Julia Turc
@juliarturc
Followers
1,738
Following
454
Media
46
Statuses
360
Co-founder of @StoriaAI (YC S24) | Ex-Google Research
San Francisco / Bay Area
Joined September 2015
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Ohio
• 491447 Tweets
Springfield
• 330867 Tweets
Kendrick
• 281054 Tweets
Haitians
• 280873 Tweets
#AppleEvent
• 259666 Tweets
EP1 U STEAL MY HEART
• 253659 Tweets
Wayne
• 201339 Tweets
iPhone 16
• 173728 Tweets
Beyoncé
• 110593 Tweets
Tim Cook
• 107729 Tweets
Super Bowl
• 53830 Tweets
Jets
• 47335 Tweets
Wales
• 46033 Tweets
Draghi
• 45971 Tweets
ايفون
• 36601 Tweets
Fassi
• 33954 Tweets
Catherine
• 28001 Tweets
Culiacán
• 27283 Tweets
Talleres
• 23496 Tweets
Tapia
• 22601 Tweets
Hakan
• 18971 Tweets
Camera Control
• 18215 Tweets
AirPods 4
• 16632 Tweets
PS5 Pro
• 15410 Tweets
Birdman
• 14962 Tweets
$INSDR
• 13592 Tweets
対象作品
• 13058 Tweets
Belgique
• 12379 Tweets
iOS 18
• 11894 Tweets
Gladys
• 10347 Tweets
USB-C
• 10240 Tweets
Merlos
• 10050 Tweets
Efficient BERT models from Google Research, now available at !
We hope our 24 BERT models with fewer layers and/or hidden sizes will enable research in resource-constrained institutions and encourage building more compact models.
2
197
628
New from Google Research:
To build multilingual NLP systems, a successful recipe is to pre-train on a multilingual corpus, and then fine-tune on labeled data in a single transfer language -- usually English. But is English best?
4
73
308
I've given in to the peer pressure and relocated to SF / Bay Area. DM me if you'd like to grab a coffee and hear me bicker about how unwalkable and bagel-less this place is. I can be positive too, but that might require wine instead of coffee.
9
3
82
Whenever I mention to my batch mates that
@dflieb
is our YC partner, the first reaction is: "I LOVE his talks. He always offers these data-driven frameworks of thinking that are insightful and actionable".
The magic of YC is that all partners are former founders, which takes BS
How do you actually know if you’re making something people want?
One of the best ways to measure successful growth is with cohort retention, which tracks the fraction of new users that come back time and time again to use your product.
14
58
411
1
4
60
We built to better understand how
#GenerativeAI
can amplify long-form content creation.
Here are 7 learnings from this experience (🧵👇).
Read the full article at
2
18
42
There’s so much OSS out there, and I’m like a kid in the candy store. I want to try it all, but there are only so many hours in the day. Sometimes I just want to ask some high-level questions before deciding whether to invest in a repo.
We made Code Sage to scratch that itch:
3
8
39
I wish the training scripts on
@huggingface
weren't so full of boilerplate. It's so hard to gauge what the training algorithm is when most of the file is defining flags and setting up logging. Am I too nit-picky?... Is there no better way to organize machine learning scripts?...
8
4
37
After spending 8 years I Google, I really thought we had the best engineering in the world. But now when I constantly need to make choices between AWS and GCP,
@awscloud
wins EVERY. SINGLE. TIME. By far.
7
2
34
Crafting effective prompts for text-to-image models like
#dalle2
takes a lot of tinkering. We created a text prompt generator by fine-tuning
#gpt2
on queries from
@midjourney
users.
Model:
Dataset:
Collab w/
@gaurav_nemade15
0
6
28
One of the most frustrating failure modes of generative image models is unwanted pseudo-text. We're open-sourcing Detextify, a Python library that removes it. Contributions are welcome!
I'm super excited to announce Detextify, a tool
@juliarturc
and I built to clean up images created by your favorite generative AI models.
The backstory:
1
2
17
3
6
27
Text-to-image models can illustrate concrete descriptions of visual scenes, but struggle with more abstract concepts. We fine-tuned
#dalleMini
from
@craiyonAI
on blog post articles and images from Medium. Try it out!
Demo
Context
2
6
23
@levelsio
OCR to mask the text regions + edit via DALL-E with a prompt that asks for something specific enough to not be in-painted with text again.
3
1
24
Our BERT miniatures were pre-trained directly with MLM loss. They are competitive to more elaborate pre-training strategies involving MLM distillation (). Our models can be fine-tuned for downstream tasks via standard training or end-task distillation.
