17 comments

[ 96.8 ms ] story [ 1955 ms ] thread
Ternary neural networks and possibly new innovations in attention are the end of trillion dollars NVIDIA.
These are MLPerf training results. I think current ternary quantization research is focused more on speeding up inference?
Not necessarily, if the training is more efficient, you just throw more data at it for better models which was not possible before
But can it solve ARC?
exactly. this is my go to question with any AI claims now :)
The original AI reasoning challenge (ARC) that predates the one recently discussed here on HN by only a year has been solved. And I have no doubt that the new one will also be solved at some point in the near future. Every time someone comes up with a supposed "AGI test" and eventually gets beaten, they just move the goalpost instead of admitting that perhaps human intelligence is not so special after all. It will probably take something truly superhuman that vastly outclasses even the smartest humans in every conceivable way before the remaining majority of humanity will dare to admit that they are now the lesser intelligence.
Your opinion is based on hype, not actually understanding the problems.

The original ARC is “grade-school level, multiple-choice science questions” which is not much different than existing benchmarks.

The new ARC is different because it’s resistant to memorization and, as a result, modern AI algorithms do terribly even though the problems themselves are very simple.

We are clearly missing something and acknowledging that has nothing to do with admitting we’re the “lesser intelligence”. Frankly your comment sounds cultish.

Lots of people seem to have forgotten that not long ago, these questions were practically unsolvable even by the best models. Similar to how a few years before that it was impossible to come up with a model that could reliably distinguish between between cats and dogs. The frontier is moving all the time and no concrete evaluation has stood the test of time so far.
I never said it's impossible to beat the ARC challenge. It seems like you're making up arguments that don't exist.
You said

> Your opinion is based on hype, not actually understanding the problems.

That's pretty strong wording, easily tantamount to saying it's impossible/improbable

They said ARC should be beaten in the near future. I don’t think it’ll be as easy as they say since we’ve been using transformers for awhile and they still fail pretty bad at the test.
Well that comment didn't age well. But even I didn't expect such a significant improvement in one week.
(comment deleted)
(comment deleted)