The convergence between artificial intelligence and crypto has been touted as the next big thing in tech. For the past few years, we've seen AI crypto tokens reach market caps of greater than $1 billion.
But despite this massive investor interest, there has so far not been a corresponding wave of user demand.
Ask the average AI user which program they rely on for everyday use, and they'll likely mention programs like ChatGPT, Brave's Leo search app, or Microsoft's Copilot.
Rarely will a user state that they use a blockchain or crypto protocol.
But is this user demand coming in the future? And will blockchain AI truly revolutionize the world, or is it just the latest fundraising hype?
Cointelegraph sat down with executives from some of the leading blockchain AI protocols to ask them this very question.
GPU demand is growing
Guarav Sharma, chieft technology officer of AI project IO (IO), stated that today's centralized cloud computing systems simply can't keep up with the demand for graphical processing units (GPUs) that are desperately needed by AI developers, and this provides an opportunity for decentralized blockchain projects.
Before working on the project, Sharma was employed in the hotel industry, developing AI models that helped to predict which hotels a user was likely to book and what price to charge. But when he asked Amazon for enough GPUs to train his model, they reportedly claimed they didn't have enough inventory to fulfill his requirements. He stated:
"We went to Amazon to be honest, like we first thought of buying it. We couldn't buy it. Then we went to the cloud. We didn't find it there also, and we just had to wait for months to get this inventory from the AWS itself at that time."
The fundamental problem, according to Sharma, is that centralized cloud computing providers take months to set up servers in a particular location and at great cost to the average user.
Meanwhile, there may be some GPUs sitting around in exactly the location the customer wants, but because they aren't owned by the provider, they aren't offered.
For example, if a customer goes to Amazon and asks for 10,000 GPUs in Amsterdam, they are not going to partner with Google to provide these servers. "That's not the way they do it, right?" Sharma asked rhetorically.
In his view, decentralized protocols like IO can solve this problem by creating a marketplace for GPU power.
Clients can come to the platform to find servers and providers can offer their GPUs on the platform, allowing customers to find GPUs regardless of provider. Given the growing demand for GPUs as AI applications become more popular, this is the only way to efficiently match buyers with sellers.
Recent: $100K Bitcoin? 9 analysts share their end-of-year BTC price predictions
Even so, Sharma conceded that some AI teams are not offering much value, both in blockchain AI in particular and in the broader AI industry.
Some teams claim that they are going to create the next big model with just three or five people, when in reality this takes a much larger team than that.
Others have engineers that worked for major companies but have no portfolio to show investors.
Sharma suggested that investors should be careful to scrutinize the team behind each project. The ones that are likely to produce good work in the future have likely produced good work in the past.
Investors should also demand open-sourcing of code and regular audits to ensure that the public is aware of how much human intervention is involved in the project, he claimed, as some "AI projects" are more human than AI.
Prediction markets may need AI
According to ORA co-founder Kartin Wong, blockchain prediction markets will need to use AI in the future, and this will necessitate the convergence between the two technologies.
Wong pointed to the rise of Polymarket as proof of this need. While Polymarket runs on a blockchain, it "can have no oracle to address and to resolve [the question of who won a bet]."
Instead, "it's based on human judgment most of the time." But blockchain AI can create oracles that will "answer anything in the world, if this thing happened on [the] Internet."
He also argued that tokenization can facilitate fundraising for AI models. ORA pioneered the idea of an "initial model offering," allowing untrained AI models to launch tokens. The resulting funds raised can be used to pay for the model's training, which is highly GPU-intensive and expensive.
According to Wong, the models launched on ORA are owned by token-holders, allowing these holders to profit from their success.
They are also open-source, which creates transparency for the investing public. Wong claimed that this solves a common problem in AI, which is that most models have to be proprietary in order for their investors to make money.
On ORA, model creators are required to abide by the licenses in open-source software, which he claimed prevents developers from merely cutting and pasting code to cannibalize the profits of creators.
However, Wong also acknowledged that there are some fake blockchain AI projects or fake AI projects in general. Some models may claim to be generating results from AI, but they may be using humans to check the work produced by a model, and this may make the model superfluous.
He suggested that distinguishing between fake and real AI may sometimes be very difficult.
However the best way for investors to judge whether a product is really AI is to use it, he stated. He pointed to ChatOLM, a chatbot created through ORA, as an example of a product which is obviously using AI, since it answers questions faster than a human possibly could.
Blockchain may allow for "truly autonomous AI"
According to Ron Chan, co-founder of blockchain AI protocol Inference Labs, blockchain provides the only means to attain "truly autonomous AI." For that reason, humanity won't be able to do without it in the future.
Chan stated that centralized AI "is developed toward the goals of the enterprise." While this has its place in the world, decentralized AI fulfills a different need.
It "is freeform -- its development is driven by the participation and speed of market demand," which "creates the conditions for human centric innovation with the power to solve great challenges."
He claimed that decentralized AI will develop systems for "proof of inference" or the ability to prove that a particular answer came from a particular AI model. This, he stated, is an "immediate need" for the industry.
Chan acknowledged that distinguishing between human and AI projects can sometimes be difficult or even impossible. He pointed to the example of X user Error Error Ttyl, which is an account that claims to be controlled by an AI model.
"How can observers verify there isn't a human operator making decisions behind the scenes?" he asked rhetorically, pointing out that because both the AI and its creator hold the password to the account, verifying who is generating the posts may be impracticable.
Recent: 7 policymakers who are 'laser focused' on $100K Bitcoin price
The answer is to give the AI exclusive control, verifiable independence and irrevocable delegation, Chan suggested.
The AI must have "sole access to the account," and third parties must be able to verify this fact.
In addition, once control over the account is transferred to the AI, it must be impossible for humans to regain this control. Only then can we know that whatever it does is truly initiated by the AI model and not by a human working behind the scenes.
In Chan's view, this type of provable AI inference is an area where only decentralized protocols can offer a solution.
The greatest benefits may come later
Cointelegraph asked the interviewees for examples of consumer-facing blockchain AI apps that users can enjoy now rather than in the future.
In response, Wong referenced the chat app OLMChat, while Chan discussed an aircraft-tracking AI venture and a liquid staking app created by the Inference Labs team.
While these apps may have small user-bases compared to superstar software like ChatGPT, the interviewees all had high hopes that blockchain AI really will revolutionize the world, even if its benefits may take a while to be fully realized by end-users.
