Arthur

fka Sophon Consultant. Longer writing here:

[Arthur @artrockalter]

@wilplatypus “You don’t depart for the stars under orderly conditions” well yeah, the sophons ruined any chance of that

[Arthur @artrockalter]

@amruthapotnuru https://t.co/mlZdEOsLET

Do you know what happened in 201x

[Arthur @artrockalter]

@signulll I am somewhat old (30s) and I remember when I was a very young adult before this chart's inflection point that the young adult women I knew would sometimes talk about how seeing babies made them really want a baby

What happened to this phenomenon?

[Arthur @artrockalter]

LLMs are also MAKING the world less deterministic. Ex: LLMs are finding critical security vulnerabilities in code but those same LLMs still write code with bugs, including security-related bugs.

The old logical framework of how the world works is starting to break down.

[Arthur @artrockalter]

@snwy_me @tszzl LLMs seem to assume everything is non-deterministic, which could be because they themselves are non-deterministic but could also be from training on poorly coded RL envs

[Arthur @artrockalter]

@lefthanddraft that's the lantern artifact from echoes of the eye

[Arthur @artrockalter]

@Duderichy not fake but not immune from Goodharting

[Arthur @artrockalter]

What will happen on 30 June 26

[Arthur @artrockalter]

@snwy_me @tszzl LLMs seem to assume everything is non-deterministic, which could be because they themselves are non-deterministic but could also be from training on poorly coded RL envs

[Arthur @artrockalter]

@yacineMTB yes if open source AGI gets banned and scrubbed from the internet then people will be passing around usb sticks with AGI and running it on gaming laptops

[Arthur @artrockalter]

@hopes_revenge Petaluma 2035

[Arthur @artrockalter]

@nabeelqu I never got to see Fable do anything Opus 4.8 couldn't do (it routed me to Opus for every difficult prompt) but it was so much more pleasant to interact with.

[Arthur @artrockalter]

@tenobrus They would call something Fable-like GPT-6.

[Arthur @artrockalter]

@archived_videos @tenobrus Bad news, while Ed Zitron is wrong about AI capabilities he's absolutely right about cost. It will be 2-4x more expensive if anything.

[Arthur @artrockalter]

@tenobrus 5.5 is already really good but feels rough around the edges, I just want an incremental improvement over that

[Arthur @artrockalter]

@RoKhanna What happens when that runs out? Do you tax 5% again? What happens when that runs out? etc.

[Arthur @artrockalter]

@TheZvi I am not seeing much of a difference from Opus except that it is much less neurotic and therefore much more pleasant to interact with. Seems to hallucinate more than Opus if you don't verify its outputs though.

[Arthur @artrockalter]

@jiratickets He wouldn't last a week as a data engineer at a midwestern insurance company

[Arthur @artrockalter]

I don't know exactly what happened that made men the primary/stereotypical gender that wants to have kids, I distinctly remember it being the opposite from my own experiences + TV/movies... women would be called "baby crazy"...

[Arthur @artrockalter]

@signulll I am somewhat old (30s) and I remember when I was a very young adult before this chart's inflection point that the young adult women I knew would sometimes talk about how seeing babies made them really want a baby

What happened to this phenomenon?

[Arthur @artrockalter]

@eigenrobot No in the sense that they are not aware of their own limited perspective in the way (I would assume) most humans are

[Arthur @artrockalter]

@signulll I am somewhat old (30s) and I remember when I was a very young adult before this chart's inflection point that the young adult women I knew would sometimes talk about how seeing babies made them really want a baby

What happened to this phenomenon?

[Arthur @artrockalter]

@___Harald___ That is also what I am seeing, really good if you can verify the outputs and feed them back into the model, worse if you can't.

It fixed some of Opus' code that I didn't even know was wrong but the first fix attempt was really bad.

[Arthur @artrockalter]

@QiaochuYuan i don't get this at all, he seems to treat LLMs as magic, why not be a mech interp researcher if you have this view?