Max Woolf
@minimaxir
Followers
19K
Following
18K
Media
4K
Statuses
29K
Senior Data Scientist at @BuzzFeed in San Francisco // AI content generation R&D
Joined May 2010
New year, new blog post: I had a random question, what happens when LLMs are prompted to write better code, again and again? Do they actually write better code? The answer is yes*!
31
84
672
I forced a bot to look at every Pokémon and told it to generate its own. Here are the results. (this isn't a joke, that's actually how I made these)
1K
10K
44K
The Reddit strike should go on forever just because of this dialogue chain.
257
1K
14K
@ZoeSchiffer This is likely to prevent employees from sabotage but this just seems like it'll be more likely for them to do sabotage.
122
283
9K
Is it a Duck or a Rabbit? For Google Cloud Vision, it depends how the image is rotated.
42
2K
4K
Everyone joking about the price of the Mac Pro is missing one very important thing:. It comes with a *free* Lightning cable!
100
340
3K
Wow, you all really, really like these AI-Generated Pokémon!. As a thanks for all your support, how about ANOTHER BONUS BATCH?! 😁
60
472
4K
Broke: prompt engineer Apple Intelligence to reveal its system prompt. Woke: just search for a .txt file containing the prompts lol
43
255
3K
At this rate of growth, the human feces can raise a Series A.
77
439
3K
CAPTCHAs in 2021.
Question: Do people who own horses typically walk around with a photo of themselves and their horse?
20
678
2K
I gave up on data visualization projects after I made this because there is no way I'd ever be able to top it.
Wow, this is cool: The duck-rabbit illusion apparently works on Google Cloud Vision. The system interprets the image one way or the other, depending on its orientation. Animation by the excellent Max Woolf (@minimaxir).
9
223
3K
I trained an (actual) AI on the titles of BuzzFeed YouTube videos and it generated some *interesting* results. 1/3
55
543
2K
It's funny that this is now the Hello World to identify overly-RLHFed LLMs.
36
176
3K
@SwiftOnSecurity why even have a "DO NOT PRESS THIS BUTTON" sign. that's just tempting fate.
12
1
2K
I hate to address the elephant in the room but…. Is building on the ChatGPT API now a liability?.
299
114
2K
By popular demand: here's more AI-generated Pokémon, but this time the model is trained on *only* Gen 1 Pokémon (the 1995-era Sugimori art of the original 151). Those who thought this would make the generated Pokémon less bizarre were very wrong.
77
390
2K
OpenAI released a Jupyter Notebook which demos a Q&A workflow using ChatGPT API as a base (similar to the Agent/LangChain workflow)
13
264
2K
Here's my new experiment with AI-generated Pokémon: I finetuned an AI on *only one image* of Pikachu and had it generate new images of Pikachu. Merry Christmas!
70
371
1K
Not sure what is a bigger statistical crime: the fact that one assumes that people associating with Elon's Twitter account is a representative sample of Twitter users, or the idea that you need millions of users for statistical significance.
4
14
1K
By popular demand: here's more AI-generated Pokémon, but this time the model is trained on *only* Gen 1 Pokémon (the 1995-era Sugimori art of the original 151). Those who thought this would make the generated Pokémon less bizarre were very wrong.
26
105
1K
The only ChatGPT-generated text detector that works 100% of the time.
42
24
1K
OpenAI just released new branding guidelines for GPT-based apps, which will kill 99% of the AI hustler apps.
34
93
961
Not sure what is a bigger statistical crime: the fact that one assumes that people associating with Elon's Twitter account is a representative sample of Twitter users, or the idea that you need millions of users for statistical significance.
9
61
854
Given the timeline, I’m still baffled Sam Altman tweeted “her.”. That’s just makes plausible deniability go away for a random shitpost.
38
32
713
Looks like YouTube fixed it; the search query now has the Top News header, but the video doesn't appear after that section either.
7
43
655
The AI used here is a fine-tuned ruDALL-E on the official Pokemon images (i.e. it is *not* VQGAN + CLIP or Wombo Dream). Unfortunately it's not very portable/easy to use otherwise I'd set up an interactive Colab.
16
37
717
I just released my new Python package: simpleaichat, an open-source tool for working with ChatGPT/GPT-4 with minimal code yet max flexibility!. I built simpleaichat out of sheer frustration with LangChain and aim to make it the easiest way to make AI apps.
19
109
740
New blog post up: so, you've probably seen all the tweets about GPT-3. GPT-3 is objectively a step forward in the field of AI text-generation, but the current hype on VC Twitter misrepresents the model's current capabilities. GPT-3 isn't magic.
