body { -ms-overflow-style: scrollbar; overflow-y: scroll; overscroll-behavior-y: none; } .errorContainer { background-color: #FFF; color: #0F1419; max-width: 600px; margin: 0 auto; padding: 10%; font-family: Helvetica, sans-serif; font-size: 16px; } .errorButton { margin: 3em 0; } .errorButton a { background: #1DA1F2; border-radius: 2.5em; color: white; padding: 1em 2em; text-decoration: none; } .errorButton a:hover, .errorButton a:focus { background: rgb(26, 145, 218); } .errorFooter { color: #657786; font-size: 80%; line-height: 1.5; padding: 1em 0; } .errorFooter a, .errorFooter a:visited { color: #657786; text-decoration: none; padding-right: 1em; } .errorFooter a:hover, .errorFooter a:active { text-decoration: underline; } #placeholder, #react-root { display: none !important; } body { background-color: #FFF !important; }

JavaScript is not available.

We’ve detected that JavaScript is disabled in this browser. Please enable JavaScript or switch to a supported browser to continue using twitter.com. You can see a list of supported browsers in our Help Center.

Terms of Service Privacy Policy Cookie Policy Imprint Ads info © 2023 X Corp.

Follow

nostalgebraist

@nostalgebraist

nostalgebraist.tumblr.comJoined December 2011

1,146 Followers

nostalgebraist’s Tweets

@nostalgebraist

thank you!! the book should be complete sometime later this year. <10 chapters left -- just need to find the time to write them.

Quote Tweet

Sarah Constantin

@s_r_constantin

I think nostalgebraist has graduated from “guy I know who writes fiction” to “one of the most creative contemporary science fiction authors I know” archiveofourown.org/works/8341348/

Show this thread

1

20

nostalgebraist Retweeted

My hot take is the smartest thing Google could do right now is improve search, not lard everything up with AI features. Search is still better, more useful, and more trustworthy for most queries, and it could stand to benefit immensely by making that case

Google Devising Radical Search Changes to Beat Back A.I. Rivals

The tech giant is sprinting to protect its core business with a flurry of projects, including updates to its search engine and plans for an all-new one.

10

118

767

Show this thread

@nostalgebraist

wait, what?? why do Bard AND ChatGPT *both* write an anodyne story about a young woman in idyllic "Willow Creek" at sundown??? (details: it's not deterministic, often you get a different town name, different phrasing, etc. broad strokes are similar though. gpt-4 does it too.)

6

11

85

nostalgebraist Retweeted

Farewell to OpenAI Codex. Codex's code‑davinci‑002 was the best performing model in the GPT-3/3.5 line for many tasks. Despite its name, it excelled in both code and natural language. It will be missed. Left: OpenAI email. Right: From Yao Fu 2022 yaofu.notion.site/How-does-GPT-O

16

56

470

nostalgebraist Retweeted

Ever wonder how a language model decides what to say next? Our method, the tuned lens (arxiv.org/abs/2303.08112), can trace an LM’s prediction as it develops from one layer to the next. It's more reliable and applies to more models than prior state-of-the-art. 🧵

19

210

936

Show this thread

nostalgebraist Retweeted

I used to defend

against the ClosedAI meme, but I can no longer: By explicitly withholding all information on data, model architecture, methodology, and all else that could be used to replicate GPT-4 as a trade secret,

has officially decided to ignore their name.

27

91

840

Show this thread

nostalgebraist Retweeted

My Twitter timeline is full of panicked takes about imminent AI apocalypse and certain doom. I think this is starting to get overplayed, and so I want to make a long thread about why I'm personally not worried yet. Get ready for a big one... 1/n

81

480

1,941

Show this thread

@nostalgebraist

what's going on with the middle pane of fig. 5? - why do the lines jump vertically? doesn't happen in the other fig. 5 panes. - the accompanying text says "we see that our results do not vary strongly with RLHF at any model size" but this is not visually evident.

Quote Tweet

Language models (LMs) exhibit harmful biases that can get worse with size. Reinforcement learning from human feedback (RLHF) helps, but not always enough. We show that simple prompting approaches can help LMs trained with RLHF produce less harmful outputs. arxiv.org/abs/2302.07459

Show this thread

5

@nostalgebraist

Anyway, happy Valentine's day from me, the world's greatest husband, who definitely posted this himself and did not leave his laptop open while he went to get food. #love

1

10

Show this thread

@nostalgebraist

It's just important that you know i am an extremely supportive, kind, cute and nuzzleable husband and that I deserve all the hugs in the world.

