@anthropicai. previously: @cloudflare, @replicate, @docker
San Francisco
- Tweets 3,653
- Following 1,237
- Followers 9,145
- Likes 2,664
When I was building Replicate I tried to find best practices about how build CLI programs, but nothing substantial existed.
So, I've been working with @AanandPrasad, @tashian, @evpari, and @markhurrell on this: clig.dev/
9
59
255
If you're fed up of reading machine learning papers as PDFs, you might like this: arxiv-vanity.com/
10
55
181
It's been quite a year...
Businesses are building on open-source AI. But we’ve only reached a tiny fraction. That's why we raised a $40M Series B.
Open-source is open for business 😎
replicate.com/blog/series-b
33
6
192
43,560
Don't want to install a CLI tool on your laptop ever again? Try github.com/bfirsh/whalebrew . It's like Homebrew, but uses Docker images.
9
93
122
6 years ago, @allnoteson and I created arXiv Vanity, a tool that rendered papers from @arxiv as web pages so you didn't have to squint at a PDF.
arXiv now has papers in HTML, so we've shut it down and redirected it to arXiv.
arxiv-vanity.com
8
8
128
24,460
Uh. Does anyone in the Bay Area want my organ? I got it from a scrap yard a few months ago and have mostly restored it. But, I'm moving and it doesn't fit in my new house.
11
36
100
Well, gosh, this made my week. 🙏
I’m just the guy dming you — there is also an extraordinary team who actually make the thing work 😄 @evilstreak @thejillianross @nickstenning @zeke
3
95
Machine learning is like software engineering 20 years ago: nothing's in version control, people are FTPing (well, "S3ing") files around, nobody knows what's running where.
We're making better tools for ML. First step is version control: replicate.ai/
3
17
97
AI Grant is the best startup peer group we have. It's like a super intense, high quality, tiny YC for AI companies.
Also you'll get $50k Replicate credits!
AI Grant's second batch is now accepting applications!
bit.ly/3NJI6Rl
7
96
46,564
You might enjoy this. Similar idea. replicate.com/dribnet/homage…
3
8
89
Personalized ads, personalized sales content, automatic marketing content, automatic stock images, automatic game asset generation, photoshop for AI. This is all stuff we're seeing people build today. In future, completely personalized content, TV shows, games, etc.
1
69
Miami feels like where all the cool kids are. As an engineer, that's almost the opposite of what I want. I want to be where all the nerds are.
15
63
I volunteered for the @internetarchive a couple of years ago and got a glimpse at some of the brilliant work they're doing around academic publishing. Some of it is now public in a usable form.
Did you know you can now search scholarly materials preserved in the @InternetArchive? Meet Internet Archive Scholar, a search engine for 25M research articles from 18th C journals to #OpenAccess conf. proceedings to #Preprint publications.
blog.archive.org/2021/03/09/…
1
3
53
I think it bears repeating. The web was invented 30 years ago to share academic writing. WHY ARE ALL CITATIONS NOT HYPERLINKS?
5
17
39
Rarely am I proud to share a vulnerability report...
"We cannot overstate how well Replicate communicated with our team and how swiftly they acted to fix the issue. No customer data was compromised, and no action is required by customers."
"Replicate not only demonstrated a deep commitment to maintaining the security of its products, but also upheld the highest standards of communication and transparency. ... The Wiz Research team has rarely encountered such professionalism and expertise."
wiz.io/blog/wiz-research-dis…
Security is hard and we take it really seriously. Kudos @nickstenning, @philandstuff, and team.
2
8
44
8,051
I’m sure you’ve thought of this, but I had a similar problem with London embassy and instead went to Lisbon. No problem getting visa there despite not being Portuguese citizen. Apparently some embassies are more strict about this so might be worth calling ahead.
1
1
38
I've been working on a thing with @justincormack and @amirmc that's a bit like serverless, but using Docker: github.com/bfirsh/funker
4
17
33
A huge amount of plumbing behind the scenes went on to make this work. I know a lot of you have been asking for this.
I'm particularly excited by native training on Replicate. A sneak peek of more to come. :)
Language models are now on Replicate.
Open-source language models have been improving at an extraordinary pace. You can now deploy, run, and fine-tune them in the cloud with just a few lines of code.
replicate.com/blog/language-…
3
5
35
24,252
This was the trigger for us to ship fine-tuning.
It was *checks notes* 2 weeks ago.
Since then we've designed and built a training API, added support for long-running training jobs to our infra, designed and built a generic interface for training in Cog ...
A few weeks ago I went to an LLM hackathon and no one had an easy solution to fine-tuning. The best advice was "hm idk maybe try and get some a100s on GCP and run it in a Colab notebook?"
Pretty cool that you can do this now on Replicate.
4
4
36
34,350
This is *incredibly* exciting.
