• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI
Reaction

Image-generation frustration fuels a jab at AI superintelligence

A user called current AI “definitely not ASI”—artificial superintelligence—saying image generation had been “dumb as bricks” during an attempt.

Lucas Beyer (bl16)LB
(((ل()(ل() 'yoav))))👾('
MMitchellMM
53 Sources, 13d ago, first seen 13d ago

TLDR

After quoting the line “what we have now are not LLMs” (large language models), a user followed up with a complaint about image generation. In that reply, they called current AI “definitely not ASI”—artificial superintelligence—because the image generator had been “dumb as bricks” while making an image.

Combined views

450.6K

53 Sources, first seen 13d ago

4.2K likes257 comments689 saves344 reposts

Combined views

450.6K

53 Sources, first seen 13d ago

4.2K likes257 comments689 saves344 reposts

Sentiment

Positive19.1%80.9%Negative

Based on 189 sentiment-bearing replies from 157 accounts across 10 conversations.

Sentiment

Positive19.1%80.9%Negative

Based on 189 sentiment-bearing replies from 157 accounts across 10 conversations.

Related

Bug fixes, timing and headlines in a short exchange

One person asks whether bugs they found and fixed partly explain the situation. The other says it’s much better for “it” to come out now than six months ago, though the headlines were probably a bit extreme even then.

LLM internals and the ‘stochastic parrot’ debate

A post argues that looking inside LLMs with mechanistic interpretability could help settle the debate over whether they’re ‘stochastic parrots.’

Roon Details Monitoring Shortfalls in AI Safety

OpenAI researcher explains infrastructure downtime and false positives as key limits on monitoring systems.

53 Sources

bbuchsbaum@Brad_Buchsbaum@MelMitchell1 The larger problem with the stochastic parrot paper is that it ornithomorphizes LLMs.13d
Andy Masley@AndyMasley@MelMitchell1 Bender's FAQ here mostly leaves me thinking that current models are still very limited for the basic reasons she'd outlined in the paper. She specifically says the framing is "extremely relevant" to current models.13d
Aidan Clark@_aidan_clark_“what we have now are not LLMs”13d
Boaz Barak@boazbaraktcs@_aidan_clark_ Can someone add the marker of when we changed from the "LLM" to "post LLM" era in this graph?13d
Melanie Mitchell@MelMitchell1@_aidan_clark_ It's true, unless "LLM" has been redefined. Stochastic Parrots paper was using the original definition: a language model is a statistical model of language, not of all the stuff today's AI systems have been post-trained on.13d
Tomek Korbak@tomekkorbakIt’s still modeling language, just from a slightly different distribution! But the architecture and objective (for a fixed dataset) are basically the same. If you’d add RL rollouts to your pretraining data the effect wouldn’t be different. Also the stochastic parrot paper didnt make this caveat about which training distribution counts and what doesnt, its claims were about the whole paradigm of predicting the next token in text13d
Tal Linzen@tallinzenStrange to be on the side of the stochastic parrots paper, a paper I've never been a fan of, but it is what it is... The meanings of words change over time and I guess it's fine if now everyone wants to call coding agents with all of their harnesses and tools and whatnot "language models". But the term "language model" has a particular history and was used in a specific sense where the stochastic parrots paper was written. It's anachronistic to criticize that paper based on the performance of the systems you *now* call "language models". (Again, there's plenty of other things to criticize it for.)13d
Jason Wolfe@w01feI got excited to join OpenAI about 4 years ago when it became clear to me that RL on top of GPT 3++ calling tools (what I was building at the time, minus RL) would probably become AGI … and I agree I 100% did and still do think of that RL model as an LLM.13d
Alex Nichol@unixpickle@MelMitchell1 Doesn't matter. People still use the phrase and reference the paper. It has been incredibly harmful.13d
Steven Hansen@Zergylord@MelMitchell1 You can still run LLMs today without a harness (so no 'complex software system') and they absolutely do not act like stochastic parrots.13d
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    Andrew LampinenBoaz BarakOpenAI

    53 Sources

    bbuchsbaum@Brad_Buchsbaum@MelMitchell1 The larger problem with the stochastic parrot paper is that it ornithomorphizes LLMs.13d
    Andy Masley@AndyMasley@MelMitchell1 Bender's FAQ here mostly leaves me thinking that current models are still very limited for the basic reasons she'd outlined in the paper. She specifically says the framing is "extremely relevant" to current models.13d
    Aidan Clark@_aidan_clark_“what we have now are not LLMs”13d
    Boaz Barak@boazbaraktcs@_aidan_clark_ Can someone add the marker of when we changed from the "LLM" to "post LLM" era in this graph?13d
    Melanie Mitchell@MelMitchell1@_aidan_clark_ It's true, unless "LLM" has been redefined. Stochastic Parrots paper was using the original definition: a language model is a statistical model of language, not of all the stuff today's AI systems have been post-trained on.13d
    Tomek Korbak@tomekkorbakIt’s still modeling language, just from a slightly different distribution! But the architecture and objective (for a fixed dataset) are basically the same. If you’d add RL rollouts to your pretraining data the effect wouldn’t be different. Also the stochastic parrot paper didnt make this caveat about which training distribution counts and what doesnt, its claims were about the whole paradigm of predicting the next token in text13d
    Tal Linzen@tallinzenStrange to be on the side of the stochastic parrots paper, a paper I've never been a fan of, but it is what it is... The meanings of words change over time and I guess it's fine if now everyone wants to call coding agents with all of their harnesses and tools and whatnot "language models". But the term "language model" has a particular history and was used in a specific sense where the stochastic parrots paper was written. It's anachronistic to criticize that paper based on the performance of the systems you *now* call "language models". (Again, there's plenty of other things to criticize it for.)13d
    Jason Wolfe@w01feI got excited to join OpenAI about 4 years ago when it became clear to me that RL on top of GPT 3++ calling tools (what I was building at the time, minus RL) would probably become AGI … and I agree I 100% did and still do think of that RL model as an LLM.13d
    Alex Nichol@unixpickle@MelMitchell1 Doesn't matter. People still use the phrase and reference the paper. It has been incredibly harmful.13d
    Steven Hansen@Zergylord@MelMitchell1 You can still run LLMs today without a harness (so no 'complex software system') and they absolutely do not act like stochastic parrots.13d
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet