Gytis Daujotas
@gytdau
Followers
1,164
Following
245
Media
61
Statuses
449
making computers do my bidding
San Francisco
Joined April 2015
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
APO IN THE NEXT ALTITUDE
• 181888 Tweets
Baba
• 119147 Tweets
Charlotte Teaser1 TheEnd
• 97424 Tweets
Raila
• 87363 Tweets
Dialogue
• 84442 Tweets
Hobi
• 66415 Tweets
乱舞音曲祭
• 62880 Tweets
ディズニークルーズ
• 44254 Tweets
BEST YOU EVER
• 26158 Tweets
Bangtan
• 26156 Tweets
TPO OUT NOW
• 26089 Tweets
オリエンタルランド
• 23243 Tweets
WE LOVE YOU JENNIE
• 22311 Tweets
#神業チャレンジ
• 11121 Tweets
Pinned Tweet
Excited to release a research preview of Feature Lab, a playground where you can sculpt images by editing their features. It uses a SAE to extract the sparse features from an image embeddings model, and allows us to finely sculpt and tweak images.
15
37
271
Surprised at the how stunning illustrations in mech interp papers are. It must take so much sweat and toil to not only do incredible and novel research, but to choose to go beyond, and to make it beautiful too.
6
27
323
We built a better way to read ebooks in browsers. Just finished processing every Project Gutenberg book on it! Please check it out and let me know what ya think:
16
66
293
By training sparse autoencoders on generative image models, we discover an interpretable palette of world concepts
6
17
230
How does self-correction affect problem solving?
In a toy transformer model that was trained to solve mazes, I found that performance reliably improved (!) by inserting mistakes and self-corrections into the training data.
14
34
204
Model interpretability techniques allow us to understand and control what models do. Here's how Sparse Autoencoders give us general fine grained control on image models:
9
20
191
Sculpting an image by mixing and tweaking features - a research preview I'm testing in private beta:
10
14
127
In my upcoming work, I trained a Sparse Autoencoder (SAE) on CLIP embeddings and found that they were highly steerable and interpretable. One of the best tests of feature interpretability is to change the feature activations and see what comes out the other end. Beyond labelling
3
2
44
Impressive how fast OpenAI APIs are becoming undifferentiated commodities
Today at TransformX, we announced a huge step forward for the open source ML community: we are partnering with
@StabilityAI
to release the first large language model trained with human feedback. 1/4
5
79
492
2
3
36
This is exciting in images, because unlike language, images tend to be much easier for humans to parse and understand. Also unlike language models, many edits can be stacked in a way that avoids failure modes like gibberish or noise. These edits can be surprisingly deep, working
1
3
35
It’s still very early days. Not all features make immediate sense or are meaningful to humans, e.g. features to do with camera configuration, or the hundreds of features that represent text in the image. For this to work in a product, we’ll need to find ways to sort and filter
1
5
35
Building an IDE for interpreting SAE feature activations. Really enjoying the near real-time streaming of activations while editing samples:
1
0
30
Many of the features I found were surprisingly interpretable, and could be added or removed from images while maintaining generation quality and coherence. I found features for tabby cats and men with hats. There’s features for compression artifacts, background settings, and
1
2
35
You can try this research preview today at this link. For more details, there's also an associated deep dive post. (And of course, the weights are open source too!)
1
1
31
Moving to San Francisco has so far been excellent. Everyone is kind and open - and I would recommend it, even if only for a while.
1
0
22
Compared with standard prompt-based control, this allows for exciting user interfaces where a palette of concepts can be tweaked and added. It allows us to both make very fine adjustments to images at varying levels of granularity. And even more importantly, we get to understand
1
1
22
Daily gratitude for being in a world where scrolls that look like this can actually be read through the sheer will of human ingenuity
0
1
17
Wonderful writeup! Really enjoyed playing with the Prism demo too - had a lot of fun doing precise semantic editing of language. Very intriguing possibilities!
🌈 Research blog:
Text embeddings contain semantic and structural features that we can now automatically discover with SAEs, and use to build rich, interactive new information interfaces.
