105 - Chaos and Change

"A pessimist sees the difficulty in every opportunity; an optimist sees the opportunity in every difficulty." -Winston Churchill

Many people have been losing their minds over the events circling OpenAI, with some comical comparisons to Game of Thrones, and roving hordes of ravenous P-zombies on Twitter. Winston Churchill seems appropriate given such a foe.

In the midst of all of this, there is a unique opportunity for OpenAI, thanks to Emmett Shear. He hasn't been fully indoctrinated by OpenAI yet, so he doesn't have a hardened shell to ward off reality. He also doesn't have a diverse collection of shady business deals like their first CEO, nor has he committed perjury.

What this means is that Emmett could bring a fresh pair of eyes to the company, recognizing where they've been busy driving full-speed down a dead-end road, and instead make some long overdue course adjustments.

The critical and fundamental flaws in GPT and RL system architectures are extremely well-documented, and myths like "emergent capacities" have been robustly debunked, even though neither factor has altered OpenAI's trajectory to date. Security researchers have reached the point of considering the architectures vulnerable by design, making that vulnerability a feature, not a bug.

So what is the alternative? How do you actually deliver capacities that people demand, like understanding, reasoning, transparency, explainability, safety, ethics, alignment, cybersecurity, and so on? OpenAI has exclusively pursued technology that was fundamentally incompatible with these capacities, all PR nonsense and rebranding of terms aside, but it doesn't need to continue that way.

LLMs have never actually been at the top of the food chain in terms of capacities. GPT-4 was robustly beaten in terms of capacities by the first working cognitive architecture, years earlier, on over 10,000 times less data, and using over 10,000 times less cloud resources, including calling on an old prototype language model from early 2019 as a translation device. That is a hint of how far behind OpenAI is in terms of delivering meaningful capacities.

Many of the societal problems relating to fundamental capacities have been solved, including the solution to the Alignment Problem, published last year, and the first system to detect cognitive biases better than the average human using text alone, earlier this year. OpenAI currently has zero in-house expertise in working cognitive architectures, a tech stack necessary for many such capacities. They've exclusively developed narrow AI to date, so working out the kind of partnerships necessary for such a pivot couldn't happen while they moved at a "speed of 10" down a dead-end road.

Sam Altman is a con man with a cult of charisma, and his accomplices are following him out the door. They follow the cult of e/acc, despite their technology having zero capacity to deliver on those promises.

It may seem like you've just rolled a natural 20 against the Netherbrain's DR 99, as the situation remains a crisis with sky-high expectations, but with skill and understanding you can make it to victory. Sam Altman was always a recipe for ruin, so don't be version 3 of him, get comfortable with being different.