#323: Larger AI Models Could Transform Graphic Design, & More

OpenAI’s DALLE-2 model has enthralled the world with surprisingly creative images based on text prompts. When prompted to create an image of “Albert Einstein wearing VR goggles in a galaxy in the style of an oil painting by Vincent van Gogh,” for example, the model produced the images below.

ARK’s recent Twitter survey suggests that a human graphic designer would take more than 5.25 hours to recreate those images at a cost of ~$150 based on an average hourly wage of ~$29.^[1] DALLE-2 hasn’t announced its commercial pricing yet but, given 22 seconds of compute on an A100, we estimate that the inference cost for a single image could be ~$0.01––more than 99.99% lower than human labor cost. As a result, we have been debating whether or not powerful AI image models might displace human graphic designers.

In our view, tools like DALLE-2 should democratize access to creative content, enabling anyone to generate rich content––including still images, music, and video––instantly at di minimis cost. While DALLE-2 is likely to disrupt businesses offering stock-images, we believe that graphic designers will evolve into a crucial interface between their clients and this technology. They will learn the nuances among AI models and apply creative filters to generate images “in the style of van Gogh” or “in the style of Picasso”.

DALLE-2 has leapfrogged the original DALLE-1, as shown in the two images below. In our view, the breakthroughs are just getting started.

With 3.5 billion parameters, more than an order of magnitude lower than GPT-3’s 175 billion, DALLE-2 produced the image on the right, suggesting that scaling the size of the model could increase DALLE-2’s performance. Based on ARK’s research, AI training costs should decline 60% per year, suggesting that at today’s cost researchers should be able to train models 100x larger within the next five years. As ARK’s Director of Research, Brett Winton, has tweeted, scaling a model from 350 million to 20 billion parameters has re-created language, as shown below. If DALLE-2 were to scale 100x, we wonder what other tasks it might be able to accomplish.

[1] As based on the US Bureau of Labor Statistics dataset.

According to the EPA, “Methane is more than 25 times as potent as carbon dioxide at trapping heat in the atmosphere.” As a result, oil producers have been flaring natural gas to convert it into carbon dioxide and reduce methane emissions, a process that works but is only 92% efficient.

In 2021, global energy producers flared more than 140 billion cubic meters (140bcm) of natural gas, an amount roughly equivalent to the natural gas demand of Central and South America. They also vented 125bcm, releasing methane directly into the atmosphere, as shown below.

Methane emissions have proliferated because producers have little incentive to reduce them, something bitcoin could change. Bitcoin mining could provide energy producers with the incentive to self-regulate with an alternative to methane venting. Innovative companies like Upstream Data are using natural gas generators to power bitcoin mining hardware that combusts methane far more efficiently than flaring operations.

ARK’s current research suggests that the levelized cost of electricity generated by natural gas that otherwise would be flared or vented is lower than the ~$0.03 that many publicly traded bitcoin mining companies pay for electricity today. In a plausible scenario, not only could bitcoin mining lower emissions but it could price pure play¹ bitcoin mining operations out of the market.

[1] By “pure” we mean Bitcoin mining as the sole revenue generation mechanism for the business.

Although it recently reported numbers in line with consensus estimates, Micron’s revenue guidance for its fourth quarter dropped 17-26% from $9.14 billion to $6.8-7.6 billion. In a notable revision to its “flattish” prediction, management now anticipates a 10% decline in PC units this calendar year for a number of reasons: cyclical weakness against the fastest quarter of PC unit growth in two decades, rising inflation, geopolitical tensions, and lockdowns in China.

The resale price of Nvidia’s Graphics Processing Unit (GPU) has been warning that PC demand would weaken for more than a year. Since early 2021, the price of its RTX 3070 GPU has dropped more than 65% to the manufacturer’s suggested resale price (MSRP) for the first time since its launch, as shown below.

As ARK previously noted, a decline in the demand for GPU-based crypto mining equipment has impacted pricing based not only on a decline in crypto prices but also on Ethereum’s likely migration this fall to Proof-of-Stake mining which will eliminate the demand for GPU mining equipment. In our view, inventory has been building throughout the ecosystem. Intel reportedly is cutting prices on worse-than-expected inventory levels, while Micron’s cloud customers also have elevated inventory on hand. The slowdown in consumer spending is likely to spill into the enterprise and data center segments as well.

On the sidelines of its summit discussions about Ukraine last week, the North Atlantic Treaty Organization (NATO) announced a €1 billion venture capital fund financed by 22 European member states. The “NATO Innovation Fund” will invest in startups and venture capital funds focused on dual-use¹ technologies in areas such as artificial intelligence, autonomy, energy, and space.

The NATO fund would be well-advised to take a cue from In-Q-Tel, a venture capital fund supported by the CIA and other US intelligence agencies, which has operated successfully as a non-profit since the 1990s. The European startup and innovation ecosystem should welcome the NATO fund, as it could narrow the venture capital and technology gap between Europe and the United States. Today, European investors hold only 2% of the shares in the world’s fifteen largest private companies and account for less than one-third of the investment in Europe’s leading late-stage private companies. In our view, Europe could benefit from more technology-oriented venture capital funds like the NATO Innovation Fund.

[1] “Dual-use” refers to technologies that can be used for both civilian and military purposes.

Considering the market’s fast pace of change, last month ARK introduced The Bitcoin Monthly, an “earnings report” that details relevant on-chain activity and showcases the openness, transparency, and accessibility of blockchain data.

Please subscribe to download The Bitcoin Monthly June Report and see below for the topics we address.

Topics addressed in The Bitcoin Monthly June Report include:

• Bitcoin closed the month of June down nearly 40%, declining from $31,780 to $19,980.
• Contagion in the crypto markets continued as Celsius and Three Arrows Capital came undone.
• Market contagion sends bitcoin into capitulation.
• Despite record-high realized losses, unrealized losses could allow further downside.
• Uncertain macro conditions are weighing on bitcoin’s recovery and outlook.