The Current State of AI x Blockchain, and What’s Next

Dominic Williams, Dfinity’s founder and Chief Scientist, recently sat down for a fireside chat with Tracy Trachsler, who joined Dfinity as Head of Institutional Relations in May. The topic was AI, specifically what Dom thinks of the current conversation around the intersection of blockchain and AI, what the Internet Computer has achieved so far, and what to expect next. Here’s a summary of the chat.

The GPU dilemma

One of the most important decisions when running AI Models is where to host them. The large language models (LLMs) that users have adopted over the last couple of years, like OpenAI’s ChatGPT and Meta’s Llama 3, run on servers controlled by their developers. The problem with relying on centralized servers, as we explored in DeAI: Shedding light on AI’s black box problem, is they’re vulnerable to cyberattacks, such as sleeper agents, which can be programmed to produce inaccurate outputs under certain circumstances.

Here at Dfinity, we believe AI on the blockchain is crucial to ensure society fully benefits from this game-changing technology. But not all decentralized models are created equally. One approach that has gained traction is to use GPU, required to train models and run inference, sourced from Decentralized Physical Infrastructure Networks (DePINs). However, DePINs will struggle to compete with centralized cloud providers like Amazon Web Services, which can leverage their infrastructure (co-located GPUs in giant data centers) to offer richer features. Centralized cloud providers also benefit from bulk buying power when investing in new hardware (Elon Musk recently purchased $500 million worth of chips from Nvidia) and can pass those savings on to customers, making it difficult for DePINs to match their pricing.

AI as a smart contract

Running AI in smart contracts overcomes many limitations of centralized models. They become:

• Tamper-Proof- canister smart contracts are immune to cyberattacks because the Internet Computer is highly decentralized, and its consensus mechanism leverages chain key cryptography. It has no single point of failure.
• Unstoppable- once a smart contract deploys, it’s guaranteed to run. The Internet Computer never goes offline, never needs to be rebooted, and doesn’t rely on a data center that could experience an outage.
• Autonomous- a smart contract can be managed by a DAO or even without a control structure, in which case nobody can circumvent its rules or change its code.

Most blockchains can’t operate AI fully on-chain because their smart contracts have access to a limited amount of memory, which explains why some rely on centralized cloud servers. As a result, they run AI WITH the blockchain. The Internet Computer is different because our smart contracts use a 32-bit WebAssembly (WASM) virtual machine, giving them access to 4GB of memory. That means we run AI ON the blockchain.

Of course, running LLMs in a smart contract is the ultimate goal. While the compute power required to achieve this presents considerable technical challenges, we’re making progress. The Internet Computer is in the process of moving to a 64-bit WASM, which will quadruple our smart contracts’ memory to 16GB, enough to deploy a light version of Llama 3.

The blockchain effect

So how does the Internet Computer’s version of AI on the blockchain safeguard against the threats to centralized models?

Let’s start with a real-world example. Imagine a legal firm employing 100 lawyers. Each lawyer might need several weeks to prepare for a trial, which is expensive for their clients. The lawyer could also miss vital details.

Now imagine the firm has an AI oracle that ingests all client communications, legal filings, historical case law, and texts. If the lawyer had access to this oracle, the time required to prepare for a trial shrinks to a matter of days, and there’s less chance of making a mistake.

If this model runs on a centralized server, it becomes a target for hackers due to the sensitive information it stores. It’s the equivalent of a hot wallet for client data, and anybody who holds crypto understands how vulnerable they are. Hackers could use this information to ruin the client’s legal case or life. Alternatively, the IT infrastructure could suffer an outage just before the trial, leaving the lawyer without research required to argue the case.

However, as described in the previous section, neither of these scenarios happen to a model running on a smart contract on the Internet Computer due to its tamper-proof and unstoppable properties.

Likewise, dapps can leverage AI on the blockchain in innovative ways. For instance, digital wallets can be smart contracts, which offer a range of benefits, including enhanced security. As smart contracts are autonomous, nobody can modify the code and steal the user’s assets. These wallets can also incorporate customizable features, such as an AI model trained to receive instructions to execute trades under certain conditions, like swapping half of a user’s bitcoin holding for USDC if the BTC price hits $100,000.

Lots of use cases can benefit from autonomy. Another example is an AI model that audits Solidity code on Ethereum to detect vulnerabilities and demonstrate trustworthiness. The audit’s legitimacy rests solely on the autonomous nature of the smart contract containing the model.

Something momentous on the horizon…

Dom finished the conversation by teasing a significant development that will introduce a new dimension to the intersection between AI and blockchain. He wouldn’t go into further detail with Tracy, but he said the team hopes to roll out the beta version at the start of September. Watch this space!

To see AI on the blockchain in action on the Internet Computer, check out this facial recognition demo that Dom recently recorded.