TechTakes

On each step, one part of the model applies reinforcement learning, with the other one (the model outputting stuff) “rewarded” or “punished” based on the perceived correctness of their progress (the steps in its “reasoning”), and altering its strategies when punished. This is different to how other Large Language Models work in the sense that the model is generating outputs then looking back at them, then ignoring or approving “good” steps to get to an answer, rather than just generating one and saying “here ya go.”

Every time I've read how chain-of-thought works in o1 it's been completely different, and I'm still not sure I understand what's supposed to be going on. Apparently you get a strike notice if you try too hard to find out how the chain-of-thinking process goes, so one might be tempted to assume it's something that's readily replicable by the competition (and they need to prevent that as long as they can) instead of any sort of notably important breakthrough.

From the detailed o1 system card pdf linked in the article:

According to these evaluations, o1-preview hallucinates less frequently than GPT-4o, and o1-mini hallucinates less frequently than GPT-4o-mini. However, we have received anecdotal feedback that o1-preview and o1-mini tend to hallucinate more than GPT-4o and GPT-4o-mini. More work is needed to understand hallucinations holistically, particularly in domains not covered by our evaluations (e.g., chemistry). Additionally, red teamers have noted that o1-preview is more convincing in certain domains than GPT-4o given that it generates more detailed answers. This potentially increases the risk of people trusting and relying more on hallucinated generation.

Ballsy to just admit your hallucination benchmarks might be worthless.

The newsletter also mentions that the price for output tokens has quadrupled compared to the previous newest model, but the awesome part is, remember all that behind-the-scenes self-prompting that's going on while it arrives to an answer? Even though you're not allowed to see them, according to Ed Zitron you sure as hell are paying for them (i.e. they spend output tokens) which is hilarious if true.

I don't toil in the mines of the big FAANG, but this tracks with what I've been seeing in my mine. I also predict it will end with lay-offs and companies collapsing.

Zitron thinks a lot about the biggest companies and how it will ultimately hurt them, which is reasonable. But, I think it ironically downplays the scale of the bubble, and in turn, the impacts of it bursting.

The expeditions into OpenAI's financials have been very educational. If I were an investigative reporter, my next move would be to look at the networks created by venture capitalists and what is happening inside the companies who share the same patrons as Open AI. I don't say that as someone who interacts with finances, just as someone who carefully watches organizational politics.

I'm terrified for the future, and not even on hater shit. The public numbers are bad, and barring some extremely surprising reports locked behind a wall of NDAs, the private numbers don't seem much better - even Saltman, perpetual cheerleader he is, doesn't have much to offer except desperation to keep the party going, barely even a week after their big model drop.

So if all the big tech players know that this is garbage, the continual doubling down on this either points to: 1. scrambling for the pie while it's there because they need it to stay afloat, or 2. everything else they have to offer is even worse somehow? And in either case, the aura of being a tech company instead of a company is lost, and I don't know what happens in the fallout. The probably best case scenario is that only tech workers like myself have to eat the blowback, but I suspect things won't play out so cleanly.

I also fear that said collapse could be ruinous to big tech, deeply damaging to the startup ecosystem, and will further sour public support for the tech industry.

Yes... ha ha ha... YES!

Hallucinations — which occur when models authoritatively states something that isn't true (or in the case of an image or a video makes something that looks...wrong) — are impossible to resolve without new branches of mathematics…

Finally, honesty. I appreciate that the author understands this, even if they might not have the exact knowledge required to substantiate it. For what it's worth, the situation is more dire than this; we can't even describe the new directions required. My fictional-universe theory (FU theory) shows that a knowledge base cannot know whether its facts are describing the real world or a fictional world which has lots in common with the real world. (Humans don't want to think about this, because of the implication.)

Microsoft is making laptops with dedicated Copilot buttons.

I think they'd rather burn their company to the ground, all the while telling their customers that they just needed to wait a little while longer, rather than admit that they got it wrong.

I worry this is going to leave a crater in the software industry that will take a decade to fill.