#443: Google’s Gemini 2.0 Flash And Project Mariner Could Turbocharge Browser Activity, & More

Last week, Google released Gemini 2.0 Flash,¹ the first of its Gemini 2.0 family of models. Like Anthropic's Claude 3.5 Sonnet’s² improvement over Claude 3 Opus, Gemini 2.0 Flash is smaller and faster than Gemini 1.5 Pro, but more performant on most benchmarks, importantly the autonomous coding benchmark SWE-bench Verified,³ on which it scored 51.8%,⁴ slightly higher than Claude 3.5 Sonnet’s 49% and just shy of Amazon Q’s 55%.⁵

Leveraging upon its impressive benchmark performance, Google is building new and updated products around Gemini 2.0. Project Mariner,⁶ an AI agent that sees and interacts with browsers, uses keyboards and cursors to chain together actions and complete complex tasks like shopping. 

Project Mariner can browse the web autonomously⁷ to collect the contact information of companies listed in user-provided spreadsheets, for example, or identify unique color combinations in the art of the most famous post-impressionist painter⁸—Mariner chose Van Gough—before adding painting supplies to an Etsy shopping cart. Though early in development, Project Mariner already is demonstrating how AI agents can accelerate worker productivity and potentially facilitate⁹ trillions of dollars in online spending.

Last week, Google released¹⁰ details on Willow, a quantum chip that has made a meaningful step forward in quantum computing. Previous quantum systems have been plagued by error rates that result in inaccurate computation. Typically, the larger the system—measured in qubits—and the longer computation time, the larger the error problem grows. Benefiting from advances in its design and fabrication, Willow now has bragging rights to a quantum computer that has crossed a critical error correction threshold, so much so that its error rate improves as the system scales to larger numbers of qubits.

Crossing the error correction threshold is one critical milestone for quantum computing, because error rates are likely to improve as the number of qubits per chip grows over time. That said, to achieve “true” fault tolerance, Google’s quantum team still must improve the error rate by several orders of magnitude—from 10^-3 to 10^-6, a daunting challenge for a chip that has increased the number of qubits from 53 to only 105 over the past five years. While Google’s tests suggest that Willow can complete in five minutes a task that would have taken a classical computer septillion years (10²⁵) to complete, commercially relevant applications are out of reach for modern quantum computers today, as shown below.

Note: The chart above was produced by Google and is based on their research, not ARK’s research. Source: Google 2024 as of December 13, 2024.¹¹ For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency.

Even so, Google’s progress is reigniting debate in the crypto space, as quantum computers ultimately could break the cryptographic schemes used in both transaction authorization and Proof-of-Work mining. While quantum-resistant cryptography exists today, and networks like Ethereum have committed¹² to upgrading their codebase to be quantum secure, Bitcoin stakeholders face a dilemma: how to maintain Bitcoin’s ethos of ossification¹³ while addressing material risks to the network’s core functions—especially the adverse implications associated with the ~4.6%¹⁴  of original BTC supply still held in the wallets of Bitcoin’s founder, Satoshi Nakamoto. Fortunately, long-time Bitcoin developers like Jameson Lopp¹⁵ are exploring and explaining the risks, suggesting that a solution is likely before it is too late.

In a milestone for AI-driven drug design, Relation Therapeutics recently signed a landmark partnership with pharmaceutical giant GlaxoSmithKline to develop next-generation treatments for fibrotic disease and osteoarthritis. Possibly the largest ever for a seed-stage biotech company, the deal could generate hundreds of millions of dollars in milestone payments per target. 

Relation’s unique "Lab-in-the-Loop" approach integrates patient-derived multi-modal data, advanced perturbational experiments, and cutting-edge machine learning—including graph neural networks and generative models—to accelerate the discovery of more efficacious, targeted therapies. Partnered with NVIDIA through a seed round in its venture arm, Relation combines high-performance computation and biology to transform drug development timelines.

Relation is one of a handful of companies aggressively pursuing an AI-driven drug design strategy that could transform the economics of drug development. Today’s drug discovery process is notoriously time-consuming and costly: including the cost of capital and failed programs, commercial development typically requires $2.4 billion in capital per drug over ~14 years, as shown below.

Note: 10% discount rate. Assumes demonstrated 70+% reduction in phase 1 failure rates moderates to a 50% failure reduction in phase II and 25% failure reduction in phase III. Assumes pre-clinical efficiency is similar to what Absci has demonstrated. Assumes that licensing timeframe reduces to 12 months from 18 months. Source: ARK Investment Management LLC, 2024. This ARK analysis is based on a range of sources as of 12/13/24, available upon request. Forecasts are inherently limited and cannot be relied upon. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

Advances in AI are about to change the situation for those companies harnessing its potential aggressively. According to ARK’s research, the cost of AI-designed drugs that have passed the toxicology phase of human clinical trials successfully could be ~$1.5 billion, or 40% less than the cost associated with conventional methods.  By using AI to design better molecules up front, companies have been able to get molecules to human trial more quickly and with less likelihood of failure. Across AI-designed drugs that have entered toxicology testing, just 12.5% have failed, a 4x reduction relative to traditional drug development.¹⁶ Recursion Pharmaceuticals is an early mover that has already advanced drug targets to this stage.

As shown in the bar on the right below, our research suggests that AI-driven drug design will continue to increase in efficiency. Progress in the preclinical process could accelerate, cutting in half the time to human trial and saving at least 4 - 5 years across the entire process. Moreover, the success rates upon entering human trials also could improve. The goals and expectations disclosed by public companies like Absci and Recursion, as well as private companies like Relation, are consistent with that rate and magnitude of progress. Altogether, the cost to commercialize a drug could drop by 75% relative to the status quo, enabling AI-first companies to bring drugs to market for roughly $600 million.

If our research proves out, the return on research and development (R&D) dollars for these strategies could increase significantly from their woeful ~4% average today. To justify $2.4 billion in development costs, drugs must clear ~$600 million in peak annual sales. With similar revenue opportunities, AI-designed drugs in clinical trials today could triple the return on R&D investment relative to the industry average today, and those in preclinical trials could improve returns up to 8x. 

Source: ARK Investment Management LLC, 2024. This ARK analysis is based on a range of sources as of 12/13/24, available upon request. Forecasts are inherently limited and cannot be relied upon. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

While compelling, these estimates do not capture the superior revenue generation that we believe is possible as AI-designed drugs commercialize earlier in their patent lives, enabling longer cashflow runways before generics and biosimilars erode their potential. As these benefits become more apparent, partnerships should accrue to the benefit of companies harnessing AI’s potential aggressively. Most important, patients should have access to better, more precise medicines targeted against a wider variety of diseases and disorders.