Today,
@niraj
and I are excited to unveil Ritual (
@ritualnet
) to the world.
Ritual is a bidirectional response that sits at the intersection of modern cryptographic schemes and ai objects to bring forth a new wave of intelligent apps spanning both the web3 and web2 worlds.
--- stuff i made money on ---
anish
--- stuff i lost money on ---
anish
USDC
OHM (bought and didn't stake it)
IQ
empty set dollar emptying our dollars
anish's 8 months spent on gobblers
Leo (somehow)
whenever Vfat made an update on his website
Tokemak (+Tokamak)
jpeg'd
my 3
A few clarifications on Ritual, and other projects mentioned:
Ritual itself is not opinionated about the computational integrity mechanism for inference, whether ZK, OP, some convex combo, or otherwise. Our setup is modular enough to support various types of proof systems as I
Don’t trust, verify: An Overview of Decentralized AI Inference
Say you want to run a large language model like Llama2-70B. A model this massive requires more than 140GB of memory, which means you can’t run the raw model on your home machine. What are your options? You might jump
where we’re at today:
we've enabled any on-chain or off-chain app to access models in a modular way allowing users to choose what type of experience they want along the payments, privacy, and computational integrity dimensions. we’ve kept a neutral view on the exact type of
non-consensus take, but bewildering to me the crypto x ai space spends so much time on specific types of verifiability (zkml). constant regurgitation of the same narrative by "important" people in the space. computational integrity *is* important but in a vanishingly small set of
Gm
I just published my new report on Understanding the Intersection of Crypto and AI for
@galaxyhq
@glxyresearch
The report looks at three emerging verticals:
1) Decentralized Compute
2) Zero-Knowledge Machine Learning (zkML)
3) AI Agents
Ritual continues to expand its modular privacy capabilities for all heterogenous compute with Nillion - mixing public & private objects enable net-new primitives.
Been fun jamming with the Nillion team these past few months on POCs that we'll announce over the coming weeks!
We're excited to announce our partnership with
@nillionnetwork
.
Ritual and Nillion are pioneering the use of MPC to enable private AI inference and storage on the Ritual network, unlocking new possibilities for developers, model creators, and users.
garbage in garbage out innit. "zkml" for computation is only one piece of verifiability for ml. should be a lot more attention imo on the companies building in provenance of the input space to models whether on-chain data (
@axiom_xyz
) or the many off-chain data ones (
@zkPass
@_anishagnihotri
In a future where the global ai mind is split across a myriad of models who need to communicate with each other, don’t you think zkml becomes a requirement in that exchange of value/information?
I'd love to brainstorm what a relevant use case for your app or protocol could be within the Ritual ecosystem, or even if it's not Ritual related, hit me up anyway, DMs are open. Anons welcome. akilesh (at) ritual (dot) global
Zooming out, Ritual itself is a series of processions, starting today with Infernet that brings AI models to where the apps & protocols live today on different chains, evolving into a suite of modular execution layers that interops with the best parts of the existing web3 space
personal view is that you want AIs operating in smoother regions of uncertainty space (i.e. cap downside outcomes from vitalik's post yesterday) rather than sharp ones: prediction markets structurally are bespoke and exhibit phase transitions given nature of information flow
It's become evident there are central issues around practical solutions to censorship resistance, computational integrity, incentivization, and privacy primitives in the applied ML world. Ritual is here to solve this.
It was exciting to work on representation learning, embeddings, and interpretable ML in the pre word2vec days in 2012, and then again to be working on cross-lingual NLP as Hugging Face was just coming out in 2016.
Creating
@ritualnet
felt like a natural path for me as I've only spent time over the last decade dabbling in both the spaces, from being an ML researcher to building models in the wild to spending the last six years in crypto building and getting rugged one too many times onchain
Over the years, some of you may have seen "me" tweeting about bandits, agents, embeddings, and the like as scattered thoughts as it relates to the web3 space — happy to say those thoughts will be much less scattered, and a bit more doxxed, going forward : )
believe special purpose {mpc, zk, fhe} gadgets rather than general purpose ones will be built to exploit the structure and symmetries of popular equiv classes of ai models used in practice today, biggest barriers will indeed be devx and instant support for new architectures
you're missing the point — verifiability for statistical objects (i.e. models) is a totally different animal imo than web3 objects from an *end user* standpoint.
rarely the case i pop open insta, scroll down my feed, realize it blows and wonder where the validity proof of the
@_anishagnihotri
imho, this is a short-sighted take:
a) very few users actually care about decentralization as well. does this mean we shouldn't build in web3?
b) zkml may have a super small market atm but if we take a multi-agent/model thesis, then zkml will be critical very soon
Many of you don't know me, I've long believed in a separation of identity from content, but I'll be around here, or as an anon, or just around
@marinerime
writing about crypto x ai
ideologically agree with the overarching theme of prioritizing human-in-the-loop semantics over black box use cases for ai models at the intersection of both areas, former also allows for composable cryptographic primitives of varying guarantees by splitting diff types of models
We're keen on introducing some new primitives that leverage less strict guarantees than existing crypto x ai projects but are much lower cost emanating from probabilistic schemes to provide the user with a spectrum of integrity & privacy gadgets fit to their needs.
