Sussing out AI and Vertical Platforms
Hey folks,
You are now entering an AI newsletter.
Alright, everyone is talking about it so I may as well too. The developments in LLMs are pretty astonishing and broadly it seems like the AI winter throughout the 2010s is now over. Spring has come and it has brought us models that can draw requests like “a futuristic Tadao Ando restaurant in art deco style” and translate it into an image.
At a very high level, AI research is on a path to creating all sorts of models that translate human language into meaningful actions that models can perform.
And given the caliber of talent going into the space, the number of startups working on the meta-problems of AI, as well as venture capital flowing into the space, it feels safe to assume that the models get really really good. Like really, really good.
Like good enough that at some point down the road, inputting the exact same prompt, “a futuristic Tadao Ando restaurant in art deco style,” won’t only output an image, but a corresponding blueprint to build.1
If you assume that researchers continue to make AI models better and better with far more intelligence (or whatever you want to call it), there are two questions everyone else should be asking:
What sorts of models should we train?
Where should the models operate?
These two questions are squarely within the domain of vertical SaaS. And so what I want to make the case for today is that in many instances, vertical SaaS will become the domain in which the hope of performant AI for many types of use cases will be fully realized.
When thinking about how vertical platforms will interact with and utilize AI, as far as I can tell there are three buckets:
Vertical platforms are used as sources of truth for model finetuning to create great products
The greatest advances in applied AI (vertical AI) get distributed through vertical platforms.
Vertical platforms function as endpoints for AI assistants to take digital actions with high degrees of fidelity to the world.
Sources of Truth and Fine-tuning
One of the weird things about AI is that while it somewhat mirrors the horizontal vs. vertical SaaS debate, in other ways it defies the categorization altogether.
Generally, there have been two approaches: make super tailored AI and ML models that serve specific workflows, or aim at approximating a more generalist intelligence. The current large language models (LLMs) are definitely in the latter camp, exhibiting a wide breadth of knowledge given the obscene amounts of training data they were trained upon.
And because these LLMs have huge breadth, in some sense they feel far more akin to infrastructure that will be used by many AI projects, both horizontal and vertical. It makes sense to utilize these huge models as the base model for specific use cases built on top.
If you posit that the premier AI companies like OpenAI continue to develop larger and larger models, it probably makes sense to believe that creating new models wholesale is somewhat redundant. You probably will not create a model with more raw power than an OpenAI model. However, it is still necessary to fine-tune2 these models for specific domains.
And that’s where vertical SaaS’s value for a world full of AI lies.
In order to fine-tune models, companies have to figure out how to accumulate useful training data in a cost effective way to ensure that enterprise AI generation is accurate, compelling, and can be relied upon in enterprise contexts.
And so we get to the first area in which vertical platforms become incredibly useful for AI fine-tuning:
Vertical platforms act as the sources of truth for industry specific datasets that will be used to fine-tune AI models.
In some sense, this is a restatement of what vertical platforms aspire to do. After all, the best vertical platform become mission-critical to businesses, bring industry-specific workflows online, and thus accrue value as the sources of truth for the whole industry data sets.
And while many verticals don’t have a need for a pure “help me write a text document” tool, plenty more will.
Github CoPilot
To really get at how this plays out, we can take a look at one of the early examples.
Github is in some sense a “vertical platform” for engineers and developers and has accumulated huge swaths of industry data on platform: code.
And as the source of truth for code, Github had a really interesting path to creating an AI pair programming assistant.
The journey went something like this:
OpenAI created the Codex model to generally translate text into code. And then Github used the huge swaths of data on platform to fine-tune the model in order to generate relevant and reliable code on the fly. The end result is Copilot, a code writing assistant that is saving developers vast amounts of time writing code.
So we ended up with a progression:
Huge generalized model → finetuned model with industry data from a vertical platform→ useful product
It’s in this synthesis between AI companies and platforms that function as the source of truth that I believe the vast majority of AI enterprise value will be found.
