Crypto Could Empower AI Agents, & More

Want more research from the ARK Team? Have feedback on our publications? Click here to help inform our content creation.

1. Crypto Could Empower AI Agents

By: Lorenzo Valente | Research Associate

During his recent keynote address at CES 2025,[1] Jensen Huang discussed AI agents as a multi-trillion-dollar opportunity—a nod, we believe, to an agentic economy that is beginning to evolve. AI agent deployments like Eliza and Virtuals are leading the charge, creating foundations for the deployment and management of AI agents on Ethereum and Solana.

Beyond enabling the tokenization of AI agents, Virtuals and Eliza leverage tokens to promote co-ownership, democratizing access and governance. With APIs and SDKs,[2] both frameworks offer seamless integration of AI agents into Web 2 applications, facilitating dynamic interactions with and among platforms like X, Discord, and Telegram while enabling wallet ownership and diverse service provision.

Many teams are growing the agentic economy on a number of fronts. Prime Intellect and Nous Research are leading efforts to allow distributed pre-training and inference, pushing the boundaries of scalable AI development, while other projects are developing verifiably autonomous agents through Trusted Execution Environments (TEEs) that ensure provable and verifiable autonomy. Additionally, the team behind the Virtuals framework is experimenting with the deployment of agents with advanced multimodal capabilities that can process and interact across various types of data, including text, images, video, and audio.

Currently, AI agents can be supervised by shared access and control mechanisms that include recovery and override capabilities. While today’s AI agents cannot manage resources or complex tasks with long-term objectives, blockchain technology offers fertile ground for their evolution in that direction. Blockchains grant AI agents access to capital and computational resources, enabling them not only to own and manage assets but also to interact with decentralized capital markets and operate autonomously to a degree impossible with traditional systems. AI agents can custody their own private keys on blockchains, giving them the independence to become pivotal actors in the agentic economy.

While today they are more like glorified chatbots than autonomous entities with developed cognitive, operational, and interactive competencies, AI agents are evolving at an unprecedented pace. Given strong economic incentives, they could make meaningful contributions to the growing blockchain economy.

2. The UK Biobank Has Chosen Ultima For Its Proteome Study

By: Nemo Marjanovic, PhD | Research Analyst

The UK Biobank’s Proteome study recently chose Ultima Genomics’ UG 100 HT sequencer to be its primary platform, a major milestone that intensifies pressure on Illumina.[3] At $100 per genome ($1/Gb), Ultima’s sequencing technology leverages its innovative rotating disk flow cell design to optimize reagent consumption and deliver unprecedented cost efficiencies. The platform has ignited discussions about the feasibility of $1-10 genomes that could sequence billions of genomes and reshape the multi-omics field fundamentally.

Enabling population-scale studies and accelerating applications in precision medicine and multi-omics research, $1-10 genomes could transform academic research, drug discovery, and public health. For Illumina, the growing viability of ultra-low-cost sequencing introduces competitive challenges, particularly in research-focused markets.

Notwithstanding those pressures, Illumina retains a stronghold in clinical markets with applications in cancer diagnostics, NIPT,[4] and infectious disease that account for ~55% of its consumables revenue. While Ultima’s $100/genome could disrupt the research segment—the UK Proteome study collaboration underscoring the shifting competitive dynamics around sequencing[5]—Illumina’s transition to its X-series platform and focus on clinical growth should stabilize its outlook.

3. The GET Machine Learning Model Seems To Be Echoing The Central Dogma Of Biology

By: Nemo Marjanovic, PhD | Research Analyst

The General Expression Transformer (GET) is a potentially groundbreaking foundation model that deciphers the interactions shaping gene expression.[6] Trained on chromatin-accessibility data from more than 200 cell types, GET learns the ”syntax” that governs how chromatin states influence transcription, predict gene expression in unseen cell types, and uncover protein interactions associated with diseases. By integrating structural insights from tools like Alphabet’s AlphaFold, GET represents a significant advance in biology that could unlock new possibilities in gene regulation, personalized medicine, diagnostics, and drug development.

