Are AI’s Reasoning Skills More Capable Than Once Imagined?, & More

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

1. Are AI’s Reasoning Skills More Capable Than Once Imagined?

By: Brett Winton

For years, skeptics have dismissed AI models as nothing more than "stochastic parrots"[1] that regurgitate and interpolate training data with no real understanding or reasoning capabilities. A groundbreaking study[2] by Christian Lewis shatters that view by demonstrating that Anthropic's Claude model is “Turing complete,” capable of performing primitive logical functions that combine and execute all possible computer programs. In other words, given enough time and memory, a Turing machine could complete any well-defined task.

Recent debates have centered on whether AI models are Turing complete. ChatGPT, for example, relies on external coding and computation in responding to complex logical queries. Lewis’s research shows that the Claude Opus LLM (Large Language Model) is Turing complete natively when instructed appropriately, directly challenging the idea that AI models are inherently limited.

Computer scientist Yann LeCun has argued that the autoregressive nature of transformer-based language models is a critical weakness that limits them to interpolation within the boundaries of their training data. If AI models can execute logic functions as Lewis's research demonstrates, they also should be able to correct errors sufficiently to extend performance beyond the boundaries of training data.

Importantly, Lewis's study also characterizes so-called “errors” in AI models. In his view, many of Claude's apparent logical errors are not genuine logical failures. Rather, they are implementation errors caused by copying previous outputs incorrectly, suggesting that minor adjustments to their training protocol could improve the ability of LLMs to follow complex logical chains.

The implications are profound. AI models could become a new class of adaptive computers guided by prompts that function like software. Exploring the potential of Claude Opus and GPT-4 models has just begun; yet, even if AI training progress were to stop today, those models could be engineered to exhibit complex chains of logic and serve as powerful agents on our behalf.

According to Lewis’s research, far from approaching a performance ceiling, artificial intelligence is in the earliest stages of explosive capability.

2. As Tesla Evolves Its Supercharger Strategy, Waymo Hits A Milestone

By: Tasha Keeney, CFA & Daniel Maguire, ACA

Last week, notwithstanding recent layoffs and plans to slow growth, Elon Musk announced that Tesla will invest ~$500 million in its Supercharger network this year.[3] Seemingly contradictory, those developments suggest that Tesla is shifting the placement of Superchargers for personal vehicles to companies like BP PLC while deploying its own capital for charging infrastructure optimized for robotaxis.[4]

Although automakers like Ford and General Motors have thrown their support behind it,[5] ARK’s research suggests that Tesla's Supercharger network will have little impact on its revenue and earnings growth during the next five years. If the Supercharger network were to grow at recent rates and its utilization were to increase from ~11% on average to the ~34% typical of gas stations,[6] the financial impact on Tesla would pale in comparison to the scaling of its robotaxi business. According to our estimates, its autonomous taxi network will account for two-thirds of Tesla’s enterprise value in 2027.[7]

Interestingly, this week Waymo announced that the number of paid trips per week on its network has increased ~150% in the last four months, from ~20,000 in December to ~50,000.[8] According to our estimates, Tesla is collecting real-world autonomous driving data at a rate 110x that of Waymo, as shown below.

In our view, robotaxis will transform Tesla's business model—from one-off vehicle sales to a sustained recurring revenue stream—turning every car into an AI-powered cash-flow generating machine. For further analysis, stay tuned for ARK’s updated Tesla valuation model.

3. Bitcoin Supply Growth Has Dropped Below 1% At An Annual Rate

By: David Puell

On April 19, the Bitcoin community celebrated Halving Day.[9] At the beginning of April, the network produced 900 bitcoin on average per day, increasing supply at an annual rate of 1.8%. Post-halving, the network produces only ~450 bitcoin on average per day, increasing supply at an annual rate of ~0.9%. By the end of April, miners had minted 93.7% of all the bitcoin that ever will be produced, as shown below.

Since its inception, bitcoin has halved four times, each halving a precursor to upward price momentum. While historically its price performance in the first few months after halving has been negative-to-muted, as shown by the black square on the chart below, one year after the halving bitcoin’s price has increased 3x to 70x+, as shown by the purple rectangle.[10]

While a 3x increase in the price of bitcoin a year after this halving could be optimistic, the chart above highlights that bitcoin limited supply growth should increase its scarcity value over time. For further details on bitcoin’s dynamics during April, stay tuned for ARK’s next Bitcoin Monthly.

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

[1] Bender, E.M. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” FAccT '21: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.

[2] Lewis, C.J. 2024. “Universal computation by attention: Running cellular automata and other programs on Claude 3 Opus.” X.

[3] See Musk, E. 2024. “Just to reiterate: Tesla will spend well over $500M expanding …” X. Musk, E. 2024. “Tesla still plans to grow the Supercharger network, just at a slower pace…” X. See also Ewing, J. 2024. “Tesla Fires Many on Charger Team, Raising Doubts About Expansion.” The New York Times.

[4] See Alake, T. 2024. “BP Keen to Buy Tesla Supercharging Sites for US Expansion.” Bloomberg News.

[5] St. John, A. 2024. “One of the Last Major Holdouts Says it Will Adopt Tesla’s Charging Technology.” Business Insider.

[6] We think it’s unlikely that superchargers will need to achieve gas station level utilization because of the ability to charge vehicles at home and where electricity is available, as opposed to the limited availability of on-site gasoline. See Keeney, T. 2016. “Supercharger: A Charge Could Cost Half the Price of Gas.” ARK Investment Management LLC.

[7] See Keeney, T. et al. 2023. “ARK’s Expected Value for Tesla in 2027: $2000 Per Share.” ARK Investment Management LLC.

[8] See Waymo. 2024. “Our safe and deliberate approach to scaling.” X. See also Field, H. 2024. “Waymo Is Full Speed Ahead as Safety Incidents and Regulators Stymie Competitor Cruise.” CNBC.

[9] The Halving Day celebrates when bitcoin’s issuance gets cut in half programmatically every 210,000 blocks (approximately every four years). In addition to the rate of issuance of bitcoin being cut in half, the rewards for successful bitcoin miners are also cut in half.

[10] Arithmetic mean is commonly referred to as an “average,” as it is calculated by dividing the sum of a collection of numbers by the amount of numbers in the collection. Geometric mean is an average calculated by taking the product of a collection of numbers and finding the root of the product. Harmonic mean is the reciprocal of the average of the reciprocals of the numbers included in a collection. The harmonic mean is always the least of the three means, while the arithmetic mean is always the greatest of the three and the geometric mean is always in between.