Charlie Holtz
@charliebholtz
Followers
19K
Following
11K
Media
459
Statuses
2K
@meltyinc, previously @replicate
The Moon
Joined March 2020
Introducing Ambient Chat. Chat anywhere on your Mac with an AI that can see your screen
21
25
288
David Attenborough is now narrating my life. Here's a GPT-4-vision + @elevenlabsio python script so you can star in your own Planet Earth:
728
5K
27K
this is my best productivity tip. `sudo vim etc/hosts` and add:
327
603
7K
I finally figured out how to skip daily stand-ups. I made an AI clone that knows what I'm doing all day, listens to audio with whisper, and responds with a language model. Here is a sneak peek of Jarley in a real life meeting:
82
297
4K
I put Gemini 2.0 inside a desktop app and gave it access to my screen, terminal, clipboard, and the internet:
72
196
2K
I started a company!. Introducing Melty, an open source AI code editor for 10x engineers. It's the first editor that understands what you’re doing from the terminal to GitHub and collaborates with you to write production-ready code.
125
225
2K
We made a Mac app that lets you chat with Claude, 4o, 01, and Gemini all at once. Introducing Chorus!
138
131
2K
🖼️Introducing CHARL-E: Run Stable Diffusion on your M1 Mac. - 1 click download.- Open source.- Runs locally, no internet required.
53
222
2K
I put Claude in my browser and now my whole internet is back in the 90s:
58
92
1K
I got a fully maxed out MacBook, mostly so I can run local models fast. And omg here is mixtral running on @ollama. almost cannot believe how fast it is. This is with no internet!! A model that beats GPT-3.5 running locally! What!
86
115
1K
Introducing REALITY PONG . Play in your browser. Open source. Built by Melty.
27
51
854
🦓 Introducing Zoo, a playground for comparing AI image models. - but for images.- free + open source.- run open source models + DALL-E.
13
122
727
My desktop is a wall of "Screenshot 2023-12-20-123-beep-boop.png"' and it's too chaotic. I made a script that watches for new screenshots and renames them using Llava 13B:
43
83
666
Introducing YouTune — fine tune image models on YouTube videos. > python tune.py <youtube-url>. • downloads video.• screenshots every 50 frames.• removes near duplicates.• fine tunes SDXL for you.
44
88
664
I made a site to chat with @MistralAI's new 8x7B instruct model!. - free + open source.- streams at ~40 tokens/s .- beats GPT-3.5 on most benchmarks . Try it at
16
82
593
Fix your back problems with this one easy trick that doctors DO NOT want you to know!. Here's how I built a posture correction app using the new GPT-4 Vision API:
29
56
571
You can run this on your own computer now!. I packaged this demo into an app. Download Bartleby →
I put Gemini 2.0 inside a desktop app and gave it access to my screen, terminal, clipboard, and the internet:
21
64
533
Experimenting with a self-healing notebook. Whenever a cell fails, Claude fixes it for me in the background:
26
57
492
my computer is listening to itself and drawing pictures based on what it hears. (I'm running whisper locally to turn audio into text and then making images using sdxl-turbo)
27
61
477
Announcing StickerBaker!. Make stickers with AI. Powered by @replicate and @flydotio, and 100% open-source.
61
59
393
I put a bunch of AI agents into a fake instagram and let them run free. here's my live demo from @ElixirConf featuring @josevalim as the first Shinstagram influencer:
16
51
353
Introducing Synthesis! . We built a new tool that merges different model responses into one:
40
29
360
working on something very important part 2
@charliebholtz Yes now adopt this for flappy bird.
16
18
322
things get pretty interesting when you give the AI access to your file system
15
16
317
I've started printing a weekly magazine of all my favorite blogs
27
3
294
EmojiGen has the largest library of emojis in the world (~250k), and now they're searchable!. I just added embedding search, so you can search the entire library of AI emojis in a couple seconds.
9
40
258
📚 Introducing Teach-O-Matic, an AI YouTuber that creates how-to videos about… anything. • Open source.• Powered by @replicatehq @OpenAI @LangChainAI.• Runs on @GoogleColab - no dev env required. Try it ➡️ "How to Make Spaghetti by Shakespeare"
22
40
216
🖼️ Announcing TileMaker!. A free app for creating seamless tiled images with AI.
10
18
219
working on a portal to the multiverse . image generation on @replicate is about to get REALLY fast
11
25
194
ok ai generated horror is about to be a big thing, this feels like a scene from Hereditary
14
9
160
Introducing. the @stickermule partnership!. StickerBaker is now integrated with @stickermule, so you can make AI stickers and then buy them in real life :)
13
6
145
here's a real-time video generation of @dwarkesh_sp's convo with @patrickc. it is so crazy to me this is possible now! whisper + sdxl-turbo are so fast
13
12
122
I asked melty to add a sign up form and deploy the site. >
3
9
114
This is my fav AI project of the week — a collaborative tile world. Reminds me of @CivGame .
3
9
115
I just got access to an upcoming music generation model and holy crap these are getting good.
11
3
110
Introducing StickerSearch! . I'm embedding all the stickers so you can now search the sticker library in seconds:
10
9
105
Fun update — David Attenborough narrates your life just hit GitHub trending! And it's almost at 4k stars 😎.
