It is impossible to explain how much artificial intelligence has changed the world in a very short time. Generative AI has become a vital part of my toolkit, to the point where ChatGPT is now teaching me design well enough that I can analyze interfaces and explain why they work.
The technology also has serious problems, well-documented by scholars. Hallucinations are a constant concern with the technology itself. Environmental concerns are paramount (as they should be) since generative AIs can consume truly massive amounts of power. And its adoption has forced education (and even job interviews) to change much faster than they are likely capable of changing.
But still, the technology has been amazing and this report outlines some statistics on how quickly generative AI has spread throughout the world. Consider some points from the report:
- Improved algorithms improved model performance by 200% per year for a decade.
- ChatGPT reached 365 billion annual searches 5.5 times faster than Google.
- The patent market exploded between 2023 and 2024 just like it exploded after the Netscape IPO.
- Generative AIs handily beat humans now in the MMLU (Massive Multitask Language Understanding) benchmark.
- Nvidia’s developer ecosystem grew 6x in seven years.
- Google Gemini’s ecosystem grew 5x YOY, hitting 7 million developers.
This Meeker report also picked out some long term trends. The most interesting ones are that LLMs are becoming infrastructure, just like the cloud or a database or something. And the race is evolving from the race to build the best model to the race to build the best product that is powered by a model.
In other words, it’s a builder’s market again. However, I am old and was in my late teens the last time that we experienced a builder’s market. And that meant that I was in my early twenties during the first dot com crash when that builder’s market came crashing down in a sea of irrational hype. Consequently, I want to add a little more context to the growth numbers stated.
It is objectively true that ChatGPT reached 365 billion annual searches in two years, compared to Google’s eleven. However, the world is vastly different now than it was in 2009 when Google hit that milestone. More affordable smartphones and widespread broadband connectivity have dramatically expanded the market, particularly in the developing world. Heck, Google’s growth to 365 billion searches took from 1998 - 2009. In 2000, global internet connectivity was at about 4% of the population. Today, we’re at roughly 68%.
Moreover the internet is a dramatically different community now than it was in 1998. Today we have influencers, gifted content creators and ubiquitous social media networks so products can launch to a userbase with an existing viewership.
This is in no means meant to diminish OpenAI’s contribution - ChatGPT’s growth to 365 billion searches a year is truly astonishing and it represents mindblowing technical, marketing and financial accomplishments. The last three years will be studied in MBA courses for the next thirty simply because of the wealth of information that could be used to write excellent cases. And OpenAI’s contributions will receive particular attention because they deserve to.
But if we are going to look at the past to help inform the present, let’s make sure that we assign hype correctly and quantify it appropriately. As a product, ChatGPT managed to catch lightning in a bottle - that is not only extremely rare but a sign of extraordinary talent. However, its growth was accelerated by the times.
With that context, I feel that we change the types of strategies we employ going forward.
First, it’s clear we’re in a builder’s market again. The last time we were in a builder’s market, there was too much hype and so valuations got too high. The hype didn’t materialize so the valuations came crashing down and made a giant mess of the economy for a number of years. If we assume that generative AI is even slightly overhyped, it’s doubly important to conserve money and grow quickly — but do so in a way that could sustain itself indefinitely without outside investment. If history repeats itself (only bigger and faster), the companies that survive the winter will be household names.
This is an area where generative AI can be extremely helpful. Whatever your AI of choice, you can offload a lot of tasks that you would have had to pay for just five years ago. So the proper use of generative AI can slow your burn rate dramatically.
Second, in a builder’s market process has to be much more important than product. Process important enough that in a builder’s market, it makes sense to build products instead of a product because building products is an excellent way to refine and standardize a process. Just as you can extract a framework from multiple projects, you can extract the knowledge to build a great process through simultaneously working on multiple products.
Again this is an area where generative AI can be extremely helpful, though it also requires excellent managerial skills to expand that process beyond a core group. Building a process is not just about building a process. It requires a lot of thought, technical leadership, testing, extreme levels of staff buy in and even more extreme levels of control in the hands of staff. Process has to be flexible enough that any employee can raise problems and have them addressed. But simultaneously, process must have enough oversight that those problems are only considered if they’re actual problems.
Third and most importantly, interfaces will change dramatically over the next 18 months. But this may necessitate changing how we deploy and commercialize software. As one example, you could currently solve a lot of CRUD problems with a chatbot that can control what gets written to or retrieved from a database. But no matter how powerful the model you use, its performance will ultimately be limited by how good the prompts are. And if a chatbot can write to or retrieve from a database, there is a point where a bad prompt can do toxic things to a database.
So you quickly end up down a path of multiple models, where an edge model can help refine prompts before they go off to a more powerful model to return answers or actions for the system to perform. Perhaps those need to be verified by an edge model too or perhaps they are fine to go as is.
Figuring that out though will require different takes on testing and on logging. It’s starting to look like desktop applications will make a lot of sense again and many companies will run models within their own environments again. I genuinely believe that we will see a return to in office server rooms and sysadmins with skillsets that look more like ML Ops…only LAN-based. This will largely be driven by the types of new interfaces that generative AIs will provide.
Concluding points
The latest Meeker report is exceptional; I’d consider it a must read. AI has dramatically increased the rate of change, and it looks like this growth will only accelerate in the future. And since we’re in a true builder’s market again, the opportunity is unprecedented. It is not without risk - because of generative AI, your competitors can implement your innovations quickly while they continue to innovate on their own. But with a strong focus on process and sound financial management, the companies that come out the other side will be poised to become some of the biggest and most successful product companies of all time.