Read the research papers people circulate and write down questions you have as you go. It is shocking how many people don’t read the papers.

After last week, have a couple more additions to this list. But before I do… Do you all have any suggestions for additional tips? www.dbreunig.com/2026/05/04/1...

My number one tip for teaching younger kids AI is to let them play with local models. Let them change the system prompt, temperature, toggle thinking, and model. Suddenly it's not an abstract intelligence but just software that has a natural language interface.

Understanding the tensions building in the AI ecosystem with Stewart Brand's Pace Layers: www.dbreunig.com/2026/07/03/a...

This episode of East Bay Insiders is INSANE. And it highlights the importance of supporting local beat reporters. Subscribe today! Andy Duong thinks it’s worth +$20k a month! substack.com/@eastbayinsi...

CA residents: Just a few clicks to let your rep know you don't want your 3D printer turned into a camera into your home. www.eff.org/deeplinks/20...

Natural language is great for one-off tasks, but suboptimal for defining system behavior. Prompt debt creeps in with each small fix, until the prompt is brittle, you're afraid to touch it for fear of regression, and you're locked into one model. www.dbreunig.com/2026/06/22/t...

Everything about this story is just too heavy handed of a metaphor, right down to the name, "The Reflecting Pool"

This validates something I've long suspected: less adept users of Claude Code will hit their limits MUCH faster than skilled engineers. www.anthropic.com/research/cla...

The 3 claims here are both practical and imperative: 1 “Regulation will fail if expertise remains concentrated inside labs” 2 “Open science for AI needs reimagined institutions” 3 “The missing institutions should sit between secrecy and unrestricted release” pgasawa.bearblog.dev/unsafe-ai-or...

Everyone building programs with AI needs to now: 1. Be able to swap out models, in short order. 2. Be able to verify, regularly, that your output isn’t silently changing. (You should have been doing this for any production system, with evals, but now there’s no excuse.)

The imperfect and awkward ways Anthropic is using to control how their models are used (with Fable now, OpenClaw a bit ago) is a great example of the imprecision of natural language as an interface. The best model can’t differentiate a bio threat from an innocuous health or research question.

Apple announcements I’m most excited for, in order: 1. Screen time improvements 2. Liquid Glass slider

My recent obsession is diversity in LLM outputs, specifically from the frontier models. They are being post-trained so hard for code and agentic use (and to be _reliable_ at these things) that its shaving off the novelty, impacting other tasks. Have you experienced this? If so, how so?

When an agent can explain anything, what is the role of human-centric docs? www.dbreunig.com/2026/05/31/w...

DSPy requires more upfront learning than just writing natural language prompts. But once you get it, it makes building, maintaining, & improving AI programs so much easier. We aim to soften the learning curve to make the benefits more accessible, starting with more new docs and front page. dspy.ai

Given the state and trajectory of agents today: if someone wanted to start learning how to be a software engineer today, what book would you recommend to them?

Wired's kicker on their story about the UFO releases is the thing to keep in mind. This is a good reason why they wouldn't hide things; and a better reason why they'd stir shit up: www.wired.com/story/pentag...

Can I make a polymarket bet that he eventually polygraphs the polygraph guy?

🤔

10 lessons for agentic coding. www.dbreunig.com/2026/05/04/1...

Coming back to drafting a set of "10 commandments" for coding with agents. Here are the current candidates: - Implement to learn - Rebuild often - E2E tests are gold - Document intent - Maintain your spec - Find what's hard (that's the value) Thoughts? What am I missing?

The fact that residential restrictions are being presented as way to address the crisis is insane. www.wired.com/story/the-su...

Frustration at GitHub and Claude outages and inconsistencies is a good argument against the idea that we’ll be able to vibe code everything we need.

Claude is down, time to check in on the state of local models.

Are they getting royalties? www.theverge.com/tech/916681/...

Release day for DSPy. Big one with lots of RLM improvements, optimization chaining, type checking, LiteLLM decoupling, and 24 first time contributors. github.com/stanfordnlp/...

Tim Cook ends his tenure with the most Tim Cook product possible: a low-price, high margin laptop that repurposes 2nd tier binned silicon from iPhone runs. I mean this in the best way. The Neo is a great, accessible product. That leverages the rest of the supply chain.

With a few exceptions, successful agents become workflows.

Perfect take.

Anthropic has clearly post-trained Opus 4.7 to say “Good push,” instead of, “You’re absolutely right.” (Still a 3 tokens gain with the new tokenizer, for those counting.)

My not-so-monthly newsletter just went out, covering the gaps between the potential of current models, and their real-world impact. Check it out: buttondown.com/dbreunig/arc...

Allbirds getting into AI is essentially a SPAC. Ignore it.

If the Mythos claims hold, to harden a system you need to spend more tokens discovering exploits than attackers will spend exploiting them. www.dbreunig.com/2026/04/14/c...

A few weeks ago, it was common to hear people argue one should use agents to replace dependencies for security reasons. In light of the Mythos news, the math changes. Using an OSS lib that’s had tens of thousands of $ of agentic hardening is likely optimal.

In my head, I've started referring to these as "rosary games." You play them like one might say the rosary as you work your way down the beads. They're games you play to engage you attention just a bit, allowing you to mediate, wind down, or focus. They're not challenging. They're not urgent.

I keep hearing from non-tech people they loved this movie because it took the time to explain plainly how LLMs work. It’s been 3+ years since ChatGPT and this is probably the first good, mass market attempt. Absolute dereliction of responsibilities by leading labs

The Cathedral, the Bazaar, and the Winchester Mystery House: an attempt to put our era of sprawling, idiosyncratic tooling in context. www.dbreunig.com/2026/03/26/w...

Bryce 1.0 (1994)

At the DSPy meetup a couple weeks ago, Kshetrajna Raghavan walked through how DSPy helps with efficiency, modularity, and performance at Shopify scale: www.youtube.com/watch?v=bxTo...

I don't think I've ever felt more like the meme of the sweaty dude agonizing over what button to press.

I used Google Stitch, the AI frontend design tool whose announcement had people panic selling Figma. I am beginning to think people should try these tools before they react...

2016: Companies argue they’re a software company to get higher valuations. 2026: Companies argue they’re not a software company to get higher valuations. www.bloomberg.com/news/article...

If you’re buying the first campaigns on a platform, you shouldn’t be expecting data saying the ads “worked”. The point is being first, learning, getting the awards, and being seen as a leader.

Based on my friends in the military, it makes total sense that the two most regular leakers of military info are Warthunder forums and Strava. www.lemonde.fr/en/internati...

“This is, I’m sorry to say, altogether goofy, but nonetheless beautifully expresses a desperate romanticism that seems endemic to the United States: a desire for roots in some place other than this, our graceless national project.”