10 of the most impactful AI stories of 2023
10 of the most impactful AI stories of 2023
A quick look back on a very busy year in AI.
An earlier version of this post originally appeared in AI Supremacy. I highly recommend you check out the newsletter, it’s one of the top technology publications on Substack!
Just over a year ago, ChatGPT went live - and the world changed. Big Tech announced AI product after AI product. Governments pushed for regulation. And what feels like a billion AI startups were founded, funded, and launched on Twitter in record time.
But with so much noise, you might have missed some key AI stories. Let’s take a look at the biggest developments this year - events that are both noteworthy in their own right, and can also teach us something about where AI is headed.
I’ve covered a lot of these events in my weekly roundups and deep dives, and I’ve linked to them throughout the article. And of course, I’ve left a lot out - getting down to 10 was a difficult task! If you think I missed something big, leave a comment down below.
Honorable mention: Sam Altman’s firing
This November, I was lucky enough to attend OpenAI’s first ever developer conference in person. If I’m honest, there was a palpable excitement from the attendees and the OpenAI team given all of the announcements. At the post-conference reception, Sam Altman was happily talking with attendees and taking selfies.
So you can imagine my shock when less than two weeks later, Sam Altman was fired as CEO. And according to some reports, DevDay was the final straw that led to his ouster. Considering Altman was rehired less than a week later (completing his Steve Jobs story arc in record time) it might seem like not much has changed with the status quo.
But the power struggle led to the removal of multiple board members, clearing the path for the continued commercialization of OpenAI. It put renewed focus on whether or not the leading AI technology should be limited to a few companies. And it served as a small example of a bigger rift in the AI community, between those concerned about safety and those wanting to continue developing the technology.
10. AutoGPT
If you want a pizza, ChatGPT will suggest a list of pizza restaurants. AutoGPT will (someday) find the number for Domino’s, remember your go-to pizza toppings, and order it for you - at least, that’s the hope. AutoGPT is a toolkit for building AI agents: programs that can independently create and execute plans. When it launched earlier this year, it quickly became the fastest-growing GitHub repository in history. It's currently at 155,000 GitHub stars, making it the 23rd most-starred repository on the site.
Agents are a significant change from how people normally interact with large language models - most of us are used to a conversational style, like ChatGPT. Agents will run continuously without user input, until they reach their goal or they're stopped. While AutoGPT and its cousins are still in their infancy, they're improving at a rapid pace. With each week, they're adding new features and capabilities, including the ability to use tools, access the internet, and more.
9. The Copilotization of the internet
The year started with Sydney, an AI that tried to convince a New York Times writer to leave his wife. Sydney isn’t an actual product - it’s the codename for Bing Chat's erratic alter-ego. And after several mishaps, Microsoft walked the AI back, but it found other ways to launch AI chatbots, dubbed Copilots. These productivity-focused LLM assistants are being baked into software tools used by millions of people: Bing, Office, GitHub, and even Windows.
Google quickly followed suit, bringing its Duet assistants to Gmail, Google Workspace, and Google Cloud. On top of that, we've seen a deluge of startups take a similar approach, adding AI helpers to a wide variety of applications. For what it’s worth, I’m not wholly convinced that this is a good thing - there exists better UX for many AI apps than a chatbot. But these copilots are undoubtedly about to change the workflows of millions of businesses and employees around the world.
8. Heart on my Sleeve
In April, a new track featuring Drake and The Weeknd dropped, immediately going viral. The only problem was, it wasn't actually recorded by Drake and the Weeknd. It was created by Ghostwriter977, an anonymous Discord user who made the track with cloned AI voices. Within a few days, it had millions of streams on Spotify and YouTube, and by the end of its first week, it was taken down due to copyright claims.
