Is the future still ‘one, giant AI’ after Coinbase ‘fine-tunes’ its frontier models?

Frontier AI models are powerful generalists. However, they are expensive, slow, and suboptimal for high-stakes company-specific tasks.

Even so, AI models are slowly expanding and cementing the belief that the future is not one giant AI, but many specifically fine-tuned models. Anthony Pompliano backed this view after remarks from Coinbase CEO Brian Armstrong concerning their newly trained model.

AI companies fine-tuning their own models

Brian Armstrong outlined how they improved Coinbase‘s frontier AI models by training to cater to specific tasks on their platform. The models are better than frontier ones, which are generalists.

In his post on X, Armstrong wrote,

We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at Coinbase, and it outperformed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).

As these models achieve better results than the old ones, Armstrong said they were designed to meet specific tasks such as payment risk review. This new architecture could entirely replace some backend services at Coinbase.

To ascertain this concept, post-trained Qwen3.5-9B eclipsed other models in terms of fraud detection, precision, recall, F1, and dollar-weighted recalls. The model also had reduced latency.

Anthony Pompliano backed it, citing Silvia [Nasdaq: SVIA], an AI CFO built by a leading research lab for personal finance.

Silvia was ranked first in tax, mortgage, and credit card benchmarks, beating the likes of ChatGPT, Claude, Gemini, and Grok.

However, some critics believe that custom-trained models only have shelf lives of a few weeks. They also claimed that while Armstrong was comparing the post-trained model to Opus 4.5, a better 5.5 version was already in the market.

All about Mark Zuckerberg’s prophecy

With it becoming evident that the future is not a giant AI, Mark Zuckerberg is being proven right. Three years later.

Zuckerberg expressed this belief on Lex Fridman’s podcast in June 2023, when he spoke about the future of AI at Meta, Facebook, Instagram, and WhatsApp. At the time, he claimed that a couple of AIs would be developed for specific businesses, brands, and needs.

For instance, creators would develop AI to proxy themselves and let it interact with their fans. The scenario is the same for businesses and humans when they need companionship, mentorship, education tutors, functional relationships, and others.

Final Summary

• Coinbase CEO Brian Armstrong released the latest AI for the company, outlining that they trained the models for specific tasks. 
• The idea of many specific AIs rather than a giant one is winning, proving Mark Zuckerberg’s prophecy from three years ago.