What OpenAI should have been

OpenAI is fucking awful. We all know this. But I want to offer a vision of an alternative, better future — what could have been, had they not been a techno-cult of privileged power-hungry tech bros totally divorced from reality, but instead people genuinely dedicated to the project of making "AGI" that benefits all of humanity.

Imagine, if you will, a non-profit foundation, incorporated in a jurisdiction that holds non-profits legally to their charter, with a board representing a wide variety of diverse (intellectually, culturally, as well as the standard metrics of diversity) scientists and engineers as long as they were interested in the project of making generally useful AI.

Imagine that it was dedicated to the developent of, not general machine learning — as there are already organizations for that — but specifically computers that can reason, understand, and act as assistents for all of us (the underlying vision of AGI and what LLMs have partially delivered on). Not with the goal of replacing human labor, but with the goal of augmenting each and every one of us, to make it easier for all of us to learn, to find information, to program, to write, and do all the other things that such an AI might one day help us do. And not with the nebulous, religious goal of "AGI" either, but with the specific, concrete goal of expanding the map of cognitive tasks that computers can help human beings with.

Imagine that this organization did all of their research out in the open, reporting regularly on their progress and how they achieved it, the specific specifications and training methods they used, the energy and water they consumed and its emissions, and the data they trained on, as well as working with the global scientific community to put together objective benchmarks to reliably determine how useful the projects they were working on were. Imagine that they trained a single line of foundation/frontier models, with public specifications, weights, and benchmarks and training data, so that anyone could benefit from this encapsulation of all of the knowledge and intellect of humanity — and the work of tens of thousands of RLHF workers — allowing anyone to take the model, distill it, fine-tune it, and use it however they wished, and preventing the duplication of work that we have now, where competing companies are all training their own immensely resource intensive frontier models, separate from each other, with all the knowledge they gain from that siloed off.

Imagine that they functioned as a research institution similar to Xerox PARC, Bell Labs, or the MIT AI Lab, with a healthy atmosphere of various teams trying out all sorts of different approaches, cross pollinating, helping each other, but also competing and criticizing each other's work, as iron sharpens iron, instead of the entire resources of the entire organization being sucked into a single approach. Not only exploring state space and diffusion models, but neurosymbolic approaches as well, and even different ones I don't know of.

Imagine that it combined the dedicated research into explainable AI and being able to tweak model concepts and weights in detail of Anthropic (which they could actually use to de-bias models, instead of just censoring them or making us safe from some nonsensical AGI doomsday), and the focus on training powerful models of OpenAI, with the attention to efficiency of DeepSeek and Google.

Imagine that this non profit, similar to the Linux Foundation, was funded by all the various corporations that thought they could benefit from taking the technology it would develop and integrating it into their products or modifying and tweaking it in various ways.

Imagine that this organization wasn't bound by the reckless logic of "scale at all costs" and the belief that they always had to be the first, so they were able to take their time — reckless scaling may be the quickest way to increase model capabilities, but it is certainly not the only way; they could focus on increasing the efficiency of model training and inference, with respect to energy and water, but also data itself, as well as taking the time to notice things like diminishing returns and the fact that you need more data than model parameters to ensure models begin to generalize instead of memorize, meaning that medium-sized models, not vast one trillion parameter ones, are probably the way forward.

This willingness to wait would also let them try different approaches to getting the data annotation work they needed, and work with communities to build their datacenters in ways that benefitted them, and wait to do so until cooling technologies and power were available that allowed such datacenters to be less extractive.

Yes, the development of AI as useful as current frontier models under this paradigm may have taken much longer, but it would have been better.