Building an AI Fortress in the Age of GPT-4
Does the acceleration in AI make your SaaS company more or less defensible?
Last week, I looked into the question if AI is a platform shift as disruptive as the move from on-premise to the Cloud. My conclusion was that Generative AI is an asteroid impact event for B2B software, comparable to the launch of the iPhone or the Internet itself. But unlike most on-premise software companies in the 2000s, the SaaS leaders of today aren’t doomed. If they can adapt fast and embrace the shift towards much smarter B2B software, they are in a good position to survive and thrive.
A related question is whether the arrival of Generative AI, and the fact that it’s become 10x easier to build AI-powered software, is a net benefit for AI-first SaaS startups. If Google (or at least some people at Google) think, “We have no moat” (referring to open source competition to its LLMs), how can a startup use AI to become more defensible?
A few years ago, Louis wrote an excellent article titled Routes to Defensibility for your AI Startup. If you haven’t read it, stop reading this post and head over to Louis’ piece. :-) What I’d like to do is revisit the question in light of the latest advancements in AI, specifically GPT-4.
1) Not every AI feature will enlarge your moat (and it doesn’t have to)
Given how easy it’s become to add AI features like video transcription, text summarization, sentiment analysis, writing assistance, image generation, or data analysis, lots of companies will use general models like GPT-4 and a limited amount of fine-tuning to create real customer value. In these cases, adding AI will improve your product and make it stickier, but you’re not creating a deep moat, and that’s fine.
In these cases, companies will use LLMs just like other building blocks like open-source databases, file storage providers, and all kinds of open-source components or APIs. Using a MySQL database is not going to increase your defensibility, and neither is plugging in an AI API for, say, text summarization.
You should, of course, aim to build a product that is extremely hard to copy. But many AI features will be bricks in your fortress, not something that greatly expands your moat, and again, that’s OK.
Here’s how Morten Primdahl, co-founder & former CTO of Zendesk, put it in a recent conversation:
“For someone like me who prefers to build business logic, using GPT is no more of a moat than using MySQL is. I believe we’re witnessing the productisation of what we used to rely on lots and lots of hard earned ML efforts for, those teams will now n̶e̶e̶d̶ ̶t̶o̶ ̶r̶e̶i̶n̶v̶e̶n̶t̶ ̶t̶h̶e̶m̶s̶e̶l̶v̶e̶s be able to focus more on business specific needs.”
2) Creating deep moat from AI requires proprietary data
To create a substantial AI moat, you have to be in a situation where (a) you have access to proprietary data that is impossible to obtain for others and (b) that data leads to superior model performance. This is conventional wisdom, so I won’t elaborate much on this point. The reason why I’m mentioning it is that the bar for what is truly proprietary data has increased. GPT-4 has already been trained on an insanely large corpus of text documents, and OpenAI and its competitors will train future models on every potentially useful piece of text, video, or audio they can get hold of (plus potentially even larger amounts of synthetic data). This may include corpora of data that one would typically consider proprietary, such as a law firm’s legal agreements archive or decades’ worth of a hospital’s patient data. I’m not suggesting that e.g. OpenAI will do anything illegal. My thinking is that there’s a vast amount of data out there which doesn’t lose its usefulness for training if you remove personal information from it.
3) Early movers in AI can have first-mover advantages or disadvantages — it depends
In some sectors, companies like Intenseye or SuperAnnotate from the P9 portfolio, which have started building AI software for many years and have dozens of excellent ML engineers on their teams, have an enormous advantage over new entrants that moved into AI more recently. Intenseye’s workplace safety solution, for example, leverages AI to analyze real-time video feeds and predict potential safety hazards. They use proprietary data, algorithms, and models that have been refined over years to get to a point where they can detect anomalies with an extremely impressive degree of accuracy.
Both Intenseye and SuperAnnotate were able to accelerate their roadmap using the latest AI advancements e.g. in image segmentation. But without everything they’ve built over the last years, including sophisticated enterprise workflow integrations, data collection and annotation processes, comprehensive models, and a deep understanding of the problem domain, they wouldn’t have been able to harness the power of new technologies as effectively.
In other areas, earlier AI movers may see the value of their tech erode. Imagine you’ve spent the last years building an AI for audio transcription, text summarization, translation, sentiment analysis, intent recognition, or any other tasks GPT-4 is so good at. In these cases, the introduction of AI services like the GPT-4 API could mean that years of work and specialized expertise become less valuable — turning what might have been a first-mover advantage into a first-mover disadvantage.
I’ve just listened to this great episode of the Future of Life Institute podcast, where Nathan Alan Labenz, the founder of Waymark, an AI video creator, described his experience when he got access to GPT-4:
“You can see me go on this journey of just having my head totally explode over the course of a couple of hours, where it’s like, OK, this thing that I’ve just spent the last year on, curating data, fine-tuning, finding the edge cases, patching the data, to get something workable, it can do it out of the box.”
4) The mid-term future
Looking 5–10 years ahead, there’s an active debate (in the industry and inside P9 😄) on how good foundational models are going to get at highly specific use cases, which up to now are served by a combination of particular workflow software and specialized ML models that have been trained for the specific tasks.
On the one hand, you can argue that a model that’s been optimized for a specific task will always be better than a general model, similarly, perhaps, like for any given job, a human expert who specializes in this task will beat any generalist. On the other hand, high-IQ humans can get pretty good at a wide range of tasks with comparably little “training data”. For example, it may take an AI millions of examples to accurately extract and process all of the data from an invoice. A human with average human intelligence can arguably accomplish this after a few days or weeks of training and many orders of magnitude fewer examples.
So the question is: If, until now, there’s been a trade-off between better general performance and better performance in highly specific domains, at what point will general AI be so good that it can easily acquire new skills in specific domains on the back of its general cognitive capabilities?
LLMs work very differently from human brains, so trying to predict advances in AI by drawing parallels with human learning and cognition comes at the risk of over-anthropomorphizing. Nevertheless, looking at the progress of the last few years and what people in AI are currently working on, I have little doubt that the level of general AI will continue to increase at a rapid pace. The startups best positioned to succeed in this fast-changing landscape are, I think, those who have the deepest industry expertise and are on top of the latest developments in AI, constantly trying to deliver more value to customers.
