AI can't learn without asking

I ran across this great video from IBM Technology on "Can AI Think? " It points out that LLMs (and neural networks in general) perform Probability Pattern Matching. It mentions the development of Inference-Time Compute (aka Test-Time Compute), in which Agentic LLMs become more iterative and have them check and refine results by spending longer "thinking." The results of this approach are quite impressive. YouTuber Matthew Berman demos DeepSeek R1, showing R1 planning and chaining 'reasoning' together to create the game Tetris from a 1-sentence prompt.

?This advancement increases the scope of what GPT AI can tackle. OpenAI's paper from this week describes how their Inference-Time Compute improved the quality and robustness of their o1-preview model. To oversimplify, agents run a prompt through a sequence of LLMs, enabling them to check and refine the output multiple times for a better result.? Another outcome of this approach is better protection against adversarial attacks, such as intentionally malformed prompts (i.e., soft-token.)

?However, one thing that continues to be glaringly missing is the ability of these agentic models to ask questions. ?Even in the Inference-time compute, the model is essentially 'asking' other models (which are always limited) for further input and analysis. Using the Socratic method to learn is completely missing from AI Models today. At this time, any Socratic dialog between humans and AI is one-directional only. The human can learn more by asking questions of the GPT chatbot. But it doesn't go the other way, and it should.

To more fully simulate human intelligence, AI models need to learn by asking questions. They need to know when to ask someone else a question, and more importantly, they need to know how to find the right person to ask. Then, they need to incorporate what they've learned (and probably verify it) into their model.

Yes, some AI models ask questions, but this is typically only to help clarify gaps identified in the prompt. Sometimes, the LLM model may generate the questions only because of its probability pattern matching.? The one method I'm aware of that comes closer to truly 'learning' AI is Reinforcement Learning/Training.? But some are looking in the right direction. A just-published paper titled "Collaborative Framework for Dynamic Knowledge Updating and Transparent Reasoning with Large Language Models" looks at how LLMs can be combined with Knowledge Graphs that can update dynamically. Google DeepMind released a paper on how a 'Boundless Socratic Learning" approach could enable AI models to be self-improving.

Until an AI agent can dynamically ask me questions like? "Why do you want to know?" or "Who is the doctor you're mentioning?" AI remains a helpful information savant but is still far from being able to reason or think. Anyone telling the press we're close to AI actually thinking or even the singularity is only trying to cash in on the hype or delusional. Or both. There's still a lot of (human) work to move AI forward, but that's what keeps it interesting. I’m looking forward to it.

?

?#ArtificialIntelligence

#LLM #LargeLanguageModels

#LargeConceptModels

#AIResearch

#MachineLearning

#NaturalLanguageProcessing

#TechInnovation

#AITrends

#SemanticAI

#FutureOfAI

?

[Rick Munoz started working in AI at Symbolics, Inc. in the 1980s. He went on to incorporate AI components like Expert Systems, Natural Language Processing, and Fuzzy Logic into multiple systems. He currently designs and implements large cloud-based applications that include AI capabilities.]