#271: Artificial Intelligence Models Are Surpassing Benchmarks at Accelerating Rates, & More

At the first Artificial Intelligence (AI) workshop in 1956, researchers dreamt of a day when AI would outperform humans. Today, modern researchers use benchmarks to measure the performance of AI models relative to humans.

In 1997, 41 years after the first AI research workshop, IBM achieved a significant milestone when Deep Blue check-mated world champion chess player Garry Kasparov. While machine learning models used to take decades to achieve human-level performance on benchmark tasks, newer AI models are routinely developing superhuman skills within just a few years, as highlighted in a recently published research paper.

With an increasing supply of human and financial capital focused on AI, we believe that innovation in models, compute, and training techniques will continue to accelerate the rate of advancement. Research from Facebook AI on Vision Transformers and self-supervised learning, for example, should spur a step-function advance in specific computer vision tasks. On the compute front, Google recently announced TPUv4, its fourth-generation hardware chip purpose-built for AI training. TPUv4 is 2.7x more performant than its predecessor and, when configured in a 4,092-chip pod, generates one exaflop of processing power, equivalent to that in 10 million laptops. This week Google also announced MUM, a transformer-based search model capable of generating responses to complex queries such as, “How should I prepare to hike Mt. Fuji in February?”. Google claims MUM is 1,000x more powerful than BERT, the last generation transformer-based search model.

ARK believes that base editing- the chemical change of one DNA base pair to another – is a powerful tool in the gene editing toolbox.

This year at the J.P. Morgan Healthcare Conference, Verve Therapeutics, a private cardiovascular gene editing company, presented interesting pre-clinical data on its use of Beam Therapeutics’ base editing technology. In healthy non-human primates, Verve’s scientists discovered that inactivating the PCSK9 gene lowered the level of low-density lipoprotein (LDL), or bad cholesterol, reducing the probability of heart attacks.

Intravenous base editing could inactivate the PCSK9 gene permanently by changing a letter in the “spelling” on one DNA base from adenine (A) to guanine (G). This week, Verve’s scientists published updated data in Nature showing that one treatment of in vivo base editing inactivated the PCSK9 gene in non-human primates effectively, durably, and safely: it reduced LDL by 60% for at least eight months post infusion without any off-target effects.

This updated data suggests that base editing could correct other “misspellings,” or point mutations, in base pairs of DNA, potentially addressing large unmet needs and curing a wide range of diseases.

Previously, we demonstrated a use case of how Bitcoin mining might encourage solar + battery systems to scale economically and provide an increasing share of grid energy. In that exercise, we used one year of data to illustrate that a grid ecosystem including bitcoin mining – powered by solar + battery systems – will generate an increasing percentage of electricity from renewable carbon-free sources over time.

Updating our model use case with various bitcoin prices and hash rate timeframes, we have reached the same conclusion. We show that Bitcoin mining can incentivize additional solar and battery installations whether bitcoin is in a bull market or a bear market. Interestingly, during the setback in bitcoin’s price this week, Talen Energy announced plans to raise $800 million and build two bitcoin data centers, including one to mine cryptocurrency, for this purpose.

In the next iteration of our model use case, we will add another dimension – the household – to this ecosystem. Please find a link to our open-source model here.

In a recent interview, Klaus Zellmer, Board Member responsible for Sales, Marketing and After-Sales at its passenger car division, said that if Volkswagen launches an autonomous service, it could charge vehicle owners 7 euros per hour and hit profitability. If the cars on its platform were to average 30 miles per hour, vehicle owners would pay 29 cents per mile for the autonomous service in addition to the 70 cents per mile to drive personally owned gas-powered cars today, the two costs summing to roughly $1 per mile. While less expensive than the average taxi today, according to ARK’s estimates robotaxis at scale will be priced profitably at only 25 cents per mile.

Charging vehicle owners per hour is a telling go-to-market strategy for VW’s autonomous taxi platform. ARK believes that personal car ownership will dwindle, if not collapse, as autonomous taxi networks proliferate and prices drop to only 25 cents per mile. Rarely do taxi rides last an hour or more, suggesting that VW is likely to transition to a per mile price if its autonomous strategy meets with success.

While it has had some false starts, Volkswagen now seems to be making strides in transitioning from a legacy auto manufacturer to a software enabled, autonomous electric future. ARK will continue to monitor VW’s progress.