7 Hours for a Two-Field Form

Simon Carver

[warmly] Welcome to the show. Lachlan, I keep coming back to this one little number: 7 hours. Because that was CJ Murphy trying to use advanced AI tools to build what sounds almost insultingly simple -- a GitHub-hosted web form with two fields, password-protected submission, and an admin view. Two fields! Not a moon landing, not a trading engine... a form. [short pause]

Lachlan Reed

[deadpan] Yeah, two fields is backyard-shed stuff. That’s not “assemble a spaceship with a butter knife.” A decent human developer could knock that over in, what, 20 minutes? Maybe a bit longer if the Wi-Fi’s acting like a galah. But CJ used ChatGPT Atlas, switched to Claude for debugging, kept prompting properly, kept correcting it, and the whole thing turned into a SEVEN-hour wrestle. [exhales sharply]

Simon Carver

[reflective] And that’s the part I find so revealing. The promise of AI, at least the sales pitch, is compression. Less time, less effort, fewer bottlenecks. But CJ’s story sounds more like supervising a bright student who keeps making tiny mistakes. Not catastrophic, not absurd, just... dozens of little misses that pile up. A wrong path here, a broken assumption there, one fix creates another problem. [pauses]

Lachlan Reed

[curious] That professor analogy is bang on. Like, the student’s keen, listens well enough, even sounds confident -- but you wouldn’t let ’em wire your house unsupervised. And that matters because people hear “AI can code” and imagine you chuck it a prompt, go make a coffee, come back, job done. CJ’s example says the opposite. The work becomes: inspect, correct, re-prompt, test, debug, repeat till your eyeballs dry out.

Simon Carver

[questioning tone] And I think the sharp question is this: if the system is truly capable in the way we’ve been told, why does a 20-minute task become a 7-hour chain of supervision? Because 7 hours isn’t a rounding error. It’s not “a little slower than expected.” It’s a totally different kind of work. [long pause]

Lachlan Reed

[responds quickly] Exactly -- 20 minutes versus 7 hours is the whole ball game. That’s not automation; that’s overhead. It’s like saying, “Good news, mate, I’ve invented a self-cooking barbecue,” and then you spend the entire arvo stopping it from setting the deck on fire. Sure, technically the barbecue’s doing something. You’re still stuck there babysitting it.

Simon Carver

[laughs softly] Right. And I don’t even think this means the tools are useless. That’s too easy. The interesting thing is that CJ seems to know what he’s doing. He wasn’t just flailing around with vague prompts. He used strong prompt engineering, moved between tools, used Claude specifically for debugging when one path wasn’t working. So this isn’t the story of a novice blaming the hammer.

Lachlan Reed

[skeptical] Nah, and that’s important. Because whenever one of these stories pops up, someone goes, “Well, you just prompted it wrong.” Maybe sometimes, sure. But if a bloke who knows the tools, swaps models, keeps tightening the instructions, and STILL burns 7 hours on a two-field form... then the friction is in the system, not just in the operator. [short pause]

Simon Carver

[softly] There’s also a psychological trick in it. AI often feels like it’s almost done. It gives you something plausible-looking, and plausibility is a dangerous kind of seduction. You think, “Oh, we’re close.” Then you test it, and no, this part fails. So you patch that. Then the patch breaks something else. It’s like walking toward a horizon line that keeps moving.

Lachlan Reed

[chuckles] Yeah, it’s the old “nearly there” trap. I’ve done that restoring bikes -- you think one last bolt and she’ll fire up, then suddenly you’re elbow-deep in a carburettor wondering where your Saturday went. Only with CJ, the weird bit is the AI is meant to REMOVE that pain. Instead it introduced a fresh layer of fiddling. So the real question for me is: are we saving labour, or just changing the shape of it? [pause]

Simon Carver

[calm] And once you ask it that way, the hype starts to look a little thin. Because the value isn’t just whether AI can produce output. It’s whether that output is reliable, repeatable, and trusted enough that a human doesn’t have to hover over it the whole time. CJ’s form story says we’re not there. Not even close, really.

