Where to play

[This post is meant for readers who are early in their thinking on LLMs and the corresponding B2B venture opportunities that arise. Founders who are farther along might not find a lot of new ideas here.]

Knowing that you’re ready to start a business is just the first step. Navigating the idea maze to find the right problem statement to commit to is the hard part. Always a long-winded and often a very stressful journey (mine here). Of course, with GenAI, it’s easy to say we are still in very early days of this transition – we don’t yet know which part of the stack will most value accrue to, how should we think about defensibility (every week we seem to be blowing up some previously held belief on moats in this space), do we have enough right talent in India (related, how much AI experience do you really need to build companies here).

Having said that, we have seen time and again that a lot of the rewards are also reaped by folks who are early — we’re already seeing that with OpenAI, Pinecone, Cohere, and several others that were building for this future before the mainstream adoption became obvious. And so if you’re looking for where to get started, here’s a quick list of areas we’ve been thinking about:

  • Vertical/ use-case specific foundation models. The business case is quite straightforward. As opposed to a ‘God model’, they should be cheaper to train and run on, have lower latency, lower memory requirements and be easier to retrain and keep up to date.
  • Private LLMs for enterprise. There are several reasons why foundational models by themselves are not adequately suitable for use by large (or small) organizations – including but not limited to privacy, compliance, efficiency, accuracy, and outdatedness. Compliance and privacy regulation in fact might be stronger in some verticals vs the others (eg. BFSI) and might be an interesting forcing function for local AI deployment in many countries creating a path for regional winners.
  • Testing. LLMs produce non-deterministic and unstructured output, different output across multiple runs, and require newer, more qualitative ways to evaluate performance. As AI deployment moves out of being just a data scientists’ domain and becomes an every person use case, we need testing methods that accommodate the variety of users, use cases and machine output we are about to encounter.
  • AI safety and regulation. There’s a lot said on this topic and the need is pretty clear. As governments, as well as enterprises, navigate this new world – balancing productivity unlocks while avoiding bad actors and unintended consequences (bias, guidance on illicit activities, skill redundancy) is a non-straightforward challenge. There’s an opportunity to work closely with decision-makers on building the guardrails and mitigating measures for an increasingly uncertain world where force amplifiers exist for both good and bad objectives.
  • Reskilling for the new economy. We have started hearing anecdotal evidence of large-scale reskilling in organizations where a lot of the junior engineering staff is now being trained to take on more complex projects. Simplistically, you need fewer people on L1, and the experience threshold for L2/L3 reduces with smarter tools.
  • My partner Dev wrote about Generative Saas here and will soon cover ‘Desi LLMs’.

You can never do enough customer calls. While this is a ‘top-down’ list, it’s usually more helpful to think about problems ‘bottom up’ while starting out. And the most useful first step is to talk to early adopters and existing customers of companies in the category you pick to understand what’s missing, what could be better, who is the decision maker, what budgets could we access over time, and most importantly – how sharp is the pain point. It’s also more common than you would think to talk to a very specific kind of customer segment and end up with false negatives or false positives on the validity of the product proposition – so make sure to speak to teams across verticals and stages. The best thing about building B2B solutions right now is that given the rapid adoption and excitement around GenAI, it’s relatively easier to get that first conversation with potential customers. We’re all AI-curious.

Will do a similar post on consumer-tech opportunities soon. If you’re thinking about starting up in these areas (or beyond…VC imagination is very limited!) we’d love to hear from you. I’m on shuvi@lsip.com.

Lightspeed Possibility grows the deeper you go. Serving bold builders of the future.