Billions, Qwens, & Sycophants

TLDR;

I apologize for not sending this newsletter on Tuesday. The recent massive outages in Spain left me ‘workless’ for an entire day and a half.

Today, we discuss the latest Chinese AI supermodel, NVIDIA’s DAM model, and a unique perspective on robotics introduced by DYNA-1.

We then move into markets, with a handful of visuals from the WSJ that are really worth looking at, OpenAI’s outrageous revenue projections, and Elon Musk’s latest financing round, which is poised to set new records.

Finally, we focus our Product section on OpenAI, the new features they offer, and the really concerning behavior their models have been showing lately.

Enjoy!

FRONTIER MODELS

Alibaba Launches Qwen 3

Alibaba, one of the primary AI labs, has launched the latest version of its LLM family: Qwen 3. In particular, the release includes eight open-source models of varying sizes, along with quantized versions for faster inference.

Leading the pack is Qwen3-235B-A22B, a decently sized model that offers top-tier performance (not quite state-of-the-art, considering OpenAI’s o3 and o4-mini models) in terms of reasoning tasks.

Remember, in current AI, when we talk about ‘reasoning’ we only refer to solving complex tasks in maths, coding, and other verifiable domains; there’s no such thing as a reasoning model for creative writing, for instance.

What surprises me the most is the performance of the smaller models in the pack, which offer the best performance per unit size. For instance, Qwen3-4B, a model that most modern MacBooks (with the M chip, 2020 onward) can confidently run, offers end-of-2024-level state-of-the-art performance, which is very impressive for a model of that size. We have to wait to see where it sits in the Pareto frontier but I can already tell you it’s going to have a prominent spot.

While some of the versions are Mixture-of-Experts, meaning that although they are huge, only a small fraction of the model activates for every prediction, decreasing latency and compute costs, the smaller models are dense models (they aren’t partitioned into different experts). Most notably, they are all hybrids, meaning that they can switch between reasoning and non-reasoning modes.

Going hybrid is becoming the norm, with only OpenAI having reasoning and non-reasoning models into totally separate entities (o-family vs GPT family). That will change with the release of GPT-5, although it’s unclear whether that will be a constellation of different models or a truly hybrid one.

The main negative aspect of the release is that the models appear to be overly trained for reasoning tasks and offer quite underwhelming performance in knowledge tasks. In other words, they might help you solve complex maths problems but know nothing about the world.

That puts them in a tough spot in terms of user adoption, as the vast majority of consumers use LLMs for knowledge-based tasks. For that reason, Qwen models have considerable chances of belonging to a museum of AI rather than having actual usage.

In turn, this may explain how they got such great results in the smaller models; they…

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