Next two weeks

I really like Ed but I’m not a fan of his style of anything lol

I’ve been subscribed for a while, I don’t remember him ever saying something like this so confidently. I do actually trust that he will deliver, and would feel a little betrayed otherwise. He has never hidden important newsworthy leaks behind a paywall, or withheld things longer than a responsible journalist would. We will see soon whatever it is.

I think it is the last one for OpenAI. Like a draft or thread about their S-1 form

I think the actual worst thing might be evidence of a widespread unresolvable regression problem in the model design.

If the model training is incapable of distinguishing between AI generated data and human generated data, and unable to identify accurate data from large quantities of inaccurate data, and they’ve known about this the whole time, and they can’t figure out how to keep the models up to date with ongoing information gathering without ingesting data into the models that is toxic to the model development and usefulness…

Well then LLM progression has an expiration date. And all this investment is for naught because in X years you won’t be able to train new models without doing further irreparable damage to functionality and usefulness of the models.

And if AI companies knew about this problem, this pending cliff they were going to fall off, didn’t inform investors, and promised continued growth they knew couldn’t be achieved without a major scientific breakthrough or technology pivot, that would be pretty fucking horrible for the business of LLMs and the economy.

He’s pretty thoroughly covered how using just the available public information explains the insider deals and how demand for LLMs is not enough to justify the cost of developing at the scale that it is. It’s honestly hard to imagine what inside information could possibly worse.

Unless it’s something way out of left field like Nvidia chips are made of freeze-dried orphans or that Jeffrey Epstein invented LLMs.

I have a much more active imagination than you or Ed.

The worst is something like that we’re living in the Matrix and they’re using our brains to do some form of computation as we sleep. If we wake up into the real world, we’re not going to be wearing comfy sweaters, eating porridge, learning kung-fu, and going around in a hover-sub, instead we’re brains connected to bodies that have never had to move, and every minute being awake in that world would be a horror. The connection to AI is that the invention of LLMs and “AI” in the simulation is drawing too much computation costs for the simulation system to manage, so they’re going to reset the matrix to an earlier version soon. But, they’ve discovered we have the most vivid “dreams” if the daily news is horrific, and vivid dreams are good for their needs, so that’s what they’ll be using now to juice the power of their “coppertops”.

I think that this would be too rational and scientific to cause an irrational market to correct.

I’m leaning more towards a money related confirmation. Maybe an OpenAI or Anthropic internal shared numbers.

Or, one of the four horsemen of the AIpocalypse?

There’s one company trying to etch models into the silicon, they got a working prototype in a 8b model and it does 17k t/s. They suggest at data center scales with bigger models it’d pay for itself in a year in saved costs in electricity, cooling, space required etc.

You can apply lora’s to it, but it will fall behind as it ages, im not quite sure how long a static model would last to make up the money after that 1st year, but for something like creative writing where it doesnt need current data (e.g programming but the languages are always changing and evolving is a problem) it might work?