Search @code_star | teapawt

Cody Blakeney retweeted

Unfortunately, with access to an abliterated DeepSeek V4.1 model, there is little value to KYC holding back the frontier models now for cyber. In fact, making it hard to secure systems w extra friction in American Enterprise is hurtful to national security.

70

2,086

210,064

Cody Blakeney retweeted

@DanielDugas1420h

Soon this will be a question in MLE interviews and everyone will have to pretend like they would obviously have solved this bug from first principles

United States

Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:

3

725

61,909

So at this point I just have to assume all software has basically been a wet paper bag of security for years. government agencies must be so pissed at all their exploits getting exposed.

2

49

1,086

Reporting a problem in the company's shared Agent memory files repo Call that a skill issue

4

23

798

I don't have a problem with Corweave buying wandb, but I'm never going to call it Forge just like I'm never calling this place x.com

X. It’s what’s happening

From breaking news and entertainment to sports and politics, get the full story with all the live commentary.

2

20

1,069

Georgia football driving sim

scary ruby innes@rubyinnesOct 23

"open world football game" what the fuck do you mean what the fuck are we doing

1

11

690

@code_starOct 23

Replying to @kalomaze

the one after Caffe

1

6

222

@code_starOct 23

Replying to @zstcsz

can you link it 🙏

1

49

Cody Blakeney retweeted

@jiraticketsOct 22

United States

the internet didn’t implode so now I still have to go to work. thanks for nothing

6

175

4,036

@code_starOct 22

Replying to @HououinTyouma

The poor engineering students do

3

55

@code_starOct 22

Replying to @cthorrez

That being true doesn’t make me feel less old.

1

53

@code_starOct 22

Replying to @xeophon @_vatsadev

Or you start a heavy drinker and open up the data. Hard to say.

1

1

30

Cody Blakeney retweeted

@leothecuriousOct 22

beggin @arcee_ai to drop a nemotron-3.5-super-shaped coding agent RL-ed to death inside prime-agent. just a little something for the gpu poor out there.

3

34

3,871

@code_starOct 22

Yes, but the better question is why. Is it good and a product of squeezing our performance or an unintended artifact of some training stage.

Thomas Hikaru Clark@thomashikaruOct 22

Replying to @thomashikaru

Our hypothesis: LLMs have a pressure for information density, leading to compressed language, novel ngrams, and a preference for short alternatives in linguistic alternations, even when this is not humanlike - humans have a pressure for successful communication with other humans

3

10

919

@code_starOct 22

Replying to @leothecurious @arcee_ai

@leothecurious @arcee_ai

6

111

@code_starOct 22

Replying to @ConjecTech

Depends on if those H100s are under contract

75

@code_starOct 22

Replying to @AllanatrixQ @DimitrisPapail

I’m listening

1

1

39

Cody Blakeney retweeted

@code_starOct 22

It would take like 30 minutes to explain why I'm getting images sent to me at work on slack, but it is pretty sick.

3

23

941

@code_starOct 22

Replying to @stochasticchasm @osoleve

Well @stochasticchasm ? What are the bangers?

1

4

81

@code_starOct 22

I mean who isn’t these days

@diz__zeeOct 15

lmao john schulman just basically blurts out that thinking machines is working on a continual learning model

2

29

1,725