With the new year looming ahead of us, many founders will wonder how the growing AI movement will affect their entrepreneurial chances. Will this groundbreaking new technology make the impossible possible? Or will it “take our jobs?”
Today, let’s look into the crystal ball and see a few opportunities, challenges, and threats that AI systems may pose for software entrepreneurs and creators.
We have seen a Cambrian Explosion of AI tooling and progress in 2023. I believe that this will be trumped significantly in 2024, and I don’t see it slowing down anytime soon.
The Proliferation of Large Language Models
Let’s recap where we are right now. OpenAI has released GPT-4-Turbo, Meta’s LLaMA has been adopted by the open-source community, the French open-sorce model Mistral is making waves because it’s so compact yet can already beat GPT3.5, and Apple is rumored to have been working on models that can easily run on their current iPhone models.
Experience this article as a podcast or as a newsletter:
The trend with this tech is obvious: more players are joining the game, models are getting better, more available, and smaller. And, maybe most importantly, they’re becoming more and more multi-modal, which is worth looking into.
Multi-Modal AI
A multi-modal Large Language Model is capable of dealing with not just text, but also audio, images, and video. It can interpret and generate any of these formats If there is structured data of some kind, you can bet that the models currently being trained and tuned are more and more capable of understanding and manipulating all kinds of media.
We’re already seeing some of these systems in their early stages in the wild: Otter.ai is a tool I use to create transcripts for my podcasts, but they can do so much more: their Zoom integration summarizes ongoing video calls as they happen and presents actionable notes and meeting outcomes seconds after the meeting has concluded, sometimes even right after a decision was made in the call. Future systems will already start working on the problem or dispatch the proper instructions to the people who will work on them, while the meeting is still taking place.
Notes will be filed and analyzed, short clips will be auto-generated, transcribed, and sent to the people who need just that part of the conversation.
In the podcasting world, this is already commonplace. It’ll become normal in less tech-literate industries very soon, and people will expect to find this kind of assistive technology in the software tools they use. If they work with audio, they will expect transcripts, summaries, and being able to ask questions of any kind of data.
Working with AI vs. Working against AI
Software offerings without these capabilities will quickly feel outdated. Clearly, the early adopters expect AI-centric features in their edgy tools already. But the shift towards the mainstream will add even more pressure to facilitate these features.
Either you embrace the technology, or your competitors will use it to chip away at your margins. AI-as-tooling will become as commonplace as using databases, analytics, and social media avatar pictures.
Fortunately, there are a few trends that will make this easier for you as a founder. I’ve been making good use of one particular trend for my own SaaS podline.fm. The magical ingredient is running these incredible AI models inside your own infrastructure. “Local AI” was almost impossible a year ago. Now, it’s two installation commands away. You can run a significant number of small-but-powerful language models on consumer hardware, without even needing a graphics processor (which these things tend to run on). Podline has a miniscule server with 16GB of RAM which can easily spin up a language model like Mistral7B and get a ChatGPT-like response in just a few seconds. There is even example code out there that lets you spin up your own API, ready to be used by your backend services. Other than the cost of a small server, running your own AI system is free.
And that is massive, because there is an open-source community out there training AI models for many different tasks. Every day, new models are uploaded to HuggingFace, a kind of model repository, where you can download and use these models for free. Most of them are licensed in very business-friendly ways, too! From speech recognition, text-to-image, classification, summarization, translation, and video generation, you’ll find a model that will take a previously manual task and do a superb task at automating it.
And even if the results aren’t perfect, you can fine-tune all these models yourself over time. None of this will be prohibitively expensive, and even though we’re still in the early days of this tech, it’s already very easily integrated into our tech stacks.
AI Platforms & Dependency Risk
There’s just one risk here, and that’s depending (too much) on AI platforms. It’s pretty easy to get a local AI going. But it’s much easier to just implement a few API calls to OpenAI and have their massive server cluster do the work much faster. Pricing is currently quite low —and I expect it to go down even further as competitor pressure increased on AI platform providers— so it makes sense to just pay a few cents for tasks like summarizing an hour-long conversation or generating a video.
My rule for this kind of feature is: there has to be a local fallback for each API call I make to an AI provider. With podline, I have implemented just that: when an audio file gets transcribed, it either gets sent to OpenAI or I run whisper.cpp on it locally. Summarization? Same thing, just with llama.cpp spinning up an instance of Mistral on the backend. If I can avoid it, I’ll never fully depend on someone else’s platform for an AI feature.
AI-as-a-Feature
And that’s what AI should always be. A feature of your product, not your product itself. AI can be an interface to your users’ data or a means of transforming that data into shapes that are more useful to your users.
”AI-as-interface” can be the good-old chatbot, which will likely morph from conversational novelty to a real “agent” that will not differ from a personal assistant you send tasks via email. Imagine your SaaS product not just being a website your users visit, but a virtual person they can reach out to, ask questions from, and have execute service-specific tasks. Your SaaS will truly become a “service” that builds a transactional relationship with its users, one interaction at a time — a bidirectional exchange you or your support staff can always jump into when needed.
Obviously, customer support will be massively impacted by AI systems, if it hasn’t already. First-level support can (and should!) be handled by always-on, always-focused AI systems that both have insight into the details of the customer’s account and have the whole history of all known problems and solutions with the product in their constantly-updated training data. This is a job that even the most experienced human customer-support agent will struggle to live up to, mostly because they need to sleep. For this kind of immediate help, AI systems are great.
But even in the background, quietly summarizing, transcribing, analyzing, and generating data, these systems will massively impact your business in the years to come.
Consider that AI tools such as the large language models of today are the pioneering systems of automation that can massively lower your operational costs while at the same time do the work that previously either took a long time to process or needed human ingenuity.
Embrace AI without depending on it. That’s my AI adoption approach for 2024.
