I think most of Google's deep research projects were done in the pursuit of pure science, not monetization or productization. In hindsight sure, it looks like they missed an opportunity. But not everything needs to be about money.
Doesn't seem that surprising or terrifying to me. Humans come equipped with a lot more internal biases (learned in a fairly similar fashion), and they're usually a lot more resistant to getting rid of them.
The truly terrifying stuff never makes it out of the RLHF NDAs.
We ought to be terrified, when one adjusts for ll the use-cases people are talking about using these algorithms in. (Even if they ultimately back off, it's a lot of frothy bubble opportunity cost.)
There a great many things people do which are not acceptable in our machines.
Ex: I would not be comfortable flying on any airplane where the autopilot "just zones-out sometimes", even though it's a dysfunction also seen in people.
>Ex: I would not be comfortable flying on any airplane where the autopilot "just zones-out sometimes", even though it's a dysfunction also seen in people.
You might if that was the best auto-pilot could be. Have you never used a bus or taken a taxi ?
The vast majority of things people are using LLMs for isn't stuff deterministic logic machines did great at, but stuff those same machines did poorly at or straight up stuff previously relegated to the domains of humans only.
If your competition also "just zones out sometimes" then it's not something you're going to focus on.
The transients were pretty easy to replicate yes. The nuclear testing stuff was pretty inconclusive but they have a much better curated collection of plates that aren't available yet.
They're there before the tests though, and potentially more frequent around nuclear testing calendar days. The argument has never been "these only showed up after a nuclear test."
Two things here: radiation exposure could explain this, since there's a period after exposure and before developing where you can get radiation exposure.
Second, and perhaps more importantly, is that there's a detailed criticism of this line of research available, including evidence against the argument that these are more likely ±1 day of nuclear tests. See https://arxiv.org/pdf/2601.21946, and also https://arxiv.org/pdf/2402.00497 for a study of plate defect issues.
I think the current paper continuing this line of research should be read cautiously. I don't love discounting ideas out of hand, as these folks clearly have put effort into the analysis. But the rebuttals read as at least as high quality analysis, and "it's aliens" requires a lot of evidence for me to take it seriously.
Nope, they're surprisingly hard to get ahold of. So I've resorted to being extremely noisy online.
Temporal accumulation: imagine you're observing a signal through a narrow window and you can only see a partial, noisy snapshot each time. "Temporal accumulation" is what happens when you let the observer remember previous snapshots and use them to improve its prediction of the next one. The persistence advantage P measures how much that memory helps, the difference in prediction error between an observer that accumulates across episodes and one that only uses the most recent snapshot.
For a black hole shadow (EHT), P ≈ 0: each snapshot already contains the full picture, memory adds nothing. For gravitational wave strain (LIGO), P is large and positive, the chirp evolves across snapshots, so memory is essential. The question the papers ask is: what determines how much memory helps? The answer turns out to be spectral entropy of the waveform, not mass.
I've been doing this for about a month. I also have wildly complicated ML pipelines working similarly in parallel. When Karpathy's 'autoresearch' came out I was surprised by how novel it was treated.
I don't think it works as well as this one.