Objective-Driven AI

Where we are going (years?)

About a month ago Yann LeCun , the chief AI scientist at Meta, gave a talk at MIT titled "Objective-Driven AI: Towards AI systems that can learn, remember, reason, plan, have common sense, yet are steerable and safe." But let me start with the optimistic statement that LeCun made at the end of this talk:

AI will bring a new era of enlightenment, a renaissance to humanity.

LeCun doesn't think we will get there with the current large foundation model approach -- with an auto-regressive generative architecture. What LeCun is highlighting is the problem with the current generative AI architecture in producing reliable outcomes. The current solution to this limitation is a human in the loop for steerage and correction which can then moderate the output of these models -- as LeCun points out, while the performance of current models is amazing, "...they make stupid mistakes: factual errors, logical errors, inconsistency, limited reasoning, toxicity..." LeCun envisions a new set of fundamental breakthroughs that will eliminate this requirement for human moderation and thus make AI much more valuable.

What LeCun envisions is a multi-component architecture for the next generation of AI systems. From his presentation --

• Perception: Computes an abstract representation of the state of the world (possibly combined with previously-acquired information in memory)
• World Model: Predict the state resulting from an imagined action sequence
• Task Objective: Measures divergence to goal
• Guardrail Objective: Immutable objective terms that ensure safety
• Operation: Finds an active sequence that minimizes the objectives

There are a number of advantages that LeCun proposes would come from this approach. First, it could eliminate one of the downsides of the current models, the need for reinforcement learning from human feedback (RLHF). This work can be quite harrowing for the humans that do it since they are reviewing the potentially toxic output with the objective of restraining the model from generating that content. It also is imperfect and can still result in toxic and incorrect generative output from the models. LeCun envisions a system that can plan toward objectives which he proposes would be inherently safer.

LeCun's full presentation slides explore this topic in detail although without the voice over some will be impenetrable, especially if you aren't closely following the field. There are two important take aways to think about even if you don't have an interest in the specific technical details:

First, yes all the critics of LLMs are right that there are some severe limitations to the way that LLMs operate. While I believe that with the right human supervision, these tools can create enormous value, we will need to continue to develop new approaches to improve and correct for these limitations.

Second, this research is happening - LeCun's paper is just one example of research work that is focused on developing new architectures which will continue to improve the capabilities of these tools. We should expect advances over the coming months and years which will rapidly increase the capabilities and reliability of AI systems.

I for one hope that LeCun's optimism is correct and already feel that we are in a new period of renaissance for humanity.