Kristian Muñiz

28 Mar 2025

At this point I'm convinced that 4o image generation is not purely autoregressive. My guess is 4o generates image tokens or latent representations in sequential patches which are processed by a tighly integrated diffusion model.

28 Mar 2025

Ah, hint from Greg Brockman himself. Seems like the "powerful decoder" here is a diffusion model.

Increasingly, large multimodal models are becoming more and more powerful and one of the first ways we can optimize them is by simplifying their I/O and writing powerful, thick encoders/decoders.

Should an omni-model that is purely autoregressive be able to pass through an image in a semi-lossless way? I understand that it depends, to some extent, on post-training and the non-stochastic nature of sampling, but I'm having trouble with consistency using 4o's image generation feature.

*of sampling the next token. Had to cut some characters.

And it's not structural or semantic consistency, but some information gets lost in the process. Perhaps it's safety mechanisms preventing certain behaviors like using people's likeness.

I want to understand the training and inference economics of autoregressive image generation. There's obviously latency implications but in my opinion, at least anecdotally, it makes for up for it in output quality.

What I know so far, autoregressive models are more expensive to run than diffusion models – of course slower too, latency correlated with cost. I'm still surprised that resolution is so good. It's almost too good. Could it be a hybrid Transformer + Diffusion approach?

Could that be a plausible solution? Using GPT-4o to generate initial image representations and passing these representations to a diffusion model component that specializes in creating high-quality, high-resolution visual outputs?

14 Mar 2025

The Response API is what this LLM APIs should've been from the beginning.

27 Jan 2025

DeepSeek R1 thinks too much

21 Dec 2024

ffmpeg is so much better with an LLM on top; I feel like I can do anything

Line go up...

It's been one week and I'm already seeing anti-vaxxers and conspiracy theories in Bsky. This is an uphill battle huh?

24 Nov 2024

Optimist: The cup is half full Pessimist: The cup is half empty Software Engineer: That's a flaky test

Jordan DeVries@jordandevries.com

Optimist: The cup is half full Pessimist: The cup is half empty UI: People are having a hard time understanding which part of the cup icon is full and which part is empty

18 Nov 2024

A big mistake LLM products make is not explaining to their customers how they work behind the scenes. Then they complain that people don't understand them.

1. Log into the bird app 2. Search for "bsky.app" 3. Toggle "People you follow" 4. Profit 🦋

Engineer in the streets, designer in the skeets

Printloop + @tldraw.bsky.social's API. Here's a video showing Printloop's TypeScript features in an infinite canvas. Type hints, autocomplete, npm support, immediate build. All happens in-browser.

cc @steveruizok.bsky.social canvas demos are way more challenging in 🦋 so much to show in 1m 🤣

Printloop + @tldraw.bsky.social's API. Here's a video showing Printloop's TypeScript features in an infinite canvas. Type hints, autocomplete, npm support, immediate build. All happens in-browser.

Here it is for the Python functionality. Much more limited and kinda janky, but pretty cool. Matplotlib, numpy and other packages available. All happens in-browser as well.

16 Nov 2024

I'm obsessed with Silo. Such a good show.

16 Nov 2024

Boricuas en Bluesky 👀

16 Nov 2024

Cereal is breakfast nachos.

Kristian Muñiz reposted

Peter Swimm

@peterswimm.com

12 Nov 2024

Hey folks - I’m on the job market! Need product help with conversation ai and understanding the power of humans in the loop? Let me know!

washingtonpost.comThe rise of AI fake news is creating a ‘misinformation superspreader’Artificial intelligence is automating the creation of fake news, spurring an explosion of websites that can disseminate false information about wars and elections

30 Dec 2023

This has been LLM Concern # 1

The Washington Post@washingtonpost.com

Artificial intelligence is automating the creation of fake news, spurring an explosion of web content mimicking factual articles that instead disseminate false information about elections, wars and natural disasters.

10 Nov 2023

Oh wow, Bluesky app on the web is so snappy

26 Oct 2023

Ugh, I'm really interested in LLM as a topic but having a hard time finding where the good conversations around LLMs are happening in Bluesky vs X

krismuniz.comBringing Language Models Closer to the Editor Through Modal EditingA blog post showcasing QQ on printloop.dev, my proposal for a new editing modality that allows the user to make quick consults with a language model within the editor.

6 Oct 2023

You can see headlines in Bluesky, what an innovation krismuniz.com/posts/qq-in-...

2 Oct 2023

I've always found it very hard to see today's technology as permanent. Every digital medium has gone through so much change that it makes me think that it's all ephemeral and flimsy.

13 Sept 2023

I need more Bluesky friends, ya'll I promise I'm fun in specific circumstances!

15 July 2023

Maybe social media should just be people maintaining their own forums

6 Sept 2023

Revisiting this idea: no.

I'm back on my bs

Bluesky, that is

To you, what's the best cure for a bad day? What do you do to turn your day around?

Serene

Gonna share these news exclusively here: I moved to Massachusetts!