The audio and video are in Spanish. These notes are a translation.
Today I interviewed Andrés Monroy Hernández, a research professor at Princeton University, where he directs the human-computer interaction laboratory that studies how we interact with technology.
Before leaving Silicon Valley for academia, Andrés founded Snapchat’s augmented reality team and spent years at Microsoft Research.
His central thesis is that the debate about whether AI is going to take our jobs distracts from something that is already happening: millions of people already have a boss that is an algorithm, one that decides how much they get paid and what tasks they do. And that the underlying problem is not that jobs disappear, but dependence on a single centralized entity.
We talked about why he believes that almost every profession, from a doctor to an accountant, is going to end up working on Uber-type platforms, and about what alternatives exist so that does not mean making the work more precarious.
Structure and Key Highlights
Where His Perspective Comes from
- His parents were political activists focused on labor issues in Mexico; they moved from Mexico City to Saltillo to organize workers at the assembly plants that were relocating from the United States
- He grew up with the technological optimism of the 80s: science fiction as a desirable future
- An uncle of his, a doctoral student in mathematics, would answer basic questions by letter (why -1 × -1 is positive) and opened the door to research for him
- Why that optimism eroded: technology companies started out as alternatives, and from ~2010 on they became companies like any other
How to Criticize Technology Without Becoming Cynical
- Use it yourself: be an early adopter, from social networks to LLMs
- Know up close the people who build it, who in general have good intentions
- His critique of academia: someone who never connected with industry produces critiques that are superficial, or that end in “the problem is capitalism,” and that is where the conversation ends
Why He Left Industry
- After a decade at Microsoft and Snap, there was no internal incentive for projects with a social focus
- At Microsoft Research (2012-2013) they tried to create, with unions, an alternative to Uber, but the unions distrusted them: “why would someone from Microsoft want to do this with us?”
- Inside a company you have flexibility over your own time, but it is hard to convince colleagues to work on something that does not help the bottom line
The Real Problem: Centralization, Not Unemployment
- What is already happening today: the algorithm as boss, with no direct relationship to the employer or to coworkers
- Gig work as a preview of how work in general will be
- “If you are General Motors or Ford and 20% of your internal process depends on Anthropic or on OpenAI, that is a big risk”
- The parallel with the cloud: large companies split their infrastructure across providers instead of depending on one
- At a personal level, he uses several models on purpose, so he does not get used to only one
Protocolize Instead of Regulate
- Bluesky and AT Protocol as a proof of concept: the same user experience as Twitter, without a single company controlling the data, the feed, or moderation
- His laboratory is building an open protocol for food delivery, with three actors: restaurant, courier, and customer
- A protocol is first a public document; then someone implements open source software on top of it, and anyone can make a compatible fork
- Eurosky: a European alternative to Bluesky, hosted and operated in Europe for political autonomy and for data autonomy
The Network Effects Objection
- The classic counterargument: network effects require enormous capital to get started
- He answers that the internet itself was born from network effects without depending on one company. HTTP, HTML, DNS, and SMTP are the proof
- “Companies like Uber, Airbnb, and Facebook want to make us believe that the only way to do that is the model they have”
- The nuance: what Facebook has is not a real network, it is a single node; a real network would have several entities interoperating
- The incentive problem: nobody invests in HTTP because nobody owns HTTP. It is the same debate today between closed models and open source
- And a real difference in cost: writing HTTP was drafting a document; training a model costs a huge amount of capital
Why Bluesky’s AI Feed Failed
- Bluesky tried letting the user describe, with AI, the feed they wanted, and the community did not adopt it
- One of his female PhD students is researching why
- What he sees behind it: the launch was misaligned with the values of a community that arrived there by a political decision and with reservations about AI
- There were misunderstandings, such as believing that their posts would train an LLM
- In in-depth conversations the opinions are more nuanced: many people accept LLMs if the model is local or does not belong to a centralized lab
What Changes and What Intensifies with AI
- “AI” as a term is not very productive: it is like asking whether you are against “transportation” without distinguishing bicycles from airplanes
- The chatbot as an interface is temporary: he does not believe that in 10 years we will still be using them this way
- What comes back is ubiquitous computing, the vision of the 90s: computing spread through the environment and barely visible
- On his students: using an LLM for everything is like paying for the gym and staying home watching television
- He is optimistic that in 5 or 10 years we will do a better job of identifying the limits of when to use these tools
Status, Inequality, and the “Addictionless Economy”
- Enzo’s hypothesis: not using technology becomes a status signal, and the person who has to work uses it out of obligation
- Andrés confirms that it is already happening: on his first day at Snap, the retreat started by handing in your phone on a tray and eight hours disconnected
- He worries that the cognitive effect and the effect on work fall on whoever has fewer resources
- The parallel with organic food: from inaccessible luxury to a Walmart shelf, by way of social change plus regulation
Agents: The Utopian Version and the Dystopian One
