4 questions on AI

Hopefully I can breathe some life back into this newsletter. And the topic du jour in Silicon Valley and the broader tech ecosystem is AI – more specifically the changes spurred by OpenAI, Stability, Microsoft, and other researchers and engineers around the world. Rather than play Nostradamus, I wanted to outline the open questions on my mind about the present and future state of AI & ML.

Here’s how I would frame the opportunity for startups and de novo projects (as many wonder if value will just accrue to incumbents – OpenAI especially). First, read the AI Grant manifesto (especially the last section), “AI: Startup vs. Incumbent Value” from Elad Gil, look at some of Dan’s recent tweets as examples, and of course the 2022 State of AI report from friend Nathan Benaich. My thinking goes to – what new product experiences are uniquely enabled by generative AI or LLMs as a form factor? Or which existing products that solve a need can be made 100x better? I’m reminded of just *how* mobile-native Uber, Snapchat, and Instagram were – they simply couldn’t exist in a previous paradigm.

How does text-based prompting supplant or augment GUIs?

Some of the language models’ ability to infer natural language (ChatGPT) is unwieldy so there has been a lot of scrum about how text-based prompts could completely replace graphical user interfaces. My sense is that this won’t happen overnight as GUIs give way to higher fidelity or correctness with the current advances in models – you really have to know how to prompt engineer to get a desired output in some cases. The historical analog to this transition is likely DOS to Windows in the late 1980’s, when Microsoft moved operating systems from a command-line centric world to a graphical one.

What happens to our understanding of the software business model?

Traditional software has a nice business model – relatively fixed upfront cost coupled with de minimis marginal cost. In their current state, models (and even applications) break this a bit. The upfront development and training costs are higher (due to the extremity of the compute), and it costs more to continuously serve the models (inference). This seems to be headed in a positive direction, but we will see how it continues to evolve. Rapidly decreasing cost curves rule everything around me.

Who owns distribution?

My friend Peter makes an interesting point on the debate of whether GPT can kill Google Search. To summarize, even if the search could be better with GPT, the Google search bar is highly coupled with the distribution of Chrome, which Google owns. If one is building a new mega-platform off of a technology transition, owning the end distribution with users (aggregation anyone?) matters a lot. I’m curious to see how this plays out with potential new networks for consumers and tools for businesses.

What is “old” that becomes new again?

There are common substrates of problems that every generation of startups continues to solve, made novel by new enabling technologies. Born in the modern web, Airbnb is a marketplace for underutilized fixed assets that eBay was a cohort prior. Shopify, Stripe, and Square continue the commerce infrastructure story for small merchants that PayPal started a decade before. Salesforce, ServiceNow, and Workday are the cloud-native children of Siebel, Peregrine, and PeopleSoft. I’m most interested in marketplaces (pushing out the extreme of search and optimization problems), communities (how can agents and NPCs exist alongside and augment connectivity?), analytics & customer data infra (Mixpanel, Optimizely, Heap, Amplitude, LaunchDarkly rehashed), and tools to supercharge entrepreneurs (will we finally see the 1-person unicorn?)

As always, this post was an exercise in fleshing out some of my thinking more publicly. I’m sure my opinions will continue to evolve and adapt, and I always learn more from conversation. If you are building in this space or have strong opinions – please get in touch aashay [at] haystack [dot] vc.