Depending on who you ask, AI is either the greatest invention since sliced bread, or the next pump and dump scheme offering little enterprise value - very much like the crypto wave before it. Even experts and news media seem lost - with contradicting articles in the same news feed decrying AI's lack of value while proclaiming its revolutionary nature. So what is going on with AI?

The Incontrovertible Truth - AI had value, has value, and will have value.

AI had value: At Emissary, we often say - generative AI hasn't created AI value as much as its helped key decision-makers visualize AI value. Companies like TikTok, Facebook, Netflix, Google etc have already created immense enterprise value rooted in conventional AI (machine learning). Insurance companies, healthcare companies, autonomous driving companies and almost every industry has done the same in their own capacities.

AI has value: But this isn't just true for AI of the past. Generative AI too has clearly demonstrated value - this is most evident in ChatGPT usage and retention metrics, and the rapid adoption every ChatGPT wrapper gets. Economic value IS being created via AI - both old school and new school. A significant fraction of the workforce IS having AI do tasks. Even more importantly, a generative interface has helped every decision-maker visualize potential AI value and unlocked meaningful budgets to do so, unlike the pre-chatgpt world, where visualizing AI value was a capability largely limited to calculus-comfortable decision makers, artificially suppressing budget allocation for AI capabilities.

AI will have value: This is where AI differs from a crypto-like future - it created and realized enterprise value pre-hype and continues to have value during-hype and we have no legitimate reason to believe it won't in the future.

BUT the concerns of froth are not without merit. The underlying criticism of AI conflates the difference between value and ROI. The real question to ask is - will AI be ROI-positive at radically smaller scales? Till date, the bulk of AI value has been harnessed by massive corporations staffed with hundreds of MLEs.

And this question is what spurs justified gut feelings of froth - the open question of whether willingness to pay for AI experiences can exceed the costs incurred in building and maintaining AI.

But AI today is frothy - here's why.

1. It really is day zero

We're in the calculator/flashlight app-equivalent era of AI. That's the closest equivalence of AI model wrapper businesses. Familiarizing ourselves with a technology as novel as AI is like adding a new appendage, and most capabilities today are more about getting familiar with these appendages than actually putting them to use - think grasping things with your hand to understand what you CAN grasp, not grasping because you NEED to. A vital process in the journey to capturing economic value through your capabilities, but in of itself, not value generating.

2. Misplaced dollars and attention

A big chunk of our perspective on a technology are dependent on where capital and attention for that technology heads. Your feelings about AI are similarly contingent on fundraise announcements and popular media, even if the value could be in a whole other world. Nowhere is this more evident than AI infrastructure. In the year post the launch of ChatGPT, capital rushed to the 'infrastructure layer' and 'picks and shovels', often from sources still getting familiar with the technology. As a result, most investments here were simply repurposing existing infrastructure for new use-cases, instead of a truly novel layer of infrastructure.

More specifically, dollars flowed into data infrastructure for AI - AI gateways, databases, observability platforms despite the fact that highly reliable incumbents like Kong, Postgres, and Grafana had to make minor additions to their product to establish their dominance over the spaces. There are real startup infrastructure opportunities for AI - but they lie in envisioning a truly novel layer of intelligence infrastructure, a world where companies run on AI just like they run on data today, a journey we are just starting to take.

3. Static, data-driven intelligence transactions

Here's a secret: most first AI capabilities suck. They're truly horrible. The first iteration of the famed TikTok recommendation model was unimpressive, to say the least. Reels circa '21 was no different, as we all vividly remember. Pre-2022, we understood clearly that the purpose of developing the first model was to kickstart an iterative learning process, a two-way continuous intelligence transaction. The new GenAI world is still coming to terms with that reality - resulting in teams rushing to production, largely freezing development once an AI feature is launched, and consequently a steady stream of underwhelmed customers and NRR challenges.

Parallely, in the current world, AI systems are treated as static sources of intelligence garnered through memorizing data. True AI value will be realized when we start creating feedback loops that continuously improve AI systems through user action/behaviour. Note: this is NOT (solely) RLHF, iterative finetuning works more often, in fact. But simply creating an AI system and exposing it to users will almost always result in wide-spread dissatisfaction. Think of your first prompted AI feature as a solution to the cold-start problem, not an end result.

4. Minorly differentiated AI + Focus on Software Moats.

This is a core consideration from a willingness to pay perspective. While a distant future with AGI might render finetuning obsolete, in the present and near future - creating prompt wrappers on top of third party models at worst, upper bounds their willingness to pay at their unwillingness to switch tabs, or at best leaves you with a software moat. In a world where we expect differentiation at the AI layer, and end up primarily with software moats, it's no surprise that folks feel a clear sense of frothiness.

Conclusion

There is very real value in the AI world, but there is just as much, if not more, froth. In a world with high noise to signal ratio, you might be inclined to sit by the sidelines, but massive gains will be established far before the fog clears. Navigating this journey is about setting a framework - establishing clear metrics to determine cost and willingness to pay, asking yourself if you can repurpose existing tools for a given problem, and outlining not just what your moat is, but what your compounding AI advantage is.