Saturnin Pugnet on AGI, Worldcoin and Proving Human Identity

Matthew Berman interviews Saturnin “Sat” Pugnet, a founding member of the Worldcoin team, about AI scaling, AGI, open models, digital personhood and cryptocurrency. Recorded in 2024, the discussion contains forecasts and project claims that should not be mistaken for established outcomes.

Asked about the state of AI, Pugnet says deep learning’s progress has growing economic and geopolitical implications. He believes intelligence can continue to scale and sees no obvious fundamental reason for progress to stop. This is his assessment of a contested research trajectory, not a proven scaling law for all future systems.

Berman asks about a possible shortage of human-created training data. Pugnet expects a combination of algorithmic improvements and synthetic data. He argues that researchers may focus more on efficiency as easy access to additional data becomes harder, and points to humans as an example of learning from comparatively limited experience.

Pugnet explains that synthetic data need not be generated by an AI: simulated environments and physics engines can produce computationally generated examples. For visual and spatial tasks, he imagines using large quantities of simulation and then real data for fine detail. His numerical characterization of Tesla’s training mixture is not verified in the interview.

Berman asks whether language models can achieve spatial intelligence or whether a different architecture, such as that pursued by World Labs, is required. Pugnet says it is hard to know whether language models are the final architecture, but expects multimodal and end-to-end learning to extend into robotics.

Pugnet predicts that models could become good enough for practical robots within a few years. He uses unloading a dishwasher with an end-to-end system as a personal test of useful generalization. His claim that success would imply broad replacement of physical labor is a speculative extrapolation, not an established result.

Discussing OpenAI’s o1 and test-time computation, Pugnet emphasizes scaling across data, model size and modalities. He believes larger systems may eventually conduct research and improve their own efficiency. He describes this as a path he expects, rather than evidence that the transition is guaranteed.

On chain-of-thought methods, Pugnet says empirical progress has outpaced theory: developers discover useful behaviors through experimentation without fully understanding model weights. Berman asks how far extra inference-time reasoning can scale; Pugnet says he cannot give a confident mechanistic explanation.

Berman quotes Pugnet’s view that AGI could be humanity’s last major technological invention. Pugnet imagines copying capable digital agents, having them research better systems and using robots to build infrastructure. He forecasts a potentially rapid feedback loop and abundance, but neither unlimited intelligence nor resources follow automatically from today’s systems.

Pugnet says such a transition could be either very beneficial or disastrous, and could disrupt society abruptly. He proposes three areas for preparation: survive the period before AGI, enable its development and infrastructure, and protect humanity afterward. This is his framework, not a consensus on what will happen.

For pre-AGI survival he cites peace, defense, health and longevity. For enabling AGI he lists research, energy, data centers and chips. He worries that a transitional period could concentrate value before abundance is broadly available, and presents Worldcoin as infrastructure intended to help society navigate that period.

For post-AGI survival, Pugnet points to alignment, limiting dangerous applications and neurotechnology. Berman asks for a timeline. Pugnet prefers a probability distribution and estimates a 40% chance of human-level AGI within four years, 80% within 12 and 90% within 20, as of the 2024 conversation.

Pugnet’s working definition of AGI is a system with which one can interact and delegate tasks as with a capable human, including memory and reliable follow-through. He imagines superintelligence exceeding the combined intellectual ability of humanity, perhaps two to four years after AGI. Both his definitions and numerical timelines are personal predictions.

Berman asks whether superintelligence might prove impossible. Pugnet argues from the existence of biological intelligence that machine intelligence should be physically possible, while allowing that particular architectures could fail. He suggests alternatives, including whole-brain emulation, although mapping a brain or simulating aspects of its activity is not the same as a functioning full emulation of a human mind.

On open versus closed AI, Pugnet says releasing every model immediately could be dangerous if a future system is highly capable and misaligned. He favors public awareness and transparency, but not unrestricted access to the most powerful weights. He is more comfortable with openness at current capability levels than near a hypothetical self-improving AGI.

Berman offers the counterargument: open systems can be inspected and hardened by more people, progress may be incremental, and private companies are not perfectly secure. Pugnet agrees that a slow transition would make openness more attractive, but expects a faster one and proposes government protection of highly sensitive systems with public disclosure of their capabilities.

Pugnet stresses that he sees recursively self-improving systems as a qualitative risk and worries about malicious actors using highly capable agents. His suggestion that self-improvement would mean effectively infinite capability is a speculative claim, not a demonstrated property; physical and operational constraints would still matter.

Berman asks whether creativity, moral judgment or other work will remain uniquely human. Pugnet believes machines may eventually perform those tasks too, but says people would still decide what they want from a more abundant world. He imagines humans choosing collective and personal goals rather than being defined by compulsory work.

His examples range from space exploration to swarms of drones building objects from a person’s thoughts. These are imaginative illustrations of his optimistic vision, not engineering demonstrations or evidence that the required resources, interfaces or capabilities will be available on his proposed timeline.