Lachlan Reed

[matter-of-fact] And that rolls into the bigger pattern. The issue isn’t one annoying form. It’s repeated errors, weak reliability, and this weird lack of memory. You correct the thing, start a new session, and half the time you’re back at the same pothole. If these systems are so clever, why aren’t they getting better at common problems? Why’s the human still hauling the whole ute out of the ditch? [sighs]

Simon Carver

[reflective] That’s the myth of learning, isn’t it? People talk about these tools as if they’re steadily becoming your trusted colleague. But CJ’s experience suggests something blunter: these aren’t autonomous systems. They’re assisted output generators, and the assistance runs in one direction. The human provides the context, the correction, the standards, the judgment. The machine produces drafts that may or may not survive contact with reality. [short pause]

Lachlan Reed

[questioning tone] So let me try and say it back -- slightly wrong, maybe. We’re calling it intelligence, but in practice it behaves more like a very fast pattern machine with no real ownership of the task. It doesn’t actually know when it’s cooked the admin permissions or broken the logic; WE know that when we test it.

Simon Carver

[responds quickly] Yes -- that’s exactly the missing piece. Ownership. Accountability. The system doesn’t experience the consequences of being wrong. CJ does. He’s the one losing the afternoon. He’s the one switching from ChatGPT Atlas to Claude, trying to salvage momentum. He’s the one deciding what “working” even means.

Lachlan Reed

[frustrated] And then you get his second example, which is even messier because it leaves the neat world of a coding task. He tries to automate podcast distribution to Facebook and LinkedIn using Zapier. Sounds tidy on paper, right? Hook the platforms together, push the episodes out, done. But instead he runs into broken workflows, dodgy integration design, and unclear permission models. After 10-plus hours, the fix comes from CJ’s own reasoning -- not from AI magic. [exhales sharply]

Simon Carver

[curious] The “10-plus hours” is the thing that sticks for me. Because again, that number changes the story. If the automation layer creates more coordination work than it removes, then what exactly are we calling efficiency? And with Facebook, LinkedIn, Zapier -- those are named platforms with real constraints, real permissions, real governance issues. This isn’t some abstract future problem. It’s happening at the level of everyday operations.

Lachlan Reed

[skeptical] Yep. This is where the hype mob goes a bit quiet. Because the failure isn’t just the model spitting out a wonky answer. It’s the whole stack being patchy. APIs don’t line up cleanly, permissions are murky, workflows break in annoying little ways, and there’s no proper standard holding it all together. Bit like the early internet before the roads were paved -- everybody had a map, nobody had the same street names. [short pause]

Simon Carver

[thoughtful] That comparison matters. Before standards converge, progress feels magical in demos and chaotic in practice. And organizations tend to hide that chaos under the language of transformation. But if you strip away the shiny language, what CJ is describing is a governance vacuum: fragmented integrations, weak interoperability, unclear rules about who can do what, and humans constantly stepping in to interpret the mess.

Lachlan Reed

[leans in][serious] Which brings us to the real reframe: AI isn’t wiping out human work. It’s creating a supervision economy. More validation. More correction. More orchestration. We’re not replacing engineers -- we’re turning them into machine wranglers. Babysitters, basically, but with better laptops and worse posture. [chuckles]

Simon Carver

[laughs, then softens] And that sounds funny until you feel the weight of it. Because supervision is still work. In some cases it’s harder work -- more cognitive load, more vigilance, more responsibility without the satisfying clarity of doing the thing cleanly yourself. The machine may generate the first draft, but the human becomes the last line of defense.

Lachlan Reed

[calm] And maybe that’s the skill that actually matters now. Not just coding, not just prompting -- controlling systems that don’t behave predictably. Knowing when to trust, when to check, when to pull the plug and do it the old-fashioned way. That’s not as sexy as the robot-replaces-everyone headline, but it’s a lot closer to what’s happening on the ground.

Simon Carver

[reflective] Yeah. I think the danger is that leaders hear “automation” and picture subtraction: fewer people, lower cost, cleaner process. But CJ’s stories point to addition: more oversight, more judgment, more invisible human effort stitched into the gaps. So maybe the real question isn’t whether AI will replace us. Maybe it’s whether we’re honest enough to admit how much of its apparent intelligence still depends on human beings standing behind the curtain, catching it before it falls. [long pause]

Lachlan Reed

[warmly] That’s the bit to sit with, I reckon. Anyway -- good yarn, mate.