- Utopian: each worker with their own agent that defends them, turns down jobs below their minimum, and learns their preferences
- A future in which we are all gig workers, from a doctor to an accountant
- Why he believes it: paying a fixed salary is economically less “extractable” than paying per minute worked
- What he fears for the region: Latin American entrepreneurs turning the informal economy into a gig economy, without improving conditions
- The counterexample he likes: Empower, in Washington DC, an alternative to Uber where the driver pays a monthly fee. The platform’s customer is the worker, not the other way around
Where Total Autonomy Breaks
- People like working on something: someone who retires takes up ceramics or a vegetable garden
- The utopian version of autonomy is discovering what you like to do and delegating the rest
Who Holds the Power in the Era of Models
- The model gets commoditized; the power moves to the interface and memory layer with the user
- Apple’s contract with Google so that Gemini powers Siri, as a signal: the user does not care which model sits behind it
- That is where the harness comes in, the layer between the model and the person who uses it
- The shutdown of access to Fable 5 by a government decision, as an example of what happens if you depend on a single model
What Not to Copy from Silicon Valley
- Do not accept as the default either centralization or business models based on advertising
- What interested him about blockchain was not the cryptocurrencies but the protocols designed to decentralize power
- He combines three vectors: regulation, companies with alternative business models, and hobbyist communities
- Open source coding agents as a renaissance: the great obstacle for open source was capital, and the agents reduce it
- On regulating: the GDPR set out to protect the consumer and ended up strengthening the incumbents, because only the large companies can pay for compliance
The Conversation Ahead
- The cognitive impact of technology, discussed in academia today, is going to reach family tables
- He predicts a return to the mystical and to the search for community, in the face of the question of why we are here
- And also pro-AI movements with an almost religious character. Enzo points out that some labs already write in those terms
Silicon Valley Glossary
| Term | Meaning | Context in the episode |
|---|---|---|
| Human-computer interaction (HCI) | The discipline that studies how people interact with technology | The area of the laboratory Andrés directs at Princeton |
| Gig economy | Work by the task, without a traditional employment relationship, intermediated by platforms | His object of study: Uber, DoorDash, and DiDi as the present of the future of work |
| Protocol | A public document that defines how the pieces of a system communicate with one another | His central proposal: protocolize work platforms the way it was done with email |
| AT Protocol | The open protocol created by Bluesky | It lets other entities build compatible alternatives |
| Decentralization | Spreading control of the data, the feed, and moderation instead of concentrating it in one company | The principle his laboratory is trying to carry from social media into work |
| Network effects | The product is worth more the more users it has | The classic objection to his thesis, and the point he argues: Facebook is a single node, not a network |
| Regulatory capture | When regulation ends up favoring the companies it was meant to limit | His reading of the GDPR in Europe |
| Ubiquitous computing | A vision from the 90s: computing spread through the environment and barely visible | What he believes comes back, in place of the chatbot as the dominant interface |
| Harness | The layer that wraps the model and defines how it is used | Where he locates real power, more than in the model itself |
| Model commoditization | Models become interchangeable and the value moves to another layer | His reading of the Apple-Gemini agreement |
| Frontier models | The most advanced closed models from each lab | Set against open source in the debate over who can invest in what |
| Bottom line | The company’s economic result | Why it was hard to sustain social projects inside Microsoft or Snap |
Books, People and References
People Mentioned
- Andrés Monroy Hernández: research professor at Princeton, former Microsoft Research and former Snap
- Enzo Cavalie: host of the podcast
- Jonathan Haidt: professor cited for leading the movement against smartphones and social networks
- Dario Amodei: CEO of Anthropic, cited for his essays about a future in which AI coordinates resources
- Elon Musk: referenced for the exodus of users from X to Bluesky
- His parents: union activists in Mexico, the origin of his interest in work
- His uncle: a doctoral student in mathematics who answered his questions by letter from Canada
Companies and Organizations
- Princeton University: where he directs his HCI laboratory
- Microsoft Research: where he worked before Snap
- Snap: where he founded the augmented reality team
- Bluesky: a social network on top of AT Protocol, his main reference for decentralization
- Mastodon: a decentralized alternative in social media
- Eurosky: a European version of Bluesky, hosted and operated in Europe
- Empower: an alternative to Uber in Washington DC, where the driver pays a monthly fee
- Uber / DoorDash / DiDi: gig work platforms he studies
- Airbnb: cited in the discussion of marketplaces and network effects
- OpenAI / Anthropic: the labs whose centralization he considers the structural risk
- Apple / Google (Gemini): the agreement to power Siri, as a signal of where the power sits
- AWS / Google Cloud: the precedent of dependence on infrastructure
- Linux and the Free and Open Source movement: his example of technological development outside the economic incentive
Concepts and References
- HTTP, HTML, DNS, SMTP: open protocols that hold up the internet, his proof that network effects do not require a single company as owner
- GDPR: the case of regulation that, according to him, strengthened the incumbents
- Star Trek: the stated inspiration of the person who founded Microsoft Research
- Fable 5: the shutdown of access to a model by a government decision, cited as a warning about depending on only one