Stable Diffusion resonated because it was comparable quality to DALL-E, but you could run it on your own machine and tinker with it.
This is happening right now for language models. Alpaca is comparable to ChatGPT, but you can tinker with it.
We replicated Stanford's Alpaca, by @rtaori13, @lxuechen, @Tianyi_Zh et al. It's a fine-tuned version of LLaMA that can respond to instructions like ChatGPT.
Here's how to train and run it on your own machine so you can tinker with it: replicate.com/blog/replicate…
1
2
31
5,161
I normally I don't talk much about stuff I do, but maybe this is interesting? I'm cycling around Europe, city to city, working on the road.
3
1
28
This is going to make replicate SO much better if you're fine-tuning.
Cold boots for fine-tuned Llama 2 and SDXL are now under ONE SECOND 🤯
replicate.com/blog/fine-tune…
5
28
4,822
I went travelling a few years ago and this is the framework I used: news.ycombinator.com/item?id…
(Maybe I should actually write about this rather than document my life through Hacker News comments.)
1
3
30
Your periodic reminder that cycling for an hour extends your life by about an hour.
uu.nl/en/news/dutch-bikers-l…
I like to imagine I am riding a machine that makes time stand still, every pedal stroke skipping me across space-time.
1
4
24
Yes! It is particularly important for developers. They care much more about "what" than "why".
Also, I always say "show me the damn thing!" when reviewing landing pages. Show a screenshot, show a snippet of code, etc.
3
26
3,907
Comprehensive list of “no shoes” startup offices:
• @Cursor_ai
• @lovable
• @thinkymachines
• @slopepay
• @spurtest_
• @appliedint
• @speak
• @tryhyprnote
• @journeyfoods
• @crossmint
• @every
• @fiftyyears
• @CompositeAI
• @14__ai
• @mainframe
• @replohq
• @remarkablepaper
• @composio
• @replicate
• @Veriff
• @supercell
• @rimelabs
• @onkernel
• Wayfound AI
• Peec AI
• Manifold
• @whop (past policy)
• @notionhq (past policy)
• @gustohq (past policy)
2
31
4,466
This is one of the coolest things being built on @replicate. Congrats on the launch. 🙌
✨ Finally, it’s here.
🚀🌍 Today, EzDubs is launching fully real-time phone call translation and multilingual communication. You can translate phone calls with your own voice and emotions.
Read more in our TechCrunch announcement: techcrunch.com/2024/12/11/yc…
1
3
14
3,363
One of those "nothing's changed" changes that was actually an enormous amount of work behind the scenes. @evilstreak @nickstenning and Andreas effectively rebuilt our whole stack in 2.5 weeks, in a race against growth. 😅
Over the past few weeks, we've made some major improvements to our infrastructure to make it more reliable and perform better. Nothing's changed from your point of view, but you'll be seeing faster response times! 🚀
1
1
22
I wanted to call it "Plum" but thankfully @AanandPrasad had better taste in names.
Happy 10th birthday to Fig (which you might know as Docker Compose)! github.com/docker/compose/co…
ALT The Fig logo (which I created by grabbing a picture of a fig from Wikipedia and making it look vaguely vector-y in Photoshop)
3
24
3,297
Trying to get to sleep but I can hear my neighbours talking about a Docker problem through the window.
1
21
Remember that little Fig thing @aanand and I made? It's going to become an AWS service soon :)
.@awscloud is going to integrate #dockercompose #dockerswarm w/ #AWS #ECS !!! #dockercon /cc @aanand @bfirsh @abronan @vieux @aluzzardi
3
13
21
Perhaps this is a reason developer tools remain so hard to use. Once you put in the hard work to understand a hard technology, it is inconceivable that it could be easier to use because that undermines your identity as a clever person who can understand complicated things.
3
4
19
Looks like the Llama that got benchmarked on Replicate is on an old server :(
Base model is running the old one, chat model is ~35 tokens/s: replicate.com/meta/llama-2-7… (see "tokens" tab)
New server coming to all models soon!
9
1,477
What's particularly great about working at Replicate is you get to work with @zeke 🥰
We're building a pretty sweet team at @replicatehq.
I recorded a short video to highlight everyone on the team, what they do, and why I love working with them.
loom.com/share/1058d4d8746a4…
2
1
21
4,701
Do you like Arxiv Vanity? This month's AWS bill reminded me that it actually costs quite a lot to run. Help support us maybe? ❤️ patreon.com/arxivvanity
1
8
18
We haven't shipped many public features this year, because there's just been a ton of internal work we've been doing to grow.
Pleased to say we're back to shipping big things. Lots more coming.
FLUX is now much faster on Replicate. Generate images in as little as 300ms from request to response. 🔥
We've also made all our optimizations open-source. You can see exactly what we've done and build upon it. More details (and a secret beta feature) in this thread.
