OpenAI, Google, & DeepSeek Walk into a Bar...

TLDR;

This week is packed with news, as AI labs are accelerating the release of new models, including OpenAI, Google, and DeepSeek. There is also a new AGI benchmark, a study of AI and loneliness, and more.

Enjoy!

FRONTIER AI

DeepSeek v3 0324, What a Model

DeepSeek has launched a new version of its v3 model (its non-reasoning model). There were no flashy announcements or preposterous benchmarks that impressed no one (besides a humble X tweet); they simply dropped it and let the community do the talking.

According to first impressions, this could very well be the best non-reasoning model across many tasks (probably except for GPT-4.5 and Gemini 2.5 Pro (introduced below)). Artificial Analysis also reaches the same conclusion, defining it as the first time an open-source model is at the top of non-reasoning benchmarks.

Notably, it introduces two significant changes:

MIT license, meaning that anyone can download the weights and do whatever the hell they want with them, even develop commercial applications.

It’s a 4-bit model, meaning the model’s weights occupy 4 bits per parameter. This drops the memory requirements from 671 Gigabytes to 335, making it available to run at scale in a desktop Mac Studio M3 Ultra (500 GB of DRAM memory).

And even if you don’t have a Mac Studio, this model runs comfortably on a single NVIDIA H100 node for less than $20/hour.

That may sound expensive, but you have 300+ GBs for your KV Cache, so you can serve a very decent amount of users with just one node. A human customer support agent hourly rate is well above that, and with this server you can have dozens or even hundreds of parallel customer support agents running on a single DGX server.

I’ll put it another way: this model, which appears to be a considerable improvement over its previous version in areas like coding (it is the best non-reasoning model in the Aider benchmark), archives amazing results that can be run at home and at 20 tokens/second, a speed that is more than enough for most use cases.

For those with consumer hardware to run this, the reasons for running proprietary models are decreasing exponentially, and, as VentureBeat puts it, this is a nightmare for these companies.

TheWhiteBox’s takeaway:

I would bet a lot of money that Sam Altman and Dario Amodei (and all of their investors, of course) are having daily nightmares with the word ‘DeepSeek.’

While they are running billion-dollar budgets at a loss, DeepSeek is releasing models that are 80% or even 90% as good (or, for non-reasoning models, even better) for free. I’m sorry, but there isn’t a business case in the world that is sustainable that way.

If we compare apples to apples between both APIs, DeepSeek v3’s new version, which would compete with OpenAI’s GPT-4.5, is 277 and 136 times cheaper for inputs and outputs, respectively.

Wild.

The only way these companies will make money is if the DeepSeek app is banned in the West, which appears…

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