2024 AI Forecast: After Claude 3, What Else Can We Expect? | Yunqi FutureScope

In the early hours of March 5, Anthropic dropped three "blockbuster" announcements in a single, terse X post: three models at once — Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus, ranked from lowest to highest capability. According to Anthropic's benchmarks, the Claude 3 series has surpassed GPT-4 across reasoning, math, coding, multilingual understanding, and vision, "setting a new industry benchmark."

Currently, the "extra large" and "large" sizes of the Claude lineup — Opus and Sonnet — are already available through the Claude API. A day later, reviews from across the industry have largely come back positive.

A year on the ground, a day in AI. Faced with AI's breakneck pace of "involution," what can we still expect in 2024? Seven faculty and researchers at Stanford HAI once offered their predictions. HAI is the "Human-Centered Artificial Intelligence Institute" initiative led by Stanford professor Fei-Fei Li and former Provost John Etchemendy.

Technology ultimately delivers value through "people." From this angle, this edition of Yunqi Capital's "FutureScope" shares the following with you — and we welcome those actively embracing AI to exchange your insights and predictions with us.

This article is excerpted and translated from Stanford Human-Centered AI Institute

1. Agent, Some Real Agent

Peter Norvig, Distinguished Education Fellow, Stanford HAI

I think there are two things to expect. First, the rise of Agents — AI that can directly connect to other services to actually "get things done." In 2023, we could all chat with AI, but the interaction was always you input something, it inputs something back. In 2024, this interaction will expand again. We'll see Agents capable of completing tasks for you: booking flights, planning itineraries, connecting to other services.

Second, multimodal media is a definite direction. Right now, our model training is very "intentional." People write pages and paragraphs about what they find interesting and important. When someone presses the shutter and points the camera, believing something is happening, they take a photo.

But with video, there's far more to train on. People make films, tell stories the same way as with text. But some cameras record around the clock — they capture "everything that happens," without any filtering, without any intentionality, achieving a better understanding of "everything that is happening."

2. Clearer, More Specific Policy and Regulation

Fei-Fei Li, Co-Director, Stanford HAI

The policy direction for AI in 2024 is especially worth watching. So far, the most progress came in 2023. In July, a bipartisan, bicameral bill called the CREATE AI Act was introduced in the U.S. Congress, providing students and researchers access to AI resources, data, and tools. It gained broad support because it promised to broaden the channels of AI development. In late October, the U.S. President signed the Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, showing that the government is committed not only to fostering a vibrant AI ecosystem but also to harnessing and governing AI technology. I hope that in 2024, we see more action. We need more legislation like the CREATE AI Act, compliance with the elements required by the new EO, and greater investment in the public sector, integrated with the values we uphold.

3. Beware DeepFakes Already Happening

James Landay, Associate Director and Faculty Director of Research, Stanford HAI

People in videos are "saying" things they never actually said — new multimodal models, especially in video generation, are making this easier and easier. Legislation and regulation must step in forcefully. Currently, we're seeing the EU enter its final stages of crafting broad AI rules.

In 2024, we'll see more companies releasing the next bigger models and various new features. We'll also see plenty of debate over "Is this AGI?" and "What is AGI?" People don't need to worry about AI taking over the world, but we should worry about more practical, currently happening harms — misinformation and deepfakes. In 2024, I'm afraid we'll see quite a few incidents like these.

4. GPU Shortage: Challenge and Opportunity

Russ Altman, Senior Fellow, Stanford HAI

Global GPU processors are facing a supply shortage. Big companies (and more big companies) are all racing to bring AI inside their organizations, so GPUs are somewhat in short supply. A few companies manufacture these processors (NVIDIA is the standout among them), and their production capacity may already be maxed out. This is a competitive advantage for companies, but also for entire nations that don't want to miss out on AI innovation.

This not only puts enormous pressure on increasing GPU production, but also enormous pressure on innovators to propose cheaper, easier-to-manufacture and use hardware solutions. There's extensive work underway in electrical engineering at Stanford and elsewhere on low-power alternatives to current GPUs. Some of my colleagues, including Kunle Olukotun and Chris Re, are making efforts in this area. Additionally, one of Stanford HAI's Hoffman-Yee projects focuses on this direction. While mass adoption is still some distance away, these ideas will bring new inspiration for alleviating the GPU shortage.

5. Knowledge Worker Productivity Gains

Erik Brynjolfsson, Professor and Senior Fellow, Stanford HAI

I expect that large-scale enterprise adoption of AI will bring some productivity gains we've long been waiting for. It will affect knowledge workers — the people who were essentially untouched by the computer revolution of the past 30 years. Creative professionals, lawyers, finance professors — their work will change significantly this year. If we embrace it, it should make our work better, let us do new things we couldn't do before. AI is unlikely to fully automate any job — but it will certainly augment and expand what we can do.

6. Confront the Hardest Questions, Find the Most Practical Answers

Ge Wang, Associate Professor at CCRMA (Center for Computer Research in Music and Acoustics), Senior Fellow, Stanford HAI

One of my hopes for 2024 is that we have enough capacity to keep asking some thorny questions, critical questions: What role do we want AI to play in our lives, our communities, our education, our society? I don't think we've ever seen a year like this. More and more kinds of AI technology will embed themselves in our work, entertainment, and communication. What does this year make us feel about ourselves?

We need to give ourselves the time and space to articulate what we think is permissible, and where we should draw limits. Back in February 2023, Springer (the academic publisher) issued a statement saying that large language models could be used in drafting articles, but not as co-authors on any publication. They cited accountability as the reason, and I think that's extremely important. This doesn't mean Springer is locked into this position forever. But the critical point is this: to seriously put something out there, understand what your reasoning is, and say this is how we understand it now — we may add more nuance to these policies in the future. I think institutions and organizations must have this perspective, and strive in 2024 to get guidelines down on paper.

7. Companies Will Begin Grappling with Complex Regulations

Jennifer King, Privacy and Data Policy Fellow, Stanford HAI

The focus of AI regulation in 2023 was mainly on the EU AI Act. However, by mid-2024, two U.S. states — California and Colorado — will pass regulations addressing automated decision-making in consumer privacy. While these regulations are limited to AI systems trained on personal information or that collect personal information, both states provide consumers with the right to opt out of AI systems.

So companies must start thinking about what it means when customers exercise their rights, especially collectively. If you're a large company using AI to assist in your hiring process, and hundreds of potential employees request to opt out of that process, what happens? Do we revert to human review? Can AI guarantee a different or better process? We're only just beginning to tackle these questions.

These questions raised from a "human" perspective explore, in more practical and grounded ways, how we can better "use" AI. We'll continue publishing in-depth AI reflections in "FutureScope" — see you next time.