RESEARCH

–

27 January 2024 20h57

Natural selection is simple. A complicated story then happens, and human beings come out the other end in a way that doesn't match up to the simple stuff natural selection optimized for. I could indeed make this story more complicated; this doesn't help the safety-is-easy case.

Complexity in AI Alignment: Why Precision Matters for Safety

By

–

27 January 2024 20h56

From my perspective, the more complicated you say these relationships are, the worse for anybody trying to build them out in some precise way that doesn't killeveryone. Rockets are complicated and that doesn't make them easier to notexplode.

→ View original post on X — @jxmnop

ROPE and Positional Embeddings: Sine Cosine Functions

By

@jxmnop

–

27 January 2024 19h20

i guess ROPE and some positional embeddings use sine and cosine as well

→ View original post on X — @jxmnop

Understanding Transformers with High School Mathematics

By

@jxmnop

–

27 January 2024 19h20

one exciting observation about transformers (and most modern deep learning) is that you can understand them using high school math. really just multiplication, division, sums, and exponentiation, many times, and in a strange and initially hard-to-grok order

→ View original post on X — @fchollet

ML Research and Applied ML Require Different Skill Sets

By

@fchollet

–

27 January 2024 19h18

The skills needed for developing new ML techniques have little overlap with the skills needed for applying ML effectively. A bit like how chip design has little overlap with software engineering.

→ View original post on X — @oriolvinyalsml

Knowledge Distillation Paper Achieves 17k Citations Despite Skeptical Review

By

@oriolvinyalsml

–

27 January 2024 19h10

Fun game. Clocking 17973 citations: "Distilling the knowledge in a neural network" @geoffreyhinton
, @OriolVinyalsML
, @JeffDean Reviewer 38 (NeurIPS 2014): "This work is incremental and unlikely to have much impact even though it may be technically correct and well executed."

Utility Function Terminology and Human Evolutionary Misalignment

By

–

27 January 2024 19h01

I think that "utility function" in retrospect is a mathematical word of power that I should not have expected lay computer scientists to understand, so let's drop that. With humanity, the outer optimization criterion was inclusive fitness; our inner preference was not aligned.

Intelligence and Outcome Matching: The Core Problem Beyond Utility Functions

By

–

27 January 2024 18h48

It isn't about "simple" utility functions or "monomania". The problem is just any sufficiently smart system whose work, on some level, can be viewed as matching up outputs and results, and learning.

Learning Reality and Selecting Outputs for Desired Outcomes

By

–

27 January 2024 18h46

The problem that 'learn how reality works' and 'select outputs which, when they interact with reality, lead to X happening' is a simple great way of doing Y for a lot of possible Y. For example, with humans, Y is inclusive genetic fitness and X is all the stuff that humans want.