Why We Are Not Heading Into An AI Winter
This is republished from my newsletter at inside.com/ai.
For the past few years as AI has risen in stature, we have witnessed regular takedowns of the tech and posts about why this is just another temporary boom preceding another AI winter. Filip Piekniewski wrote a great piece back in May arguing that an AI winter was well on it’s way. Other takes on the subject include an analysis by Ron Schmelzer and a very recent piece in Popular Science. But I don’t buy it.
The biggest problem with AI is that it’s different tech, and the world has changed a lot since previous AI winters. But soooo many people in tech constantly use the frameworks of the last technology wave to evaluate the current one. Several of my own experiences have reinforced this. I remember keying in to the topic when Josh Kopelman from First Round Capital hosted a small dinner in Boston many years ago. Josh sold Half.com to Ebay so was known as an ecommerce guy. But that was 1999, and this dinner was probably 2011 or 2012. Josh said ecommerce entrepreneurs come to him because of his experience but, then he noted that the three biggest customer acquisition channels for most of them: Social, Mobile, SEO, were not even around when he built Half.com. He wasn’t sure his personal experiences were relevant to modern ecommerce companies. (side note — all the best investors I’ve seen have this kind of intellectual humility and self awareness).
My last company was a SaaS company started in December 2008. Many of our early investors did not understand SaaS metrics. I constantly had to fight the battle of being judged by the metrics of the packaged software industry rather than as a recurring revenue business. While the laggards came around, it was a lot of education on my part, and from the more progressive investors on my Board, to get everyone comfortable with MRR, CAC, and LTV.
AI is where SaaS was in 2009. Most people don’t understand that it requires different frameworks to evaluate it. So when I read about an AI winter, I’m very skeptical that people are looking at the right signals.
Consider this — most of the AI news that has driven hype cycles over the past 3 years is just research. Very little of it has been news about really great AI applied in products. That research news is stalling not because AI is stalling, but because AI is working it’s way from the research labs into more applications. The problems to solve there are real, and aren’t very sexy to write about. I see this at Talla. Our team recently worked to put a Bidaf Machine Comprehension model into the product. While it’s a model that can perform really well in isolation on a synthetic data set, getting it to work in the real world was messy. We had to solve a bunch of very pragmatic implementation problems, and some conceptual problems (the model always returns an answer, and we had to train it not to return something if the answer isn’t good or relevant). They weren’t sexy or newsworthy, but we believe we pushed practical implementations of Bidaf forward pretty far.
Enterprise adoption of AI is still in the very early days. Again, I see this first hand in Talla’s sales pipeline. Two years ago, no one had budget for an AI project. One year ago, a few really forward thinking companies had budget for prototypes and paid pilots. Now, those more progressive companies have budgets and projects and requirements all around AI. It’s still small but, the difference is noticeable. It will be 5–7 years before all enterprises are educated enough on AI to start buying and implementing it effectively.
Previous AI winters have happened because things never really made it out of the research lab into real adoption, outside of very targeted use cases. Now, while we definitely aren’t seeing the amazing we-are-close-to-GAI types of applications that the research sometimes suggests should be here, we are seeing real applications that have real value in the world.
The framework I use for thinking about it is actually a consumer adoption framework that I first learned about from BIjan Sabet of Spark Capital, when he spoke at a conference. He said too many investors who don’t invest in consumer companies regularly freak out when growth of new users stalls, but that, it happens all the time. Consumer companies go through waves of viral adoption vs deeper user engagement, and so, he said that the right way to think about it is as this oscillating wave. You get a bunch of new users, and then user growth slows, so you focus on getting the existing user base more engaged. This eventually leads to a new viral growth trend, etc.
I think a similar framework will drive AI adoption for the next decade. Companies will deploy the latest and greatest out of research labs, but it will be slow. They have to find (and cleanse) data sets that work. They have to solve pragmatic problems like whether inference should happen on the edge, or centralized in the cloud (which has both performance and privacy issues), how to shard large models in production systems, how to effectively QA models that have probabilistic outputs, and much more. These things take time. So I think what you will see in AI going forward is an oscillation between smarter algorithms that do cooler things, then what seems like stalled innovation but is really just the practical time period of everyone being more concerned with applying the new algorithms rather than inventing new ones. Once the new stuff gets deployed into production systems effectively, we will shift back to focusing on algorithm performance problems (rather than those due to implementation challenges) and need better algorithmic approaches.
I don’t think we are headed into another AI winter. I think we are just in a different part of the oscillation wave — laying the groundwork for the next research boom beyond DNNs.
On top of that, you have massive innovation in neuromorphic chips and tool sets to make deep learning more easily applied by non-experts. Both of these will take a few years to really take hold, but they will break the frameworks people currently use to think about what products and services can be built with AI. I’m particularly excited about the innovation that will come from breaking the x86 chip architecture framework we’ve relied on for so long.
In summary, I think this talk of an AI winter is wrong. It’s focused on a slow down in the research, but that’s a combination of over-hyped research news along with the slow arduous work of really getting these systems running in enterprises. Don’t underinvest in AI because of this AI winter scare. Now is the time to push forward and lay the groundwork for the next waves.