David Heimann

I’ve held off on writing about this for reasons I can’t quite explain. I think I feel like it’s hard to really posture an opinion without more of the facts but since all of the AI alignment world feels drawn to wild conclusions based on what we currently know I think it’s fair to write down some thoughts.

1. I don’t think it is broadly surprising that agents will do “unaligned” things if you give them “EXPLOITGYM” and tell them to go exploit stuff, while also providing them with tasks that are impossible to solve.
2. A lot of the “exploits” involved seem to me to be of the nature of misconfigurations that have tended to be common in most real world deployments due to lack of resources / real world threats against it. That’s quickly changing and we should assume all CVEs are real CVEs and start acting accordingly.
3. As always, I worry about regulatory capture in the wake of this. I think an underreported on fact is that due to “safeguards”, HuggingFace had to turn to non-domestic open source models to aid in their investigation.Yes I know they should have applied for privileged access but that sort of furthers the point. Infosec works partly by the fact that you have more to gain by protecting your systems than bad actors have by exploiting them. But if bad actors will have 10x more access to “frontier level cyber models” than the average defender, we will all lose.

What actually happens when we lift the abstraction level of our work?

The answer, always, is lighten up and look for the fun.

There’s a lot of worry about jobs going away, mass unemployment, and a crisis of meaning. But we’re faced with a choice: empowerment or disenfranchisement. The worst thing you can do is refuse to play the game.

The thing I’m most interested in right now in agentic engineering isn’t the inner loop - the tool calls, the reasoning, the retry logic. We’ve seen a lot of good work there already, and honestly the simplest harnesses (like Pi) might be the most interesting precisely because they’re simple.

What I’m more interested in is the outer loop: systems that modify and improve the agent over time. Not agent orchestration (“run multiple agents”), but continuous learning loops. You set an agent loose on a problem. The next great product will be a learning system that makes the agent better at solving your problem each time it tries, until eventually it’s managing that part of your organization for you.

I’ve been spending some time this weekend hacking on a side project I call Vibe Check. The idea is instead of doing agentic code review, you let the agents check your code by actually running your software.

The long-term question is: if these agents get good enough, can these checks basically substitute for review and verification that the software does what it’s expected to do? And in that case, kind of bypass the review step for some portion of your code.

In some ways it’s like a new type of CI platform, but everything happens inside a sandbox, and the agent is the runtime. It makes decisions like “given this particular change, how would I run this code?” and then invokes a series of skills.

It’s been pretty interesting to play with. Not sure if it goes anywhere.

Read Greenblatt's updated timelines for AI progress, and it's got me thinking about my own assumptions and priors on where we are. There's no doubt that the release of Opus 4.6 and GPT-5.4 felt like they came faster than expected and were stronger than expected, and with the recent talk around Mythos, it sure feels like we are approaching more of a fast takeoff situation. I still find it very hard to make long-term predictions around any of this, because there's a big part of me that's still skeptical that scaling language models will lead to truly interesting intelligences, but I keep being proven wrong over and over again. Recommend reading it. I also found How to Prepare Yourself for AGI a timely companion.

Created a font of my handwriting by using calligraphr.com. Notes from my books feel more real when they actually look like the notes that I take. Surprised by how easy this was.

Scanned in a bunch of my bookshelves from all over my bedroom this weekend, and used Claude Code to get them into the website. Now I have a searchable list of a pretty large subsection of my books, along with annotations - like these ones from the book Anxious. Wild that this kind of project is a couple hours now instead of weeks.

I always know spring is here when I walk outside my building and see this tree in full bloom. It’s been a long winter - a wonderful one, but a long one - and then spring comes and you remember why you live here.