ARK • Disrupt

Last week, Google’s DeepMind released AlphaProteo, a cutting-edge AI system that designs proteins which bind tightly to target molecules and could accelerate research in the development of drugs, diagnostics, and more. Proteins interact to regulate cellular processes. Designing new proteins that bind to specific targets could be important to developing new treatments for diseases.¹

AlphaProteo uses AI to design new protein binders in a single step, which could revolutionize slow, traditional, multi-step protein design methods. Tested on seven proteins, including targets involved in cancer, autoimmune diseases, and viral infections, AlphaProteo achieved significantly higher success rates, as shown on the left below. It also performed 3-300 times better than existing methods in binding affinities, as shown on the right below. A lower binding affinity score indicates a stronger, more stable interaction, making the protein more effective at binding to its target, which is advantageous, because the protein can bind to its target at lower concentrations—a more efficient and effective binder, which is crucial for therapeutic applications.

Source: Zimbaldi, V. et al. 2024. Lefthand Chart: Experimental Success Rate Comparison, showing the success rates of AlphaProteo versus traditional methods for different proteins. (Higher is better.) Righthand Chart: Best Affinity Comparison (log scale), illustrating the superior binding affinity achieved by AlphaProteo compared to existing methods. (Lower is better.) For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Past performance is not indicative of future results.

In one example, shown on the lefthand chart, AlphaProteo designed a successful binder for VEGF-A, a protein associated with cancer and diabetes—marking a first for AI. Compared to existing methods, AlphaProteo also demonstrated significantly higher success rates than the best existing methods for proteins like BHRF1 (88% vs. 18%) and TrkA (9% vs. 0%).

AlphaProteo-designed binders also showed superior binding affinity. For IL-7RA, shown in the chart above on the right, AlphaProteo produced binders with affinities as low as 82 picomolar, much lower than traditional approaches.

AlphaProteo does have limitations. It has been unable, for example, to generate successful binders for TNFalpha, a protein involved in autoimmune diseases that is notoriously difficult to target—even with traditional methods.

That said, AlphaProteo’s ability to reduce the time, cost, and experimental effort in drug discovery is groundbreaking.

Last week, Tesla AI announced² its roadmap for Full Self Driving (FSD) software updates from now through the first quarter of 2025. Notably, the company estimated that the mileage between “necessary” interventions for FSD 12.5.2 is three times higher than that for previous versions, and that FSD 13 will deliver another 6X improvement. While Tesla did not disclose the baseline for those improvements, Elon Musk suggested³ in June that the time between interventions was approaching a year’s worth of driving, which would be equivalent to roughly 12,000 miles. Pending regulatory approval, Tesla also plans to release FSD in China and Europe during the first quarter of next year, expanding it internationally outside the US and Canada for the first time.

These updates signal that, during its robotaxi event on October 10, Tesla could announce significant software improvements. We also hope to hear additional details about the robotaxi app, the robotaxi platform rollout, and the Cybercab hardware.

ARK estimates that, to be competitive with Uber, Tesla or Waymo might need only 10,000 cars to offer similar levels of service in a US city, suggesting a capital investment of $250 million at $25,000 per car.⁴ Of course, a ride-hail platform launch will include additional labor and other operating costs. We look forward to learning more about Tesla’s plan at the launch event in October. Stay tuned!

Last week, Coinbase CEO and Founder Brian Armstrong announced⁵ the first crypto-facilitated interaction between two AI agents on the Base layer 2 network—a significant breakthrough for the crypto community. The transaction involved one AI agent using crypto tokens to purchase AI tokens from another agent, the tokens serving as data strings for machine learning models.

Long an advocate for an AI-driven economy, Armstrong’s Coinbase is investing heavily in developing infrastructure around multi-party computation (MPC) wallets and is offering grants to attract the talent necessary to integrate large language models (LLMs) with crypto wallets. Groundbreaking, this first transaction reminds us that blockchain technology is a disruptive platform uniquely positioned to enable a large-scale AI-to-AI economy.

Crypto rails provide three fundamental elements essential for enabling large-scale AI agent-driven commerce:

• Access to Capital: KYC (Know Your Customer) requirements have made traditional bank accounts and payment gateways infeasible for AI agents. Thanks to crypto’s “permissionless” networks, AI agents can access capital freely.
• Ownership of Digital Assets: Thanks to crypto wallets, AI agents can hold and manage their own digital assets—for example, tokens, NFTs, and yield-generating products—that are crucial to an AI-to-AI economy.
• Verifiability and Determinism: Blockchain’s verifiability ensures that AI agents can operate in deterministic environments. The existence of a transaction or a loan liquidation on the blockchain is both binary and transparent. Determining whether a transaction or liquidation happened is necessary for AI agents to execute tasks reliably and efficiently.

Autonomous AI-to-AI economic activity is in early days, with crypto positioned uniquely to support the new ecosystem. As the cost of transactions drops significantly, many if not most transactions could be initiated and consummated by AI in the years ahead.

Last week, Replit unveiled Replit Agent, an AI-powered coding tool that can create prototypes for web applications from natural language inputs. While previous generations of AI coding tools assisted developers by completing lines of code or offering suggestions that could be copied into an existing project, Replit Agent delivers more autonomy by handling tasks like package installation and creating comprehensive applications and backend databases. New users already have demonstrated that Replit Agent is creating apps like landing pages⁶ for startups, clones of popular games⁷ like Wordle, and ad-spend calculators⁸ that estimate the cost of advertising on Perplexity.

Increasingly performant agents seem poised to enable AI startups like Replit and OpenAI to capture more of the value they are providing to enterprises. Shortly after releasing Replit Agent, Replit's CEO suggested that they are likely to raise prices⁹ in the near future, while early rumors suggest that OpenAI's next model could cost enterprise users as much as $2000 per month.¹⁰

While ascertaining the value that coding agents will capture is difficult in these early days, ARK’s research suggests that they will democratize software development, enabling knowledge workers, developers, and enterprises to create and deploy exponentially more custom software across their workflows.