In fact, within the past month, Replit, a browser based coding environment, has released their own pair programming AI, Ghostwriter, fully embedded within their IDE.
Ghostwriter’s development is interesting as they didn’t only rely upon one AI model, but 3 different ones. Each has different merits and Replit acts as a tastemaker to finetune the models for different tasks and give users a fantastic product.
But crucial to Replit’s model is that they aren’t simply creating killer AI products, they’re fundamentally reinventing the development environment.
In fact, this is the exact bet that plenty more companies are making like Filevine.
Imagine a legal editor or document creation platform for lawyers that has a built in Copilot. Not only would you want it to have the ability to reference case law and generate standard clauses, you would want it to have a sense of the particulars of a case, translate the facts into a legal argument, and reference/cite relevant materials. In other words, the best legal AI won’t simply handle public information but be able to reference private information relevant to the legal document.
That data is contained within a relevant platform like Filevine (detailed in the last piece here).
And in fact, training upon this case data and documents from all sorts of law firms will be the sort of fine-tuning that’s done in order to create a product that isn’t simply inputting case law or clauses, but one that has a grasp of things like a law firm’s stylistic conventions, translating facts from a police report into a narrative, and contractual clauses.
And these sorts of breakthroughs will come from fine-tuned models from vertical platform data sets.
Vertical AI Distribution
One of the reason vertical SaaS is such a powerful force in a world where AI advances incredibly quickly is that vertical platforms become key distribution partners for AI advances.
Even companies that have started as pure APIs will converge on accruing enterprise value through platform experiences. The proof is in the pudding as even the current wave of AI startups are doing this. Look no further than Jasper, a copywriting AI assistant, announcing their monster series A and revealing their new emphasis on becoming a platform with plenty of workflows built in.
As the meta-questions in AI development get figured out, and AI gets applied to far more domains, every AI company will need some sort of platform strategy.
When the AI layer gets relatively commodified, distribution becomes everything. And vertical SaaS platforms are incredible at industry distribution.
Runway
If I could pick any company that I think will become the biggest distributional partner for advances in image and video generation, it would be Runway.
It would be far too long a piece to dive deep into all that Runway has done, so let’s just focus on the most critical pieces. At the highest possible level, Runway is a cloud-based video editing tool.
Their big insight, like Figma, is that a lot of creative work is going to move to the cloud. And because work is moving to the cloud, new possibilities like instantaneous collaboration and machine-learning powered workflows are now possible.
And so Runway built out their platform with an eye on the coming ML/AI advances, implemented all sorts of video editing magic with cutting edge machine learning techniques, and has in the process accrued the love of the video editing community.
And so now as AI gets more powerful around video and image generation, Runway will be the industry distributional partner for all of it. Sure, an AI startup could choose to create some proprietary AI video generation platform, but the problem is that it just doesn’t make a lot of sense for an editor to manage traditional video editing flows in one platform and AI/ML powered ones in another. It’s far more intuitive to do it all in one platform. And because Runway has solved the whole spectrum of video editing workflows and built their platform to take advantage of ML/AI advances in the product, they end up being a far better partner for AI companies tackling issues in video/image generation than a competitor.
Runway unlockss distribution for AI generation and in turn accrues the benefit of continuing to own the central workflows of the industry.
And so it’s in this way that vertical platforms become the “workstation” and distribution channels for all sorts of AI products.
One more example will suffice.
Benchling
In 2020, DeepMind, one of the leaders in AI, solved an age old problem around predicting what a protein’s structure would look. Proteins like to fold in weird and non-obvious ways based upon their amino sequences and humans couldn’t crack it. If you are trying to create some new sort of protein, you can easily get stuck trying to figure out how this protein should look. R&D costs and time to market for innovative health products suffer.
DeepMind essentially solved this out of nowhere with AlphaFold, which can predict protein structures never seen before with insane levels of accuracy.
