Now is a disorientating time to be building a product in the AI sphere. A new class of applications / interfaces is emerging under the GPT banner and OpenAI (among the other major players) are shipping new products almost weekly. Products which have capabilities far beyond what a single indie hacker would be able to create.

The name on everybody's lips is, of course, OpenAI. They have been almost dormant for years, quietly amassing the world's best engineers to work on both the research and applied ends of AI. Since the first release of ChatGPT, which rumbled the world almost two years after GPT-3 [RLHF and chat interface did the trick?!]…

The release of ChatGPT Plugins has frightened many builders in the AI game, because an extendable ChatGPT which can browse the internet, connect to any API, and execute code it writes itself, is clearly powerful. Let's double down: surfing the web, connecting APIs together, and executing code is the backbone of any software engineer's job. No wonder we are all so disturbed by this tech: it's a competitor not only to our startups and products, but it does what we do, and often it does it better.

Many a YC-launched startup must now reconsider their wrapper around GPT, because OpenAI, who have heretofore focused on research…

Big tech is shipping like start-ups

It's not just OpenAI. Microsoft (inevitably) is adding new features to Bing daily. Google is launching Bard. Adobe just announced Firefly, its own text-to-image content generation service. Canva did the same. GitHub is levelling up its Copilot extension.

Some AI hype artists are describing this GPT wave as a 'platform shift'. Bill Gates recently described it as comparable to the invention of the Windows GUI. The most compelling evidence for this is surely the behaviour of big tech, which has been fairly quiet for roughly a decade. (Can you think of a big new product Google released in the last 10 years? Colab might be it.) Now, like start-ups chomping at the bit to find an edge with AI, they are shipping at terrifying speed. After all, they have the expertise, the infrastructure, and the engineers to do this.

So what's a developer, who wants to build in this space, to do?

But… not even the big names know where the value is

Be an employee

The rate of progress in both AI capabilities and (productionlisation) right now is unprecedented. I don't think a single builder can reliably ship products at a high enough calibre or speed right now for it to be worthwhile.

For the last few months I was working on a social network for ML people, but HuggingFace just released 'Paper Pages'–exactly what I'd built, except theirs is a natural extension of an existing ecosystem. I then looked into retrieval-enhanced LLMs and what it would take to launch an app where developers could easily integrate in-context learning to any model (think langchain in the cloud). But then, along came ChatGPT plugins…

There are plenty of generative AI start-ups hiring right now. It's a safe play, sure, but, joining a company with a mature approach to this crazy market is a way to ride the wave without much risk. It means going with the flow, when the flows are strong and unpredictable. Unless you have some phenomenal insight or expertise, it's better to accept that you alone probably won't be the developer of a revolutionary AI product. With a team, however…

You might even consider joining an established company which is branching into generative AI; this way you can be creative and enjoy the fun of launching new things, but without the stress of pivoting.

Go where big tech won't

I don't foresee OpenAI making religious chatbots (like JesusGPT), nor will Microsoft support adult video/image generation anytime soon. Anything to do with warfare is likely out of the question too. And yet, two things are true: there is massive, massive demand; and AI can be valuable in these markets.

If you let your imagination roam, you'll think up more exciting applications. You might imagine a dating service where 'matches' are just avatars connected to a NSFW fine-tune of Stable Diffusion. Or a language model used for wartime propaganda.

As always, you can make some money by leaving morals outside.

Choose a super-specific niche

'ChatGPT for language-learning' is not specific enough. You'd be competing with all the other language-learning apps which will inevitably add ChatGPT. At which point, you'll have a feature and not a platform. Try 'fine-tuned Stable Diffusion for teaching Japanese to English speakers'. That would be fun, unique, and most importantly, it's not 'general' enough for the larger apps to focus on (for now).

Or how about 'convert chat to biography for the elderly'? I assume there are millions of older people out there who are bored and have interesting stories to tell. Beyond virtual companionship, chat models could provide a way to distill their tales and experiences into sort-of autobiographies for others to read. Or maybe the virtual companion side of ChatGPT is more interesting to explore.

My two examples are B2C, because I think the B2B space will be heavily dominated but the major players.

Start-ups like to keep their target audience wide while they figure out who exactly is willing to pay the most for their offering. In a way this is a 'start big and get smaller' strategy, when the mechanics of start-up growth are the opposite. With so many indie hackers descending on the APIs, you need to validate an idea instantly and grow it outwards.

A way to think about AI

It has been noted how, for a long time, the story used to be that basic, repetitive tasks would be automated away first, and now, the opposite appears true. GPT-4 is a top-tier developer; Midjourney a ridiculously talented artist and designer. These were the jobs we thought AI would steal last!

Recently, OpenAI published a paper describing the professions they thought were least at threat. They include woodworking etc. But importantly, these jobs are deemed safe for the time being because there isn't much training data out there, not because they are inherently un-copyable tasks.

Although they might know less, humans are capable of behaving like GPT-4. In the olden days, you would consult a specialist who could relay information to you (in the format you requested). You could ask humans for legal advice in the style of the King James Bible, or for rhyming recipes, and depending on the creativity and knowledge of the human, you would get what you wanted in response. I don't think GPT-4 has 'done something with its mind' which a human hasn't been able to do yet, although I would be very interested to know if I'm wrong.

The same is true of Stable Diffusion, right? There are artists out there who could do a fantastic job drawing an astronaut riding a horse, or a polar bear wearing a hat in the style of Dalì.

There are two main differences: speed and cost. AI is capable of doing these things much faster and much cheaper than a human equivalent. So the question is: 'in what fields do we care about getting results faster and cheaper?'

The answer seems to be 'all of them'.

Block out the noise

On Twitter, I have to mute thinkbois posting threads about how to make money with ChatGPT every day. Likewise, sexy demos showing off an incredible new product can be inspiring, but they can be distracting too. It seems wise to look at them with a careful eye: they can help calibrate the relevance or value of what you're building and they can help you identify ways to improve your knowledge and your product. But to me, at least, they are a source of stress as I see how quickly others are putting their apps forward.

Suhail, founder of PlaygroundAI, recently tweeted 'Run your own race.' He's encouraging builders to knuckle down and persevere, irrespective of what others are sharing (mainly on Twitter). And continually making improvements is a sure, if gradual way, to end up with an excellent product on your hands. An excellent product which has been iteratively upgraded over time is a worthy contender, until ChatGPT drops 'Plugins'.

The problem is that 'your own race' – a race with nobody else in it – is not really a race, and losing sight of the market or operating at a slower speed than the indie hackers with whom you're competing, does not sound like a winning strategy.

Figure out your edge

The arrival of high-quality generalist transformer and diffusion models is going to flood the internet with products, especially when there are great public APIs ready to go. Low-hanging fruit is low-hanging fruit. Somebody will take it and make a quick buck, sure, but it's not sustainable and it is stressful.

Of the 1000 highest-performing people in the world, I would imagine that over 25% of them are involved with AI in some capacity. Your competition is the best of the best: the researchers who are able to format abstract challenges as neural network architectures, and the lightning-fast developers who can build great things fast. The scary part is that these two populations are usually united at the big tech companies, and people wanting to build apps are competing with these guys.

So you have to have a long, hard think about what edge you have. No doubt this is pretty generic advice, used for startups in general, but it seems even more pertinent now when the major companies are capable of vacuuming up all the capital in the AI space.

We still need to figure out what the advent of AGI means

Developers now feel what the artists felt when midjourney came out