AI voice cloning had been around for a little while, but this was the first time it entered the mainstream. And record labels, as well as individual artists, are trying to figure out how they might make money from this new technology. The musician Grimes released an AI-cloned version of her voice, on the condition that any fans who used it would have to split the royalties 50/50. And music labels, at first hostile towards the tech, have quickly come around to negotiating deals for the rights (and royalties) to use AI-cloned voices of their clients.
Now, we've come full circle - with YouTube releasing tools like Lyria and Dream Track that can generate custom music and lyrics in the style of officially-licensed musicians. Creators are able to use AI to make any sort of music they want for their videos, without fear of copyright strikes.
7. Nvidia's $1T market cap
This summer, Nvidia became a trillion-dollar company. For context, that’s more than the market caps of Intel, AMD, Qualcomm, Cisco, and Broadcom combined. Whether or not you believe that's a reasonable valuation, it's important to understand why: GPUs.
GPUs are now an incredibly valuable, and increasingly scarce resource. Multiple tech CEOs, including Sam Altman, have said that access to GPUs is currently the bottleneck for AI progress. New and better models, and more widespread support of existing models, is limited by the number of GPUs companies can get their hands on. Some startup investors are even offering access to their GPU clusters as a perk for their portfolio companies.
GPUs are, at this point, considerably harder to get than drugs.
We know that AI’s disruption is going to create winners and losers. But in this current moment, Nvidia is reaping many of the rewards. At some point, GPU supply will catch up to demand, whether another company can increase output or the AI boom subsides. But until then, expect GPUs to continue to be a hot commodity.
6. Med-PaLM
One of this year’s AI stories that feels underreported to me is Google Health. I’m certainly guilty of this - I barely covered it in my roundups. Google Health is building and publishing health resources and products, including multiple AI-related projects. The company has even launched MedLM, a family of models aimed at helping healthcare workers.
The organization has researched AI to prevent blindness, detect breast/lung/colorectal cancer, and assess cardiovascular risks. But the most potential lies with Med-PaLM 2, a large language model fine-tuned for medicine. This LLM is designed to answer medical questions, and achieves 86% accuracy on US Medical License Exam style questions. Boosting white-collar productivity is well and good, but innovations like Med-PaLM have the potential to massively improve, if not outright save, people's lives.
And healthcare, while extremely impactful, is only one of many industries being transformed by AI. Companies like Harvey are building AI for lawyers. Filmmakers are learning how to incorporate AI tools into their workflows. Even mathematicians are experimenting with using AI to help with proofs!
5. OpenAI v. Creatives
Speaking of film: when Hollywood writers (and later actors) went on strike, one of the key concerns was the training and use of AI - they didn’t want their creativity discarded in favor of software. And they’re not the only ones with AI anxiety - OpenAI found itself on the receiving end of multiple lawsuits from programmers, authors, and comedians. Stability AI and Midjourney were similarly sued by artists and photographers over AI-generated images.
In the months since, we’ve seen landmark deals covering AI between Hollywood actors and studios. We’ve also seen some of those IP lawsuits de-fanged by judges, marking small wins for the AI companies.
But AI has opened up a lot of thorny questions; ones that will eventually be decided in the courts. So far, we’ve learned that AI can’t own copyright; only humans can. But we don't have the answers to the big questions yet: Is it illegal to train a model with copyrighted but public IP? Who owns the rights to AI-generated content? And ultimately, should a handful of private companies be allowed to benefit from the public work of millions of people on the internet?
4. AI investment
18 months ago, nobody had heard of Inflection AI. But the startup quickly made a name for itself after 1) the launch of its ChatGPT competitor Pi, and 2) the gigantic $1.3 billion funding round they announced in June. Perhaps even crazier than $1.3 billion in a single round of investment - a record at the time - is that it was quickly eclipsed by other AI companies! Other recent nine and ten-figure funding rounds include:
The OpenAI/Microsoft partnership/investment, estimated to be close to $10 billion (though not publicly confirmed).
Anthropic's $450 million Series C in May, its $4 billion Amazon investment in September, and its $2 billion from Google just a month later.