In this write-up on my SAE experiments and a prototype called
10
63
473
1
1
16
In no particular order, thanks
@DavidMcsharry3
,
@MatthewWSiu
,
@thesephist
,
@mehranhydary
,
@maxkriegers
,
@NoaNabeshima
, and many others for the feedback, advice, and support :)
1
0
17
There's a couple of advantages compared to interpretability on language models. In SAEs trained on transformers, the sparseness penalty discourages features that code for the overall shape and structure of the work. In this domain, image embeddings contain all the information in
2
1
14
This prompt-free steering can help us discover more about what's inside our models, and have industrial applications in providing new forms of user interfaces for humans to use. We can also use features to apply human-composed rules from a few examples, like this guardrail that
1
1
13
You can read more of the writeup on my blog (in my bio)! Thanks
@DavidMcsharry3
,
@NeelNanda5
, and others for early reviews of this work.
2
0
10
This IDE streams from a cloud GPU using a SAE that I've trained on a NanoGPT instance that learned to solve mazes.
I've started by working on transformers that play simple games - to figure out what techniques might transfer over to games that we understand imperfectly, like
1
0
7
I trained probes on the model, and found the residual stream has some representation of mistakes in the residual stream, but not linearly. Future work could focus on other ways of levering down the mistake-making "feature" to make a model which is better all-around.
2
0
8
@BrendanEich
@ivanbuncic
@runhappylife
@getify
That's Javascript Weekly, though it probably should instead be called JS Weekly.
0
1
8
Wonderful documentary on Robert Caro and Robert Gottlieb, his editor, and how they have worked together, quarreling and waging wars against one another, for over 50 years:
0
2
8
Over 200k words in my journal and counting. Next step is to train a language model, and start with a prompt like "Journal, today I discovered something incredible..."
1
0
7
Enjoyed this post from Daniel Dennet:
"These days I almost always outsource the hard work of comprehension when I encounter difficulties, and the policy works wonders—for me. Distributed understanding is a real phenomenon, but you have to get yourself into a community of
1
0
6
Working on better ways for people to use language models. Here's an interface that allows you to backtrack and wander through the space of creative possibilities:
1
0
6
0
0
6
The previous Fitzwilliam meetup in Dublin was wonderful, and now -- look! -- an opportunity to attend the next in the series:
I am hosting a meetup next Friday the 23rd at 7:30pm in Parnell Heritage Pub in Dublin – hope you can make it!
1
6
21
0
1
6
@thesephist
Agreed! I am curious how we can get back some of the high-level features in autoregressive transformer LLMs. I am pleasantly surprised how easy embeddings are to manipulate, e.g. quite easily supporting hundreds of feature edits.
I have a hunch that important features (like
0
0
5
Bored of having to say 'social media platforms' every time. Cooler, piratey sounding name: the shallows. Has mystique, sense of danger, lawlessness
1
0
5
@dogmadeath
Here's a hilarious example of this, where two Claudes were composing music, and Claude 2 got a little bit too into it for Claude 1's taste:
1
0
3
This is a 1-meter sized mirror used to make chips. If you were to blow this mirror up to the size of the Earth, the largest aberration would be the width of a human hair. (!!!)
1
1
5
Wow: "The world’s largest, most valuable tech companies are dependent either directly or indirectly on the steady output of TSMC’s fabs. If [they] went offline [...] it would immediately devastate the global economy. "
2
5
5
@andy_matuschak
A good friend recommended this high CRI floodlight, which I much prefer to the regular cheaper ones on Amazon:
0
0
5
One small thing so far: top activations can be surprisingly deceiving! One example: this feature seems like it represents first moves, but it actually encodes location information too. Maybe still in superposition?
1
0
4
Two instances of Claude collaborating on a Chopin-inspired barbershop tag:
1
1
4
0
0
4
@DavidMcsharry3
@MatthewWSiu
@thesephist
@mehranhydary
@maxkriegers
@NoaNabeshima
(Got the wrong Mehran - thanks
@mehran__jalali
!)
0
0
4
Something that has been unexpectedly good: getting a very cheap thinkpad, disabling the wifi drivers, and using it exclusively to journal. Makes getting into the flow of writing mysteriously far easier than using any other computer!
0
0
3
Keep burning myself by switching to promising but immature software. I'm going to just switch back to text files for note taking and see if I can make a more stable workflow work.
0
0
2
@tszzl
@nickcammarata
Maybe the visual system is going crazy spamming every primitive (edge, corner, colour) at every position, which is then interpreted by the higher order face detectors, explaining why the faces look so weird and impossible. But why faces out of all the other objects...
3
0
3
My
@livecodingtv
T-shirt just arrived. Pleasantly surprised, I thought it was just going to have their logo. Neat!
http://t.co/bi1Xczm94G
1
2
2
This counterintuitive performance gain remains even at varying model sizes, and remains even though the model has to pay a penalty for learning to make mistakes.
