercial — don't work. The C-suite has a steering wheel. They're diligently trying to steer. What they don't realize is the steering wheel is a prop from the Jungle Cruise ride at Disneyland. It's not connected to anything."

The evidence is on the record. Companies spent tens of billions of dollars on enterprise supply chain software. COVID hit and all of it fell over in two weeks. A paper tiger. The only IT investments that actually kept companies running through the crisis were Zoom and Teams. Elon Musk built Tesla and SpaceX's entire manufacturing software stack from scratch because every commercial off-the-shelf product he evaluated was, in his words, garbage. The only commercial package he still uses is the general ledger.

Sankar's core abstraction is worth internalizing. A military kill chain — sensor to shooter — and a commercial value chain — supplier to customer — are the same thing: a decision chain. Every decision constrains downstream options. Optimizing this chain is the meta-problem of the next decade. A computer scientist looking at both would see the same structure.

Palantir's Apollo platform runs 5,000 microservices across 1,000 production environments, including 100 air-gapped installations, pushing 100,000 upgrades per week with container image lifespans of 40 to 72 hours. When Log4j hit, most enterprises did not even know which versions of which libraries were running where. Palantir knew across all 1,000 environments and had a patched library deployed inside 24 hours with no manual intervention.

That is not a nicer IT department. That is a new category of software.

Operator read. The bar is no longer "does it work?" It is "does it work in crisis?" Compliance, security, and operational readiness belong encoded into the software, not into human process. The next decade of enterprise wins goes to teams that design for failure modes the incumbents never tested against.

Shift 05 — The Intelligence Manufacturing Loop

Anjney Midha runs one of the largest AI infrastructure funds in the world. His Stanford lecture laid out the thesis that ties everything else together: we have discovered a predictable machine for converting hardware dollars into software revenue.

Every time a frontier lab brings up new compute, capabilities jump 60 to 90 days later. Revenue follows on the same lag. One dollar of AI chips is producing roughly ten dollars of software revenue downstream. Hard assets — land, power, shells — trade at three to four times revenue. Software revenue trades at thirty to forty. The output is ten times more valuable than the input. That gap is what is fueling the single largest capex cycle in modern business history.

The five largest tech companies will spend more on infrastructure in the next three years than they spent in the preceding thirty years combined. $300B. $600B. $1.2T. Year over year. This is printed in their public earnings reports. It is not a forecast.

Every general-purpose technology runs the same cycle: invention, hoarding, panic, standardization, golden age. Steel took twenty-eight years. Fiber optics took six. AI compute is running on a three-to-five year cycle, and we are in year two. The companies that laid the fiber in 1995 went bankrupt. The companies that built on top of that fiber — Google, Amazon, Facebook — became trillion-dollar businesses. Same pattern, ten times the scale.

Operator read. Pick your position on the loop. Compute. Capabilities. Revenue extraction. Your proximity to the flywheel determines your leverage. Being adjacent and passive is the most expensive place to stand right now.

The Synthesis

Read the five shifts as one picture and a coherent thesis emerges.

The entire internet — and every business sitting on top of it — is about to be consumed by autonomous agents that do not behave like humans (Shift 1). The arbitrage is not at the base-model layer — it is in domain-specific post-training pipelines (Shift 2). Voice is being plumbed as the primary interface to those systems (Shift 3). Enterprise software is being rebuilt from first principles to survive real-world failure modes (Shift 4). And behind all of it, we are in the hoarding phase of the largest infrastructure cycle in human history (Shift 5).

The operator question is simple. Where on this loop do you sit? If the answer is "adjacent and passive," you are paying rent on someone else's flywheel. If the answer is "I own a domain, and I can encode it into a post-trained, voice-accessible, agent-readable, crisis-resilient system" — you are early to the golden age.

The window is open. Stanford CS 153 is telling you the truth about it. The companies moving now will be the trillion-dollar businesses of the next cycle.

Sources

The full Stanford CS 153 lecture series is public. Go to the primary sources.

<Agents Don't See Ads — and Four Other Shifts From Stanford CS 153 That Are Already Rewiring AI

Isomorphic Labs Raised $2.1B. AI Drug Design Is Now Real Medicine.>

Changelog