The Problem with OpenAI and Open Feedback Loops

This article was sparked by the New York Times article, “How OpenAI Uses Complex and Circular Deals to Fuel Its Multibillion-Dollar Rise,” by Jacqueline Gu and Cade Metz that highlights the interaction between OpenAI and many of its major investors (see table). OpenAI and others are spending billions on building giant data centers.

Electronic Design covers artificial intelligence and machine learning (AI/ML), but from a technical perspective. We still address issues like the electronic supply chain and major business events in our space, like Skyworks buying Qorvo, so this discussion isn’t too far out of our realm.

One thing we do understand and talk about is feedback loops.

AI Investment and Positive Feedback Loops

Engineers use feedback loops — typically negative feedback is more useful, but positive feedback is needed for oscillators. Negative feedback in an amplifier circuit, which has a negative loop gain, tends to reduce perturbation effects. Positive feedback systems increase the changes from an incoming signal that may cause a runaway effect.

In the case with OpenAI, developer of the large language model (LLM) ChatGPT, we usually have a positive feedback loop. That loop could lead to an exploding AI bubble if the circular investments continue without generating sufficient profits to cover the CAPEX costs of building and running the cloud-based AI training and services required by generative AI solutions.

At this point, the returns aren’t keeping up with the investments, which have been significant, although it’s not like the hyperscalers such as Microsoft, Amazon and Meta are betting the house on AI. The amount may seem large, but they’re still a small percentage of those companies’ worth.

The other thing to consider is depreciation costs. The AI market and technology is changing rapidly. Companies like NVIDIA are coming up with newer and faster hardware each year. Swapping out the electronics in a data center year after year isn’t economically feasible. Even trading out electronics after five years can be a challenge because performance is more than doubling every year.

Still, the electronics are only part of a data center’s cost. Data center power requirements are growing to support the AI infrastructure. On the plus side, raw power technology isn’t changing as fast, though GaN and SiC are making power conversion significantly more efficient and cost-effective.

Are You Addicted to AI?

The challenge with AI is that term covers a lot of ground and things in the AI space aren’t created equal. Generative AI, LLMs, and chatbots tend to be how most people are exposed to AI these days. The ability to ask the system to generate feedback based on text prompts hides both the underlying complexity as well as the source of information.

AI-generated output has gotten many into trouble. Lawyers have been fined by judges presented with citations from hallucinated AI responses, and they often make up excuses when they’re busted. Code generated by LLMs, aka “vibe coding,” has been found to contain major security risks. The code duplication and other issues with AI-generated code is increasing technical debt that translates to long-term costs.

It's also important to note that experienced developers using AI tools do better than their peers with less experience because they’re not taking the results verbatim. Such developers can more readily identify issues and address them.

Agentic and Physical AI

Moving to agentic AI allows these tools to make changes without immediate human oversight and insight. It’s interesting how quickly hooks necessary to support agentic AI are being added to software. Previously, the addition of APIs to provide developers access to this type of manipulation was restricted or banned. For some reason, many have greater trust in new agentic AI support.

Most agentic AI users aren’t at the developer level. They simply use it under the assumption that it’s been vetted by a third party.

Physical AI takes agentic AI to the extreme. Physical AI are essentially robots that we’ve had for decades. But now they have the underlying AI models, which are relatively new. Likewise, the idea is to make these robots work more closely with humans. The term cobots is often used to differentiate robots that need to be isolated, like welding robots on an automotive assembly line, and a delivery robot that can be seen rolling through a hospital or down the sidewalk.

Now the question is: “Are you becoming dependent upon AI and is it actually saving you time and effort?” Simply generating images, videos, or text ad infinitum is child’s play for LLMs, but simply filling the bit bucket isn’t useful if the results generate more work. Likewise, how will agentic AI and physical AI play in your home and work environment?

In closing, remember the adage, “It is on the internet so it must be true.”

Now it’s become, “(name of your favorite AI system) said/made it, so it must be true.”