This may seem surprising to many people — if LLMs can score above human level on all those hard human exam benchmarks, why couldn't they do something as simple as ARC tasks, most of which seem trivially obvious to humans? It's because these tasks are mostly *new* — you won't
@fchollet
-
Formal Definition of Intelligence and Benchmark for LLMs
By
–
Here's a paper you can read about it. https://
arxiv.org/abs/1911.01547 It introduced a formal definition of intelligence, as well as benchmark to capture that definition in practical terms. Although it was developed before the rise of LLMs, current state-of-the-art LLMs such as Gemini -
Memorization is not intelligence: skill requires novelty and uncertainty
By
–
That memorization (which ML has solely focused on) is not intelligence. And because any task that does not involve significant novelty and uncertainty can be solved via memorization, *skill* is never a sign of intelligence, no matter the task.
-
Keras with PyTorch matches vanilla PyTorch performance when compiled
By
–
That's because Keras + torch is not compiled by default whereas the native torch models are compiled. Keras+torch has the same performance as vanilla torch.
-
Framework Comparison: JAX, PyTorch, TensorFlow Development Experience
By
–
I agree with your point that people hours are worth more than GPU hours (except for large-scale training runs or inference workloads of course…). Having been developing in all three frameworks (torch, TF, JAX) extensively over the past year, I've personally found JAX to be the
-
Keras torch backend performance overhead compilation optimization
By
–
Keras adds virtually no overhead over torch actually… the torch backend numbers are much worse simply because Keras + torch is not compiled by default, whereas the native torch models were compiled with the best compilation config we found. Keras + torch has essentially the
-
XLA Compiler and JAX Integration for Optimal Performance
By
–
The XLA compiler — and the fact that JAX was designed for XLA, making the two work really well together.
-
AI Industry Investment and Hidden Costs Analysis
By
–
The new AI: 60B/year in investment, 4B/year in private revenue, 10B/year in externalized public costs?
-
Photoreceptor Density Does Not Equal Information Intake
By
–
If the photoreceptor density of your retinas were 1,000x higher firing 10x faster, your wouldn't be "trained" on 10,000x more "data" (and not just because your brain isn't a curve fitted on data via gradient descent). It would change almost nothing to your information intake.