Jonathan Pryde

∃

[Jonathan Pryde @Jonathan_Pryde]

@PINTO03091 Such a permanent solution to such a transient fool

[Jonathan Pryde @Jonathan_Pryde]

@yacineMTB 🤌🤌🤌

[Jonathan Pryde @Jonathan_Pryde]

@yacineMTB Try harder

[Jonathan Pryde @Jonathan_Pryde]

@W4pmqbkgchK @Joeoqbu @jun_song @grok actually a pretty good question

[Jonathan Pryde @Jonathan_Pryde]

@staysaasy There might be very rare but significant little things humans could beat ais at but we’ll never know because the ais take over the human’s prerequisites

[Jonathan Pryde @Jonathan_Pryde]

Still don’t understand calling them key query value, too clean, it seems weirder to me “pick four matrices, doesn’t matter what numbers just toss em in, multiply one by two of the others, multiply those by each other, then multiply that output by the last matrix. Oh and make sure the first one represents tokens. Also make sure numbers don’t get huge. Now PUNISH THE MATRICES!” And it works??

[Jonathan Pryde @Jonathan_Pryde]

did it now we don't
better "will, then won't"
few now will persist
others may exist

not aliens what are you some kind of kook, it's a riddle

[Jonathan Pryde @Jonathan_Pryde]

as aesthetic as this idea is ````` it is a rapidly moving target and has a poor shelf life

[Jonathan Pryde @Jonathan_Pryde]

I like the idea of working at the highest level of abstraction possible—and at times: at my leisure
diving all the way down

[Jonathan Pryde @Jonathan_Pryde]

@MishraArya5819 @OpenAI Try it

[Jonathan Pryde @Jonathan_Pryde]

@MishraArya5819 @OpenAI I could /goal that one maybe

[Jonathan Pryde @Jonathan_Pryde]

@MishraArya5819 @OpenAI Does the problem need a lot of compute

[Jonathan Pryde @Jonathan_Pryde]

Shtudying

[Jonathan Pryde @Jonathan_Pryde]

@justindross I wonder if they’re “going to sleep” because they’re building a model that actually does

[Jonathan Pryde @Jonathan_Pryde]

Elon: I do everything, if you’re a supplier I want to replace you with my own company

Jensen: I want to do as little as possible, as much as necessary, if something can be done by someone else we’ll stop doing it and use them…

And both strategies work??

[Jonathan Pryde @Jonathan_Pryde]

We could have a program/database indexed for codex such that I can tell it to “use my compute to do valuable novel research” and it actually does, in a way that isn’t redundant; many sessions from different people could be stitched together, arxiv et al should be working on this

[Jonathan Pryde @Jonathan_Pryde]

1.4m parameters

[Jonathan Pryde @Jonathan_Pryde]

replicated results of ABA data augmentation paper, this model was only trained on digits up to 999

[Jonathan Pryde @Jonathan_Pryde]

@SpaceKoala @code_mom_ It’s 10% cheaper to compress natural gas in Canada than Texas because of temperature