#44: AI Agents - Hype, Complexity, and What Actually Works

Are AI Agents really a new breakthrough, or just a new hype word for something we already discussed a year ago? What could be potential challenges when adopting an agentic approach?

I recently had a splendid conversation with Demetrios Brinkmann from ML Ops Community about what is up and down in AI and AI agentics, and how that looks from a more traditional AI ops perspective. The episode is a great discussion where we probe each other for more perspectives and I am super happy to share it with the audience today!

Read more about Demetrios at the bottom of this newsletter.

As always, I am very happy that we once again managed to make an episode that explores more than it explains ??

Listen to the episode here:

Apple: The Only Constant on Apple Podcasts

Spotify: The Only Constant | Podcast on Spotify

Spreaker: The Only Constant (spreaker.com)

What made this episode special to me?:

When I decided to do The Only Constant, I had as one of my primary goals - and this may sound weird, but hear me out - to have my guests pause for a moment. To stop momentum and think. Be surprised, be out of standard answers and - explore. I wanted to learn during my conversations, and I wanted more than anything that my guests also learned, and that we together with our listeners would create this weirdly curious, inquisitive, thoughtful point in space time and our ears that would be The Only Constant. And I actually think that it finally happened in this episode. Demetrios thought for so long in silence on my question of "what is the only constant", that I actually removed a couple of seconds of silence twice in postproduction. The question simply did not make sense to him - and when it seemed it did, he again went back to no, it didn't. And let us in on his thoughts.

This pausing to reflect, which occured often throughout the episode, reflects exactly what we need after two years of massive, relentless hype. A hype now infused with ASI and agentics over the last couple of months... which is what we actually pause to consider the ups and downs of in this episode. And who better to have on such a chat that the founder of the machine learning OPS community - this is not about getting something buzzing in the corner for a second. This is about long term value. Predictable outcomes. Stable executions. And perhaps - just perhaps - a bit less license to kill for our AI agents...

My key post-recording deliberations are:

• How can we move beyond the sheer promises of AI agents to start assessing real world tasks performed by AI with consistent performance? And is the hype - so great at kicking off the discussion - starting to stand in the way of actual value generation by having us focused on the wrong deflation messaging (tomorrow it will be cheaper and better)?
• Agent on agent on agent - Demetrios calls out the challenge of avoiding compounding errors from small hallucinations in multiple agents in a chain, and calls it a house of cards. Does this mean we will be working more in AI bubbles instead of chains? Like... this is a kitchen bot, but not a garden bot. Or even narrower - a sauce bot. If that is the future, how do we keep track of all the sudden bots we will be managing, and will managing them even make our lives easier on a whole? We get more stuff, but we are still occupied with boring tasks.
• AI is good a the small coding tasks - but we all want to solve complex needs, right? Put together with the compounding errors, this is one to watch...

Demetrios is a brilliant chat, and I loved having him on the show - and I hope everyone will follow him and his insights and enjoy listening to the episode!

About Demetrios Brinkmann :

Demetrios is one of the main organizers of the MLOps community and currently resides in a small town outside Frankfurt, Germany. He is an avid traveler who taught English as a second language to see the world and learn about new cultures. Demetrios fell into the Machine Learning Operations world, and since, has interviewed the leading names around MLOps, Data Science, and ML.

Since diving into the nitty-gritty of Machine Learning Operations he felt a strong calling to explore the ethical issues surrounding ML. When he is not conducting interviews you can find him making stone stacking with his daughter in the woods or playing the ukulele by the campfire.