If I were in Sam Altman's shoes
Sam Altman argues that memory sitting idle in personal computers could be used more efficiently in data centers. His reasoning is that instead of everyone keeping memory on their own computers to run artificial intelligence, the same resources could be used more intensively in the cloud.
The first thing that came to my mind when I read this view was the variables left out of the equation.
There is no memory crisis today that requires us to give up on personal artificial intelligence. I do not find it correct to seek a solution by gathering all resources in data centers today, based on a potential bottleneck that might occur in the future. How do we know that production, memory management, models, and data representation formats will remain the same?
One of the most common mistakes when calculating the future of technology is to assume that today's production and usage methods are fixed. If more artificial intelligence is used, more memory is needed; therefore, more memory must be produced and placed in data centers. Such a calculation can be made. However, if you exclude all the engineering possibilities in between, the result you reach is predetermined.
A memory architecture that goes beyond traditional production methods can be developed. Memory management can change. A compression method that represents data in a way that takes up less space could render a portion of the capacity we deem necessary today obsolete. Models can learn to do the same work with fewer resources. The impact of each development on speed, quality, and cost is evaluated separately. However, by ignoring all these parameters, we cannot calculate future needs simply by applying a multiplier to today's memory consumption.
Moreover, these are not all possibilities belonging to the distant future. Google's Turbo Quant study is a concrete example that aims to reduce memory requirements by compressing data in the KV cache used by artificial intelligence models. Such research alone does not solve all memory problems. However, it shows that alongside the number of chips to be produced, how existing memory is used will also be a determining factor in solving the problem.
Let's recall the period when Gmail emerged.
Storage space was one of the primary problems for users in email services. The mailbox would fill up, and one had to delete old messages to receive new ones. Google introduced Gmail in 2004 with 1 GB of space. In 2005, it increased the capacity and placed a constantly rising storage counter in front of users.
While that counter showed that capacity was growing, it also gave users confidence: The task of providing space was the responsibility of the service provider. The user would not have to think about which email to delete for every new one received.
We cannot underestimate the role of this approach in Google's acquisition of users. By providing a solution to a significant capacity problem in the sector, the company created a strong reason to be preferred.
Email storage and artificial intelligence's working memory are different technologies. The lesson I draw from the Gmail example, however, can be applied to both: Today's capacity limit is not an immutable law of nature that defines tomorrow's product. What is expected from a technology company is not only to build a business model around this limit but also to develop ways to exceed it.
I believe that the direction in artificial intelligence will also increasingly expand toward user devices. People will want to work with their own documents, their own data, and on their own computers. They will want to perform every operation without sending it to another company's infrastructure; they will want to be less dependent on connection, access, and service conditions.
I am of the opinion that this will weaken the weight of data centers within the current structure. Especially in the task of generating responses for the user, as the share of local hardware and more efficient models grows, the power of centralized infrastructure to be the only option will diminish.
I think Sam Altman also sees this direction. The fact that OpenAI has released gpt-oss models that can be run locally is a step that suggests the company has realized this need. My suggestion is for him to turn that step into a product that everyone can use.
If I were in Sam Altman's shoes, I would immediately release a local ChatGPT.
The user would download the application, install the model suitable for their hardware, and work on their own computer with their own documents. They would handle updates, installation, and daily use with the ease they expect from ChatGPT. They would clearly see the process taking place on their own device versus the process sent externally, and they would make their own decisions.
Thus, OpenAI would become the company that meets the user's desire to run artificial intelligence on their own computer. Going beyond just publishing model files, it would make local artificial intelligence usable for the masses.
There is also a sectoral consequence to trying to stand in the way of this trend. The concentration of access to artificial intelligence in the hands of a few large data center operators paves the way for monopolization. The opportunities the user has to develop alternatives, use their own hardware, and change their service provider when necessary would narrow.
While advocating for the widespread adoption of artificial intelligence, one must also advocate for the widespread ability to run it. The user's control over computing power is a value that should be included in the efficiency calculation.
I do not believe this movement can be stopped. More efficient models, different memory usages, and users' desire for independent work are all moving in the same direction. The choice before OpenAI is how early it will be a part of this transformation.
If I were in Sam Altman's shoes, instead of debating where the memory on people's computers could be used more efficiently, I would develop the ChatGPT they would want to run on that memory.