I teach blockchains at Princeton university. We are just wrapping up "Elements of DeFi", a very hands-on course (each class is a lecture+lab on ETH Sepolia).

After all this dust settles, my lasting legacy in crypto will be that I pitched blockchains to such brilliant minds (and founders), led by @sreeramkannan and @hstyagi and @dntse and @yangl1996.

Thanks to @sreeramkannan for teaching the last lecture that summarized the entire storyline: combining all elements of DeFi into a super-organism (Eigenlayer). Kind of like combining basic elements of life (hydrogen, oxygen, carbon) into the sentient, flexible, homo sapiens.

Pleasure and privilege to welcome Ed Felten back to Princeton and interview him on his journey as a professor, CTO of the United States (Obama White House) to starting Arbitrum. @EdFelten #Arbitrum

All lecture slides, videos, labs are online at the course website: The lecture by @sreeramkannan can be seen at:

1/ As a long time educator, I see immense gulf between the hunger college & high school kids have and the ability to connect with demand for their ideas. This is especially exacerbated in our era of AI, where tech is largely shake-and-bake and very accessible (see figure):

Most pressing scientific agenda of our times. Most pressing engineering and societal agendas too. We wrote a research monograph on OML, a breakthrough native AI and native cryptographic library this is the protocol and big picture around it.

2/ The goal of Sentient is to bring ownership rights to open AI development. By inventing science and technology that enables anyone to build, collaborate, own and monetize AI products, the  birth of an era of AI entrepreneurship.

@sandeepnailwal Thanks @sandeepnailwal The key thing about DePIN is P (“physical”) and decentralized trust here is the ability to do “consensus on physical state”. Blockchains run consensus on “digital (and objective) state”. Physical state is both NOT digitally native and is inter-subjective!.

There is huge diversity of DePIN projects (networking, storage, energy, sensors, batteries). Correspondingly the “Physical state” also comes in many forms. At witness chain we focused on creating proofs of the most basic physical states.

The impact of AI on all of humanity is amazing even for those inside the AI-bubble. For the crypto world: @sreeramkannan was into AI+Biology before he got into @eigenlayer (coauthoring a Nature ML paper with Nobel laureate David Baker)

A technical comment here: the cryptography needed to enable OML is reminiscent of the early days of cryptography (late 70s to the 80s) when public key cryptography and hash functions were invented. We call this AI-native cryptography. Hallmarks are:.

Human learning agents (aka children) play games naturally. This is how they learn ("scientist in the crib" by Prof. Gopnik). Our thesis is that AI agents are similar. The goals of this agi-thon are to discover trajectories of learning improvements via such a tournament.

Ax = \lambda x. The fundamental theorem of linear algebra. \lambda = $eigen.

INTERNET CONNECTIVITY is next most basic. gives you bandwidth test but we need decentralized solutions. We build cryptography primitives inside classical network telemetry libraries (ping, iperf) in our Proof Bandwidth protocol

@sentient_agi @foundersfund @peterthiel @PanteraCapital @hiFramework 3/ Excited to join forces with @hstyagi and @sewoong79 (20+ years of my research collaborations with them) together with the inimitable @sandeepnailwal on this journey.

Academic tenure is such an extraordinary privilege and empowers the recipient to explore freely. Kind of like improving one's swimming times in a specific stroke in the pool vs open water swimming! Great catching up and swimming together in the open waters (of the Aegean Sea).

Where is Waldo? Now we know. Trust-free.

We wrote a research monograph on OML: Open, Monetizable and Loyal AI last fall: We studied what seems an impossibility: how can an AI model be open (so easily downloaded) and *yet* be "controlled" by its community? Below is the real world instantiation.

LOCATION is the most basic physical state. Traditional localization is through triangulation. But we need Byzantine resistance. We invented “adversarial trigonometry” (geometry the ancient Greeks would’ve liked) in our proof of location protocol

Thanks for the vision @karthik_r_n -- "natural language evolved for multi-agent collaboration and communication" is a great insight. Turns "natural language processing (NLP)" on its head to "processing via natural language (PNL)"! @sentient_agi Look forward to @openagisummit.

Information theory has influenced foundational design of wireless networks (and entrepreneurship around it -- e.g., Qualcomm). It is also a great background for blockchains.

Thanks for bringing up this distinction @sreeramkannan -- proof of inference is rather distinct from loyalty. "Being open and yet loyal" seems to be an oxymoron a priori. As @valardragon suggests, TEE is an approach (not really open).

@sentient_agi @foundersfund @peterthiel @PanteraCapital @hiFramework As a long time educator, I see immense gulf between the hunger college & high school kids have and the ability to connect with demand for their ideas. This is especially exacerbated in our era of AI, where tech is largely shake-and-bake and very accessible (see figure)

@sreeramkannan @tarunchitra @NotoriousPtG Aggelos Kiayias at IOHK proposed to work on fungibly mixing resources (e.g., work and stake). We came up with a beautiful protocol (with @sreeramkannan and my graduate students and IOHK researchers): Minotaur. This was way before Eigenlayer was even conceived.

3/ Excited to join forces with @hstyagi and.@sewoong79 (20+ years of my research collaborations with them) together with the inimitable @sandeepnailwal on this project.

@sentient_agi @foundersfund @peterthiel @PanteraCapital @hiFramework 2/ The goal of Sentient is to bring ownership rights to open AI development. By inventing science and technology that enables anyone to build, collaborate, own and monetize AI products, the  birth of an era of AI entrepreneurship.

3/ Eigenlayer @eigenlayer and Sentient @sentient_agi and Babylon @babylonlabs_io and Witness chain @witnesschain and FedML and MegaETH are all blockchain startups by information theorists. Happy to pay back Information theory via sponsoring ISIT.

@tarunchitra @soona Crazy. What’s the connection? @sreeramkannan (the restating guru, or just guru more generally) and I wrote a paper a while ago connecting the Lovasz extension and wireless networks. And I’ve been interested in RL-driven credit markets (papers at AFT). But those are just partial.

Cooperation and competition via language (and rules of biochemistry) led to human intelligence. These campaigns are evolution in a microcosm. The seeds of AGI. @sentient_agi.

Look out for our OML research paper (part whitepaper) tomorrow where we propose to use, and invent, novel AI methods themselves to create this new cryptographic primitive OML for this era of AI-entrepreneurship (ownership rights to AI builders).

@tarunchitra Hilariously good. Not many people who this joke can be appreciated with.

2/ In wireless, variation and interference abound: every leaf of every tree bounces the radio waves. In blockchains you have adversarial and incentive driven actors. In both cases you need robust, reliable communication. Information theory connects them both.

NETWORK SERVICE is third most basic. Your carrier can list your data usage but we need decentralized solutions. We invent trusted two-sided network measurement and integrate into open source 5G wireless stack (Magma) in our Proof of Service protocol.

Finally #witnesschain platform allows community to contribute their own proofs to build the library of physical states. Thus the physical state verification substrate is born. Combine with #EigenLayer security and tokenomics we have the first open DePIN collaboration platform.

FHE is another direction, but as @sreeramkannan points out, it is a long way away. At @sentient_agi, we thought a lot about this topic and call this OML (open, monetizable and loyal) -- a new type of file format for AI models (and more).

3/ Approximate: cryptographic security means even a single bit off is fatal. AI models are embedded in continuous spaces and OML can tolerate sliding scale security. Eg: OMLized GPT4 degrades performance to GPT3.5 when used without authorization.

AI-native cryptography: 2/ Average case: cryptographic guarantees are for the worst-case. For OMLized AI models, we dont mind if some queries escape monetization tracking.