My latest column in @IEEESpectrum. (I prefer the print headline coming in July: "The Other Side of The Innovator's Dilemma.")

@prem_k @burkov Agentic refers beyond LLMs. Those are part of the building blocks. Agents are more like an AI system

Yes. It's probably splitting hairs, but I think we all have a responsibility to remove the veil & show people that LLMs cannot "do" anything other than output text (image/audio/whatev). A separate component then uses typical if-then-else to run what's in / based on the output. Any LLM based app, including so called "Agentic" ones, are only adding a natural language interface to traditional apps. 👇🏼 But in a very expensive manner. I do wish that something more efficient is invented (or discovered) that brings the linguistic capabilities to our apps, without having to throw in the world knowledge into the models.

I regularly receive direct messages from people wondering what's wrong with the so-called agents and agentic frameworks. Here's my answer. The main topic of my PhD was agents and multi-agent systems. What they currently call "agents" (LLMs that were instructed to do something) aren't agents. LLMs hallucinate too much to be trusted with any important task, even if you have 100 "agents" to do the job and 100 to validate it. And even in this case, you can simply use an LLM directly, without using any "agentic" framework, and you will get the same deplorable result. LLMs are only good under two conditions: 1) they are used on data similar to the Web data (which literally means the input must be some Web data) and 2) their output is always used as a recommendation to a human expert (which means that they cannot be programmed to work autonomously, as these framework creators want you to believe). If you only apply agentic swarms under conditions 1) and 2), you will quickly realize you don't have much use cases and you don't need agentic swarms.

Text knowledge is not sufficient for scientific discovery. Language models lack physical grounding. They only have a high level of understanding but cannot simulate physical phenomena. Image and video models focus on "looking good" vs. being physically valid. We’re teaching A.I. physics. We are building physical AI that can model, simulate, design, and control in the real world. Due to the massive amount of data available on the internet, the first wave of generative AI has been digital, but the next wave will be physical. This means that the notion of digital twins needs an upgrade. They need to be trained on laws of physics to have the right foundation. Having physically accurate digital twins means faster design cycles, leading to a huge reduction in R&D costs. Watch my TED talk