Things I'm Thinking About: Generative AI
Some (quietly distributed) thoughts on "the current thing"
Hey everyone — Switching up the pace here a little bit. While far from being an expert in this field (and most fields, for that matter), I’ve written +1,000 words on generative AI and how it might alter the technology landscape.
Generative AI is a subset of machine learning that leverages enormous datasets to produce new, computer-generated content like text, images, and video. In recent months, tech-savvy consumers have grown familiar with a new suite of products born out of this technology, including Large Language Models (LLMs) like ChatGPT, Text-to-Photo models like Midjourney, and Text-to-Video models like Runway. In all cases, it’s worth highlighting that the mechanics of generative AI resemble that of a matching process, rather than a critical thinking process. Said another way, Generative AI models deduct results not on account of their own intelligence but through a probabilistic matching function. Asking ChatGPT what the color of the sky is will produce an output reading “Blue”, but only because the mountains of data in which the model is trained on tends to produce that very fact. This results in a familiar dynamic within the tech sector: Generative models are only as strong as the data it's trained on. Consequently, the most resourceful technology companies (Apple AAPL 0.00, Microsoft MSFT 0.00, Amazon AMZN 0.00, Alphabet GOOG 0.00, Meta META 0.00) are likely the best equipped to capitalize on this trend, given their access to massive data sets and cheaper compute costs. The quasi-marriage of OpenAI and Microsoft only serves to bolster this competitive advantage, particularly as the ChatGPT product finds soft market-fit in the Search and Messaging verticals. Nevertheless, smaller disruptors have also emerged onto the scene as the broader tech ecosystem licks their wounds. Profitability is paramount for today’s public market investors. Broadly speaking, Big Tech’s renewed focus on cost discipline and efficiency has created a vacuum for new startups to tackle this market. Outside of OpenAI, prominent operators in this market include Stability AI, Midjourney, Runway, Jasper, and many more.
Perhaps unsurprising to many, the frenzy surrounding generative AI has provoked a mix of reactions across all varieties of the economic spectrum. Creative types — particularly artists, writers & musicians — have admonished the technology, highlighting that generative content (specifically art) is simply piecemealed from countless uncredited sources, infringing on intellectual property without proper compensation or acknowledgement. Others warn of the consequence that generative AI will exacerbate misinformation and spam on the internet, as it takes mere seconds for a generative model to produce content ready for mass distribution. These concerns are worth heeding – the marginal cost of generating content has shrunk to nearly zero. A deluge of AI generated content has the potential to crowd out authentic, human-generated content – and the probability of such an outcome is higher than most would like to admit. What would this look like, in practice? Generative models trained upon content produced by other generative models could render this technology ineffective, while also obliterating trust in the internet economy.
The Global Capitalist is free for readers. To support my work, please consider sharing or subscribing below
Technologists, on the other hand, are ecstatic about the onset of this technology. In their view, Generative AI will be integral towards boosting labor productivity, a figure that has languished in the aftermath of 2008. They insist that jobs will not be replaced by this technology. Rather, information (white collar) workers will employ this tech in their Day-to-Day responsibilities, saving time and costs as a result. Asking a generative AI model to publish 1,000 words on banking regulations, for example, will cut out the time sink and grunt-work of essay structure. The model should help address key terms and provide context for the essay, while the end-user will be responsible for tightening up the model’s work, adding specific details to the argument, providing deeper explanations of key facts, and adding a personal style that GPT models are inherently forced to borrow. Most writers, if not all, appreciate that it is often easier to edit than it is to write from scratch. The process behind this kind of workflow has become known as “Sandwich Work” as generative models provide the buns of a product, while the end-user adds the meat to the piece. It wouldn’t be difficult to picture similar processes taking place across the Film, Art, and music industries as the cost of generating new content moves closer and closer to zero.
As highlighted earlier, Big Tech has been the first industry to bring this product to market, albeit with mixed responses. Microsoft’s partnership with OpenAI felt like a veiled declaration of War against Google’s Search Monopoly, with CEO Nadella taking implicit jabs at Alphabet’s profit margins. Meanwhile, Google’s rollout of Bard, their own language model, was met with disappointment as their highly-anticipated Demo Day featured an incorrect response on a question regarding the Webb Telescope. Google fumbled the optics of their Go-To Market efforts, leading some to describe ChatGPT and competing LLMs as “Google Killers”. That being known, it’s not entirely clear to me that generative models, at least in their current state, are appropriate substitutes for search platforms. The workflow is clunkier than traditional search functions and there is no guarantee of valid results, despite the confidently presented answer. Thus far, the data reinforces this view, as most Search users have yet to migrate over to the new GPT-infused Bing, although that has yet to deter Google from rolling their language models out to market.
On the other hand, I see enormous market potential in the creative and personable elements of Large Language Models. I see iterative improvements over existing AI-assistants like Alexa (Amazon), Siri (Apple), Cortana (Microsoft), and Watson (IBM IBM 0.00) thanks to the conversational nature of large language models. In addition, I see generative AI anchoring the tool set for metaverse developers – particularly as games like Horizon Worlds, Roblox (RBLX 0.00), and Minecraft leverage bots or daemons to create authentic social experiences in digital spaces. Finally, the loneliness epidemic is real, and has worsened since the pandemic, with surveys painting dour stories about the social lives of American adults and their children. The presence of a personalized AI model could help alleviate this sense of loneliness, as demonstrated in the disturbing case with Replika.AI, while also helping people reduce anxiety in other social settings. I’m certainly less sanguine about the broader implications of this use case, although I believe this market serves as a potent addressable opportunity for tech companies, particularly in the communications and gaming spaces.
Finally, I am specifically excited about the generative AI applications that will be brought to market by digital advertisers like Meta and Google. Meta and Google can specifically leverage generative models for the benefit of their advertising customers, generating fresh ad campaigns while allowing the company to iterate and edit where they deem fit. In Meta’s case, the company is navigating the integration of Reels into Facebook and Instagram feeds as short form video grows more popular in consumer apps. While the Reels product has been successful and popular among users, it does not monetize as well as Feed or Stories. One reason being that Meta is trying to incentivize advertisers to buy up new ad inventory: Newly introduced advertising slots are value-priced to accelerate adoption amongst existing customers. Then again, the novelty of the Reels product means that advertising teams are 1) Not as familiar with the Return-on-Ad-Spend (RoAS) from this product and 2) Not equipped with the tools to create enticing short-form video. This is clearly a focus area from Meta’s management, as video now comprises roughly half of a user's time spent on Instagram. Looking forward, I think Meta will begin to generate short form video ads and other pieces of content for their customers as they tighten their grip on the digital ads market.
Thanks for reading! We will revert to emerging markets in the coming weeks with a feature on evolving power dynamics across Europe and the Middle East.