AI is changing so fast that it is important to anticipate changes before they arrive. If you want to know what to expect for GPT-5, all signs point towards a virtual brain. When I hear Sam Altman mention the brain I perk up, as this is one of the main lenses I use to think intuitively about AI. Altman’s reference to a “virtual brain” may sound like marketing hyperbole, but I can explain why it likely isn’t.
OpenAI have been aware for a long time about how insights from neuroscience can be used to develop their models. Now everyone else is catching on, and we are seeing agents have a ChatGPT moment. For a long time, researchers have used our understanding of the brain to help AI to listen to you, learn languages, identify cars in images, among many other things. These intelligent capabilities were developed separately but now we are seeing them rapidly converge. However, one fundamental piece of intelligence has yet to be fully integrated into the best of AI: planning.
Enter: Reinforcement Learning (RL) – one of the most successful areas of Artificial Intelligence outside of Large Language Models (LLMs). RL happens to be an area of AI that was borne out of neuroscience and psychology over the last few decades. It is still the most successful approach for developing many kinds of AI, especially robots. Applying a new kind of reinforcement learning algorithm is what brought us ChatGPT from the lesser known GPT-3, triggering the modern AI hype.
Attempting to fully merge RL with LLMs has been an active area of research in AI recently, especially relating to agents. However, to date, no one has fully pulled it off. It is only a matter of time before someone does, and a lot of what Altman is quoted saying in this article makes it sound like OpenAI may have just done it. This would track with what people have speculated about OpenAI’s mysterious Q* model that may have gotten its CEO, Sam Altman, fired.
If you’re interested in how neuroscience relates to AI at a basic level, here’s a quick overview: Reinforcement learning (RL) and deep RL are linked to key parts of the human brain involved in planning and decision-making. Large language models (LLMs), like GPT, are similar to the brain areas responsible for understanding language and its meaning. The processing of speech happens between these areas, and emotions like empathy are believed to be managed near the speech processing region. Therefore, it didn’t surprise me when GPT-4 began to show capabilities in speech and empathy. Just as LLMs can grasp some aspects of language understanding, speech models might also capture elements of empathy through speech patterns. Although it’s crucial not to stretch this comparison too far, since AI development might eventually take a different path from human brain functions, our understanding of intelligence—largely based on human intelligence—suggests that AI will continue to have many similarities with how we think and process information.
“Brodmann Areas.” Kenhub. https://www.kenhub.com/en/library/anatomy/brodmann-areas. Accessed on 13 June 2024.
For those gearing up for agents, be prepared for their potential to reach a whole new level as they acquire the ability to think for themselves. It may still be a year from now before virtual brains are financially viable and/or publicly understood. But I’m sure Industry will be frantically pivoting the moment such a model is released. OpenAI has stated that it is now training what will be GPT-5. I expect it will meet our expectations.
References:
Okemwa, Kevin. “OpenAI CEO Sam Altman suggests GPT-5 may work like a ‘virtual brain’.” Windows Central. 21 May 2024.
Weng, Lilian. (Jun 2023). “LLM-powered Autonomous Agents”. Lil’Log. https://lilianweng.github.io/posts/2023-06-23-agent/.
Ha, Yun. “After ChatGPT and the rise of chatbots, investors pour into AI agents.” CNBC. 7 June 2024. https://www.cnbc.com/2024/06/07/after-chatgpt-and-the-rise-of-chatbots-investors-pour-into-ai-agents.html.
“Hello, GPT-4O.” OpenAI. Accessed on 13 June 2024. https://openai.com/index/hello-gpt-4o/.
“Exploring Architectural Innovations in Deep Learning.” arXiv. 23 March 2023. https://arxiv.org/abs/2303.11366.
Eliot, Lance. “About that mysterious AI breakthrough known as Q by OpenAI.” Forbes. 26 November 2023. https://www.forbes.com/sites/lanceeliot/2023/11/26/about-that-mysterious-ai-breakthrough-known-as-q-by-openai-that-allegedly-attains-true-ai-or-is-on-the-path-toward-artificial-general-intelligence-agi/.
“Removal of Sam Altman from OpenAI.” Wikipedia. Accessed on 13 June 2024. https://en.wikipedia.org/wiki/Removal_of_Sam_Altman_from_OpenAI.
Hart, Robert. “OpenAI says it has started training GPT-4 successor.” Forbes. 28 May 2024. https://www.forbes.com/sites/roberthart/2024/05/28/openai-says-it-has-started-training-gpt-4-successor—heres-what-we-know/.
“Brodmann Areas.” Kenhub. https://www.kenhub.com/en/library/anatomy/brodmann-areas.
CyberKind 
Leave a comment