Last poster before threadlock gets a cookie (cookie thread (Part 7)) (Part 11)

Psychological disorders.

2 Likes

everyone is telling me this (you and one other person) ja

it’s just mm
…and slightly selfishly the other is better for creating characters

3 Likes

I have made a few predictions on AI and they’re been very largely true. I’ve said for a while that I think the bubble will pop & the breaks will hit before it can generate commercially viable media and the odds of that occurring looks ever more likely as time passes. And if you think AI couldn’t ever possibly generate commercially viable media, all you need to do is look at the people who are consuming slop now.

The second thing is that I predict they’ve handed the CIA a really fucking nasty vulnerability. Like in a cryptographic algorithm or in TOR. And did Anthropic (and OpenAI) really not know about nginx2shell, php2shell and java4shell before everyone else?

1 Like

I feel like this has been a much warmer November than previous years, not to jinx it.

1 Like

why do I care about it being selfish it’s what classes I’m selecting of course it’s a little selfish

2 Likes

HE SAID IT

FOUR THOUSAND INCHES OF SNOW DIRECTED TO ASH IMMEDIATELY

5 Likes

Picking up the cat and carrying her around before I go play Deadlock so she hates me and doesn’t try to get in the room

4 Likes

Well, at least you’ll be able to avoid creating Toko if you study psychological disorders

3 Likes

Tutuu

3 Likes

…she’s already meowing at the door, mission failed

3 Likes

let me in

2 Likes

I got Mina. Exciting. I will do bad

3 Likes

My understanding is the training costs are so high and operating margins are so low that the debt is amortized over a longer period than the model remains relevant. I’m willing to be wrong about that though since I don’t know how much older models get used as new ones are released. However, Anthropic’s IPO filing revealed about $900 billion in debt and they’ve made about $30-40 billion in revenue this year which does make me think they’re still a long way from having a balance sheet in the black (if ever).

1 Like

use your 2 to avoid predictable stuns/hits

you’re immune for like .6 seconds

rake is best as a kill move since it deals more damage comparably to how low the person’s hp is

1 Like

Ah, I understand your point now!

I think you’re definitely right in the significant advantage their internal models will have.

Although I think you may also be overestimating the extent to which they use specialized models for the actual difficult problem solving - and I think that advantage instead comes from their highly specialized scaffolding + specialized supporting agents (and effectively infinite compute + resources).

As I understand it the way that the these sort of auto-research tasks are done is by having swarms of agents working on the problem with various other kinds of agents as the scaffolding surrounding the swarm. Things like monitoring, validator, evaluation, or benchmarking agents. Those supporting agents are likely to have some of them as specialized agents to support the research agents, but the research agents themselves are often just the frontier models without any specialized training.

This is a notable shift from a few years back when solving these kinds of problems was done by specialized agents rather than generalist agents, but a relatively recent shift has seen better performance by the generalist reasoning agents (not sure if that’s because there are gains to be had from having stronger general reasoning skills that apply cross-discipline, or if that’s just because generalist agents have had significantly more investment in their training by at least an order of magnitude, or both).

All that said, a lot of my understanding on this comes from a conversation with the Anthropic reps who came by to advise us on how to begin our own auto-research program. And how Anthropic does their own internal research was more of a side tangent than the primary focus, so this could very well be over-simplified + I could be misunderstanding/misremembering + they would obviously leave out any details that weren’t already publicly available.

2 Likes

Oh yeah, their finances are basically betting on solving the “trillion dollar problem” of “having to pay people wages”

Like, the situation is fucked and I personally am concerned that it’s not as much of a bubble as people want it to be, because I’m worried about the possibility that they just might solve that ““problem””

But the thing that stands is that if it is a bubble, the model weights still exist, and even if they go under, the weights and ability to run the models will still exist at the current prices, and whoever acquires them can still turn around and run the models at current prices.

It’s not a comment on “will Anthropic + OpenAI continue to exist?”, it’s a comment on “will the models that they’ve developed continue to exist and be runnable at current prices?”

3 Likes

Like, I see the argument for “they don’t go under but now all the frontier labs decide to just start charging a lot more for their models because the bubble burst and they need to re-coup their losses”

But then they just get undercut by Chinese models which are also very, very good and are able to be subsidized by the Chinese government itself. (Not even to mention the possibility of the US government just bailing out these companies in response too, hurray! The taxpayers get to pay for their losses while the rich get to keep the profits! I love capitalism!)

So I don’t really see that as a terribly likely scenario

2 Likes

I’m not that worried because it’s great if you can fire most of your employees up until the moment no one can afford to buy your goods anymore. You know how there are champagne problems? I think displacing that amount of labor would be something of a guillotine problem.

3 Likes