If you’re working in AI or trying to get a job in the field, I’m willing to bet you feel behind.

And honestly, I get it. I felt it while working on AI engineering at Twitch, and I see the exact same stress with the people I coach. We’re all putting in tons of effort, and yet every day there’s another model or technique we apparently needed to learn yesterday.

Keeping up with all of it is literally impossible.

So I’m trying a new format called Chill AI News. Once a month, I’ll go through what actually matters for people building with AI, what you can use, and what you can safely ignore.

No hype-bro craziness, and no endless doomer complaining either. A lot of this stuff IS really cool! I want us to enjoy learning it and building with it without letting it make us lose our minds.

So here’s what I’d pay attention to this month:

If your system is routing support tickets, you don’t need a paragraph explaining that the customer seems upset about an unexpected charge. You need it to say “billing.”

That’s the idea behind Jev, a new model built for structured decisions, with predefined answers and confidence scores. But you can also do this with cheaper LLMs. Basically, the idea is to use a cheaper model for the straightforward cases, with uncertain cases handed off to something bigger/more expensive.

An eval is basically a test for your AI system. And if you’re still figuring out whether an idea is useful, you probably don’t need to spend three weeks to build an enormous automated testing setup.

You should build your evals at the appropriate level of complexity for where you are in a project. Just start with around a dozen realistic examples, look at the outputs, and write down what would make each result acceptable. As the project gets more serious, save the failures and turn them into tests.

Andrew Ng’s advice on matching evals to your project’s stage is worth a read.

With releases like GPT-6 Sol and Luna, it’s worth testing whether a cheaper model can now handle work you’ve been sending to something more expensive.

But compare cost per successful task, because a cheap model that needs three attempts and somebody to fix the result might not actually save you anything.

And check your prompt caching. If you keep sending the same instructions, tools, and reference documents, reusing that context can reduce your bill. Make sure it’s actually happening.

The “harness” is the software that manages an agent’s instructions, tools, memory, and execution loop. Two coding agents using the same model can still behave differently and cost different amounts, which is what the HarnessTax study explored.

Before you switch tools, look at what your current agent keeps putting into context. Does it really need the entire test log? All those tool definitions? The same giant file again?

I’d try five recurring tasks in two setups and track whether they finished correctly, what they cost, and how often I had to step in. That’s enough to start noticing where you’re wasting money.

And a few quick hits:

And for everything else, the question I suggest asking yourself is this: What decision does this actually change for me?

A dramatic benchmark headline or an announcement about a model you can’t use yet might be interesting, but if there’s nothing for you to test, replace, or change, it probably doesn’t need to get added to your TODO list.

This first edition is a test. I’m also thinking about trying a longer video version with technical breakdowns, where I’d go deeper into how this stuff works and how you could use it in your own projects. Let me know if you’d find that useful or if you have any feedback on this first round.

A good friend of mine is hosting a virtual hackathon from October 23-25th.

This event is a collaboration between the Data Science for Social Good Berlin and NYC chapters, and is a great opportunity to meet likeminded people, build something cool, and learn!

The event description is below:

​A Datathon is a fast-paced, collaborative virtual hackathon that partners up curious, clever, quantitative data people (that’s you!) with empathetic non-profit people (you’ll love them!) to take a stab at real analytical challenges that our partner non-profit organizations are facing.

During 24 hours you’ll have a chance to step away from your regular problem sets, contribute to a good cause, and meet other socially minded problem solvers. It is a fantastic opportunity to apply your technical skills, collaborate on open-source repositories via GitHub, network with like-minded developers, and contribute directly to an impactful cause.

You can learn more and sign up here.

Forwarded this email? Sign up here.

Follow me on YouTube | Instagram | X | LinkedIn | TikTok

Note: This email may contain affiliate links. If you make a purchase I may make a small commission, at no cost to you. Thank you for your support!