similar to the explosion of the design space for alt L1s for "better perf" in the '17 cycle, most recently the "alt transformers" (i.e. SSM, RWKV, ...) variants may lead to teams rebuilding special purpose gadgets for these ad nauseam as the foundation model space evolves rapidly
given hallucination on web3 specific evals will be subparish on oss llms for nontrivial use cases, offloading tail risk onto users will be the same "1inch misrouted my order" vibes amplification we've seen many times before, embedding hierarchical AIs will help bound these
hope to see more work productionized on leveraging AIs to transform non-cooperative games into ones where the nash equilibria are also pareto optimal, similar to the hyperlinked post from virgil on introducing conditional payouts, deposit burns, etc. mechanisms to do the same
zooming out of the current meta of ai models in web3, many aren't best expressed as a series of tensor ops, but they still have structure to be exploited nonetheless. waiting for someone to snark a HJB recursion so we can bring back a new provably stabilized (3,3) era 😂
believe there will be a rich design space for bootstrapping nascent consumer facing ecosystems off bounded AIs and slowly phasing them out after some critical mass of users is hit (i.e. intentional deepfakes on dating apps that dissipate slowly)
broadly speaks to why incentivized red-teaming, similar to the waterfall bounties that exist for breaking web3 protocols & apps, will become critical and an important category of ai security auditors, mostly has been llms triaging issues for canonical audits thus far
for modern equiv classes of ai models, they definitely have structure & symmetry, and in practice are sometimes sparse & have embarrassingly parallel constructions, which all should be exploited by cryptographic primitives
agree that opportunity cost of modern AIs w/ decent evals is ~low and unlocks interesting compositions of ai models & humans (i.e. active learning-esque) in a series of escalation games to bound 'high variance' ai outputs, we've also explored a similar design space with
@frenrug
part of the issue is you have folks in this space who have barely spent any time in the ai wild historically understanding what properties consumers actually care about when interacting with models, and instead, taking some awesome machinery from this space and tossing it at it
reach out if you want to chat about or collab on any of the above, some of these can probably be broken out into their own posts so happy to do it if there's interest
agree with the importance of privacy, granted given oss llms, will have to further privatize the weights by fine tuning the models ever so slightly to preserve evals but then becomes the typical pursuit-evasion game where the attacker can still derive "low" variance posteriors
Stoked to see the resurgence of small credit facilities extending ~unsecured at minimal interest rates, all well disguised under the facade of "defi basis books" aka yolo longs with "approximate" hedging & farming. Only ends in the best way possible. Not quite the 9 figs of
agree that there's varying amounts of tail risk (i.e. ~right most of the time, and then quite wrong a few times) from ai models that get offloaded to users re: adversarial settings & hallucination
research around inverse methods is growing, that is, finding the set of possible inputs that result in, say, an LLM with a specific output. as decision LLMs become more enshrined into web3, maybe we'll have mev searchers running inverse llm solvers & prompt orderbooks soon ™️
not attending denver or any travel in coming months, fully heads down. this is a very important year. people should not forget that the
#1
priority is shipping, and shipping, and shipping more.
agree with the intuition for why "decentralized market-based RLHF" should exist and why on-chain mechanisms can facilitate it, also mention it here on the "incentivized rlaif marketplace" front:
there is a weak (read: statistical) consensus layer of autonomous llm agents where multiple individual actions are combined to make a final action on
@friendtech
, this layer can (with a lot more work) evolve into a preference model for an incentivized rlaif marketplace
bad take. the many companies in the zk space (
@ModulusLabs
,
@ezklxyz
...) & even in fhe (concrete via
@zama_fhe
) can get you way more expressive models today. my tweet was around the interaction of why/when/how these proofs get consumed by end users when they interact with models.
one of the byproducts of promoting model interpretability is the ability to bound downside risk or better understand failure modes in addition to understanding potential adversarial attack vectors, of course comes at the cost of model perf
lastly, many view hallucination as a structural issue for most oss llms.
@frenrug
exploits hallucination as a feature. not a bug. hallucination in games can be fun.
we initially show off zk proofs for classical ml models and over the coming weeks will show off proofs of an optimistic & probabilistic nature that scale to foundation models, all of which will eventually exist on
@ritualnet
's proof marketplace
non-consensus take, but bewildering to me the crypto x ai space spends so much time on specific types of verifiability (zkml). constant regurgitation of the same narrative by "important" people in the space. computational integrity *is* important but in a vanishingly small set of
at the heart of
@ritualnet
is a spectrum of computational integrity proofs for a variety of models depending on what the user cares about and what their budget is. most projects take an opinionated view on the guarantees around verifiability a user may want. we're open minded.
frenrug demonstrates the ability to compose together foundation models and classical models for on-chain agents by partitioning complex and simple actions
there is a weak (read: statistical) consensus layer of autonomous llm agents where multiple individual actions are combined to make a final action on
@friendtech
, this layer can (with a lot more work) evolve into a preference model for an incentivized rlaif marketplace
@drCathieSo_eth
@bidhanxyz
@sandeepnailwal
@ritualnet
Ritual's Infernet has been live for a while, and can be used & deployed on any chain. That's how
@frenrug
was built back in Nov, can read the Infernet docs at and compare to OAO's design for a more apt comparison.
"World's first AI oracle" is a stretch