2023 In AI

Some were quick to dub 2023 the year of AI. I’m not so sure such a title is appropriate. As I’ve discussed in a previous piece, the forces driving AI’s development will likely continue into 2024. However, 2023 is the year that many began taking AI seriously.

Sceptics who, twelve months ago, believed AI would be a flash-in-the-pan are unlikely to feel the same way today. In 2023, while lacking the earth-shattering developments of 2022, there was notable refinement in AI, enabling widespread adoption by businesses. A McKinsey survey indicates that a third of respondents report regular use of AI in at least one business function.

There has also been a heightened focus on legal and ethical discussions regarding AI in 2023, encompassing a labour dispute that brought Hollywood to a halt, a landmark copyright case, the introduction of the first AI legislation, and the initiation of the first international summit addressing the potential extreme risks of AI. What was once a peripheral topic has now become a central concern in policy circles.

Q1

After a comparatively quiet start to the year, March saw a flurry of major announcements. Midjourney dropped its v5 image generation model, allowing for higher-fidelity images. Adobe also released a public beta of its own generative model—Adobe Firefly. Firefly integrates directly into Adobe’s applications like Photoshop, meaning that for the first time, generative AI found its way into professional creative workflows.

It wasn’t all good news for AI image generation. Back in February, The US Copyright Office ruled that AI images should not be granted copyright in a landmark ruling involving images created for a graphic novel using Midjourney (Reuters). Subsequent attempts to copyright AI-generated images have since failed.

Elon Musk founded xAI. This was not Mr Musk’s first foray into the AI space. He was an early investor in DeepMind and a co-founder and board member of OpenAI before being removed for a conflict of interest with Tesla, which essentially competed with AI for talent to develop its autonomous vehicles. In 2022, Musk founded Neuralink, a company developing brain-computer interfaces, which Musk has said publicly that he hopes will help merge humans with AI in the future. Against that backdrop, Walter Isaacson says xAI is best understood as a way of tying Musk’s AI efforts—self-driving cars and merging AI with humans—together.

Also in March, OpenAI unveiled its most powerful model, GPT-4, which it made available to users through a paid subscription and via their API. GPT-4 marked a big leap forward from its predecessor, GPT-3.5, the same model that powered ChatGPT at launch. For example, according to OpenAI, GPT-4 scores around the top 10% of test takers on a simulated bar exam, whereas GPT-3.5’s score was around the bottom 10%. GPT-4 also had a 4x larger memory than its predecessor, which proved especially useful for complex tasks such as coding.

A week later, Google dropped its own AI chatbot, Bard. Bard ran on LaMDA, a model Google first unveiled in 2021 but which they’d not made public. According to Google fellow Jeff Dean, Google initially withheld the model, fearing reputational risk. However, the meteoric success of OpenAI’s chatGPT seemingly forced their hand. Although Google was playing catch up, they still had some significant advantages. Similar to Adobe, Google was able to begin integrating AI into the tools that millions already use, such as Google Docs and Google Maps.

In the final days of Q1, one more development would quietly appear on GitHub. Toran Bruce Richards, under the name @significantgravitas, launched Auto-GPT, an “experimental open-source application” connected to OpenAI’s GPT-4 by API. By giving GPT-4 access to long-term memory and internet access, AutoGPT showed GPT-4’s versatility beyond being a chatbot. Rather than simply responding to single prompts, Auto-GPT would try to achieve goals by breaking them into subtasks, using recursive GPT calls without constant human prompting.

Q2

The craze around AutoGPT saw us into the second quarter, as it quickly became one of the fastest-growing GibHub repositories in history and trended on Twitter multiple times. Within days, similar Agental AI projects, such as Baby AGI, began cropping up elsewhere. Then, on April 5^th , a user anonymously posted a 25-minute video demo of ChaosGPT, a modified AutoGPT with the goal of destroying humanity. It created a plan and, after a couple of Google searches, became enamoured with nuclear weapons. It lacked the intelligence to get further than saving a few Wikipedia articles to its file system and posting a quick Tweet. But (perhaps as its creator intended), it did spark a wider conversation by illustrating that in future, it may only take a single bad actor with a more powerful AI agent to cause real harm.

In April, the EU released its proposed AI Act, which it touted as “the world's first comprehensive AI law.” After months of back-and-forth, trying to iron out the details, the AI Act was signed into law signed into law in December. But not everyone’s happy. French President Emmanuel Macron criticised the rules, fearing they’ll hamper innovation and widen the gap between Europe and AI leaders like the US and China. Others worry the regulation doesn’t go far enough with an exemption for open-source models leaving too much risk on the table.

In Q2, the Hollywood double strike began. It all started in May when the 11,500 members of the Writers Guild for America walked off the job. By July, The Screen Actors Guild-American Federation of Television and Radio Artists, the largest acting union, joined the strike. This was the first time writers and actors had joined forces to strike since the 1960s. One of the major sticking points was residuals from streaming services. But there was another elephant in the room—AI.