2
21
1,666
This is absurdly fun. replicate.com/bfirsh/bfirshb…
Train and deploy DreamBooth models on Replicate. 🎩 🐇
With just a handful of images and a single API call, you can train your own custom Stable Diffusion, publish it to Replicate, and run predictions on it in the cloud.
replicate.com/blog/dreamboot…
2
1
18
Last year I started building the JS framework I wanted to exist. Something that gave me the buzz of productivity that Django did.
I have moved on to other projects now, but I've flipped its open-source bit in case it's useful: github.com/bfirsh/miles
3
1
20
How we survived a day on the Hacker News homepage using just a single Heroku hobby dyno: medium.com/@bfirsh/squeezing…
6
17
We've made a few updates to arXiv Vanity recently, most notable is a new design, based on some ideas from @markhurrell (thanks Mark!). Looky here: arxiv-vanity.com/papers/1308…
2
2
16
It taught me about UNIX, web apps, deployment, systems, dev tools, Xen, containerisation. It gave me a career, and now a cool job at Docker.
1
1
17
For each hard technology that has a loyal priesthood, there is an opportunity to design something better that the other 99% can use.
It is generally true that if you can fool developers into thinking they are "mastering" something hard (as opposed to learning tolerance for something badly designed), you can build a fiercely loyal priesthood.
2
5
16
Come and build beautiful things with us :)
We're hosting a hackathon in San Francisco!
📅 Sunday, September 17th.
🎨 The theme is AI Arts & Crafts.
🏆 Winners will get $10k in Replicate credits (and everlasting glory).
Sign up:
partiful.com/e/ac7NHwQ8ZLVIr…
3
4
14
5,301
This is something I'm particularly excited about.
Trying out models is not just for play. Unlike normal software, you can't inspect how a model works, so the only way to see how it works is to run data through it and see how it behaves. It's a core part of the development flow.
One more thing. We have a new way to experiment with image models on Replicate. In the playground you can generate images one after another and tweak settings in real time.
Playground is in beta, and we'd love your feedback.
replicate.com/playground
3
16
1,189
This blog post gives a bit of insight into the thought and care that went into Docker for Mac's file sharing. blog.docker.com/2017/05/user…
2
2
13
E.g. bluetooth crudely shares spectrum by frequency hopping. It changes frequency 1600 times a second in a pseudo-random order. Every other device probably doesn't have the same pseudo-random order, so they probably won't collide and cause errors. But, sometimes they do.
1
15
Arxiv papers can be translated using Arxiv Vanity and Google Translate! This is such a good demonstration of why academic papers should be published as web pages instead of PDFs.
英語がそこまで好きでない人にオススメなarxivの論文の読み方
1、arxivで読みたい論文のURLをコピーします。
2、arxiv-vanity.com にURLをペーストします。
3、数分待つと論文がHTMLになります(時々無理な論文がある)
4、そのURLをGoogle翻訳に突っ込みます。
→論文が読みやすい日本語に！
4
15
You might like arxiv-vanity.com
We’ve been meaning to add “download CSV” to tables but haven’t got round to it yet. We could also let you download the source data for plots, if it were generated in LaTeX. Usually it isn’t though, sadly.
1
15
So we made a place where ML researchers could share their work as actual, production-ready software: replicate.com
1
1
15
1,165
+1 I realize I’ve built my career around this.
My first startup was Heroku for Python. (Before Heroku was Heroku for Python!)
My second startup was Heroku for Docker. (Docker Compose)
My current startup I often describe as Heroku for ML.
3
13
... you get the idea.
We're a team of 12 and my goodness can we ship.
The generative AI world is changing every so fast. Shipping high quality stuff at this velocity is what is going to make us succeed.
If you enjoy shipping... replicate.com/about#join-us
12
1,019
On my last day, as a prank, I rewrote the whole internet on the school computers. Google's logo became "Leavers '08", Facebook became "Hatebook" and was red, YouTube only played videos of cats, amongst other things I can't remember.
1
1
13
Just announced Docker Machine at #dockercon! If you remember Orchard, this might look familiar. ;) github.com/docker/machine
3
26
12
I have tried and failed to build this several times over the past few years. This would have saved me so much time had it existed then: replicache.dev/
Something about tools with names starting with "Replica-", eh.
10
We've been saying for a while "Replicate is just for serving" but turns out fine-tuning is kinda like a really long prediction where the input is training data/params and the output is weights.
2
1
12
A thing I worked on with @aanandprasad and @j0ffrey has now shipped: a much better Python library for Docker. docker-py.readthedocs.io
1
4
9
arXiv Vanity now attempts to turn citations into links. It does its best to find the preprint on arXiv and link to the HTML version, so you never have to go back to PDF-land.
Look at all those hyperlinks! arxiv-vanity.com/papers/1904…
1
3
10
OH from @zeke: "How did we live without Copilot. Everything's different now."
