US Gov vs Anthropic, AMD🤝Meta, & More

TLDR;

Welcome back! This week has it all. New models, national security drama, billion-dollar deals, new agents coming for new market sectors that should be trembling in fear, new cool demos that show us the future of coding, and more.

Enjoy!

THEWHITEBOX

The First Diffusion Reasoning Model Drops

Inception, a start-up founded by prominent Stanford alumni and professors and uniquely focused on creating diffusion Large Language Models (dLLMs), has released Mercury 2, its first diffusion reasoning model.

The model offers incredibly fast speeds, allowing it to produce outputs at speeds modern autoregressive LLMs (next-token predictors like ChatGPT) can only dream of, if not for very particular hardware implementations like Codex 5.3 Spark running on Cerebras. The difference is that, here, the model is designed to be this fast.

But what is diffusion? You probably heard the term before, as it was once (and to some extent still is) extremely popular in image and video generation models.

The idea behind diffusion is beautifully exemplified in the thumbnail above, but what’s really going on behind is a process called denoising, sometimes referred to as coarse-to-fine inference.

In it, instead of predicting the next word based on the previous one step at a time, diffusion LLMs sample a certain amount of response tokens, and progressively refine them in unison across several denoising steps (we call these denoising steps because we start from apparent gibberish and progressively “unearth” the real word).

Since we can process many tokens in parallel per round, the number of generated tokens per second is much higher.

TheWhiteBox’s takeaway:

With AI, everyone has to accept that things almost always look better than they really are.

While dLLMs aren’t something new, and they do have a place in inference due to their obvious speed, we have the issue of performance, as these models are, in terms of intelligence, only marginally better than models a year old or so, and meaningfully worse than frontier models today.

To me, thinking these models put at risk modifying the underlying architecture after all we’ve achieved with autoregressive LLMs is just shortsighted. There’s too much money in that direction to reconsider at this point.

And besides speed, they don’t seem to offer a huge advantage at the hardware level (yes, they prevent the need for a cache, but they also increase weight loading requirements for long sequences), and that’s the real deal here:

If you want to really change the status quo, you need to find something that profoundly reshapes the hardware ecosystem. And I just don’t think this is it.

In a bullish AI scenario, the sheer demand for AI products will ensure dLLMs have a place, but they don’t seem like the new architecture paradigm, at least in my limited view.

MODELS

Alibaba Releases Qwen 3.5 Medium Model Series

Alibaba’s Qwen lab has released its new medium-sized model series, Qwen 3.5. These models come in different sizes, all intended for use, ideally, in local settings.

Quite possibly the most notorious fact from this release…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.