I have been building with Web3 and, more recently, with AI in the same products, and this note records the joints between them I was actually using in July 2024.
Web3 gives you shared state, signatures, and settlement that do not sit inside one company's database, while models give you judgment over messy inputs. The combination is interesting when the model needs a payment rail, an identity check, or a compute market that is not a single cloud invoice.
Shared compute
Training and inference concentrate GPUs in a handful of providers. Peer-to-peer networks on chain try to meter spare capacity instead, and I treat them as a way to buy cycles without assuming AWS is the only meter, well short of a finished cloud replacement. If AI is going to run closer to users and communities, the hardware map cannot be only three regions and two vendors.
Identity when the other party might be a model
Once text and images are cheap to fake, "is this a person" becomes an infrastructure problem. The chain world has been iterating on that in public, with Worldcoin's iris-scan proof of personhood as the loudest example. The constraint interests me more than any biometric product: sybil resistance for airdrops, votes, and rate limits gets harder when the attacker is a language model.
Models looking at chain data
The useful AI work on chain, for me, is unglamorous and specific: auditing smart contract code for known classes of bugs, scanning the firehose of blocks and logs for fraud patterns, flows, and anomalies, and summarizing that for a human who still signs the transaction. None of it requires a "decentralized ChatGPT," only tools that read hex and return something a reviewer can distrust productively.
Live data into models
Prediction and automation get less theatrical when the model is fed a real-time stream instead of a CSV from last quarter. Platforms such as Synternet sit in that gap, taking Web3 data, shaping it, and feeding it to an agent or a dashboard, and I care about the interface (stable schemas, latency, who can lie) more than the brand.
Those four joints (compute, identity, contract review, live data) are the whole thesis I was willing to stand on in mid-2024. Broader claims about healthcare, supply chain, and entertainment can wait until someone ships them.
Two projects I was close to
Morpheus is the open-source attempt I keep watching: a peer-to-peer network of personal general-purpose agents, with the MOR token paying people who build, host, and use the infrastructure, so inference is owned by its users instead of by a hosted assistant with a Discord.
As Product Engineer at Bitlauncher, I helped build a launchpad where AI startups raise through batch auctions on chain — with AI for project context and discovery and Web3 for settlement and price discovery. Each auction has a fixed bidding window and one clearing price, and every winning bid receives tokens at that price.