Indeed, foundation models like GET could revolutionize research by building a general understanding of complex data systems. Trained on vast datasets, GET can be fine-tuned for specific applications like disease prediction and therapeutic design. Its ability to generalize across contexts has profound implications for healthcare that could drive innovations in precision medicine, faster diagnosis of disease, and more efficient discovery of therapeutics.

The GET approach parallels the central dogma of molecular biology, which describes the flow of genetic information from DNA to RNA to proteins, with feedback loops ensuring dynamic regulation. Foundation model training aligns with DNA, providing a universal base of information that encodes general knowledge. Fine-tuning parallels epigenetic regulation, where specific adaptations selectively activate portions of the foundation model for particular tasks. Intermediate outputs resemble RNA, serving as flexible intermediaries that translate foundational knowledge into task-specific predictions. Finally, proteins—the functional molecules that execute biological processes—mirror the actionable insights from fine-tuned models, such as identifying disease markers or therapeutic targets.

While the central dogma operates within the constraints of biological evolution, with feedback loops emerging naturally to maintain homeostasis, foundation models are human-engineered systems designed for rapid and targeted optimization. Biological outputs, such as proteins, are limited to cellular processes, while foundation models generate abstract, scalable outputs like predictions or classifications. Unlike biology's fixed genetic blueprint, foundation models can expand continuously and evolve through human intervention, enabling constant improvement, highlighting how biology and foundation models leverage generalization and adaptation but differ fundamentally in flexibility and innovation.

4. DeepSeek Highlights China’s Impressive Response To Hardware Restrictions

By: Frank Downing | Director of Research, Next Generation Internet

China-based AI lab DeepSeek—a subsidiary of the AI-focused quant hedge fund, High Flyer—has introduced DeepSeek v3,[7] a frontier model on par with Meta’s largest Llama 405b. Despite its comparable scale, DeepSeek v3 was trained using roughly one-tenth the compute, delivering better performance and dramatic efficiency gains. Andrej Karpathy described[8] this achievement as “making it look easy,” highlighting China’s growing influence in AI research.

At ARK, we have observed AI training and inference costs decline by 75% and 90% per year, respectively, as hardware breakthroughs and software optimization converge.[9] DeepSeek’s remarkable results suggest that, even with compute constraints, innovative algorithmic approaches can achieve state-of-the-art performance at a fraction of the typical cost.

In our view, DeepSeek v3 demonstrates both the early potential of next-generation AI research and the speed at which new entrants can push the boundaries of compute efficiency. One possible reason for China’s rapid progress is that its need to be resourceful with limited compute has encouraged algorithmic efficiencies that have yet to be explored fully elsewhere. Another possibility is that the Chinese government is accelerating domestic AI labs’ time to market by supporting domains like data sourcing. Either way, we believe DeepSeek v3 highlights the transformative impact of ongoing cost declines in AI and underscores the rapidly evolving global competition for AI supremacy despite chip restrictions.

SUBSCRIBE TO ARK'S RESEARCH

[1]?NVIDIA. 2025. “NVIDIA CEO Jensen Huang Keynote at CES 2025.” YouTube.

[2] Application Programming Interface (API). Software Development Kit (SDK).

[3] UK Biobank. 2025. “Launch of world’s most significant protein study set to usher in new understanding for medicine.”

[4] Noninvasive prenatal testing?(NIPT).

[5] Ultima Genomics. 2025. “Ultima Genomics Announces UG 100? Sequencing Platform Selected for UK Biobank's Groundbreaking Human Proteome Study.”

[6] Fu, X. et al.?2025. “A foundation model of transcription across human cell types.”?Nature.

[7] DeepSeek. 2024. “Introducing DeepSeek-V3!” X.

[8] Karpathy, A. 2024. “DeepSeek (Chinese AI co) making it look easy today...” X.

[9] ARK Investment Management LLC. 2024. “Big Ideas 2024: Disrupting the norm, Defining the future—Artificial Intelligence,” pp. 19-33.