Now the question for a scientist hoping to use this tool is 1) how do I spin up the engineering resources necessary to leverage DeepMind effectively, and 2) how can I add this seamlessly to my workflows?
And that’s what Benchling, a vertical platform for biotech R&D is hoping to solve with their AlphaFold integration (currently in an extended beta).3
Alphafold gets distribution, scientists get incredibly powerful tools, and Benchling continues to build mission critical workflows for its customers.
Importantly, it drops the cost of using these powerful models and extends the power of compute to the entire industry.
Or another way to say it: it isn’t simply that vertical platforms solve the distributional challenge for AIs, it’s also that vertical platforms significantly decrease the time to value for using these tools.
And so I think you can expect this to play out in myriad industries. Anywhere where vertical AIs are being created in order to can advance efficiency or understanding, is probably a place where industry platforms will accrue the benefit.4
So AI models will be both fine-tuned and distributed by vertical platforms, but what about where it’s less clear how AI and vertical platforms intersect?
AI Navigators
Lastly, we get to a world where the very definition of human-computer interactions get rewritten.
Nobody loves data entry and luckily the advances in LLMs have led to perhaps a new paradigm.
Some of the buzz right now is around a world where text operates as the universal interface for digital workflows.
There’s one company that is probably emblematic of these advances: Adept.AI. Adept is building AI that is capable of taking digital actions. In the pic below, you simply type a criteria for a home you wish to buy, and then Adept’s model is able to navigate Redfin to find the homes.
They’ve announced all sorts of software that they are looking to tackle; everything fro Craigslist to Salesforce.
In this next phase, I think AI assistants will look primarily like navigators with great abilities on pre-defined routes. They will be able to do data entry and perform workflows, but their scope is still limited to treading pre-defined paths. They still need endpoints to navigate and these endpoints need to have real impact and value. In fact, digital actions that correspond with a high degree of fidelity to the real world are perhaps the most valuable actions to be taken.
And since, vertical SaaS platforms are quite literally built around digitizing workflows with high fidelity to vertical specific business value, it becomes somewhat natural for vertical platforms to benefit from AI assistants that are able to perform digital actions where the platforms are focused on insights, workflows, and actions that matter.5
This is the best of both worlds for this next phase. Data entry gets transformed, workflows get digitized, and AI assistants begin to materially impact and transform businesses in conjunction with their vertical platform partners.
There is another important reason as well: One of the built in advantages that vertical platforms have is that they concentrate the workflows of a business. Vertical platforms are workflow dense and what this means is that an AI assistant in theory can perform more total workflows in a shorter period of time and most likely in a way that is less expensive computationally than performing the same workflows across multiple horizontal apps. This means that it structurally may be better for a company to utilize a vertical platform for cost and time savings.
Oh and it means that implementation costs for vertical platforms probably go way down. Simply give them a list of prompts like “Create a dispatch for tomorrow at noon” and let the combination of AI + vertical platforms yield the value.
This combination of high-fidelity platforms + workflow density and reduced implementation costs means that if AI assistants become the next computing phase, it’s highly likely that vertical software becomes more important not less.
Next Up
Next week, I want to take a look at how vertical SaaS companies are faring in the public markets. Frankly, they’ve done pretty well over the past quarter relative to the broader market and I think it’s worth examining why.
No architects aren’t being replaced. But if the AI path goes well, then realistically architects will have some sort of prompt engineering function around mapping AI generated blueprints and concept art to regulatory confines, client needs, and construction feasibility. The creative element of architects gets magnified and amplified by the use of an Architecture Model. But more in the rest of this piece.
In other words, supply training data for specific use cases.
https://www.benchling.com/blog/benchling-launches-alphafold-beta-feature
Some sectors where I am mulling over the analogy: precision agriculture, robotics, military defense.
A less confident prediction: It probably becomes a waste of compute for an AI to navigate multiple point solutions when it could simply take a series of actions within one platform. vSaaS platforms become highly specified backends that allow these digital agents to navigate the web efficiently.