ARK • Disrupt

Last week, Tesla released the latest version of its autonomous Full Self-Driving (FSD) software, v13.2, to a limited number of external customers.¹ Building on v12's transition to end-to-end AI earlier this year, v13.2 introduced several enhancements, including FSD parking at the start and the end of journeys and driving in reverse. Its Cortex cluster in Austin has increased Tesla’s training compute power by 5x,² optimizing the performance of FSD v13.2 in vehicles with Tesla’s AI4 computer chip and higher-resolution video inputs. According to user videos showcasing zero-intervention drives,³ early feedback is that v13.2 has resulted in much smoother rides.

FSD v13.2 is a major step toward Tesla’s goal to launch a robotaxi service in California and Texas next year,⁴ aligning with ARK’s estimate, as shown below.⁵ Next year, Tesla predicts that the new model will extend miles between critical interventions,⁶ achieving a rate better than the national average of ~670,000⁷ miles between human accidents—a ~700x improvement over the widely released FSD v12.5,⁸ according to our estimates. While surpassing this milestone is not a prerequisite for Tesla’s robotaxi launch, it should save lives, enhance conversations with regulators, and lower the costs associated with remote human operators. We look forward to tracking FSD’s progress as Tesla moves closer to robotaxi commercialization.

*Figures may not add to 100% due to rounding. Source: ARK Investment Management LLC, 2024. This ARK analysis draws on a range of external data sources, data as of 6/12/24, which may be provided 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.

Could the human brain hold the code to a revolution in computing? With living brain organoids—tiny, lab-grown replicas of brain tissue—Organoid Intelligence (OI) represents the convergence between biology and computational technology.

The human brain’s computational capabilities are staggering. Consider this comparison: today, the fastest supercomputer in the world, Frontier, consumes 21 megawatts to achieve 1.1 exaFLOPs of processing power, roughly 1 million times more power consumption than the human brain’s 20 watts for comparable performance. Interestingly, 20 watts is barely enough to power a light bulb. For more perspective, humans can learn “same-versus-different” tasks with ~10 examples, while machines require millions of samples and orders of magnitude more energy. Google's AlphaGo is a case-in-point: to win against the world champion in the ancient Chinese game of Go, the machine trained on 160,000 sample games and consumed as much energy in a month as a human would need to live for a decade.⁹ Please see the chart below for more comparisons.

Sources: Morgan 2023, Wikipedia 2024, Butler 2023.¹⁰ For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

To advance computing, Oganoid Intelligence aims to recreate the brain’s efficiency and adaptability in a controlled lab environment. If successful, it could pave the way for sustainable hybrid biological computing systems, revolutionizing fields from medical research to artificial intelligence. Challenges to OI include scaling organoids while preserving their functionality, engineering seamless interfaces, and addressing ethical concerns around sentience.

In our view, the question isn’t whether biology can redefine computing, but when.

Last week, OpenAI launched the 12 days of OpenAI during which time it will debut or demo a new product each business day. On Day 1, OpenAI launched the full version of its o1 reasoning model and an upgraded “Pro” version of ChatGPT that will sell for $200 per month.

After fewer than 3 months,¹¹ the new version of o1 replaces o1-preview, with notable improvements to performance on both competitive mathematics and coding exams. On the American Invitational Mathematics Exam, o1 scored 78%, much better than o1-preview at 50%. In Codeforces coding competitions, the new version scored 89% compared to 62%. o1 is multimodal, allowing the model to evaluate charts and images.¹² Thanks to its ability to evaluate questions and discern the optimal time necessary to “think” before answering, o1 also is 50% faster than o1-preview.¹³

For $200, the ChatGPT Pro tier offers unlimited usage of the model and an “o1-pro” mode that taps additional compute, enabling it to answer challenging questions more reliably and outperform the standard version of o1 noticeably when factoring in the repeatably of correct answers. To compare o1-preview, o1, and o1-pro, OpenAI ran each model through the mathematics and coding exams referenced above as well as a series of PhD-level science questions, counting answers as correct only if the model could solve the questions correctly in four of four attempts. In one example, o1-pro mode outperformed o1 in mathematics, coding, and PhD-level science by 13, 11, and 7 percentage points, respectively, as shown below.

Source: OpenAI.¹⁴ For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

The ability to answer difficult questions reliably and repeatedly is an important requirement for models executing complex, multi-step tasks, especially in academic research. Researchers already are using o1 in medical research. To spur broader adoption in the sciences, OpenAI has awarded ten ChatGPT Pro grants to leading researchers in the United States. Accelerating research and other complex tasks could give o1 and o1-pro the data and power to spur broad-based adoption of AI across consumer and enterprise workloads.