1
3
22
Building coding agents today feels so similar to self-driving cars. Most of the effort goes into getting machines to interact with an interface that was designed for humans. Feels like we should just gut VSCode (and roads) and start with a clean slate.
7
2
21
Main takeaway: choose carefully the language to annotate when the goal is a highly multilingual system, and don’t assume that English will always perform the best.
2
0
18
I have never felt more accelerated in my lifetime than now. LLMs are taking all the drudgery away from coding, so that I can focus on what matters.
I am looking forward to a future where none of us move buttons on a screen, and instead go to space or cure cancer.
0
2
17
The irony of AI today is that, at the same time, we're making HER a reality and also struggling with the most basic stuff, like installing a damn library.
I struggled with ONNX more than I would have liked to, and wrote a guide that might save you some time.
🧵
1
2
16
Surprisingly, in 5 out of our 6 experimental settings, non-English fine-tuning data leads to more effective zero-shot transfer -- even when this data is automatically translated from English!
1
2
16
😲 Less than 1M developers received a suggestion from Copilot in the last year out of 26M developers worldwide (slide by
@sqs
at
@aiDotEngineer
conference).
Code assist adoption is still low, but you only need to convince 25M people to try it once and there's no way back.
Just
0
6
15
Hey besties, I heard you're on the market for changing the text on your AI-generated tequila bottles. Maybe try ?
@theallinpod
@Jason
@DavidSacks
3
2
15
Thank you
@MadronaVL
and the judges for the recognition! Honored to be part of an event with so many strong founders.
Can't wait to grow together with
@mihail_eric
into the go-to platform for long-form content generation!
Congratulations to the winning team from Launchable: Foundation Models, Storia AI! We are so excited for what’s in store for founders
@juliarturc
&
@mihail_eric
Thank you to the builders, judges, mentors, speakers & sponsors for making this event possible.
@MadronaVentures
0
1
9
1
1
14
Thanks
@TDataScience
for the highlight, and for your support in general!
"Creativity starts when you truly internalize a concept or technology, are able to question it, improve it, or use it in a new context."
In our new Q&A,
@IuliaTurc
discusses her fascination with NLP and shares advice for early-career ML practitioners.
0
1
8
0
4
14
Really fun working with
@mihail_eric
on
#rickandmortify
, an AI playground to generate your own
@RickandMorty
storyboards.
We're already LOL'ing at the premises people are coming up with, like "Rick and Morty go to six flags but just for the tasty chicken wings".
I'm super excited to announce Rick and Mortify, a system
@juliarturc
and I built for creating never-before-seen episodes of
@RickandMorty
by leveraging the state-of-the-art in generative AI.
A longer origin story:
9
28
88
4
4
14
Generative AI shines when embedded into real-world workflows. Storyboarding is an established process in filmmaking that can be supercharged by AI's ability to quickly experiment with virtually infinite cinematic styles.
Follow
@StoriaAI
's blog to find out what we're up to.
0
3
13
Thank you
@MadronaVL
and
@mfridgen
for taking the time to chat with me about
@StoriaAI
and the challenges of being a first-time founder. Your continued support makes this journey so much easier!
Check out our Founder Spotlight on
@juliarturc
, co-founder & co-CEO of
@StoriaAI
. In this spotlight, Julia shares her journey from NLP researcher on LLMs at Google to early-stage CEO:
0
1
5
0
2
13
I spent literally days trying to find a model or API that can tell me the closest Google font for the text in an image. Couldn't find it, so we made our own. You can find it on Github, or try it out directly in Storia Lab. 🚀
1
2
12
Finally, the end of small talk is near
Waymo is now available to everyone in SF, no invite code needed. The future is here.
Now let’s make sure SF bureaucrats don’t ban it.
56
97
2K
3
0
12
New form of masochism: run stable diffusion inference without using
#huggingface
at all in your pipeline ☠️. This is going to be a long weekend.
4
0
11
We made it so that you can index your own repo on Code Sage and I can get some sleep.
Exciting update:
@StoriaAI
Sage now allows you to index your own codebase, in a fully self-serve fashion.
All you need is a public Github link, and 3 minutes later you'll have an AI expert assistant that can answer questions and contextually generate code related to your repo.
4
3
37
0
1
11
My Chrome tab groups look like a ML stack 101 course. Feels almost monetizable, given that I've been curating them for weeks. New idea: marketplace for tab bundles.
