"AI First" Startups are Just 2023 Crypto

As the dust from the 2022 cryptocrash is settling, tech investors are moving on. If you're in the loop, the latest craze is "AI First" companies. VCs will be excited since this seems like a differentiator that will give them an excuse to invest in more SaaS companies. I'm writing here to warn you if you're not in on the scam or preparing to be the first one out, you'll be stuck holding the bag when the bubble pops sometime in 2025.

To be clear, money in 2022 is tight. The feds are raising interest rates and I don't think this bubble will be as inflated (ie. detached from real utility) as 2022 crypto was. People are still going to put their money somewhere and plenty of VCs will be taken in by AI startup gurus claiming revolutionary transformation.

Red Flags

1. Any startup that doesn't outline what exact process AI speeds up. This isn't 2015, you can't just gesture wildly and say AI and expect people to take your product seriously. AI is an accelerator, not a viable product in and of itself.
2. "AI is like electricity". The speaker is throwing spaghetti at a wall and hoping something sticks. AI (and data) are not energy - they don't power anything, they make things go faster or expend less energy. If someone's trying to make the energy analogy, that's a pretty big red flag that they have no understanding of their product.
3. "AI native software". What problem does this solve? Speed is no longer an issue for React-era software. Process automation requires you to define a process. Unless you have a clearly defined AI use case, data pipelines, and product-market fit, there won't be a real product anywhere in there.

The only "AI first" applications out there will be in hardware and research - simulated physical trials, self-improving chips (and maybe batteries!), and robotics, and that's because those things weren't AI to start with. If someone's trying to convince you they're building an "AI first" software company they're pulling your leg. There's just nothing there.

But What About All This Image-Generating Software I Keep Seeing on Social Media?

DALL-E, Stable Diffusion, GPT-3, and other large models are cool exercises. They're examples of what you can do with big data. They're also not products by themselves.

Google uses their large language models to augment their search results. Microsoft might use DALL-E to make PPTX generation better. Again, these aren't products by themselves. It's AI being applied to an existing process to make it easier.

I might eat crow at some point - the one use case I thought might have merit was prompt-driven essay generation. Essay writing apparently employs 20K people in Kenya. Maybe we get prompt-driven tweet generators, illustrations, PPTX slide generators, or papers and these companies that pursue this do well. I don't believe this will be the case, for three reasons.

1. The uncanny valley effect. When artificial things get 'too real' we reject them. I think this will be the main reason writing-based prompt engines won't go too far. People whose careers are based on reading and writing won't readily accept machine-produced writing the same way trust-fund college party boys do.
2. Content for content's sake is meaningless. There's already a bot problem on Twitter, a fake influencer problem on Instagram, and more. They're treated as sideshows and not serious issues - for good reason. When artificial content replaces real content, the artificial content is just ignored and you focus on the content that matters to you. It's why 90% of corporate account tweets get 3 likes. Generators might crack the code eventually and make content people enjoy, but we're a long way from that and we're approaching that problem the wrong way (careful knowledge curation > brute force algorithms).
3. There's so many issues with existing generative models like GitHub's CoPilot when it comes to intellectual property and licenses. These are massive issues in the art, programming, and writing worlds and all these new AI first startups are running headfirst into a bunch of lawsuits (?? just like crypto).

If there's one lesson to be learned, it's that AI works best when it augments an existing process, not when it replaces one. It can't create something from nothing.

How Did This Hype Cycle Get Started?

How early AI successes are presented in the media echo the same mythology as the "20-something" dropouts that founded Microsoft and Facebook. Growing up in the 2000s, everybody claimed they would drop out of school and start their own tech company, like dropping out was the reason for Gates and Zuckerberg's success ??

It's why so many people were surprised when the HBR released (totally obvious!) findings that 50-year olds make the best startup founders because they're well connected, know the industry, and have management experience. This should've be obvious! But there was this established societal perception that 20-year old prodigies make the best founders (and you saw this belief reflected in VC funding decisions for a decade).

We've made the same mistake with AI. Investors and founders hear about the wild successes: DeepMind & protein folding - and think "I can disrupt any industry with the right algorithm!" The fact that DeepMind was applied on top of meticulous curation of medical data for nearly seven decades, perfectly formatted to machine learning, and still isn't an actual product or even part of a pharmaceutical research pipeline is ignored.

I also think a lot of people fundamentally misunderstand machine learning, especially unsupervised learning. One - I've always disliked the term "unsupervised". Despite AI experts trying their best to explain it I think it does the uninformed an injustice to try and describe unsupervised learning algorithms as learning without labeled data.

The typical paragons of unsupervised learning - large language models and pre-trained image nets are still supervised. You just didn't label the data. Large language models work by trying to predict what the next word is. By feeding the model a sentence (which has words in a specific order), you've fed the system labeled data. In the sentence "the quick brown fox jumps over the lazy dog", the is labeled as the first word, quick is labeled as the second word, etc. There's no such thing as unsupervised learning - there's just data you explicitly created and data that someone else implicitly created.

I think that's why so many people equate AI to magic - it's described that way in breathless tones by reporters and how it works gets obscured.

Where Should I Put My Money Then?

Legacy companies are elegantly aging dinosaurs (hi Microsoft) that are slow to adapt - they won't be winners in this decade either. What you should be looking at are "AI Second" startups.

Why?

Data is key. If you look at the most successful use cases of productionized machine learning, it's all companies who already had a useful product before they tried to throw AI in there. Facebook was a social media platform that employed successful targeted advertising AI. Capital One is a bank that already had a robust fraud detection unit decades before they began automating anything. Dynamic insurance pricing is built on established insurance modeling.

AI Second SaaS companies build a self-sustaining platform that doesn't use any AI. Then they take steps to make sure their system is producing legions of labeled, formatted data through interaction with the platform. Then, when all of the valuable process information is encoded, you can replicate those processes and accelerate them with AI.

AI has had some cool successes in recent years. We can do a lot with a little labeled data in ways that we couldn't before. There's a lot of interesting work seeing what knowledge can be accelerated by mixing and matching pre-trained models. There's equally interesting work being done on figuring out what steps are being considered in deep network neurons and using that to improve how we observe patterns. Focus on companies that are working on applying those developments, not ones whose use of AI is splashing around in a kiddie pool.