1
0
3
woo, go teams! Can't wait to see what they've built
🚨 Announcing: Patch Demo Day
Join us over lunch on Wednesday, August 18th to see what our latest batch of teams has been building!
Register here:
3
5
22
0
0
3
Nobody (not even your doctors) tells you that you can pay $20 to get access to the forbidden medical diagnostic knowledge to actually figure out what's going on with you for all known ailments and problems:
0
0
3
Reading translations of the classics side-by-side makes the experience feel far more rich - we should have far more tools like this:
1
1
3
1
0
2
When thinking about reliability, demos live in the 1st percentile, deployments live in the 99th percentile. Usually not much in between is useful, and the difference between the two can be surprisingly large
1
0
2
Colab is an excellent product. Something like Colab could be an entry point to cloud IDEs for the masses - with all the perks that brings (less env setup, faster artefact downloads, customizable machines, etc)
0
0
2
@mark_cummins
Profound question! Could it be that in the effortful reading style you are testing your 'world model'? i.e. when reading maths with effort, you have a notepad by your side and test your understanding via examples, counter-examples, answering your own Qs?
1
0
1
Just back from
@CoderDojo
's brilliant community hackathon at
@dogpatchlabs
. Had a great time! 😀
0
2
2
@thesephist
@DavidMcsharry3
@NeelNanda5
That’s a cool idea! We’ll need to label all the mistake moves in the training data, which is possible in a toy model, but difficult to apply in large internet scraped datasets. It would be so cool if you only needed a handful of examples of mistakes to train a probe or extract
0
0
1
@LittleJazzHands
We're building a pomodoro time tracker for creatives. We'd love to have you as one of the first beta testers 😁
1
0
1
Currently at
#DojoCon
nabbing all the free merchandise 😂 with
@ChungWung
@OisinODuibhir
@ciaranpflanagan
0
1
2
Even if you had no other information, knowing that McDonalds, an organisation which cares deeply about staying open and making more money, has chosen voluntarily to close down for an indefinite period of time should concern you greatly.
0
0
2
DCU is going to switch all of our exams to alternative assessments. This will be entertaining to see for sure.
0
0
2
Any examples of ML researcher + product builder + CEO being all the same person? Maybe a unicorn, but have heard this life plan from more and more friends!
0
0
2
@TheOisinMoran
Good luck Oisin! What's the motivating factor for dropping the coffee, may I ask?
1
0
2
This is an excellent opportunity to get very valuable creative work done with little physical distraction for 6-18 months.
0
0
2
If the next few decades come with fast AI progress, then we can expect text/image generation to get really good, which will probably do something really weird to the way we experience social media
1
0
2
In ebooks vs printed books, a big plus for printed books is that you can write notes on the paper which keep the context of your notes. Seems like a trivial issue to solve in software, but have not yet seen anywhere near a universal and adequate solution.
0
0
2
@coffee_pot
Great question! A previous paper serialized mazes using spans, but I'm not entirely sure why they did that. I found that serializing it with ASCII in a human interpretable way, just one character per tile, worked well enough for my purpose. Maybe it's not the best maze solver
0
0
2
The Socratic Method (2021, Farnsworth) is, nominally, a defense and guide to a well-respected yet rarely-practiced technique. But it is also a capable defense of serious thinking and intellect. One of the better books of the genre - highly recommended.
0
0
2
If aliens came down to earth to design user interfaces they'd add interaction sound effects. All of our apps are unnecessarily quiet
0
1
2
People can avoid fatal crashes in 99.999999% of miles driven -- really sets a high bar!
0
0
2
They should reduce capacity on public transport so that a whole bunch of people aren't standing huddled together, even if it means many will have to wait longer for another bus to arrive.
0
0
2
@FusorFusion
On further calculations, it's maybe only like two or three bags. Still, a substantial portion.
0
0
1
@Abhi__Khatri
@travisneilson
Damn it, I just signed up but I guess I won't be able to read it. Pretty disappointing. :P
1
0
1
@michaeljelly
Agree, very Deutschian! In my blog post I explore how the “fallible” model learns to use the information locally to its position on the board to figure out if it just made a mistake or not. In a way, this form of error correction seems to approximate search, by trying stuff out
0
0
1
@nickcammarata
@gwern
Tried getting myself addicted to meditation by using nicotine at the same time, and had some but limited success - did using nicotine and working out help you?
1
0
1