5
0
11
Going from text to image takes so much more than a text-to-image model. We never truly want to start from *text*, but rather from *context* -- a story, a screenplay, a podcast. It turns out that, in order to generate great visuals, you actually need great NLU.
Midjourney, Stable Diffusion, and DALL·E 2 are impressive but they run into big problems at scale.
Here are 6 vision-language grounding challenges in applying these generative models to long-form content like movie scripts, stories, and podcasts:
2
1
11
0
1
10
My primal brain is ecstatic when it sees patterns. And text-to-image models are depriving me of said patterns, even the training loss is all over the place. But I finally found a validation metric that is monotonically going down. It's... just... bliss.
2
2
11
The power of AI is not creativity, but scale. Real stories can only be written by real humans. Delivering and amplifying them is where AI can complement us.
You pour your soul into writing a screenplay,
@StoriaAI
puts its GPU cycles into illustrating it in 1000 ways.
I'm excited to share a sneak-peak of the future of AI-generated storyboarding.
A complete shot list for the script of "Women Talking" in the style of a Wes Anderson film created entirely by generative AI. Done in minutes.
1
2
23
0
3
10
6/7
For reliable GPU access, try Lambda Labs.
The three big clouds are very protective of GPU resources and prioritize lucrative contracts rather than the pennies they get from individual use. With Lambda Labs, the GPU is yours with the click of a button.
1
0
10
I feel heard. The "big corp struggle" language is so universal.
How Alexa dropped the ball on being the top conversational system on the planet
—
A few weeks ago OpenAI released GPT-4o ushering in a new standard for multimodal, conversational experiences with sophisticated reasoning capabilities.
Several days later, my good friends at PolyAI
86
305
2K
1
0
10
I'm not good with superlatives, but I'm lucky that my cofounder is. I am proud though that, in an industry moving at a *crazy* pace, we put something out there that nobody else has. Textify is just the beginning of a thoughtful, practical and intuitive experience for creating
Textify is live on Product Hunt today!
Have you ever generated an AI image and thought: "This is incredible...but what is that gobbledygook text"?
Textify:
✅ Uses state-of-the-art AI to seamlessly fix textual artifacts and blend text back into your image
✅ Preserves the
4
2
15
2
2
9
We are often told to "show, not tell". But some of us are better at "telling". So we made a tool that helps with the "showing" part.
StoriaBoard is live on Product Hunt today. Help us tell our story by upvoting, commenting, or sharing our launch page. 🚀
0
3
10
You know what would make
@paulg
's essays better? Gratuitous AI illustrations with
#stablediffusion
.
Introducing gratuitous AI into your startup is at least going to be more useful than introducing gratuitous crypto stuff.
101
144
2K
1
1
9
Our
@GrainHQ
meeting recorder is a great productivity boost and ego killer. I can only hope they trained a "worst frame" classifier for their video thumbnails. Otherwise
@mihail_eric
will just have to convince himself he's the "smart" one in his family.
1
0
8
If you're in SF today, come and join us at the AI Quality Conference. I'll be talking about serious stuff and
@mihaileric
will close the event with a nerdy comedy bit. Story of my life.
1
1
8
If you are building in the
#GenerativeAI
space, are interested in democratizing long-form content creation, or are a creative looking to speed up your workflow,
@mihail_eric
and I want to hear from you!
2
2
9
I just realized why this `webuiapi` Python library for
#a1111
makes me chuckle.
History:
1. Stable Diffusion code is all in Python
2. That's too much code, we'll make a web UI for it
3. What should we call it? Uhm `webui`?
4. Actually we kinda liked Python better
5. Make an API
6
1
9
I HAVE BEEN LIVING A LIE. Do you know what the output of `os.path.join("a", "/b")` is in Python?...
Since when is Python basically Javascript?
7
0
8
2/7
#GPT3
is not a silver bullet for text generation. For instance, Rick's voice via
@character_ai
is net superior to vanilla GPT-3 (though
#ChatGPT
might help close the gap).
In general, we expect a great unbundling of text generation into vertical products.
1
0
7
Remember the days when StackOverflow was down and we were like "I guess I'm taking the day off?". That's
@huggingface
now. I don't appreciate it enough until it's down.
0
0
7
Fun fact: today on
@ProductHunt
,
#Claude3
(which may or may not be sentient) is barely beating a TODO app on honorable position
#6
.
4
0
7
Our Italian vocabulary grows every day, from Storia to Nutella. What should be next?
4
0
7
I've been trying a bunch of agents and most of them get stuck in infinite loops😵💫...
