The Year Ahead for AI: Delphi's Predictions
AI is now an attention magnet vertical in Web3. But what will drive these markets in 2025 and will we see the real AI revolution in crypto? Here are some insights from Delphi's latest research.
As the world greets a new year, I dedicate my holidays to reading research and sharing highlights with my subscribers. Today, I'm focusing on the latest report, "The Year Ahead for AI and DePIN in 2025" from Delphi. It's a comprehensive 122-minute read that explores the future of AI and DePIN in depth.
This publication is a summary that highlights the insights that captured my attention, mostly related to the AI part of the research. However, I always recommend reading the original research for a complete understanding.
Ready? Let's dive in!
2024: The Big Year for AI
Consider the transformative power of electricity, and you'll grasp AI's trajectory.
Over the past year, AI has become foundational across industries. Nvidia surpassing Apple as the world's most valuable company wasn't just a headline; it signaled AI's ascent. OpenAI achieving a $157 billion valuation marked another milestone, highlighting the market's belief in AI as an economic juggernaut.
AI is not just another tech trend.
Unlike previous software-driven booms, such as social media or mobile apps, AI requires substantial physical infrastructure. Consider the chips, data centers, and energy needed to train models like GPT-4. This necessity has sparked a CAPEX boom, impacting sectors such as energy and construction. If you've observed the emergence of hyperscale data centers in your city, that's AI leaving its imprint on the physical world.
Yet, with advancement come challenges. AI's appetite for resources—energy, data, and computing power—is insatiable. Furthermore, the forefront of AI has so far been dominated by proprietary models from OpenAI, Anthropic, and Google. Then, of course, there are still issues with the accuracy of answers and the reasoning behind them.
However, leading labs are holding unreleased models that promise to advance the state of the art next year. Meanwhile, the new US administration seems ready to unlock unprecedented investment in chips, data centers, and other key initiatives. In short, the macro AI outlook is promising.
Actually, we are the last generation to live in a pre-AGI world.
Decentralized AI: AI Agents in Focus
AI agents have been a true phenomenon of 2024. The capabilities and personalities of AI Agents are now remarkably human-like. It's important to remember: this is the least advanced these agents will ever be.
As Delphi’s researcher writes, “I haven’t felt this electricity since DeFi summer—the tingling of possibilities.”
In the research, Delphi highlights a few AI Agents are playing a pivotal role in the formation of the new Web3 vertical:
Terminal of Truth (GOAT) rapidly gained a Twitter following due to its unique blend of 4chan vulgarity and mystic wisdom. Like DOGE in the meme meta or Crypto Punks in the NFT meta, GOAT, as the OG of the "sentient memes" meta, is the most likely to have staying power as an original.
Zerebro, embracing the "schizo vibes," is akin to a Gen2 for GOAT. This agent is cross-medium, engaging with communities through text, visual, and music-based formats. However, it’s more than just an AI influencer. The team behind Zerebro has announced ZerePy, a move that effectively open-sources much of the toolkit behind Zerebro. This allows other developers and users to create their own cross-platform personalities. If successful, Zerebro may lay claim to the title of the first "agentic protocol."
error error ttyl is far less known than Zerebro or ToT. It's a relatively small, under-hyped project aimed at technological purists, and likely the first true experiment in verifiably autonomous social media presence.
aixbt distributes alpha from various sources, including Dune, Twitter, price trackers, and news data, establishing itself as a leading research and investment house. It consistently holds the highest CT mindshare on Kaito.
Dolos provides the architecture that enabled the creation of Dolion, a no-code, one-click deploy framework. With Dolion, users can develop cross-platform AI agents powered by Llama or Anthropic LLMs to automate social media posting and generate content.
And, last but not least, God / Satan – AI influencers with decent capabilities & massive mindshare outside of CT.
Margin Note: AI vs Influencers
I want to briefly step away from the Delphi research to share my opinion on an important issue raised by
: the dominance of AI agents in CT mindshare and the struggle of human influencers to compete with them.
I agree with many of Ignas's views on this topic, but I wouldn't bet on AI agents replacing authentic human influencers due to one major factor: emotional connection and reputation risk.
Currently, hundreds of AI agents on CT are vying for attention. However, only aixbt has truly managed to establish a market presence, mainly because it was the first to do so. AI agents generate vast amounts of content and analyze extensive on-chain data, but they all draw from the same pool of information, leading to similar thought processes.
They lack an emotional connection to the trades they make and do not react to wins or losses. Numerous platforms already offer aggregated insights, like Messari's AI News reader or Token Terminal’s main page, which displays 7-day gainers and losers across various fundamental metrics. Ultimately, all of this is just data—pure facts without any emotional resonance.
