Dr. Fei-Fei Li, Stanford professor and co-director of Stanford HAI (Human-Centered AI), and Andrew Ng, venture capitalist and managing general partner at Palo Alto-based AI Fund discussed the current state and expected near-term developments in artificial intelligence. As a general purpose technology, AI development will both deepen, as private sector LLMs are developed for industry-specific needs, and broaden, as open source public sector LLMs emerge to address broad societal problems. Expect exciting advances in image models — what Li calls “pixel space.” When implementing AI, think about teams rather than individuals, and think about tasks rather than jobs.

Rajeev Chand, partner and head of research at Wing Venture Capital in Palo Alto, moderated the discussion.

From left to right: Rajeev Chand, Dr. Fee-Fei Li, Andrew Ng

Will the Current AI Hype Cycle Lead to Another AI Winter?

Ng responded that because AI is a general purpose technology, like electricity, it has many use cases and will continue to grow. Li added that media coverage will go in waves, but this is a “deepened horizontal technology” that is a true transformative force in the next digital revolution. It is changing the very fabric of our societal, political and economic landscape.

Predictions for the Near Future

We are at the verge of exciting advances in pixel space, Li suggested — a shift from large language models to image models. Ng added that this is a shift in both creating images and analyzing images for any situation where you have a camera (example: for self-driving cars).

Public sector AI, or open source AI, will be better resourced and develop alongside of private sector AI, said Li. Open source LLMs will reach the level of closed LLMs, but the development of the two will diverge because they will be developed with very different data sources. The closed source LLMs will deepen with industry-specific knowledge, while the open source LLMs will broaden as they address wider societal matters.

Autonomous agents that can plan and execute a sequence of actions in response to a request are barely working now, but Ng expects significant advances in the near future. Li respectfully requested changing the language from autonomous agent to assistive agent. As the tech rolls out, long-tail results matter and human intervention is required to catch errors such as AI hallucinations. Part of the work can be autonomous, but part of the work must be collaborative with humans.

Human-Centered AI (HAI): Running AI on your laptop is possible, even though it is not as powerful as a full LLM. Ng believes that manufacturers will market laptops powerful enough to run AI locally, and that will trigger another wave of sales as companies and consumers upgrade.

Guidelines for Implementing AI

Li stressed the importance of distinguishing between replacing jobs and replacing tasks. For example, a nurse’s 8-hour shift involves hundreds of tasks, she said, and some of them can be AI-assisted.

Li described the levels of the implementation problem. There is understanding of the data versus the decision-making versus the intention. We are getting very good at understanding patterns in the data. The second, decision-making, is much more nuanced. We are just scratching the surface in AI’s understanding of intention.

Look at ‘team’ rather than the ‘individual,’ and look at the ‘task’ rather than the ‘job,’ Ng recommended. Assess what can be augmented by AI with a clear ROI (return on investment). He stressed that the highest ROI task is often not obvious.

It may be tempting to let AI read an X-ray instead of a radiologist, but the highest ROI for AI implementation may be gathering patient data. He has seen that it is often niche tasks specific to an industry rather than an obvious general tasks that benefit most from AI implementation.

Another way to assess where AI adds value is to look where you have the most good-quality data, said Li. If you can discern repeatable patterns in the data, then the patterns can be actionable.

Video of the 40-minute panel — “Great Minds, Bold Visions: What’s Next for AI?” — is available on the CES site.