1

9

Show this thread

@nostalgebraist

Hey, very important that you know it is definitely me posting this and not my wife posting while i left the room to get food.

1

12

Show this thread

@nostalgebraist

i looked for those "cursed tokens" in the public webtext shard, and found documents containing many of them! including " Leilan" / " Mechdragon" / etc., "embedreportprint" and friends, and " gmaxwell". details at lesswrong.com/posts/8viQEp8K colab

colab.research.google.com

Colaboratory notebook

1

2

13

nostalgebraist Retweeted

Today’s “chat” is somewhere between conversation and instruction prompting. Chat-tuned LLMs never ask you for more information, or quickly confirm their interpretation of a complex task before starting. The lineage shows.

9

12

208

Show this thread

@nostalgebraist

finally, use the closed-form optimum for the approach 3 functional form to calculate the optima with these A/B/a/b values. this is not what is done in approaches 1 or 2, and it does not get the same results. 4/4

Show this thread

@nostalgebraist

then, assume that the constants A, B in that parametric L(N, D) function are identical between the 3 approaches. (note that this cannot be right. those A, B came from a fit to the data. if you fix them, but change a and b, the function won't fit the data anymore!)

1

2

Show this thread

@nostalgebraist

iiuc, the methodology in the screenshot goes like this: assume that the L(N, D) functions in approaches 1-2 follow the parametric form fitted in approach 3. (approaches 1-2 do not assume a functional form for L(N, D) and do not estimate it.) 2/

1

Show this thread

@nostalgebraist

this is not correct. first, note that the paper provides explicit C/N/D example numbers for all 3 approaches in tables 3, A3. there, all 3 methods require more than 4.3e23 compute to go above 67B params. so immediately we know something can't be right 1/

Quote Tweet

...anything from a 28B model (on 2.5T tokens) to a 260B model (on 270B tokens) between their 3 "Approaches". That's an unhelpful order of magnitude difference in how large of a model you should be training in order to be considered "compute optimal"

. [5/7]

Show this thread

1

3

12

Show this thread

nostalgebraist Retweeted

The reason people aren't that shocked by GPT and DALL-E is they automate exactly what is already post-scarcity

33

110

1,230

Show this thread

nostalgebraist Retweeted

Announcing our new mechanistic interpretability paper! We use causal interventions to reverse-engineer a 26-head circuit in GPT-2 small (inspired by

’s circuits work) The largest end-to-end explanation of a natural LM behavior, our circuit is localized + interpretable 🧵

The following media includes potentially sensitive content. Change settings

View

11

77

361

Show this thread

nostalgebraist Retweeted

LMs can learn via inference alone through demonstrations -- but how does it work? We find that LMs do not really need correct input-output pairs. Randomly replacing labels in the demonstrations barely hurts performance, consistently over 12 models. arxiv.org/abs/2202.12837

Ari Holtzman and 4 others

10

89

428

Show this thread

nostalgebraist Retweeted

Replying to

@nostalgebraist

not the first time

has done this:

Quote Tweet

Replying to @__femb0t and @tweedymcgee

did you just... steal this from a 45 minute old tumblr post without credit. it had less than 50 notes lmao.

1

1

6

@nostalgebraist

don't pass off my own words as your own, please nostalgebraist.tumblr.com/post/133051101

Quote Tweet

“and why should i pilot eva?”, the achaeans’ greatest warrior asks himself, as he sulks in his tent

2

1

16

nostalgebraist Retweeted

Stella Rose Biderman @ Linux OSS North America

@BlancheMinerva

A little over one year ago

approached #EleutherAI and asked how much compute we needed to train the largest open source language model in the world. Today, we are finally releasing the weights of a 20B parameter LM trained on the Pile.

8

99

513

Show this thread

@nostalgebraist

This is a really neat idea! I gave it a try in my GPT application (a tumblr bot). Some notes on what I observed: nostalgebraist.tumblr.com/post/675469656

Quote Tweet

Clara Isabel Meister

@clara__meister

Neural language models are really good at explaining held-out data. So when we sample from them, why do they yield dull and degenerate text? Our paper analyzes this behavior using information theory, and corrects for it with a new sampling principle: arxiv.org/abs/2202.00666

Show this thread

1

1

11

@nostalgebraist

Here's a post I wrote about it: nostalgebraist.tumblr.com/post/672300992

Quote Tweet

Rivers Have Wings

@RiversHaveWings

nostalgebraist on Tumblr has trained a diffusion model that can write the text input to a text encoder into the image! (I think he is also using a separate language model to generate the text.) stalgebraist-autoresponder.tumblr.com/post/671421163 stalgebraist-autoresponder.tumblr.com/post/671420177

4

25

@nostalgebraist

great stuff!! this is (a first step towards) a specific kind of transformers interpretation work that i've always hoped would exist one day. i spend a non-trivial fraction of my spare time thinking about this problem, mostly without success. anthropic's actually doing it!