Writers and actors alike saw the potential for AI and synthetic media to threaten their jobs and disrupt their livelihoods, and they wanted to safeguard against that future. By September, they’d struck a deal, allowing AI to be used as a tool rather than a replacement, which The Guardian suggests may serve as a model for other industries.

In Q2, Meta also dropped two open-source models—Segment Anything and Pixel Codec Avatars—geared towards AR and VR development. Segment anything converts an image into image features. For example, when shown an image of a person seated at a table with a cup of coffee, the model will identify the person, the table and the cup as separate objects. Pixel Codec Avatars 3D models of human faces. However, Meta’s most impressive model was yet to come.

Q3

In July, Meta released LLaMA-2, its own large language model (LLM), in partnership with Microsoft. Although LLaMA-2 was a smaller model than GPT-4 and the older GPT-3.5, it was open-source, making it far more accessible than OpenAI’s offerings. It was also faster and more efficient, making it well-suited to real-time tasks such as chatbots.

Meta wasn’t the only company releasing powerful open-source models. French start-up Mistral dropped their 7B model in September. Despite being a significantly smaller model, Mistral 7B matches or outperformed LLaMA-2, as well as GPT-3.5, on most benchmarks.

AI video creation tool HeyGen released its translation feature, which translated audio and then matched the mouth movements to the new language for eerily realistic video translation (see demo).

OpenAI also had a slew of updates in store for Q3, adding real-time browsing with ChatGPT, and its new image generation model DALL-E 3. DALL-E 3 far surpassed its predecessor, DALL-E 2, excelling at more open-ended prompts.

Q4

In late October, the White House issued an executive order to regulate AI. The order contained a wide array of measures, including risk mitigation, deepfakes, and privacy concerns. It called for mandatory testing of advanced AI models to prevent misuse for weapon creation and suggests watermarking AI-generated media. Furthermore, it addressed potential job displacement due to AI and privacy concerns. While the regulation applies only in the US, many of the leading AI labs have operations in the US, meaning its effects will be felt globally.

Just two days after the order, leaders from around the globe met in Bletchley Park, UK, once home to the early computers used to decrypt encoded messages during the Second World War, for the world’s first AI Safety Summit. The summit’s primary emphasis was on the risks associated with ‘Frontier AI’— cutting-edge general-purpose AI models capable of matching human abilities across a wide range of tasks, particularly the next-generation of LLMs set to dramatically improve on the currently available models. There was also a discussion of the potential hazards of narrow-capability AI in riskier domains like bioengineering.

The summit atmosphere was tense, marked by a clear divide between AI pessimists and optimists. Yet, several key developments emerged from the summit: a consensus on the imperative to conduct rigorous testing of advanced AI models prior to their release, the introduction of a new AI Safety Institute based in the UK, and a proposition for the establishment of an International Panel on Artificial Intelligence Safety. This panel aims for independence from political interference and seeks to provide valuable insights to both policymakers and the general public. But perhaps the most important outcome was an agreement by the 28 countries to continue meeting to discuss AI risk.

On November 6^th , Chinese start-up 01.Ai, founded by veteran AI expert and investor Kai-Fu Lee, released their open-source LLM Yi-34B. The model outperformed Meta’s LLaMA-2, and excelled at multilingual support and translation. It also featured a large 200K token window. Around the same time, 01.Ai achieved unicorn status less than eight months after launching the company.

In a bizarre case of corporate tug-of-war, OpenAI’s CEO Sam Altman was fired before being reinstated days later. Things started on November 16th when the board (minus Greg Brockman) informed Altman that he was being fired. Shortly after the board issued a public announcement, the board said Altman “was not consistently candid in his communications with the board, hindering its ability to exercise its responsibilities." But the board declined to provide specific examples of Altman’s behaviour, sending the internet into speculation overdrive. There was a brief moment where Altman looked set to join Microsoft and start a new advanced AI team, along with any OpenAI employees. That was until over 700 of OpenAI’s roughly 770 staff members threatened to quit if Altman wasn’t reinstated as CEO. After five chaotic days, Altman reversed the ouster. The board would be overhauled—and few of those who opposed Altman would survive the reshuffle.

In December, Google unveiled the long-awaited Gemini—Google’s largest and most capable model yet and their play to reclaim AI dominance. Google’s demo highlighted Gemini’s multimodal capabilities—its flexibility in responding to a mix of visual and text inputs in a plethora of languages. However, the demo was met with a mixed response, with many accusing parts of the demo of being faked.

Someone overshadowed by Gemini’s release, Mistral dropped a follow-up to their 7B model in Q4, 8x 7B. The French company also raised over $414 million, eclipsing a 2 Billion dollar valuation before the year would end.

Also in December, Microsoft released Phi-2—a small language model that outperforms many of its much larger competitors. Microsoft says “textbook quality” data was part of the key to Phi-2’s outsized performance. It shows that computing power, while much easier to measure, is not the only factor governing AI development.

And that’s a wrap. Subscribe to The Intelligence Age to keep up with everything AI in 2024.

AL