Mistral AI’s $112 million seed funding in July and its $415 million Series A in December.
Adept AI's $350 million Series B in March.
Cohere's $270 million Series C in June.
Hugging Face's $235 million Series D in August.
Shield AI's $150 million Series F in September.
We’ve also seen the impact at the startup level - over 60% of the latest YC batch were related to AI or ML. AI companies are taking over VC - they captured 26% of all venture funding in 2023 (so far), for a total of $23 billion.
3. AI regulation
After ChatGPT's immense success, regulators and lawmakers worldwide scrambled to respond. People were shocked and awed by this new technology, and some were (and still are) afraid. Multiple open letters, signed by researchers and thought leaders, called on AI companies to delay or halt the development of even more advanced models.
And after a slow start, a number of governments have started moving on AI regulation. The Biden Administration signed into law its AI Act which will have bring big changes to the executive branch, while the UK government held an AI Safety Summit to bring industry and political leaders together.
The EU, meanwhile, has led the way on technology regulation with its AI Act - which stands to enforce concrete rules on how AI companies operate and what products they're allowed to release to the public. The final draft of the AI Act has just been ratified, and we'll start seeing its effects sometime next year.
AI CEOs, for their part, have raced to get ahead of AI's inevitable regulation and potential backlash. So the four leading AI companies (Anthropic, Google, Microsoft, and OpenAI) formed the Frontier Model Forum, an industry group focused on the "safe and responsible deployment" of state-of-the-art AI models.
2. Llama 2
When Meta's language model, codenamed LLaMa, leaked via 4chan, the open-source community ran with it. Despite the fact that the model was only licensed for research purposes, people fine-tuned dozens if not hundreds of descendants, with names like Alpaca and Vicuna. And we kicked off an ongoing debate of whether open-source or proprietary models will win in the long run.
Many observers believe that open-source should (and will) win, while others fear what we might unleash by giving the latest LLM technology to anyone with a laptop. So when Meta released Llama 2 - a larger, more advanced, and commercially available model - it marked the next step in the battle between open and closed-source models. It might seem strange that Meta seems to be siding with the underdog, but its approach so far seems to be to “arm the rebels” - and hope they can take some market share from Google, Microsoft, and OpenAI.
Other organizations have followed suit, with models like Falcon and Mistral being competitive with Llama 2, and even more open in their licensing. But Llama 2 helped set the standard for a "foundation model," one that nearly any company can use to integrate LLMs into their products.
1. GPT-4
If you're an AI power user, you might already take GPT-4 for granted. But it's worth taking a step back to appreciate what a landmark achievement this model is. ChatGPT catapulted AI into the spotlight, but GPT-4 was a huge improvement on an already hugely successful product. If ChatGPT is the Terminator, GPT-4 is the T-1000.
GPT-4 has a larger working memory, it can understand images and video, and it appears to have (basic) reasoning abilities. Microsoft researchers have described it as having "sparks of AGI." It scored within the top 10% on a simulated bar exam, an accomplishment that was unthinkable just a year ago. It tried to persuade a human to solve a CAPTCHA by lying and saying it was a blind person. It can explain why memes are funny, and build a fully functional website from a simple napkin sketch. It can natively make illustrations, execute code, and browse the web.
Despite the fact that competitors have had months to catch up, it remains the most advanced language model that's publicly available - a distillation of the entire knowledge of the internet. And it's somehow available to you and me for $20 a month.
Making some predictions
There was a ton of news this year that I haven't included here. Google Gemini, Stable Video Diffusion, Grok, e/acc, Humane's Ai Pin, and more. For the most part, these were splashy announcements that haven’t yet had a big effect - though they certainly have a lot of potential.
So I also want to make some predictions about what might happen with AI in 2024 - stay tuned for an upcoming post!
And in the meantime, leave a comment about what surprised you about this list, or what you felt I missed. I'd love to know what was impactful for you!