Stupid Investors, AI Lab Struggles, & Agile Robots

TLDR;

Welcome back to our weekly Premium rundown! This week, we take a look at widely different news, from:

deep dives into xAI and Anthropic’s businesses and struggles,

the different approaches to robotics coming from China and the US,

investors making stupid decisions (again),

and of course, technological improvements, mainly open-source but also including the new efficiency king, Grok 4 Fast.

Enjoy!

OPEN SOURCE

Open-Source Is Killing It

Recently, the world has seen the arrival of several open-source models that are, quite literally, state-of-the-art (SOTA). Yes, you read that right, open-source is closing the gap very effectively in certain areas.

The first example is Moondream’s new version, number three, a small model that specializes in computer vision tasks (understanding images). As you can see in the image above, which is filled with numerous images, the MoonDream v3 model perfectly identifies the object, ‘fire’, that the user requested, even though it isn’t a real depiction of fire but a figure. Absurdly impressive.

Perhaps more impressively, we have dots.ocr, a new model that sets SOTA performance for OCR tasks, outcompeting frontier models too:

The most surprising thing is that we are discussing a model with just three billion parameters, at least two orders of magnitude smaller than the frontier models.

And even more shocking is Alibaba’s Tongyi DeepResearch model, a “small”, 30-billion-parameter model (perfectly runnable in a beefy laptop), with just 3 billion activated parameters (meaning 90% of the model doesn’t activate in every prediction, considerably improving inference speed).

It outcompetes frontier Deep Research models (ChatGPT, Gemini) while being not only free if you run it locally, but also free if you use OpenRouter’s endpoint, meaning that what’s probably the most essential Generative AI use case currently, generating reports using Deep Research features, is now basically free.

Personally, I’ve been using it for a few days in substitution of o4-mini-deep-research that OpenAI serves, and the results are similar, while I was spending around $20 a day (yes, a day) using the other model.

Are times changing?

TheWhiteBox’s takeaway:

The short answer is yes. Open-source models are catching up in meaningful ways, which will allow you to drop your AI costs considerably in the next few months.

I’ve been an advocate of having your own compute for more than a year, and I still believe it to be a necessary, almost mandatory decision. The reasons are threefold:

Energy constraints

Margin pressures

Reinforcement Learning

As for the former, I believe if demand continues to grow as fast as incumbents claim, we are soon going to be all fighting, as users, for the same undersupplied compute (if you don’t trust me, the next news gives a real example of this reality), meaning labs being forced to serve you worse models to cover demand or even rate-limiting you certain hours or at demand peaks.

Of course, one would assume that countries have acknowledged this and are building power as if there were no tomorrow. And the answer is yes for China, no for the rest of…

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