10 AI predictions for 2024

Hey Siri, set a reminder for 365 days.

I've always wanted to write one of these lists, and now that I have my own publication, what better way than to end the first year with some AI predictions? I'm looking forward to revisiting these in a year and being hilariously, egregiously wrong - but at the very least, it's a valuable exercise in trying to understand where the puck is headed.

Artwork created with Midjourney.

Open-source models catch up to GPT-4

One of the most impressive things about OpenAI is that GPT-4 has been out for the better part of a year, and nobody has managed to surpass its quality. But after about a year, we're finally seeing open-source models (i.e., Mixtral, Llama, Falcon) get to the level of the original ChatGPT (GPT-3.5).

As the trend continues, we'll see open-source models reach GPT-4 quality sometime next year. There will be plenty of optimizations along the way - we'll also have a GPT-4 level model capable of running on an iPhone. But I also expect said model to only operate with text; GPT-4, Gemini Ultra, Claude, and other closed-source competitors will retain their advantage with improved multi-modal capabilities, RAG, and tool usage.

Meta releases Llama 3

Of course, leading the open-source pack will be Llama 3. Meta's Llama models changed the game for open-source LLMs, and I fully expect them to keep moving forward with foundation models. Unlike OpenAI, they don't have to move the whole industry forward; they "just" have to replicate the current state-of-the-art and then open-source it.

Llama 3 will have:

• A larger context window (64K minimum)

• Multi-modal support (at the level of something like LLaVA)

• Interoperability with Meta's other foundation models

Apple's AI

Apple has been conspicuously absent from the AI free-for-all this year. Most other tech giants have made announcements in one form or another: Microsoft, Google, Meta, and even Amazon have released models, assistants, and cloud platform capabilities for generative AI. But Apple has stayed out of the spotlight, even as new iOS and Vision Pro features are clearly powered by machine learning.

But Apple has a well-known pattern of biding its time and introducing technology once it can create a refined offering. To that end, we'll see an announcement of a new, privacy-focused LLM from Apple. One that fits on an iPhone and is powered by a brand-new, AI-optimized chip. It will look like a supercharged Siri, with impressive tool/agent capabilities.

A prominent AI company goes bust

As excited as I am about AI, I also admit that the space is likely over-hyped as a whole. There will be few winners and many losers - and we'll start to see some of the latter emerge next year. I know "goes bust" is a pretty vague description, so I foresee one (or more) of three scenarios:

• A well-known AI company runs out of money or gets acquihired.

• A hot AI startup implodes due to founder and/or legal drama.

• A flashy piece of AI-enabled hardware is a complete flop.

A realistic AI-generated video goes viral

Generative video, for the most part, is still not good. However, we know how fast things can move - the differences between Midjourney outputs just one year apart are astonishing. Right now, we're just scratching the surface of what's possible with video - tiny animations using tools like Runway or Pika.

Next year, expect to see some shockingly good AI-generated video. So good that it'll become a viral hit - in the form of a short film, TikTok, ad, or music video. We're currently seeing bits and pieces of video content made with AI - but the big difference will be a piece of content that's almost entirely AI-generated rather than stitched together in post-production. Shortly after that, we'll see AI videos become indistinguishable from the real thing.

A big media outlet gets bought for training data

As the industry has grown, it’s become increasingly clear that the next big advantage will come from proprietary datasets. Access to compute, and even model architecture is becoming increasingly commoditized - meaning AI companies need to secure better data to stand out.

Scraping the internet is more challenging than ever, both because more sites are explicitly banning crawling for AI training purposes, and because we can no longer trust internet results as human-generated. With that in mind, news and media publications are ripe targets for AI companies.

Why wouldn’t Google buy Reuters or a similar outlet, to train its models on their verifiable stories? To make Street View work for Maps, Google paid thousands of people to drive, bike, and walk through millions of roads and paths. By comparison, maintaining a news service seems pretty straightforward.

AI companions go mainstream

One of the things that struck me when using the iOS version of ChatGPT was how immersive the voice features are. To be clear, it still needs some polish - the voice isn’t the most realistic sounding, and there are plenty of lags and glitches. But just a short conversation made me nearly forget I wasn’t talking to a person. Even though I understand how LLMs work, and do not in any way think they are sentient, I found myself approaching ChatGPT-over-voice like a human.

It also made me realize that we have all of the ingredients to make companions from the movie Her - we just haven’t done it in a fully integrated way yet. But many are trying, and next year, we’ll see AI companions have their Tamagotchi moment. They’ll move from the fringes and into the mainstream, with millions having some a digital friend or romantic partner whom they converse with.

A major election is derailed by AI

2024 is the biggest election year in history. Across 76 countries, between two and four billion people will have the chance to vote in an election. And virtually none of those countries have elected a new leader since ChatGPT, Midjourney, and ElevenLabs went mainstream. In some cases, national media and social platforms will proactively try to fight misinformation (Facebook, for example, is already banning all political advertising created with AI).

But not all governments are democratically inclined, and not all voters can distinguish AI-generated content. At least one election next year will be heavily influenced by AI misinformation or propaganda. Politicians around the world are already experimenting with these new tools - the intensity will only increase as we enter election season.

More creative AI monetization

So far, AI business models have boiled down to 1) charge for subscriptions/credits or 2) raise more VC money and don't worry about it. And even those charging users are still doing so at a loss - running inference at scale gets pricey, fast.

As we move into the next phase of AI products and services, though, companies will experiment with new (and old) forms of charging for AI. Video game DLC that includes characters with infinite dialogue. Commissions for AI-powered sales agents. AI girlfriends that cost money to "progress" to the next stage of relationship intimacy. And, of course, ads - right now, someone is working on a way for chatbots to 1) mention promoted products as part of the conversation and 2) charge money for each "impression."

The AI backlash gets bigger

I've said before that we're in the earliest stages of this stuff - even if we stopped all new AI development now, it would take years to fully integrate our current capabilities. And as we continue to do so, we're going to see the AI backlash continue to grow. So far, it has primarily been about safety and IP infringement. And with respect to the latter, we're going to see a landmark legal decision made sometime next year around AI and copyright. If the decisions are in favor of existing IP holders, big names like Disney and Nintendo will jump in with new lawsuits.

But we'll also see some new areas of pushback. The cost of training and running AI is an often overlooked area - what does it really cost to build these models, in terms of electricity and emissions? More will begin protesting the environmental cost of building foundation models.

And as critical events like elections are impacted by AI, more governments will look to regulate its usage - not due to apocalyptic safety fears, but more pedestrian ones like, say, protecting democracy. While regulators and legislators grapple with AI content on a national level, educators and parents will do so on a local level. We will see pushes for “human-made” labeling of content, and (ineffective) tools to detect AI generated media.

But ultimately, these approaches will be bandaids - the genie is already out of the bottle. With any luck, we can find better solutions, ones that I’m not smart or qualified enough to imagine here. I’m not naive enough to predict that the back and forth over AI will be settled anytime soon - if anything, we’re in for many more years of debates, hearings, lawsuits, and regulations.

Artwork created with Midjourney.

What's your take?

Honestly, I probably could have added another ten predictions to this list - there's so much happening in the space, and every month brings new innovations and announcements. The ones above were a mix of technology guesses, political guesses, and cultural guesses. I also wanted to make some predictions that felt a little bit of a stretch rather than an obvious win - I don't think anyone would argue with "Midjourney releases v7," for example.

But let's turn it over to you - what do you think I got right and wrong? What's your hot take for 2024? Leave a comment below!