minimaltom

Wait _how_

Very cool

Hell yeah, whats your thoughts on doing it for like a whole city?

This is beautiful! I would love if we could model out a whole city, ideally using a much cheaper model.

You can have a prediction be specific and falsifable while inside the system, its just then riskier (and less useful).

For example, that french dude that bet on a prediction market what the temperature would be, then broke into the weather station at the airport with a hair dryer. The prediction was still specific and falsifable!

Its a fair question to ask! and ive landed on the opposite answer lol

Indeed!! It was an example of an utterance that wasnt falsifiable to illustrate what falsifiable means.

That’s the crux of it!! You don’t have to agree but people do care, and I think (ignoring the obvious partisans) that’s the two dominant camps of commenters on this thread.

One pony’s trash is another pony’s treasure, so I guess here one persons pedanticism is another persons hobby.

Take a look at the thread below yours, but basically if the prediction has to be wishy-washy or just interpreted as 'direction', theres way more noise and way less signal. Why not just state the underlying trend/forces instead? Why lean into the online culture of predictions and do it badly.

Bad (form) prediction: OAI is gonna be wobbly in a bit Good (form) prediction (could be totally wrong): OAI as we know it today is going to collapse due to running out of money around/at 20xx.

And sure we can be pedantic about detail, but the litmus test is: is the prediction useful if you had a crystal ball and you could know if it was true/false a priori?

Falsifiable means able to be correct or incorrect. So if i said that my stepmom was kinda whelp, that wouldnt be falsifable, because like what does that even mean. But if I said my stepmom is going to be on her third marriage by the end of the decade, that is, because in 2030 someone can yell at me either way.

This is a spectrum of course: a prediction that OAI will collapse is probably right as _eventually_ all companies come to an end, but under that interpretation, the prediction is useless. It's more signal / useful / falsifable to say OAI is going to collapse around/at <year> due to <thesis>.

Anyway thats my two cents. Its fine to outline forces and trends, but when you make predictions there are useful (better, falsifiable) and useless.

I agree he is directionally correct, or at least is a vessel to raise important points and valid criticisms.

But predictions need to be specific and falsifable. If not, its just rag-chewing over a beer (luv that shit, but i aint predicting on taco tuesday). If they arent falsifable, then its not a prediction.

I think a lot of the HN comments generally can be described as one camp which cares about and enforces the rigor of predictions and trying to direct limited ear-time to voices which tend to get predictions right, vs the other camp that puts more weight towards directional accuracy.

Yeah, I think this is fair criticism, and I definitely felt the pointedness in the tone as well.

On the overly-literal / narrow thing, I think thats the culture around evaluating predictions overall. Like, all those posts around christmas where people make predictions and evaluate how last year went. The rigor is the norm.

You made a poor characterization of the content of the post and were rebutted with direct quotes. It happens, thats okay! Its not an attack on your character. There's no need to feel defensive!

Idk about that. A bunch of his wrong predictions were predicated on things turning sour sooner than a "year or so" than now.

Hell yeah a birdie over thirty, with a splash of tech!

(you either go the path of a burner or a birdie, I don't make the rules)

Its not in the kernel but in the userspace tool that goes from password to key (the key is handed to the kernel).

You can see the implementation here: https://gitlab.com/cryptsetup/cryptsetup/-/blob/main/lib/cry...

Oh like one of those scenic cone towers like on a nuclear power plant?

Iiuc youre saying: its more cost-efficient to waste water using evaporative cooling so thats what we'll get, not that a closed loop with a heat exchanger is technically infeasible?

For datacenters specifically I've never understood what specifically consumes the water. Arent the water-cooling loops closed, so the water just cycles around and around and around?

Thats what I thought too but then it would be s**?

Interesting, for anything more than side chats/projects I usually am watching the output generate and thinking about the problem. I have the same issue with switching back, takes a while to recall and page everything back into my context, so I try not to alt-tab away.

What is bpw?

Also whats your cutoff for 'acceptable' speed? I would have said 25tok/s.

Worth distinguishing knowledge/task benchmarks from IF / agentic. It doesn't seem out of the question that you can have a small model thats generally good at instruction following and long-horizon agentic, as usually in those cases any requisite knowledge is in the context.

Most of the benchmark improvements afaict are in agentic and instruction following benchmarks.

Architecture thread! Afaict they continue to use gated attention + delta net, which was also adopted+adapted by K3, but im surprised theres no improvements to the residual stream (deepseek are using manifold hyper-connections, kimi have attention residuals) ?

Perf improvements seem to all come from training?

I mean look at the name... this seems more like a fun quip than a product. I'm over here basking in the glow of dem vibes.

Yeah 100%! And I'm sorry if I sound a little more critical and less eager, its just thats theres a world of difference between a priori finding the bug, and reproducing it, and the impression of the article (from my read) was the former.

But please keep writing, I know its super hard to put yourself out there and make content!

Thanks for clarifying! It would be really interesting if Antithesis finds the bug when:

1. The specific bug isnt mentioned 2. (If youre game) a model with a knowledge-cutoff date before the report is used

I went clicking through to see if I could find the prompt they fed the AI to locate the issue / write the test suite.

I couldn't find it, so its unclear if the prompt was completely "make a test suite" or was lead towards finding it in the first place, which wouldn't be a fair test.

The closest I mention of the prompt I could find was:

> Then I asked it to write a simple workload which exercised the WAL insert and checkpoint code. Notably, this is a completely generic workload.

With a skeptical lens, unclear.

I'm equal parts intrigued and skeptical- I guess if the prompt doesn't lead on there is a bug there then I'm impressed.

Idk if its accurate to rote project the policies of OpenBSD to OpenSSH, yes technically its a subproject but in practice stewardship and thus effective policy is pretty much all damien.

What about mHC? I'm surprised it helped with such a small compute budget.

Was really cool to see yous use Engrams to cut down compute!

Given its basically an O(1) lookup with disk space being the main constraint, I was curious if you've tried ablating engram layers and sizes across your setup?

Also, why mHC over attention residuals?