13
234
723
a Pikachu attempts to code binary search on a whiteboard at a job interview at Google (DALL-E 2)
17
72
682
A professor failed their entire class because they put their students' essays into ChatGPT and it replied "yes I could write that"
45
57
648
I can confirm that with system prompt engineering and a high temperature, OpenAI's new gpt-4o-audio-preview model can be instructed to generate voices and any vocal style.
🔊 The Chat Completions API supports audio now. Pass text or audio inputs, then receive responses in text, audio, or both.
18
72
644
The ChatGPT API has a new "system" parameter which lets you specify the hidden rules for the model.
25
72
627
The "personality" that Elon seems to hint is the key differentiator can be trivially replicated with a ChatGPT system prompt like "You are a world-famous irritably sarcastic comedian. Never give a straight answer to the user. Always attempt to be funny even though you aren't.".
Announcing Grok!. Grok is an AI modeled after the Hitchhiker’s Guide to the Galaxy, so intended to answer almost anything and, far harder, even suggest what questions to ask!. Grok is designed to answer questions with a bit of wit and has a rebellious streak, so please don’t use.
57
39
585
The new Llama 2 model has an additional licensing requirement which prevents use if you have >700M MAUs, which is an oddly high number. OpenAI's ChatGPT hit 100 million MAUs in January, and has gone down since. It's essentially a "Amazon and Google don't use this k thx."
23
54
582
Here's another experiment with animesque AI-generated images, this time with an AI on trained Genshin Impact playable characters!
15
140
587
That's one hell of a counterargument.
18
34
569
Twitter/X is throttling link clicks to the @nytimes website by causing a 5 second delay during the redirect.
35
157
474
I am legit annoyed at how productive the Hugging Face dev team is
6
39
539
Chaotic Good web developers.
The URL where Sears was selling Ashli Babbitt shirts now reads, "yikes"
6
38
435
@MarioBrothBlog sounds like a feature that was removed from the game but actually removing it causes issues so they just made it impractical to trigger.
6
3
495
A few days ago, an AI image generated using Stable Diffusion + ControlNet was posted on Twitter and resulted with the weirdest reactions, mostly "I thought this was cool at fist but now that I know it was made by AI it's no longer cool.". I'm not sure how to process that.
85
31
513
Have you wanted to create your own ChatGPT chatbot using the ChatGPT API?. I'm releasing my Colab Notebook for a ChatGPT chatbot of the lovely GLaDOS, with fun features! The system prompt is easily hackable to create your own chatbots for other characters!
17
95
492
@ashleyfeinberg officially the antidril
who the fuck is scraeming "LOG OFF" at my house. show yourself, coward. i will never log off.
0
6
435
If DALL-E 2 didn't have limitations on human faces, we'd have Osama bin Laden getting slimed with hyper-realistic slime.
2
23
436
I'm taking a break from AI image generation. The toxicity of the discourse around the space has unexpectedly evolved faster than the actual modeling innovation itself and I don't have the time or energy to deal with it.
67
20
434
Surprise! I'm releasing ANOTHER ChatGPT API Colab Notebook!. This time, it's zero-shot sentiment analysis, where you can specify which emotions ChatGPT can classify! ChatGPT can differentiate between happy/very happy, and sad/very sad!.
14
87
465
Jesus Christ as a Pokemon, via Pokemon-finetuned Stable Diffusion
11
41
436
this is literally the one GPT-3 use case that OpenAI explicitly tells you not to do.
wow guys, i was skeptical at first but it really seems like AI is the future
8
38
416
@eatinginmycar Reddit's take is that this is reverse psychology to make people think he maybe looks like Adam Levine
1
9
404
New blog post up, and it's the hottest take I've ever made. I've been using the ChatGPT API for the past few days, and it's the real deal to the point that I can't justify using anything else for AI text generation.
24
64
427
CEO Pikachu attends a corporate retreat. Stock photo. (DALL-E 2)
2
48
389
Stable Diffusion 2.0 appears to have more legible text rendering.
9
38
383
ChatGPT's infamous "As an AI language model I can't do that" spiel appears to be natively intergrated within the model itself and not a special filter. But since the ChatGPT API allows access to the AI's rules.
10
37
389
Extremely hot LLM take: you will often get better results with few-shot prompting (with good examples) on a modern LLM than with a finetuned LLM. Finetuning was the best option for weaker LLMs with lower context windows: both problems have been solved nowadays.
37
35
376
@MarioBrothBlog I told my car insurance agent that I hit an angle-dependent invisible wall and they didn't believe me 😡.
0
12
350
@jesawyer Pepe LePew: Removed for sexual harrassment jokes that did not age well. The Droogs:.
2
1
361
So here's my postmortem after hunting for a data science job.
16
106
361
My weekend project: automatically block people in images (like in Black Mirror) using a pretrained neural network.
12
109
354