Browser Agents, Token Stagnation, & more

TLDR;

Welcome back! This week in our Premium news rundown, we take a look at the new hot product in Silicon Valley: browser agents. We also discuss Cohere’s great reasoning model, Google’s SOTA image generation model, and the growing concern that AI price reductions are stagnating, among other stuff.

Enjoy!

REASONING MODELS

Cohere Surprises

For once, Cohere has actually delivered a model I would see myself using. Command A Reasoning is their first ‘reasoning model’ (models that generate chains of thought during a response to increase chances of success by imitating human step-by-step problem-solving), and the model is surprisingly good (on paper).

It offers a very high performance-per-cost ratio, as the model is small (enough to fit inside a single NVIDIA H100 in a quantized FP8 format), even runnable in consumer-end hardware with large RAM allocations, despite offering just-below frontier performance.

Crucially, the model has been designed for agentic tasks, offering great tool-calling performance (the ability to call tools effectively, crucial for agents) and great deep research capabilities (they claim it to be the best deep research model on the planet).

→ READ FULL ANNOUNCEMENT

TheWhiteBox’s takeaway:

Canada’s champion. And for once, I am happy with their progress, as I now finally have a reason to at least try the model, especially for deep research tasks.

Let’s not forget that Cohere is solely focused on enterprise AI, which is why they optimize for performance-to-cost metrics. The fact that you can deploy the model in a single NVIDIA H100 or A100 GPU is a blessing; it’s a fast and very cheap model.

It’s also a hybrid model with user-controlled token budgets, allowing you to define the model's thinking time on a task based on its complexity, without needing to switch between models.

With that said, the license isn’t very permissive, and you can’t commercialize the model even if you can download it for free, which is a more restrictive open-source approach than those of Chinese counterparts, who are ultimately strong competitors.

This company is still trying to establish its position in this market. And this model is a step in the right direction.

RESEARCH

All That Matters is Data

An article by a Meta researcher in Substack echoes one of the main messages for the year: the acceptance that most breakthroughs in AI are actually data breakthroughs, not algorithmic breakthroughs.

I went down this rabbit hole in a Medium article (free link), but having it confirmed by an actual researcher in a top-tier lab is a great way to validate my beliefs.

The blog perfectly describes how most technical progress we’ve had over the past few years has been focused on the “systems side” (hardware, improving things like arithmetic intensity we have talked about so much in the past), while algorithmic breakthoughs, maybe except the Transformer architecture from 2017, are mostly decades old, some of them going to the early 50s and late 40s (Alan Turing, Claude Shannon et al).

As I argued in my own article, progress is primarily driven…

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