forrest2

6 ms·

1.
by forrest2 2y ago
This is largely a side effect of mimicking the distribution on the internet via pretraining. It's a good basis for setting up a model of the world since we have so much data and it's free. Post-training techniques like DPO and RLH
2.
by forrest2 2y ago
A single synchronous request is not a good way to understand cost here unless your workload is truly singular tiny requests. Chatgpt handles many requests in parallel and this article's 4 GPU setup certainly can handle more too. It is
3.
by forrest2 2y ago
Of course, we already need to be working in Spaghetti ints or prepping them will be as complex than a linear scan. Can't wait for spaghetti arithmetic. Do we have a better algo than log(n) for locating the min? I'm thinking spaghe
4.
by forrest2 2y ago
A lot of the yolo stuff from ultralytics is AGPL3 fyi. Recommend caution depending on what code or models / model lineage are used
5.
by forrest2 3y ago
Then it's an arguably arbitrary classification right? The borders seem potentially ill-defined and/or social in nature.
6.
by forrest2 3y ago
In some places, it's "gay" if two guys hold hands. On the racist bit: it can be hard to disambiguate why someone treats you in an unfavorable way -- it could be because they're actually racist or they could just be a pis
7.
by forrest2 3y ago
You can get most products on prem at a certain price & size. A lot of companies will apply resistance though unless the contract size is right b/c on-prem contracts tend to be less unit-profitable, risky, + unique annoying terms or
8.
by forrest2 3y ago
Is there a design paper somewhere? Curious how this was accomplished (and what trade-offs / failure modes it has): "System uses a consistency sharding algorithm, lock-free design, task scheduling is accurate down to the second, su
9.
by forrest2 3y ago
I think I have this: https://www.nationwidechildrens.org/conditions/auditory-proc... Might be the same thing depending on your full set of symptoms.
10.
by forrest2 3y ago
You might not see it available via self serve interfaces, but I think the h100 cards can do 256x to a single host.
11.
by forrest2 3y ago
I think people said the same thing about NNs in general before we hit a scale where they started performing magic. There could be exponential or quadratic scaling laws with any of these black boxes that makes one approach suddenly extremely
12.
by forrest2 4y ago
If you have tight interfaces that fully encapsulate the use of any, you're fine. If you want to return `any` or pass it around outside of an insulated zone, you're gonna have a bad time. `any` is a contagious/viral construct
13.
by forrest2 4y ago
You make that impossible with application logic or constraints. Just like any linked list, you need to update the other pointers Example constraint would be a unique index for the Playlist member table on the combination of Playlist id and
14.
by forrest2 4y ago
or perhaps for some applications it would be sufficient to do playlists { id song_ids [] } ------------- and just store the ordering in an array. Some postgres drivers might start shitting the bed though at some gigantic array siz
15.
by forrest2 4y ago
Or model it as a linked list and you can sidestep the limitations / complexity of some kind of numeric (or bytes / text based ordering field) ------------ playlists { id } ------------ playlist_members { id playlist_id prev_playli
16.
by forrest2 4y ago
I don't wanna speak for them, but for some people, going in deep on something is the only way to stay healthy / happy / sane. Different strokes for different folks
17.
by forrest2 4y ago
One could argue that what you're describing is one of those alternative business models. You don't care if someone comes to your page anymore for ads if the scraped page is the ad. I'm sad about what I think the stable equili
18.
by forrest2 4y ago
There is flywheel issue with generative search: content producers, information curators or outlets, etc need to get paid. People are okay-ish with Google in part because it drives traffic and traffic can be profitable. Fancy features that s
19.
by forrest2 4y ago
Anecdote: I'm a midwesterner and half of my siblings, most of my wife's siblings, and some of my friends' siblings are in their late 20s, have no jobs (or <10 hrs a week), and live with their parents. From the outside, the
20.
by forrest2 4y ago
Haven't used yet but systemd flag added a month ago: https://devblogs.microsoft.com/commandline/systemd-support-i...
21.
by forrest2 4y ago
Vespa looks pretty compelling; indexing looks like a dream. I'd recommend basically anything else over a customized ES / Solr cluster. Some of the least fun clusters to manage. Great for simple use-cases / anything you see in
22.
by forrest2 4y ago
> I am not talking about investing in companies. I am talking about how these companies invest their R&D capital in projects. Some food for thought: These companies do have economies of scale on their side for internal tech investmen
23.
by forrest2 4y ago
Yeah this is what reminded me. Figured I'd revive the "mega earthquake" while the news was buzzing about the "mega flood".
24.
by forrest2 4y ago
https://en.wikipedia.org/wiki/Megathrust_earthquake An absolutely nuts quote from the article -- "the northwest edge of the continent [...] will drop by as much as six feet and rebound thirty to a hundred feet to
25.
An earthquake could destroy a sizable portion of the coastal Northwest (2015)(newyorker.com)
121 points by forrest2 4y ago|90 comments
26.
by forrest2 4y ago
I've used an emulator in the past for GPS spoofing too. Obviously useful for testing and what not but also for location based apps / games / etc. Tangent: IIRC Pokemon GO (and probably other more professional usecases for GPS
27.
by forrest2 4y ago
Investors usually have liquidation preference. If they do, they'll have a first dibs on that $50m up to their invested amount.
28.
by forrest2 4y ago
Arguably needs to be abstracted out to some kind of multiplayer middleware; multiplayer sounds like one-size-fits-all but it's very nuanced. And that feeling that it needs to be abstracted/genericized might be the same reason why
29.
by forrest2 4y ago
TLDR: Usually you start with num_interesting_event / total_measurements This biases the result toward 0 or 1. Instead, you can bias toward .5 if you know that both sides of a binary outcome are possible: ( num_interesting_event + 1 )
30.
Rule of Succession – Estimating true probabilities with few samples(en.wikipedia.org)
1 points by forrest2 4y ago|1 comments

More ›