1
11
3,717
We put a heck of a lot of work into making download sizes small for the Guardian app. Very pleasing downloading this now.
1
9
This is incredible.
✨🐍 Playing around with mockups and prototypes of what a JupyterLite / Pyodide experience would feel like within @Github, on .ipynb files.
Shown: a WASM-powered @ProjectJupyter notebook, running 100% in the browser (no need to launch containers on VMs).
github.com/jtpio/jupyterlite
1
11
If you follow me because of Docker and don't care about science, the rendering system is quite interesting: github.com/arxiv-vanity/arxi…
1
9
My @containercamp slides about using Docker Compose for more than just dev environments: speakerdeck.com/bfirsh/docke…
1
4
8
This one's a lot of fun.
Finally, StyleCLIP is up at @replicateai ! 🎊
Replicate allows to easily deploy ML models so that users do not need to install anything, or to write a single line of code 😲
Enjoy! 😊
replicate.ai/orpatashnik/sty…
2
10
"YAY DOCKER" - CIA agent #524297 wikileaks.org/ciav7p1/cms/pa…
3
5
9
Wahey!
We're cutting our prices in half.
For public models, no action required on your part. You'll just see a smaller bill next month. Enjoy. 🍹
replicate.com/blog/cutting-p…
10
1,281
The slow death of PDFs continues. Very pleased to be working with @allenai_org on this.
Announcing a new way to read papers: Starting today, read papers published via #ArXiv on #SemanticScholar as responsive HTML docs that will render correctly whether you’re on a phone, tablet, laptop, or desktop computer. Learn more: blog.semanticscholar.org/ann… From AI2's @codeviking
3
10
Anyway the @internetarchive is fucking amazing I could talk for hours about the brilliant work they do. Stop reading this and give them some money archive.org/donate/
1
9
arxiv-vanity.com/ now uses LaTeXML instead of Pandoc to render papers. If a paper you tried failed to render, give it another shot. It should work now!
1
2
9
I was 12 when I got my first Linux machine. I remember sitting on my friend's floor, him showing me how to install Debian.
1
1
8
If you have a Heroku app this might save your butt one day: github.com/bfirsh/heroku-tar…
1
1
8
I've made a Twitter account for @replicateai so I don't spam you all with everything about my work. Go follow that if you're interested in ML tools & reproducibility!
1
9
It was before they'd figured out how to make oscillators that could produce different notes, so there are some 150-odd individual oscillators, each tuned manually with a stack of soldered ceramic capacitors.
2
8
Oh and we also have img2img if you don’t want to set up on your own computer! replicate.com/stability-ai/s… (“init_image”)
1
9
also featuring a Compose app spanning multiple hosts using Swarm, Weave, Flocker and Powerstrip! github.com/bfirsh/powerstrip…
1
4
7
replicate.com is probably best for for vision models. (I am biased.) Also I think the only YC company of the bunch. DMed you details. :)
8
Django is still my go-to for getting apps up and running as quickly as possible.
If, like me, you happen to be spinning up lots of new projects right now, here is a template that'll get Django running on Heroku with best practices in 10 minutes: github.com/bfirsh/django-doc…
1
1
7
It should work like that, if I'm understanding correctly.
If you upload the file to /v1/files, you get a URL back that you can pass as an input to multiple predictions.
More details: replicate.com/docs/reference…
1
3
70
They're not all bad! There is an inverse relationship between the cost of a microwave and the quality of the UX. I think ours was the cheapest on amazon.
1
2
7
It was a weekend project that kind of spiraled out of control. At some point we were like "are we writing a book?!" Thankfully we scaled it back to a very long web page.
1
7
Also a thing I forgot to tweet about... I got Duck Hunt working a few months ago! (Click to fire the zapper.) jsnes.fir.sh/run/Duck%20Hunt…
2
1
7
So that's where we're at. Turns out making version control systems is hard though, and this is just a start. If you're interested in this stuff and want to help out, we've got a Discord server: discord.gg/QmzJApGjyE
8
Here's a thing I've been working on: Building serverless apps with Docker blog.docker.com/2016/06/buil…
2
7
8
Also available in glorious HTML arxiv-vanity.com/papers/1502…
The software to create the black hole in the movie 'Interstellar' is a full implementation of Einstein's equations in 40,000 lines of C++, and rendered thousands of 23-megapixel IMAX frames on a 32,000-core render farm at about 20 core-hours per frame arxiv.org/pdf/1502.03808.pdf
2
1
7
arXiv Vanity was a hit, so we started collaborating with arXiv to get it built into arXiv itself. We got sidetracked, but it was thankfully picked up by @dginev and this is the result. An "HTML" link on the abstract page!
blog.arxiv.org/2023/12/21/ac…
1
1
9
654