Agentic Workflows, Finally Making AI Useful

🏝 TheTechOasis 🏝

part of the:

Breaking down the most advanced AI systems in the world to prepare you for the future.

5-minute weekly reads.

TLDR:

Leaders: Agentic Workflows, Building GPT-5 Without Building GPT-5

Research you Should be Paying Attention To: Mixture-of-Depths

Matryoshka Representations

Meta’s Theory of Cognition

Did you know about…: OpenAI’s Inpaiting feature and how it works?

🥇 This week on Leaders… 🥇

One single chart was enough to open my eyes to how powerful agentic workflows are going to be.

Source: Andrew Ng

This chart, elaborated by Andrew Ng and his team, one of the most prominent names in the AI industry, displays an amazing reality that I had never seen so clearly portrayed:

There’s so much untapped potential in our current—and previous—eras of Large Language Models (LLMs).

In fact, when wrapping GPT-3.5 in an agentic workflow, independently of what type used, it beats its big brother, GPT-4, by a landslide.

And more importantly, still when wrapping both models in agentic workflows, the difference in results is almost nonexistent, meaning that model size doesn’t really matter that much, while workflows can be a true unlocker.

This leads us to an intriguing conclusion: Are we evaluating our models correctly? Or have we been missing the point all this time, and our current LLMs have yet to display so much untapped power?

But first and foremost, what are agentic workflows, and why are they so powerful?

Understanding Our Current Bottlenecks

In the current age of AI, when referring to LLMs, the time variable is very unbalanced: We take a huge amount of time training these models, sometimes for months, but we expect them to answer almost instantly to the user’s requests.

But this is how we should evaluate humanity’s best effort at creating intelligent machines?

Well, research says otherwise.

Giving Time to Think

In our current standard approach to using LLMs, we expect our models to always act in ‘zero-shot’ mode, meaning that we don’t allow them to review their work and they have to get things right on the first try.

This is akin to asking someone to compose an essay from start to finish, typing straight through with no backspacing allowed, and expecting a high-quality result always, as Andrew himself described.

But if you look at this from the perspective of how they generate their responses, we soon realize it’s a highly unfair way of measuring their output quality.

This is best understood with the image below:

The chosen word will probably be ‘Playground’, but what if the previous context mentioned how the kid got hurt?

In that case, maybe ‘Hospital’, the least probable current response, would have been the best option.

But not only do we tend to be terrible at providing context to our LLMs, we don’t even give them a chance to reconsider their predictions. We simply expect them to work, period.

Therefore, could we be missing out on a huge amount of value creation from LLMs?

One of the research papers I’ve seen…

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