Zach Brock

"Amateurs have a goal. Professionals have a process."

@nk Got any example conversations? Happy to pass to the team cc @ericmitchellai

Chat GPT 5.4 is really frustrating to work with. Awful hallucinations and laziness. It's hard to tell if it's "smarter" when it's so difficult to steer.

Ok

@TheZvi @tszzl @kimmonismus I'll mention it to the team

@tszzl @kimmonismus Can you just skip to 5.4 so I can get some sleep.

Seriously, I dont get it. - Today, GPT-5.3 instant is being released . - The blog post states at the very bottom that 5.3 Thinking and Pro will also be released very soon. - An hour later, the official OpenAI X account tweeted that GPT-5.4 will be released very soon. ??? So in a few days we get: GPT-5.3 thinking + Pro + GPT-5.4 (???) instant/thinking/pro?

@btaylor Glad you enjoyed the post!

I’ve been trying to simulate using Codex for the next year and what will change about my perspectives on software engineering as I transition from being a computer programmer to a harness engineer. There are so many, but here are a couple that have stuck with me: Software dependencies - Large open source systems like Linux and MySQL seem like they will remain just as important, but I wonder if I will start to have different perspectives on smaller software libraries when the functionality can be relatively easily produced and tested with AI. Given the past decade of supply chain vulnerabilities and maintenance issues in open source libraries, will it become a best practice to reduce dependencies and write our own where possible? Documentation - When I built a product before, the “specification” was split between docs, Slack, Figma, and Linear — but the vast majority of behavior was specified in code, i.e., the long tail of functionality is an emergent property of the code I write. The conundrum with agent-produced code is that it’s not clear which parts of the code were prompted (i.e., specified) and which parts were “vibed” (i.e., unspecified). That seems problematic when continuously evolving a large system over time because the harness will “forget” past instructions. I don’t think replaying prompts is correct either because in a single Codex session, a good chunk of interactions are interactive and effectively transient. I have an intuition that documentation will be as important of an output of my Codex sessions as code, documenting the substantive product decisions made during my session. Those docs clearly need to be directly in the repo, versioned with the code and available as context for future sessions. The docs / context discussion in OpenAI’s recent post on harness engineering resonated with me and maps to my intuition: openai.com/index/harness-…

@martin_casado We wrote a bit about how we approached this problem here: openai.com/index/harness-…

The difference in stuff AI coding is good at vs not is getting more stark. It's very good at basic engine stuff you'd want to build on the way. Tooling, testing, basic engine design. Frameworks etc. But it's really not good at anything where runtime understanding is important. I've seen this working on a splat renderer and multiplayer back end for a game engine. In both cases, AI creates a pretty reasonable guess. But lacking actual understanding of the runtime semantics, the results are basically unusable. This produces somewhat of a dilemma. The better it is at the stuff you can derive optimally from syntax, the more disconnected we are when we actually need to design around runtime semantics. To manage this, I've started to include both schema design, notes on state consistency and runtime traces to the LLMs. It's not perfect, and I still need to be in the code a lot. But it helps to start pulling semantic dependencies like this out.

@mvanhorn @slashlast30days @openclaw @YouTube Very cool!

V2.1 of @slashlast30days is out. Now with @OpenClaw, free @YouTube transcripts and a Codex Skill. 1. @openclaw + watchlists - automated research via cron jobs on your competitors, people, and topics 2. YouTube transcripts as a 4th source 3. Works in OpenAI Codex

@jacobtechtavern @dkundel @PaulSolt We used codex cli, the app we were building wasn’t the codex app

@dkundel @PaulSolt How did they write the MVP if they didn't have codex to write it with?

It was fascinating to watch this team build with Codex without writing a single line of code by hand. Lots of wisdom and shared experience in this article and a glimpse of how the role of the engineers changed for the team

@ClayCampaigne @asymmetricinfo You should try!

@asymmetricinfo The guy says that Claude code can create a full working app (not a prototype) and test it like a human in a single prompt. Do you really have friends in tech who tell you that? I think these coding agents are amazing but his characterization is obviously dramatically exaggerated

If “AI-generated word salad” can convince a bunch of folks that AI will be the biggest disruption since man tamed fire … well, you should update your priors in favor of AI being incredibly disruptive.

@NerfThatNerd @snsf @_rake92 👋

@snsf @_rake92 @z What in the gigachad handle is @z

We shipped an internal beta of a product with zero human-written code — every line was generated by Codex agents, boosting velocity ~10x. Here are some lessons from how the team did it. Great work by @_rake92, @z and the team in pushing the boundaries of what software engineering looks like in the Codex world.

Totally normal timeline we live in to have buckled my laptop into the backseat of my SUV so it can be tethered and let 4 tmux sessions of GPT-5.2 Pro cook #Codex

📣 Shipping software with Codex without touching code. Here’s how a small team steering Codex opened and merged 1,500 pull requests to deliver a product used by hundreds of internal users with zero manual coding. openai.com/index/harness-…

Not writing code by hand has been a big shift for sure. Unexpectedly, it’s also been really fun. As @_lopopolo likes to say “we just do the fun parts of engineering on this team”

I’m very excited that we’re finally publishing a deep dive into how my team has been working for the last 6 months. @_lopopolo is a real visionary and it’s been a blast working with him. openai.com/index/harness-…

It’s clear that we’ve barely scratched the surface of what the models are capable of. I can’t wait for non-coders to Feel The AGI the way codex users do.

The biggest crowdsourced science effort of the internet was the Bodybuilding.com forums (2008–2013). Turns out millions of bros running uncontrolled n=1 experiments was better than whatever the FDA was doing.

Codex is the ultimate "one more turn" game

Maybe the real super intelligence is the friends we made along the way

A lot of people haven’t used Claude Code or Codex in the last 3 months. The cognitive dissonance they’re going through reading twitter this week must be rough.

The real Super intelligence test will be “tell me how much this surgical procedure costs at my 3 nearest hospitals”

This from @obie captures how I'm feeling about coding agents