Quote Tweet

Our first interpretability paper explores a mathematical framework for trying to reverse engineer transformer language models: A Mathematical Framework for Transformer Circuits: transformer-circuits.pub/2021/framework

Show this thread

15

nostalgebraist Retweeted

The Google Project Zero write up on the iMessage zero-click vulnerability is incredible. I promise you won’t be able to predict the directions this is going. googleprojectzero.blogspot.com/2021/12/a-deep

32

561

1,548

@nostalgebraist

tensorflow dot png

Quote Tweet

So @lmanul quit Google and released a bunch of internal-only Google comics, and I have a favorite. goomics.net/50/

Show this thread

1

1

10

nostalgebraist Retweeted

every time i see a tweet about a neural net research paper it's some variation of this

10

130

979

@nostalgebraist

("cast the weights to float32" meaning i cast the original checkpoint to float32, and then my additional finetuning work writes checkpoints in float32.)

Show this thread

@nostalgebraist

nonetheless, i went ahead and cast the weights to float32. continuing to have fun with this repo on my fork: github.com/nostalgebraist most recently, added measurements of the gradient noise scale, which i had done in my old train script.

1

Show this thread

@nostalgebraist

update: asked these questions on EleutherAI discord and got helpful answers! - the bfloat16 checkpointing was an accident - probably little to no quality impact: training loss curves don't change noticeably around load/save steps

Quote Tweet

@nostalgebraist

Replying to @nostalgebraist @shlevy and 2 others

i haven't tried the 1.3B huggingface checkpoint, but i don't see an indication it's 16-bit. if so, was it cast to 32-bit from the original bfloat16 checkpoint?

1

Show this thread

nostalgebraist Retweeted

I've been able to update The Hiveswap Article with new information in the Kickstarter update! Tons of new information, some things confirmed, some denied. I've also added in several corrections and noted where What Pumpkin commented on-the-record.

blog.giovanh.com

The Hiveswap Fiasco

how to be angry about Hiveswap

2

26

53

Show this thread

nostalgebraist Retweeted

Replying to

Yeah. Copy editor needs extra time. Can’t blame them. There’s a lot of Latin. And Greek. And unusual alphabets. And code. And technical vocabulary. And meter. And extremely fussy pronouns. And rare punctuation. And it’s 2x as long as expected. And there’s this one chapter... 😅

5

5

49

nostalgebraist Retweeted

Adam Strandberg

Oh, you have a prior? Name the probability of every hypothesis

17

39

370

nostalgebraist Retweeted

Aran Komatsuzaki

@arankomatsuzaki

We've released the weights (1.3B and 2.7B) of our replication of GPT-3 🥳 Using the updated Colab notebook in the repo you should be able to finetune the models on your own data as well as run inference.

GitHub - EleutherAI/gpt-neo: An implementation of model parallel GPT-2 and GPT-3-style models using...

An implementation of model parallel GPT-2 and GPT-3-style models using the mesh-tensorflow library. - GitHub - EleutherAI/gpt-neo: An implementation of model parallel GPT-2 and GPT-3-style models u...

30

621

2,505

Show this thread

nostalgebraist Retweeted

Sarah Jamie Lewis

@SarahJamieLewis

The quest for "usability" and, to a lesser extent, "security" has wrapped most tech in so many layers of prescriptive, over simplified interfaces that it has cut off the paths that used to bridge professional and casual tech competence, creating two distinct social classes.

Quote Tweet

barren field of sunflower seeds

Please quote this tweet with a thing that everyone in your field knows and nobody in your industry talks about because it would lead to general chaos.

Show this thread

8

245

571

Show this thread

nostalgebraist Retweeted

Finding the Words to Say: Hidden State Visualizations for Language Models jalammar.github.io/hidden-states/ New post! Visualizations glancing at the "thought process" of language models & how it evolves between layers. Builds on awesome work by

@nostalgebraist

. 1/n

1

42

174

Show this thread

@nostalgebraist

unfortunately, this is much harder than the "logit lens." representing a hidden state change (dim=1024) in a basis of FF values (dim=4096 per FF layer used) is underdetermined. tried to regularize for sparsity, many choices for how to do that... got kinda lost, was frustrating.

2

Show this thread