ARK • Disrupt

Last week, Google’s DeepMind unveiled AlphaFold3, an AI model that can predict interactions between and among proteins and other molecules like DNA and RNA. DeepMind’s technology should enable scientists to determine biomolecular structures, revolutionizing drug design and the ability to target diseases.

AlphaFold3’s updated diffusion-based architecture can predict not only structures involving proteins, nucleic acids, small molecules, ions, and modified residues, but also protein-ligand, protein-nucleic acid, and antibody-antigen interactions. Those capabilities suggest that deep learning will create models with high accuracy across the biomolecular space.

Last week, Elon Musk announced that Tesla will upgrade its AI-enabled end-to-end Full-Self Driving (FSD) software from v12.3 to v12.4 in a limited external beta release. Slated for this week, the update is likely to improve miles per intervention—a key measure of capability and safety—by 5-10x. Further, Musk expects v12.5, the next release in June, to expand Tesla’s one-stack AI capability from surface streets to highways.¹ No longer constrained by AI training compute, Tesla is expediting major software updates by leveraging its ~1.2 billion cumulative miles of FSD to date.²

According to our research, a Tesla in FSD mode on surface streets is ~5x safer than a Tesla in manual mode, and ~16x safer than the national average, as shown below.

Source: ARK Investment Management LLC, 2024. This ARK analysis is based on a range of underlying data from external sources, including CDC 2024, Kusano 2023, NHTSA 2023, Tesla 2023, 2024, and Zhang 2023,³ which may be provided upon request. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency. Past performance is not indicative of future results.

As software updates accelerate, Tesla should be able to demonstrate statistically significant safety metrics and receive regulatory approval for its robotaxi network. We look forward to monitoring FSD’s progress toward full autonomy. For more details, stay tuned to ARK’s Research Center for the publication of our updated open-source Tesla model.

During its Spring Update, OpenAI showcased GPT-4o, a natively multimodal model that spans text, voice, audio, and video with improved performance across most benchmarks. The new model will power an enhanced ChatGPT voice experience that is able to adapt its tone to different situations and facilitate use cases like real time translation. The more lifelike ChatGPT voice assistant allows users to interrupt and direct the model dynamically, setting a new standard for interactivity. Please watch the live demonstration here.⁴ OpenAI also launched a desktop app for Mac, enabling users to share their screen so that the model can respond to the content it sees, leveraging its multimodal capacity to provide context-sensitive assistance.

The event highlighted that free users of ChatGPT will now have access, previously available only to premium subscribers, not only to the new GPT-4o model, but also to advanced data analysis, browsing, and the GPT Store. OpenAI also has reduced application programming interface (API) pricing significantly, making GPT-4o 2x faster and 50% cheaper than GPT-4 Turbo—cutting the cost by 90% from $75 to $7.5 per million tokens, as shown below.

Source: ARK Investment Management LLC, 2024, based on data from OpenAI as of May 18, 2024. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency. Past performance is not indicative of future results.

The wide rollout of GPT-4o to free users and API price reductions suggest that OpenAI made significant progress in optimizing GPT-4o for inference, and also that Microsoft has increased its GPU compute capacity to support additional users. If so, ChatGPT’s monthly active users (MAUs) could scale far beyond the 100-200 million generated during the past year.

During its “I/O” event last week, CEO Sundar Pichai highlighted Google’s recent successes with agentic AI. Thanks to products like Gemini 1.5 and Project Astra, Google is integrating sophisticated, proactive AI partners into Google Search and Chrome. Importantly, the new AI tools assist, anticipate, and adapt to user needs.

A standout example,⁵ Gemini 1.5 can start and completed a retail return effortlessly after a user snaps a photo of a pair of shoes and asks Gemini to return them. Gemini 1.5 autonomously sifted through the user's Gmail to locate the original purchase order, filled out the return form, and scheduled a pick-up.