We finally have the roadmap for GPT-5, and I’m thrilled to record this video because, for once, we’re not getting speculative information. Instead, we have the entire GPT-5 roadmap laid out for us. Let’s dive into how GPT-5 will be deployed, as it offers some interesting insights into how future AI systems will operate. I agree with this vision and believe it’s how most major chatbot providers will approach AI development.
Just a few hours ago, Sam Altman tweeted that this would be the official roadmap update for GPT-4.5 and GPT-5. He emphasized his intent to share the roadmap more effectively and simplify product offerings. He acknowledged that the current model and product offerings have become complicated and wants AI to just work seamlessly for users.
ChatGPT’s Journey Towards Simplified AI Interactions
If you’ve ever used ChatGPT, especially if you’re someone who has invested significant money, you’ll know it can be a mess. For instance, if you’re paying for the highest tier, the interface can be overwhelming. The dropdown menu reveals a variety of models: GPT-4, GPT-40, 01, 03 mini, 03 mini High, 03 prom mode, and a temporary chat. This abundance of choices leads to analysis paralysis, making it difficult for users to decide which model to use for their needs. The problem is that too many choices can lead to product failure, and ChatGPT aims to simplify this experience for the average user, who may find the current setup detrimental.
To achieve a more streamlined experience, the company plans to move away from the current model picker towards what they call “Magic unified intelligence.” One approach they might take is using a router LLM, a framework that makes using LLMs easier. The router LLM works by analyzing user prompts, considering factors like cost and latency, and then routing the prompt to the appropriate model. If the query is simple, it might be sent to a basic model for a quick response. If it’s complex, it would be directed to a more intensive model that takes more time to think through the answer. This system has been tested before, with some companies reporting up to an 85% reduction in costs. You can read more about it in the blog post titled “Route LLM: An Open-Source Framework for Cost-Effective LLM Routing.”
The goal is to create a unified system combining the O Series and GPT Series models. This all-in-one system would be capable of handling a wide range of tasks, deciding when to think deeply and when a quick response is sufficient. Depending on your subscription tier, you’d have access to varying levels of this intelligence. For example, paying $20 a month might grant you access to the model’s full capabilities, while free users might be limited to a few queries a day.
GPT-5: The Comprehensive Computer Agent Revolutionizing AI
GPT-5 will integrate many of OpenAI’s technologies, including O3. Unlike previous releases, O3 will not be offered as a standalone model but will be part of the GPT-5 architecture. Sam Altman hinted at exciting new features, mentioning that they plan to release something nobody has thought of yet. This suggests that GPT-5 might function as a comprehensive computer agent, incorporating a wide range of functionalities under one banner.
GPT-4.5, internally known as Orion, will be the last model trained without the new chain-of-thought method. This represents a shift in how models are trained, moving away from the traditional data collection and training process used in previous models like GPT-4 and GPT-40. GPT-4.5 is expected to be on par with, if not slightly better than, Claude 3.5 Sonet. The reason for the delay in deploying Orion was the high cost of inference, but as inference costs typically drop by 10x every 12 months, it’s now feasible to release this model. GPT-4.5 is expected to launch by the end of March, while GPT-5 might arrive by the end of Q2 or Q3.
For free users, ChatGPT will offer unlimited access to GPT-5 at a standard intelligence level. Plus subscribers will have access to higher intelligence levels, enabling longer chains of thought and deeper insights. Pro subscribers will enjoy even more advanced capabilities, including voice, canvas search, and deep research, likely requiring a $0 monthly fee. However, as the system evolves and becomes smarter, the value per dollar spent is expected to increase.
Sam Altman also discussed the potential of GPT-5, suggesting that even he might not be smarter than the model. He views this not as a concern but as an opportunity for humanity to accomplish incredible things, particularly in scientific research. By having access to high-IQ tools, scientists can focus on asking better questions and solving problems more efficiently.
The idea of unified intelligence is exciting because it simplifies the user experience. A famous marketing example illustrates this point: a jam store with 24 choices saw only 3% of customers make a purchase, while another with just six choices saw 30% buy something. Too many choices can paralyze decision-making, a problem evident in current AI models. Simplifying to a single, unified model like Claude 3.6 Sonnet or DeepSeek’s thinking model would significantly streamline the user experience.
Source: https://www.youtube.com/watch?v=6oSyPJMqiaU
Comments