This podcast was inspired by a week of naysayers piling onto a recent issue of the Goldman Sachs Top of Mind newsletter titled “Gen AI, too much spend, too little benefit?”
In that same week, John Hwang published an article in his Substack Enterprise AI Trends called “Why AI Infrastructure Startups Are Insanely Hard to Build.” In that article, he advises founders working on AI infrastructure that they should sell to an industry leader before they lose everything.
Roman Shaposhnik, our CTO & Co-founder of AIFoundry.org and Nekko.ai, called Bryan Cantrill, CTO & Co-founder of Oxide Computing Corporation, to reply to these questions and comments. In this fun bantering session between two old friends, they explore building startups on “hard” mode, saying that creating something transformative for the industry requires being in the fight for more than just the money.
Roman and Byran start the conversation by observing that with AI, we’re in another rapid innovation cycle around a new problem domain. While end-users would like to see a simple solution and investors would like to see immediate rewards from providing these solutions, certain widely adopted foundational bits are still being developed, making it too early to consolidate this industry. This is just the ongoing pattern in the world of technology, and it's only problematic when people, such as investors, have the wrong expectations.
Roman points out that part of what has to be figured out in AI is what’s a product and what’s a feature in the eventual solutions. He brings up vector databases as an example. Are they a new class of database, or are they just another supported data format? He suggests they’re already being consolidated as a feature, as NoSQL has been added to databases.
The conversation then turns to exploring the idea that people are expecting too much out of the current capabilities of AI because we tried to anthropomorphize it. It writes & talks like a person, it must be like a person. When in reality, we should consider AI at its current state to be missing key features, and these missing capabilities are preventing even wider adoption. As an example, we start discussing the Honeywell Kitchen Computer, sold by Neiman Marcus. No models were ever sold, despite present-day evidence that using computers to help find recipes for cooking is in fact a legitimate use case, since many of us are now using our mobile devices to search and read cooking recipes and cooking techniques.
Bryan then waxes into the existential technical difficulties that foundational infrastructure startups face. He then explains that the human reasoning it took to get past these difficulties shows that AI is incapable of helping in these cases, and thus is probably not ready to take over the world. Bryan covers some “terrifying existential technical problems” the Oxide Computing team faced in a talk he gave at Monktoberfest 2023 called “Intelligence is Not Enough."
I highly recommend you listen to the conversation yourself, as I cannot do the topic justice with this summary.
If you’d like to catch our future podcasts live, please subscribe to the AIFoundry.org calendar below to get a reminder, and join us in the AIFoundry.org Discord Server.