2023 State of AI Report, GPT-4 rise, safety concerns and future trends
The State of AI Report for 2023 delves into fascinating developments in the AI industry, aiming to ignite informed conversation about AI and its prospects for the future. The report, produced by AI investors Nathan Benaich and the Air Street Capital team, has grown into a widely-anticipated annual event, offering industry and research perspectives on the latest in the AI world since its inception six years ago.
The 2023 report covers wide-ranging topics, most noticeable of which is the rise of GPT-4, credited as currently 'mastering its terrain' for topping all other Large Language Models (LLMs) in both classical benchmarks and exams designed to test human capabilities. This achievement solidifies the strength of proprietary architectures and reinforcement learning derived from human feedback. The report anticipates increased efforts to clone or surpass this proprietary performance amidst concerns that human-generated data may only sustain AI scaling trends for a finite time.
Significantly, LLMs and diffusion models continue to facilitate real-world breakthroughs, notably in the life sciences sector. Substantial developments in molecular biology and drug discovery have been marked as key achievements this year. Timely too is the report's assertion that 'Compute is the new oil', highlighting NVIDIA's record earnings and startups utilising their GPUs as a unique advantage.
The report also spotlights GenAI, which purportedly 'saves the VC world'. Amid a downturn in tech valuations, AI startups specialising in generative AI applications, including video, text, and coding, successfully raised over $18 billion from VC and corporate investors. Meanwhile, traditional community norms around openness face growing pressure from commercial necessities and safety fears.
The safety debate concerning AI has also burst into mainstream discourse, prompting action from global governments and regulators. However, the report points out this rush of activity belies the deep-rooted divisions within the AI community and a marked dearth of concrete progress towards global governance. The complexities of this are confounded as many of the suggested models of worldwide governance require cooperation amongst renowned geopolitical rivals, currently engaged in conflicts around chip production and trade regulations.
Lastly, challenges persist in evaluating state-of-the-art models, as standard LLMs often grapple with robustness. Nathan Benaich and his team at Air Street Capital remark that given the stakes, a 'vibes-based approach isn't good enough'. They emphasise that they compile this report to outline the most compelling occurrences and yearn to instigate an informed dialogue about AI's state. They eagerly encourage feedback on the report, individual perspectives on their predictions, or contribution ideas for their next edition.