INFRASTRUCTURE FRONTIERS

Hello, and welcome to Infrastructure Frontiers, where Adam Jacob, Nick Stinemates, Paul Stack run down all the AI news of the week for Infrastructure Engineers. Due to an audio issue, we're not going to get the podcast out this week, but we'll be back on schedule next week. Here's what we thought was interesting:

• Databricks wrote about managing their AI costs at scale, seeing 50% savings from lower cost models, 30% from smart routing, 10% from spending controls, and 10% from context optimization. What they didn't talk about was simple efficiency - doing more deterministically (i.e., using models to write code) rather than just throwing more and more at the problem. We think the multi-model world is definitely coming to all of us.
• Simon Willison breaks down the timeline of the OpenAI attack on Hugging Face. It's the best summary of the most interesting attack so far in the AI era. You can watch the great talk OpenAI gave at Black Hat as well. We think it's a new era for offensive capabilities in these models, and we're looking forward to what the new defensive technologies will be. Nick makes the point that this is largely not that novel, because the techniques are all things a human would do. This one is a must-read.
• Meta releases Muse Glimmer, following the trend of new open weight models with frontier intelligence. More evidence of the multi-model mixture of local and remote intelligence.
• Docker launches a sandboxing plugin, which we think makes a ton of sense. Figuring out where and how you are going to run your AI workloads remains a big part of the puzzle right now, and it makes perfect sense to us that Docker would throw their hat in the ring.
• Anthropic announces they are going to start watermarking text generated by their models. This one caused social media to really freak out. They're doing a variant of what Google already does with SynthID - essentially making the random number generator a little less random, and then tracking that signal across generated text. We largely think it's a good thing you can tell if something has been written or manipulated by AI.
• Stratechery wrote a great article about Nvidia's new vehicle for funding AI infrastructure buildouts. With everyone but Microsoft tapping debt (and Google tapping equity), Nvidia creates a novel pool of infrastructure underwriters from some huge names. The underlying question here is: will we figure out how to make all this useful before the debt comes due? Like everyone else, we're not sure - but we think the advances we see with folks in the Swamp community give us hope that we can.
• Birgitta Böckeler finds that TDD adds no value inside the agent loop. In the end it's about having good regression tests, regular refactoring, and building the machine that builds the machine to gain confidence. We think it's great that Birgitta published this work, and that it was on the blog of Martin Fowler (one of the fathers of TDD). It also matches our own experience - TDD added value sometimes, but not anymore.
• Kristen Lowe wrote a piece for Lenny Rachitsky's newsletter about founder comms in the age of AI. In an age where the hype cycle is rampant, the best way to get folks to care about what you do is to share authentically why you are doing it. We think we need more of this, and fewer AI generated ads.
• Adam gave a talk about building software factories at AI Agents Seattle. It's the first time Adam has really tried to explain how to work this way, and it was lovely to share it with a bunch of eager, smart folks. We think more and more success with AI software engineering will be tied to factory creation - now is the time to talk to each other and experiment, to learn what works and what doesn't.

In addition to getting Infrastructure Frontiers in your inbox every Monday, you can listen to it wherever you get your podcasts:

• YouTube
• Spotify
• Apple Podcasts
• iHeartRadio
• Amazon Music

Looking for a community of like-minded folks to help you on your journey? The Swamp Discord is the place for you. Things are moving fast, and the best way to understand what's working for people and what isn't is to share your experiences with each other. Can't wait to meet you there.

If you're trying to figure out how to manage production infrastructure and applications or to scale your use of AI within your organization, Swamp is here to help. Tell your favorite AI agent to use Swamp to deliver a deterministic workflow for a problem you have, and watch it build you repeatable automation you can use right away. It'll save you money and time, and make it possible to scale AI usage safely and effectively across your organization.

Read something interesting this week and want us to cover it? Tag @adamhjk on X, @adamhjk.me on Bluesky, or adamjacob on LinkedIn.

See you next Monday!

Adam, Nick, and Paul