Before: "oh no, the UI got updated so my python script is broken"
Now: "oh no, the LLM got updated so my agent is broken"
2
0
6
more AI demos to kick off the year with
@sundeep
!
this week, we focused on different tools for image generation and editing:
--
@RoomReinvented
: AI-powered interior designer
grades: JC: B, SM: B
-- Face ID Plus on
@huggingface
: photo enhancer
grades: JC: C+, SM: C
--
16
12
76
1
0
7
I wish LLMs would be more like grumpy cat sometimes, especially for code generation.
Often times the answer is just *it's not possible*. Instead they keep suggesting ok-looking things that end up wasting your time.
ChatGPT & co, please be mean to me.
1
0
7
Most services from the 3 big clouds are tomato-tomato, but that's not really the case for fully-managed model inference. SageMaker from
@awscloud
is just better on all possible dimensions than Vertex AI from
@googlecloud
. Here are some reasons 🧵
Based on her own industry experience and deep knowledge of AI product development,
@juliarturc
presents a nuanced comparison of SageMaker and Vertex AI for model inference.
0
4
16
2
1
7
7/7
For reliable text-to-image inference, try
@replicatehq
.
With generative AI, you don’t want to re-invent the wheel. Calling the
#stablediffusion
inference API via Replicate is quite affordable ($0.0023 per second). Their staff is really helpful and responsive too.
2
1
7
You know how it's hard to trust code from Claude/ChatGPT sometimes?
We're running a little experiment where we (1) heavily link model responses to docs/code, and (2) reindex overnight for freshness.
Does this increase your trust in generated code? Red pills only please!
1
1
6
5/7
The future might not be fully generative.
The holy grail is finding the right synergy between humans, traditional tools, and generative AI.
Most of the credit still goes to the talented creators of the show:
@JustinRoiland
and Dan Harmon.
2
1
6
Storiaboard is one day old and already democratizing filmmaking: "Such a powerful tool for people who don't have the means to communicate their vision. It cuts down the giant Pacific Ocean between my vision and my partners." (film writer / director)
We’re thrilled to announce the public launch of Storiaboard, our AI-powered platform for supercharging video preproduction workflows.
We’ve been in the trenches building for more than three months.
And finally we think we’re on to something...
1
1
4
0
0
6
Our friend, advisor, and award-winning filmmaker Trey Nelson spoke about Storia on a podcast episode by The Cast Station.
Whenever I talk to Trey, I feel the need to write things down. There's always so much substance and poetry at the same time.
1
0
6
While building , we experienced first hand how brittle generative models are: unwanted pseudo-text, deformed body parts, inconsistent characters.
There is still a lot of work to be done towards robustness and trust!
0
0
6
When
@mihail_eric
and I started working together, we half-jokingly decided to generate new episodes of Rick and Morty. But most images that would come out of Stable Diffusion and DALL·E at the time were full of gibberish, in a failed attempt to spell the name of the show. That
Leveraging the power of AI to edit text within images remains a challenging use case. Julia Turc offers a comprehensive overview of recent research into scene-text editing and explores promising future directions.
0
1
6
0
2
5
Startup idea (for someone else, my hands are full): Hugging Face Hub, but for vector databases.
1
0
6
By popular demand, we're adding the ability to index GitHub Issues in our open-sourced repo2vec. Coming soon to our hosted version too!
GitHub Issues now surface as context in the conversation when relevant. For instance, here I'm talking to
@huggingface
's `transformers`
2
5
21
3/7
Image generation severely suffers from inconsistency across runs.
#dalle
and
#stablediffusion
assume users want a single output, but this assumption breaks for episodic content.
Here's how inconsistent Rick can be in different contexts:
1
0
6
@vikhyatk
With all those Huggingface tabs open you have to take some responsibility for the outages we've seen recently
1
0
6
@mathemagic1an
That should be straightforward. I added an issue so you can track the feature request:
1
0
5
@GrigoryEvko
Btw, we are open-sourcing our eval work. It's still in construction, but you might find it useful already. You can use CMMD with other embeddings as well (`insightface` works well if you're training portraits):
2
1
5
Lol Copilot is autocompleting my code with banter with
@mihail_eric
// TODO: [Mihail] Remove this, it's not used anywhere. [Julia] I think it's used in the backend, but I'm not sure. [Mihail] I think it's not used in the backend either, but I'll double check.
1
0
5
4/7
In-painting needs some tweaking too.
We worked around the inconsistency problem by *in-painting* cut-outs of the characters from frames of the TV show.