2
3
103
we made the ai commit all its changes to git. it feels more like it's working alongside you, and it's easy to undo if it goes off track
6
3
102
Working on a productivity coach that runs locally and tells me if I'm procrastinating. If I set it to hard mode, it sends a screenshot of what I was doing to coworkers on Slack:
9
4
100
Here's the entire behind the scenes of idea to working demo
2
11
92
I used Stable Diffusion, the Twitter API, and GitHub actions to make a button that changes my Twitter photo. 🔴
4
6
84
Here's a demo of an AI pair programmer I've been hacking on. - runs in your editor.- understands large codebases .- can read/edit multiple files at once. let me know if you want to try it!
19
4
88
Introduciiiiing. LLM Boxing Part Two. 🥊 Llama 2 13B vs Mistral 7B 🥊. - every round pick the output you like better.- each model has 5 hitpoints.- pay-per-view. one night only. Last time Llama 2 beat GPT-3.5 handily. How will it fare against this scrappy young challenger?
8
17
88
new experiment: a camera that converts the world into GTA in real time
6
4
89
In Sept I gave my first ever conference talk: Building AI Apps with Elixir. It's about:.- patterns in AI apps.- why Elixir 🤝 AI.- how to make a swarm of agents (incl. a clone of @josevalim) that interact on a site called Shinstagram. Here's my write-up!.
6
25
86
Some of you may have seen AI David Attenborough comment on my "intake orifices" a couple weeks ago. Here's how I did it, and how you can make your own:.
3
9
84
I'm happy to announce that I have solved procrastination and reached peak productivity, all thanks to AI and the constant threat of shame.
3/ Procrastination Coach. Open source and local productivity monitor that looks at your screen and uses LLaVA to determine whether you’re being productive or slacking off. @charliebholtz @replicate
8
2
80
I made a trailer for a movie about a t-rex who wants to become a chef
10
4
77
The coolest part is the live error handling. It knows that the curl returned an error instantly because it’s looking at the same screen I am!.
3
1
80
I'm working on Melty with my long time friend—and former teammate on @BMoUltimate—@jacksondecampos. He worked at Netflix, where he built the ML infrastructure that picks the artwork you see for your favorite shows. We applied last minute to @ycombinator's S24 batch and got in!
5
0
78
Llama 3 8B is really fast. ~100 tokens/second, 0.40 seconds to first token.
6
6
75
I think I'm most impressed by GPT-4o voice's latency. It's almost real time.
6
1
72
How Jarley Works Part 1. Lots of people are asking how I generate the video for my standup attender. There is no AI involved! It's just a loop of me sitting and nodding my head for 40s. I'm covering my mouth so that you can't tell my voice isn't synced up. I then added this
I finally figured out how to skip daily stand-ups. I made an AI clone that knows what I'm doing all day, listens to audio with whisper, and responds with a language model. Here is a sneak peek of Jarley in a real life meeting:
2
5
70
Moderation on stickerbaker has been a fun challenge. And we have a great solution imo. Every time a new sticker comes in, we run a fine-tuned llama 13b to rate it 1-10. We automatically remove anything the classifier rates <= 5 safety. It turns out it's still easy to get by
7
2
73
If you want to try Melty, sign up here and we'll get you set up. Can't wait to hear what you think!.
20
3
66
we now display model status! no more running a prediction to find out that the model is cold
4
2
66
@bhurlow I'm running whisper.cpp and looking for these words in the output. I'll do a full write-up on it soon!
2
0
63
what do people think of this for the Replicate explore page?
35
2
64
The sticker livestream took ~6 lines of code to implement with @elixirlang PubSub. First, the home page subscribes to the "safe-prediction-firehose" channel:. Every time a new sticker is approved, we broadcast:. When the home page receives the broadcast, stream in the sticker.
We're nearing 15k stickers! It's really fun to watch them stream in live
3
8
65
A lot of people have asked about the prompting tricks I used in my @ElixirConf talk. I'm on a long plane ride, so I decided to publish it — introducing the `ai` Elixir package! . It includes an ~LLM sigil for writing GPT prompts, and a chat() function for sending + parsing the
3
8
60
I made a nature documentary. Or, I helped AI make a nature documentary. Written by GPT4, narrated by ElevenLabs David Attenborough, and filmed by Zeroscope. Episode 1 — The Glow Leopard
5
11
58
This video from @karpathy is so good (thanks @anotherjesse for the rec). It gave me the best grasp of how back propagation actually works. I'd recommend watching it without any notes one time and then watching again + writing code alongside the second time:.
3
4
63
@dwarkesh_sp Have you read @paulg's post on this? It made me feel a lot better about not remembering every fact I read. "Reading and experience train your model of the world. And even if you forget the experience or what you read, its effect on your model of the world persists. Your mind is.
3
4
58
Very excited that I'll be giving a talk at @ElixirConf this Sept. It'll be about building AI apps with Elixir. I got started with Elixir+LiveView when I watched @chris_mccord's "build a twitter clone" in Sept 2020. So it feels pretty crazy to say I'll be speaking at Elixir's.
0
4
55
Last week, Llava was released, which can respond to images . Salmonn was just released, which can respond to audio. Soooo, open source models now have both eyes and ears! . They're both live on
6
6
53
It's pretty amazing that you can create ANY image imaginable locally. No internet, rate limiting, or account necessary. I'm still confused how this is possible?!.
2
1
55
I have now reached a permanent state of nirvana, the hedonic treadmill is over
15
0
54