AGI might have an efficiency problem
December 7, 2025
Welcome back. Major model developers have made the lofty vision of artificial general intelligence their North Star. But if we ever get there, will we even need it? Using AGI for every task might be overkill, Ruchir Puri, chief scientist at IBM, told The Deep View.
Puri sat down with The Deep View to discuss the problems with existing perceptions about AGI and how enterprises may get more out of their AI deployments with focused, niche models. This interview has been edited for brevity and clarity.
1. AGI might have an efficiency problem
2. AGI versus “Artificial Useful Intelligence”
3. AGI’s “bag of tricks”
HARDWARE
AGI might have an efficiency problem
Nat Rubio-Licht: What do you think are common misconceptions about AGI?
Puri: In some ways, we did ourselves a disservice by labeling it “artificial intelligence.” Anytime you label “intelligence,” and you put “artificial” in front of it, it conjures all kinds of feelings. What is it going to do to humans? Is it beneficial for us? Is it not? When we give it a middle name, call it “general,” then it amplifies those feelings even further.
I care less about general intelligence than I do about making things useful. We need to make AI capabilities and technology, specifically generative AI, really work in use cases that are incredibly useful, but also narrower. I'm not looking for something generalized. I'm not looking to replace humans. I'm looking to augment subject matter experts and enterprise workers who are trying to get a day-to-day job done. That's the difference between AGI and AUI, or artificial useful intelligence.
Rubio-Licht: What are your thoughts on the promises that AI developers are making about the future of AGI?
Puri: If you focus on artificial general intelligence, it drives the progress in technology. It makes the underlying technology better and better and better. The misguided part is that I don't need artificial general intelligence to write an email.
I'll give you an example. Let's contrast artificial general intelligence with real intelligence. Real intelligence, which is our brain, lives in a 1200-centimeter cube, consumes 20 watts and runs on sandwiches. That's the efficiency of a human brain. A single GPU board of Nvidia Blackwell chips consumes 1200 watts, 60 times more, and you need tens of them, if not hundreds of them. You're talking about a difference of three orders of magnitude in efficiency.
What I'm saying is: don't use artificial general intelligence to solve very specific enterprise tasks. Usefulness implies solving a problem with the cost that I need, with the efficiency that I need and where I need it.
TOGETHER WITH YOU.COM
Successful AI transformation starts with deeply understanding your organization’s most critical use cases. This practical guide from You.com walks through a proven framework to identify, prioritize, and document high-value AI opportunities.
In this AI Use Case Discovery Guide, you’ll learn how to:
Map internal workflows and customer journeys to pinpoint where AI can drive measurable ROI
Ask the right questions when it comes to AI use cases
Align cross-functional teams and stakeholders for a unified, scalable approach
Get the Guide.
WORKFORCE
AGI versus “Artificial Useful Intelligence”
Rubio-Licht: Let’s go back to the concept of “artificial useful intelligence.” How can businesses actually get use out of their AI deployments?
Puri: The very first task is – no pun intended – choosing the right task. There's been a lot written about how only 5% of AI cases or applications are getting value. This happens because people are not thinking through what tasks to solve.
Second is evaluations. If you set the right evaluations, then you don't know what success actually means. Define the right success criteria.
Third is to optimize. Don't expect miracles in the first go. Learn and optimize your implementation as you go along.
Rubio-Licht: What’s a common mistake that enterprises make when deploying AI?
Puri: The number one mistake is a mismatch of expectations. People either choose a problem that’s too hard or too easy. You need a sweet spot to roll things out, and once you get the trust in the technology, then you can amplify it and roll it out further. Some pilots fail … and then people are already apprehensive. They lose trust in it.
Second is having the right data. Having the right knowledge for the task at hand. Don't use general intelligence as a hammer. Enterprises are trying to solve particular tasks. You need knowledge in that domain to really guide and solve that task appropriately to get the ROI that you need. This is why we at IBM are huge fans of open ecosystems, specific models, and small-to-medium-sized models that can be purpose-built for the task, versus general intelligence, which somehow miraculously replaces humans.
Finally, I think the most important one overall for this entire field is educating your workforce, building the skills to really infuse AI as part of your workflows.
TOGETHER WITH YOU.COM
AI is all the rage, but are you really using it to your advantage? This practical guide from You.com guides you through the process of identifying and prioritizing opportunities where AI can best serve your org. Explore where AI can drive the most impact in your organization, both internally and externally.
RESEARCH
AGI’s “bag of tricks”
Rubio-Licht: What is the purpose of generalized intelligence?
Puri: I do think that by focusing on generalized intelligence, it does drive progress in technology from a reasoning point of view. New techniques are emerging, new architectures are emerging, memory built into the models, statefulness – that's emerging, and it does drive progress in the underlying technology, making it stronger and stronger. Is that generalization needed for every problem that you want to solve? We certainly believe not.
You can build more specific, purpose-built models that are much more efficient to solve that problem at hand. Although those models also benefit from continuous progress on this generalized technology, you don't need the whole bag of tricks.
Rubio-Licht: Are there risks involved in the pursuit of AGI?
Puri: As a society, we have to be very careful on how we are developing this technology, and develop it in – and it's an overused and abused word – responsible ways. This technology is built upon data that reflects captured knowledge. Captured knowledge can be biased, and it may propagate certain biases in society that exist. We have to be very, very careful in putting guardrails around things. There are a tremendous number of benefits of this overall general technology, but there are tremendous “gotchas” that we need to be careful about.
LINKS
AI-powered entertainment leader Netflix buys Warner Bros. for $83B
User growth has slowed on ChatGPT, according to new report
CNN, USA Today, and Fox News sign licensing deals with Meta AI
SpaceX reportedly in talks for an $800 billion valuation ahead of potential 2026 IPO
AskCodi: An OpenAI-compatible orchestration layer for composing models on top of any LLM
Z-Image Turbo: Open-source tech-to-image generation model for photorealistic images
A QUICK POLL BEFORE YOU GO
What matters more for AI right now? |
The Deep View is written by Nat Rubio-Licht, Faris Kojok and The Deep View crew. Please reply with any feedback.
Thanks for reading today’s edition of The Deep View! We’ll see you in the next one.
“The handrails on the wooden bridge are not at the same height; perspective is off between the bridge deck, handrails, and overhead trellis.” “Look at the reflection of the top of the bridge in the fake one. It is far too jagged to be a reflection. Also, the water is just too white without a hint of movement.” “Too out of focus for AI” |
“Stones and reflection feel so unreal in [the other image]. It does not make sense.” “No!!! That stone bridge reflection is totally wrong... oh well.” “Shadows on [the other image] look misaligned.” |
Take The Deep View with you on the go! We’ve got exclusive, in-depth interviews for you on The Deep View: Conversations podcast every Tuesday morning.
If you want to get in front of an audience of 450,000+ developers, business leaders and tech enthusiasts, get in touch with us here.