Pugnet introduces Worldcoin as a project intended to support a transition in which advanced AI makes it harder to distinguish humans from bots and changes the distribution of economic value. He describes two aims: a digital proof-of-personhood system and a widely distributed cryptocurrency that could support future transfers.

Berman asks whether the system could identify AI-created images or text. Pugnet says Worldcoin focuses on people performing actions, not direct content attribution. He proposes signed camera hardware for images and doubts that text alone can reliably reveal its origin; those are his views rather than proof that any one method solves provenance.

Pugnet says World ID is designed to let an application check that a participant is a unique human without revealing their name, and suggests one-time promotions and social platforms as examples. Such a system can restrict repeated verified actions, but cannot by itself prove that content is human-written or prevent a verified person from sharing AI content.

Asked about the founding story, Pugnet says the project began in 2019 with a core team that met at Caltech. They connected with Sam Altman and refined an initial idea about infrastructure for an AI future. He says privacy, personhood and deployment challenges became clearer through several years of iteration.

Pugnet describes a long development process for the Orb: an initial prototype, further work to make it scalable and gradual deployment. He compares its iris imaging ambitions and costs with India’s Aadhaar system, but his cost, quality and operational comparisons are estimates offered in the interview, not independently verified here.

He says the Orb must detect a live adult and resist attempted impersonation in varied conditions, using sensors including thermal cameras. He describes device security as critical because compromised enrollment could undermine uniqueness. These are design objectives and company claims, not an independent security audit.

Berman asks whether an iris scan is written to a blockchain. Pugnet corrects that account and gives a simplified explanation of on-device image processing and an iris-derived representation used for duplicate checking. He says images are deleted from the device and presents the representation as non-reversible; that is his description rather than a complete privacy guarantee.

World’s own March 2024 Personal Custody announcement says data generated at the Orb, including images and derived data, is encrypted and held on the user’s device. Its May 2024 announcement also describes deleting a central iris-code database and using secure multiparty computation. These details qualify the simplified claim that only a number is kept and no images are stored anywhere.

Pugnet says verified users can receive token allocations and use an app to prove uniqueness to participating services. He gives a merchant’s one-time discount as an example. Eligibility, distribution and availability depend on the relevant jurisdiction and current project terms; his 2024 description is not a promise of present access.

He describes downloading the app, locating an Orb, scanning a QR code and using the app to approve an interaction. He says the app acts as an interface to a cryptographic wallet and that alternative apps could be built. Berman asks about lost phones, stolen keys and account recovery.

Pugnet says a user could reverify at an Orb to recover an identity to a new wallet, while acknowledging a stolen wallet may be usable before recovery and discussing possible blocking mechanisms. This is his account of the intended workflow; users should verify actual recovery and security features for their version.

Berman asks whether an agent could act on its owner’s behalf without repeated verification. Pugnet says a wallet key can authorize identity interactions and imagines additional layers so a bot need not directly hold the private key. He does not describe a completed, independently verified delegation system in this interview.

On agent-to-agent transactions, Pugnet says an identity layer could help integrate cryptocurrency with regulated financial services. He distinguishes a basic proof of uniqueness from additional attributes, such as age or nationality, that an application might associate separately. His privacy claims describe the intended system and do not eliminate implementation or governance risks.

Berman asks how the project is financed and whether its token is tradable. Pugnet distinguishes the Worldcoin Foundation, which he says has a token pool for the ecosystem, from Tools for Humanity, the company building the Orb, app and related services. He describes a long-term goal of more independent developers and hardware providers.

Pugnet says the foundation funds operations using its token holdings and pays Tools for Humanity for work. He describes the token as not being a stablecoin, since it has no price-maintenance mechanism. His account of organizational and funding arrangements is a 2024 snapshot, not a full financial statement.

He argues that a larger network could make the ecosystem more useful and illustrates the idea with network effects. Berman asks whether a token tied to identity can be sold. Pugnet clarifies that the identity credential and token are separate and says tokens can be traded independently, subject to available markets and rules.

Pugnet associates the token with a possible future universal basic income, defining UBI as income given regardless of a person’s work. He says he does not favor it as strongly without AGI, but imagines it becoming necessary if advanced AI concentrates production and wealth. That is a conditional policy proposal, not an implemented global income program.

He argues that a global uniqueness mechanism could help distribute benefits once per person and potentially reduce upheaval during a rapid shift in work. The ability to verify uniqueness does not itself supply funding, ensure universal enrollment or resolve the political and economic choices required for UBI.

Berman asks whether giving everyone more wealth would cause inflation. Pugnet argues that abundant supply, rather than money alone, would determine long-run availability and allow prices to fall. His simplified economic account is a forecast with significant assumptions about production, scarcity, distribution and demand, not a demonstrated guarantee that prices will approach zero.

Pugnet points to historical productivity gains and imagines extremely high future living standards. Berman welcomes the optimistic vision while acknowledging the risks discussed earlier. They close by emphasizing the need for public understanding and deliberate choices during technological change; the imagined abundance and timelines remain speculative.