2023 has been an exemplary year for AI.
Not necessarily just in its mass adoption (chatGPT, being the quickest, growing software product ever with 100M users in months), but in its demonstration of potential.
Having studied AI at university back in 2016 in Scotland, I remember limited AI jobs were available. The space often suffered from what’s called AI “winters” — the technology kind of worked, but didn’t fully to impress or convince masses of people.
However, the 2023 meteoric rise of AI has come at the expense of the AI spaces’ long-term profitability.
Incredibly expensive tools like ChatGPT or GitHub copilot have been offered virtually free or incredibly subsidized to consumers to the point that these are now made every day expectations that any company should offer for pretty much free.
LLMs have been incredibly commoditized in the span of a year.
It’s likely this will continue to be a money burning sector. With every GPU overheating being a symbol of money being burnt.
But in terms of making AI a household name, it has been incredibly effective and an incredible example of blitzscaling.
As a consequence a lot of the real money in AI likely won’t be made through models and software alone.
The most advantageous sectors will be those in the intersection of software and the physical. this includes of course, the obvious chip companies (Nvidia) but also a lot of legacy, physical business with productivity blockers that would otherwise deliver value, faster, cheaper, and better.
Software companies have been incredibly investable in the last 20 years because of the ease of creating quick cheap value that’s monetizable. This means it’s a low risk ROI investment.
I think AI will substantially make manufacturing and legacy physical sectors incredibly cheaper and more productive to lessen (but not eliminate) the gap with software. Not necessarily through LLM’s but through optimization.
Automation of warehouses, customer service, stock prediction, prediction of customer trends, aggregation of customer feedback, automatic modeling and testing of products before they’re produced. Even to a large extent most manufacturing machines to date have been rule-based (following a specific recipe) — What will the first truly intelligent manufacturing machine look like to automatically iterate on prototypes?
2024 will be an exciting year, it will likely be more boring in LLM terms (don’t quote me on this), but boring is cool. Boring is useful.
It will likely be the year in which LLM’s are actually embedded into the software services we use on a daily basis to improve them, instead of just be a flashy demonstration.
The intersection of AI and physical industries will likely take longer, but appears very promising within the next 5 to 10 years and much more defensible.
