Eric Fredine

@thdxr Narrators voice: there is no flywheel

they said cursors data flywheel would make them unstoppable but then claude code came out they said claude codes data flywheel would make them unstoppable but then codex came out they said codex's data flywheel would make them unstoppable then composer 2 came out

@StatisticsFTW I’d say gpt 5.4 likes pure functions well enough

Claude really likes methods on structs instead of pure functions.

I've started using Codex with JetBrains and I'm liking it better than Cursor. I'm using the same GPT-5.4 medium model in both cases. Two main reasons: - Cursor is a cpu hog that often makes it unusable. - The explicit plan mode in Cursor adds friction and confusion.

It’s mind boggling bad that Cursor is unusable while the agent is working through any non-trivial task?

“But this is the biggest change I’ve seen in my career by far. An extinction-level event for rules of thumb.” Agreed. I feel humbly uncertain like never before but also excited and curious.

My fave Chuck Norris joke: Chuck doesn't flush the toilet, he scares the shit out of it.

Something you're missing is having the agent do less work. So, for example, instead of having the agent run ESLint, you instead have a function that runs ESLint and only invokes an agent to fix the issues if issues are actually found. That's the premise of Barnum. x.com/StatisticsFTW/… This also makes it extremely easy to prevent context bloat.

@charliermarsh @tszzl Wow - congratulation!

We've entered into an agreement to join OpenAI as part of the Codex team. I'm incredibly proud of the work we've done so far, incredibly grateful to everyone that's supported us, and incredibly excited to keep building tools that make programming feel different.

The AIs are better at Ui code than writing efficient data pipeline type code. It can write efficient code but won’t do it without being walked through how to do it.

Fascinating to watch the entire industry grappling with this.

@JustDeezGuy Hmm, yes. Learn the true network structure rather than just the weights in a static network.

So… back to Bayesian belief networks?

@jamonholmgren I think this is how much we should be spending if we are taking best advantage of it.

@fredine I have Claude Code premium plan, Codex not exactly sure what plan offhand but it's robust, and a token-usage Cursor plan which I'm willing to spend a decent amount on. Probably like $500/mo but with this system I'm actually burning fewer tokens than when I was doing interactive.

Allow yourself to feel the pain. Nothing will motivate you to improve your workflow, docs, and specs like waking up to a mess.

Introducing Barnum, or... how I ship hundreds of PRs per week, burn through backlogs, and automatically fact-check documentation. LLMs are incredibly powerful tools. But when we try to use them to drive more complicated refactors or more intricate workflows, their shortcomings are quickly revealed. When their context gets full, they get forgetful, and they can't be relied upon to necessarily do the steps that you ask. They often cut corners. Put simply, having an inherently probabilistic process perform what should be deterministic work necessarily comes at the cost of reliability. And you can't build a complicated workflow off of unreliable foundations. That's where Barnum comes in. Barnum is the missing workflow engine for agents. Rather than having agents be responsible for upholding guarantees (e.g., always lint and commit your changes atomically), agents instead do just what they're good at: reading text and reasoning. Everything else is done deterministically, on the outside, by Barnum. This means that you can build bigger, more involved workflows without sacrificing reliability. Because you can intersperse bash scripts, you save on token usage. The agents performing a micro-task only receive the instructions for that specific task, meaning that context does not get overwhelmed and they don't get forgetful. And because all inputs, outputs, and transitions are validated, the agents can't wriggle out of doing the work. This workflow is essentially a state machine described in a config file. And the best part? The configuration has a JSON schema, so agents are actually really good at writing the workflow! It's already been used to ship hundreds of PRs, run automated refactors, burn through various backlogs, fact-check every statement in documentation, and build a deep-research clone! The attached image is a representation of the workflow that I use to identify and implement automated refactors. I follow this up with a separate workflow that splits each commit into a separate PR, judges the refactor, and potentially completes the refactoring (for example, by modifying call sites if the refactor changed some public API). So go on, give it a try. Check out barnum-circus.github.io, star the repository, and join the Discord! I can't wait to see what you build with it! And I'd love for you to get involved!

Great things are for stealing! I love Oxide's RFDs. We tried something similar at Sentry, but it was just tricky to do it with GitHub issues and pull requests. Earendil's attempt is now based on Google docs synced to a repo and website. rfc.earendil.com

@zombodb They could have a checkbox: “avoid left turns because of course I hate them”.

@fredine My mother goes out of her way to avoid left turns. Google could consult with her!

The Google Maps routing algorithm needs a bigger penalty for the cognitive load of needing to make a left turn at busy intersections.

Ironically, hypermedia (HATEOAS) has accidentally become a plausible API design scheme again. LLMs will robustly follow API links just like its designers hoped.

I think one of the conclusions we should draw from the tremendous success of LLMs is how much of human knowledge and society exists at very low levels of Kolmogorov complexity. We are entering an era where the minimal representation of a human cultural artifact... (1/12)

@LewisCTech It’s often implicit in the name of the field.

@fredine This has been a good communication exercise to me because I never thought of merely having an ID that points to something else be a foreign key: for me it was the contraint itself that made it foreign. I will be more precise in the future.

Foreign keys - do we really even need them? Discuss.