But to prevent the character cut-outs from getting distorted or blended into the background, we needed more hacks.
1
0
5
Interesting to see what a huge difference the sampler makes during inference. Has anyone played with swapping different samplers during *training*?
#stablediffusion
#GenerativeAI
The scheduler you use for image generation in
@StabilityAI
diffusion models makes a HUGE difference in quality.
Check out this comparison between the DDPM and Euler Ancestral schedulers:
0
0
11
0
1
5
Everyone and their mother (including me) is releasing open-source repos.
Github is the new Substack, and every GPT wrapper is a thought piece about how AI should be used.
2
0
5
@alexandr_wang
Isn't it just as touristy for
@scale_AI
(which has been focusing on data annotation for years) to announce a generative AI-based *marketing* tool?
Nothing wrong with that, just pointing out the dissonance.
I wanted to preview of one of the coolest products from the
@scale_AI
labs.
We're enabling marketers to AI-generate UNLIMITED and INFINITELY CREATIVE images of their products for:
- ad creatives
- brand campaigns
- social media
Every image in this thread is AI generated🧵
36
105
905
1
0
5
Synergistically what now? Feels like I got slapped in the face with a dictionary.
0
0
5
2/ Once you're ready for production, switch to the latest version on PyPI instead of building from source. You'll have to very carefully match various versions (OS, Python, CUDA, cuDNN, ONNX runtime). More details on how to get them to match in my article:
1
1
5
@ptsi
Reminds me of a panel at AWS Reinvent last year. Someone asked
@aidangomez
what advice he has for founders who think they're to late to the AI game, and he said something along the lines of "you should actually worry you're too early"
0
0
5
@ptsi
@jayyeh
@hnshah
@sbyrnes
@djsampath
@graiz
@tanayj
@annetgriffin
Thanks for building it public!
0
0
4
Tried
@cursor_ai
for the first time. Delightful!
"AI Fix in Chat" is my favorite feature. Normally I'd have to copy the code and error message into ChatGPT, then come up with a "pwetty please fix" prompt.
The only concern is they forked VS Code... are they going to stay on top
1
0
4
If you're running
#stablediffusion
or other GPU-intensive models, what's your go-to GPU provider, beyond Google Colab? Bought your own? Renting from one of the big 3 clouds? Some startup with a neat offering?
5
0
4
AnyText is supposed to be state-of-the-art for editing text in images. I'm trying to use their playground, but can't get a single example to work.
Here I'm trying to get "ORANGE". What am I doing wrong?...
1
0
4
Glad to know I’m not the only one cringing at “3 ways you’re promoting ChatGPT wrong” posts / TikToks etc. Learning is sweat and pain. It’s fighting with CUDA errors and repeatedly being told by the compiler you are dumb.
# on shortification of "learning"
There are a lot of videos on YouTube/TikTok etc. that give the appearance of education, but if you look closely they are really just entertainment. This is very convenient for everyone involved : the people watching enjoy thinking they are
683
3K
15K
1
0
4
I ask Claude what icon I should use for GitHub issues in my app, and it gives me this link that opens up a .png with the 💩 emoji. I almost feel Rickrolled.
Then I proceed to get stuck in an infinite 💩 loop.
2
0
6
@StoriaAI
@techhalla
@HBCoop_
@chetbff
@ciguleva
@juliewdesign_
@icreatelife
Thank you all talented prompters for supporting this cause! Gen AI is such a great tool to show that there's beauty in everyone.
1
0
4
How can someone find just as much pleasure in deleting code as she does in writing code? In theory it should make you feel like Sisyphus but it really doesn't.
4
0
4
@gneubig
@KreutzerJulia
Good question! There is evidence in the XTREME paper that fine-tuning on all *target* languages can be helpful (see their translate-train multi-task baseline). So investigating combinations of fine-tuning languages would be a promising next step for our work.
1
0
4
@gneubig
@KreutzerJulia
It is particularly interesting to ask how to allocate a restricted data collection budget across multiple effective transfer languages (that might or might not be in the target set).
1
0
4
@mihail_eric
@OpenAI
@sama
@mlopscommunity
I have some connections in the Basque Country. Or remote villages in Romania. LMK if you need a warm intro.
1
0
3
Story of my life
1
0
4
I can't believe that
#dreambooth
just worked on the first attempt. Trained on black dress. Asked for the same dress in red at inference time. Still in disbelief that I tweaked 0 hyper-parameters. Faith in OS restored.
The face could be better though.
3
0
4