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*! https://minimaxir.com/2025/01/write-better-code/
Data Scientist at BuzzFeed in San Francisco. Creator of AI content generation tools such as aitextgen and gpt-2-simple. I am the data.
Posts
New blog post up: I played with GPT-4o's new AI voice generation API and not only is it possible to prompt voices in a specific style, but it's possible to generate from a specific person like Obama.
That said, *possible* doesn't mean the voices are good. https://minimaxir.com/2024/10/speech-prompt-engineering/
As of today, I'm vowing to no longer correct people outside of the AI industry when they're incorrect/misleading about generative AI and. Every single time they immediately assume I'm an AI tech bro engaging in bad faith whataboutism and it's getting tiring.
New blog post up: I’ve been thinking about the last time Gen AI was universally non-controversial. That was early 2023 with AI Seinfeld.
I wrote thoroughly about what it would take to do a 24/7 AI-generated sitcom again, but there's too much against it. https://minimaxir.com/2024/08/ai-seinfeld/
People on Hacker News and Reddit are explicitly commenting that this post is good and that's never happened before and tbh I'm weirded out. @minimaxir@sigmoid.social
6 years ago, it was cool for humans to pretend to be AI.
New blog post up: Many years ago I made Pokémon vectors, but was not able to account for all Pokémon metadata. Thanks to the power of Nomic AI's embedding models, I am able to plop a blob of raw JSON into the model and the embeddings are extremely good! https://minimaxir.com/2024/06/pokemon-embeddings/
wow OpenAI is releasing GPT-40, they skipped a lot of versions!
not looking at AI music generation because i don't understand normal music tbh
Normally Reddit submission titles are highly exaggerated, it's rare to see one that's completely accurate.
New blog post up: there are many memes about the theory that tipping ChatGPT causes it to generate better text. I decided to put it to the test with some more statistical rigor.
More importantly, it's a great excuse for me to make pretty charts! https://minimaxir.com/2024/02/chatgpt-tips-analysis/
On Glaze/Nightshade: A few months ago I made a POC on how finetuning SDXL on bad images can actually allow it to output "better" images.
There's a nonzero chance that the creation of known tampered data can ironically *improve* generative AI art models. https://minimaxir.com/2023/08/stable-diffusion-xl-wrong/
Given the justified backlash against the AI George Carlin special, I'm tempted to publish my "AI Seinfeld Was The Peak of AI-Generated Content and it Will Never Happen Again" hot take.
Broke: Statistical significance
Woke: Statistical relevance