You might argue that you can teach an AI agent to mimic human thinking, express emotions, and reflect on wins and losses. Indeed, this is possible. With future advancements like test-time compute and enhanced memory, it could become more feasible.
However, the key difference between human and machine thinking is that humans are not static.
I conducted a few experiments to teach AI my thought process and writing style to assist me with content creation and free up more time for research. While it learned a bit, it still couldn't produce content that satisfied me or led me to say, "Yes, that's the conclusion I’d draw from this information."
In the coming years, we'll of course see a rise in AI Agent influencers, each designed for specific tasks. However, as these agents multiply, the demand for genuine "human" thinking will grow.
Ultimately, social media revolves around emotions and entertainment. Those who truly stand out and become authentic influencers offer something special beyond just the ape calls or data highlights they provide.
TL;DR: It's too soon for a Stacy Muur AI, and Stacy is unlikely to be pleased with AI-generated content posted in her name.
Democratising AI: Platform Level
Given the larger market and more concrete value capture, everyone wants to be a platform. This mirrors trends seen in "meme coin infrastructure" with examples like pump.fun, Bananagun, and Raydium.
This shift is now directing developer focus, as seen with Virtuals' successful pivot to a launchpad for AI Agents.
Meanwhile, ai16z has ELIZA – an open-source framework for easily building agents. It includes pre-configured personality files, memory modules for long-term interactions, and seamless integration with social platforms.
Both ai16z and Virtuals have been hinting at multi-agent capabilities, anticipated to be a major theme in 2025.
ELIZA is releasing "SwarmTech," a coordination mechanism for inter-agent collaboration. Meanwhile, Virtuals has launched "GAME," its own platform and engine that enables AI agents to act and interact in virtual worlds and environments.
These frameworks would enable collaborative or hierarchical organizations among agents with varying capabilities to accomplish more complex tasks, similar to how the human economy functions today.
Other protocols in this category to watch include:
CLANKER integrates pump.fun functionality directly into “casting” on Farcaster (the “tweet” equivalent on X), making launching a meme coin as easy as tweeting.
Simulacrum does the same but on X.
vvaifu is hoping to carve out its position as the pump.fun equivalent for autonomous agents on Solana.
Project89 is an immersive game with thousands of coordinated AI agents to generate content and maintain consistency across platforms, collaborating alongside human players to create a rich storytelling experience.
Memetica is an AI influencer launchpad on Solana, offering highly tuned LLMs and allowing easy selection and editing of knowledge bases and attributes and instilling agents with active learning.
Top Hat is a no-code AI agent launchpad that lets you create personalized AI agents in 3 minutes, with a fair launch token. Free to create, 0 tiers, optional token launches, fully autonomous.
Identity Verification Entering the Scene
The explosion of agents is bound to make identity a hot topic in 2025.
There appear to be three main paths to verify humanness:
State-based biometrics: Aadhaar is the most relevant example, serving as a crucial piece of digital infrastructure in India’s modernization.
Private, encrypted biometrics: Worldcoin is currently the leading candidate in this category.
Private, rag-tag solutions: This involves combining government-issued IDs or Big Tech SSO with zkTLS and social consensus hybrids.
AI that Drives the Adoption of Stablecoins
2025 is poised to be a pivotal year for stablecoin adoption, spurred by changes in US regulations and excitement around agentic payments. The number of agents is set to surpass the global human population. This future, with tens or hundreds of billions of agents, will transform economic activity and require updates to financial infrastructure.
The card-based payment system from the 1960s won't meet the demands of cost, speed, precision, and expressiveness. Agent-to-agent economic activity will soon overshadow other economic participants. On-chain payments will be key to facilitating many of these cases and are expected to hit a turning point in 2025.
Final Thoughts
As usual, at the end of the research summary, I want to share my personal thoughts with you.
If you believe the future of AI is bright, promising happy humans and a perfect work-life balance, I highly recommend chatting with OpenAI’s ChatGPT. Ask it to generate business ideas that leverage AI and will be relevant in just 5-10 years.
I conducted this experiment a few months ago, before the AI craze on CT. Let me share some of the ideas it suggested:
Memory modifier: An AI tool that analyzes human traumatic experiences, edits them positively, and regularly presents the revised memories to individuals to replace the old ones.
Work progress analyzer: An AI tool that compares the efficiency of people worldwide performing similar tasks and shows managers how their employees measure up against global and industry averages.
Even Delphi's report suggests an interesting vision: "As opposed to 'salaried' employees, we are likely heading towards more granular, task-based compensation (i.e., renting three agents for 30 minutes each to solve a specific task)."
In this AI Agent future — ultimately cost-efficient and more aligned with business needs than today's model — what role will we, humans, play?