Thinking of an LLM based Startup? …Don’t.
Five things about LLMs that you need to consider before you start on a company or a project.
Since the success of ChatGPT, two trends have taken over the world of startup and startup investing. As someone who has actually built and scaled products based on Large Language Models (LLMs), I can assure you that both the trends are misleading.
First, there is a slew of startups that are planning to bring a better solution to known problems using LLMs, most commonly ChatGPT or GPT4. It’s like saying they are taking one of a few pre-baked cakes and putting some icing on top to make it taste better. Second, the startups are baking so many of these cakes that investors have already started looking into the icing for differentiation — They are now stressing on the “thickness” of the tech layer that sits on top of LLMs. Many of the celebrated funding rounds recently tout the different colors, flavors or textures of this icing.
Unfortunately, this is not enough. In fact, far from being enough. And there are five good reasons why. I am not for one second saying that a team cannot overcome these reasons; all I am saying is that unless you do, you may be in trouble.
1. LLMs Are Too Unpredictable to Program Directly
Generative AI is notorious for being unpredictable. So much so that it’s written in the spec doc. Open AI has gone to the extent of providing a parameter (temperature) that controls the creativity of the model. Even when set to zero, and even when using the new function calling features, the model is a tad too creative to directly program downstream applications in a deterministic way. Just try asking the same question of ChatGPT again and again. Bard will give you even higher variance.
For a human this doesn’t matter. In fact the variance indicates some personality in the bot; makes it sound more human. For a computer, though, this kind of variability is difficult to manage. There could be many reasons for this — slightest change in the prompt, some edge case in the user message — or no reason at all; but while LLMs provide a lot of power to solve the original application, teams working with LLMs are often busier adapting for the formats and narratives that come out of LLMs than actually developing other aspects of downstream apps. These variations only get compounded for chain of thought prompting, the predominant way today to work with LLMs.
I have met multiple teams that write their first solution within a week. This solution is all LLM. Then as they try to scale and as this unpredictability becomes harder to manage, they start defaulting back to more traditional, structured methods for their use cases. Each such reversal takes the solution away from the promised power of LLMs.
2. LLM Providers Will Add Your App as a Feature
Since launching ChatGPT on November 30, 2022, and GPT4 shortly after that, Open AI has added multiple functionalities to its APIs like plugins, JSON building, function calling, etc. On the face of it, it appears that this flies completely against the previous point: ‘Hey,’ you may say, ‘this is making the LLMs more predictable.’ However, you are missing the point —
Open AI’s mission is to make LLMs more accessible to everybody. That’s how it is planning to develop AGI. To do this it keeps launching newer versions of GPT (4 is the latest) and keeps adding new features to each of these. For example, since the launch of ChatGPT or GPT4 barely months ago, OpenAI has added plugins, function calling, better ‘steering’ controls, and integration with Whisper models. Insiders tell me there is an ambitious and flexible roadmap ahead.
Each of the additions and enhancements makes the innovations you have been doing less necessary. In fact, I would bet that most successful things that you or other startups innovate on are going to be added to GPT in some shape or form. They are not being evil, they are just being true to their mission. The same goes for other LLM providers. In other words, as soon as you are successful, LLMs will adapt so that the next guy, including your customer, can do what you are doing super easily.
3. Your LLM Based Solution May Not Be Worth It
LLMs are an amazing invention that is set to change the pace of technological innovation. Today, however, they face innumerable problems in terms of predictability, hallucination, and many more. If you are applying LLMs as a new way to solve existing problems without reimagining the problem itself, then it does merit some thought.
Let’s concede that if you know what you are doing, then the solution will be a good one despite the two points I have made so far. With this solution, not only are you competing with other LLM based solutions that Open AIs of the world are making easier and easier to develop, but also with the solutions that already exist to the problem.
Let’s also concede that the current solutions are not solving the problem as well as you do. Still, the current solution is perhaps doing a good enough job at a fraction of the cost. If it’s working well, then there is no reason to uproot it. If it’s not working well then it has created a great expectation barrier in the customers’ mind that your marketing or sales cycle must overcome. Think of the sales barrier Watson created for this whole industry by bombing spectacularly. It took 15 years to undo it.
So unless you are redefining the problem and/or creating an experience that has no par, do take a pause. To continue the example, Open AI tried with GPT1, GPT2, GPT3, and then with GPT3.5 before it was able to solve the Watson problem in the industry. Many others, like my own startup, Coseer, simply couldn’t convince the market.
4. LLMs Look like Shortcuts but Are Not
I had written earlier that while GPT changed everything, nothing really changed. TLDR: Even with LLMs, to build a successful business you have to invest in product, engineering and science innovations as you used to earlier. This is necessary for the reasons stated above and a variety of reasons I cite in the article — workflow integration, user experience, control, costs, advocacy, tone of your client’s brand/ language, etc.
When you do invest at that magnitude, the solution that you are building is no longer an LLM based solution. Rather, it is a solution that relies on these innovations in product, engineering and science. In other words, LLM is just a co-passenger in your amazing journey, and not the destination, or even the car. You still need to put in the hard work that you put before LLMs to create a sustainable competitive advantage.
5. LLMs Come with Serious Security Concerns
Last, and nowhere near the least, is the idea of enterprise security. So far we have discussed the ideas around control of generative AI, which is bad enough. Another reason CIOs and CISOs are very worried is that everything that an LLM processes needs to go to the mothership. This is a whole different dimension added to the routine concerns about your readiness with enterprise security. B2C businesses may not have to face a CIO or a CISO, but they may have to deal with consumer perception and regulators.
LLM providers like Open AI and Anthropic have reasonable and transparent policies about security. However it is not foolproof; they don’t guarantee that they themselves will not look at the data; and given the recent news around hacks, CIOs and CISOs don’t feel comfortable. It’s the old axiom of information security — once it’s out of your firewall, you really have no control.
Another option some startups consider is to think of proprietary LLMs within their (or client’s) firewalls. That is indeed a reasonable option, but now LLMs are no longer easy enablers. You will have to innovate on LLMs in addition to innovating on product, engineering and the science pertinent to your domain.
I will repeat that this article is by no means an argument for not starting up in the LLM space or starting with an LLM project. It’s just a caution — some food for thought before you embark on your plans. Feel free to reach out on LinkedIn if you need to discuss further.
