In a time of rapid advances in cryptography, AI, and distributed systems, how do we reimagine trust? At the ETHSF “Verifiable Computation: From Zero-Knowledge to Infinite Intelligence” event, four leading thinkers came together to explore this question through the lens of proof-oriented programming — a new paradigm for building software that is not only functional, but verifiably correct.
This panel featured:
- Jens Groth, Chief Scientist at Nexus
- Albert Garreta, Cryptography Lead at Nethermind
- Benedikt Bünz, Chief Scientist at Espresso Systems and Professor at NYU
- Alex Evans, Partner at Bain Capital Crypto
Together, they explored the trajectory of verifiability, the compression of trust, and the evolving interface between cryptography and AI.
The event was co-hosted by Nexus and Nethermind, sponsored by Halliday, and in partnership with Blockchain Builders Fund and ETHSF.
Evolving proof systems
The conversation opened with a discussion of the current bottlenecks and breakthroughs in zero-knowledge proof systems.
Albert Garreta laid out the state of the art and what’s next:
“We will see 10x improvements in proof costs in the next years.”
He explained that the next wave of gains won’t come from faster provers alone, but from smarter circuit design — reducing circuit size and minimizing expensive field translations.
Jens Groth drew a parallel to Moore’s Law: if these 10x improvements compound year over year, we may soon approach native compute speeds — turning proofs from overhead into infrastructure.
Consensus meets computation
Benedikt Bünz reframed the problem from a systems perspective, outlining Espresso’s approach to bridging consensus and execution:
“The faster that the proving gets, the more useful [our protocol] becomes. Then the bottleneck is just agreeing on what to prove — which is the consensus part.”
Bünz explained how Espresso separates consensus from execution, enabling real-time commitments from rollups without waiting for full proof validation. He emphasized that scalable, distributed proving is becoming not just a cryptographic challenge, but a core systems and product design problem.
Alex Evans brought a venture lens to the conversation, highlighting the speed of commodification in zk:
“You fast forward three years… now there’s like a dozen general-purpose zk processors, and they’re all open source. It’s about as quick as I’ve seen — maybe the LLM stuff commoditizes faster — but it’s extremely impressive.”
He reflected on how investors now consider the shifting fault lines: do you go down the stack (hardware, proof networks), or up (applications, developer tools)?
Durable value is harder to anchor when each layer is moving quickly.
Formal verification
As proving systems mature, the panelists discussed a deeper layer of confidence: formal verification. Garreta emphasized its critical role in certifying circuit correctness, verifier soundness, and even the cryptographic proofs underlying STARKs.
Asked which layer he would verify if given a magic wand, he surprised the panel:
“I think I would go for the verifier.”
Bünz countered with the importance of verifying circuits:
“There’s no way these circuits are secure.”
The lighthearted exchange underscored the complexity — and subjectivity — of what “trust” really means in these systems.
Proofs and AI
As AI becomes more powerful and ubiquitous, the panel agreed: verifiability will become even more essential.
“The more power we give to AI,” said Garreta, “the more we need zk proofs.”
He sketched a future where AIs may police each other — but only if they can prove that interventions are justified.
Bünz pointed to emerging efforts to combat deepfakes with cryptographic authentication:
“You want to prove an image wasn’t altered — or that it came from a real camera. That’s where proof systems come in.”
Groth added that, unlike humans, AI agents could eventually be provably transparent — their internal reasoning made legible through formal systems.
What does verifiability enable?
In closing, each speaker offered a long-term vision.
Garreta spoke to the challenges of verifying AI training:
“We’ll never prove GPT was trained correctly — the computation is just too massive. But maybe we’ll invent smarter ways to bypass that.”
Groth, reflecting on his early work building the first SNARKs, shared the intuition that led him here:
“I had no clue this would be picked up in the blockchain space. I just had a feeling there was something useful here.”
And that instinct — that mathematical proofs could underpin new architectures of trust — is now reshaping the future.
“The vision? That all of the internet will be verified. At least the parts that matter.”
Toward a verified future
The conversation offered both a technical deep dive and a philosophical meditation on trust. As proof systems become faster and more usable, they promise not just better blockchains or more secure AI — but a deeper reimagining of how we build and verify digital systems.
At Nexus, we believe that future is within reach — and it will be proof-based.