They Were So Preoccupied With Whether They Could 🦖

"Your scientists were so preoccupied with whether or not they could, they didn't stop to think if they should."
— Ian Malcolm, Jurassic Park (1993)

Jeff Goldblum delivered that line in 1993 and it landed as sci-fi cautionary tale. A pithy warning about hubris wrapped in a blockbuster about CGI dinosaurs eating lawyers off toilets.

Thirty years later it's not a cautionary tale anymore.

It's a product roadmap.

🦕 The Park Is Open

The Jurassic Park script reads like a transcript from any AI all-hands in 2024. Swap "genetic power" for "large language models" and Hammond's island for a data center in Virginia and you've got the same conversation — same roles, same dynamic, same uncomfortable laughs.

Gennaro is the monetization guy who's already pricing the merch before the ink dries on the safety report. Hammond is the visionary founder who genuinely believes he's doing this for humanity and will not hear otherwise. Malcolm is the chaos theorist nobody wants at the table because he keeps asking questions that don't have good answers yet.

Every AI company has all three archetypes in the room right now.

Most of them are listening to Hammond.

But it's 2026 now. And the park has new exhibits.

They're not just running the park anymore.

They're replacing the park rangers.

💰 "$2,000 a Day. $10,000 a Day. And People Will Pay It."

Gennaro's pitch isn't cynical exactly. It's just pre-emptively incurious. He doesn't need to see the park to know it'll work. He has a model. He has comps. He has a slide deck and a coupon day joke ready for the optics.

This is how AI pricing, packaging, and GTM strategy is actually being built right now. Not after the safety questions get answered. Not after we understand the second-order effects. The lunch boxes get designed during the safety review. The enterprise tiers are live before the red-teaming is done.

The business case is airtight.

The dinosaurs are already loose.

🔬 "It Didn't Require Any Discipline to Attain It"

This is the line that hits hardest in 2026.

Malcolm's critique isn't anti-science. He's not a Luddite and the movie is smart enough to have Hammond call him one so we can clock the deflection. Malcolm's point is about the relationship between capability and accountability — that when you inherit power without earning it incrementally, you skip the part where consequence teaches you respect.

We trained on the internet. All of it. We took the next step. We didn't spend decades in a lab selectively breeding capability up from first principles — we hoovered up the sum total of human knowledge, found the pattern, and scaled it.

Before we knew what we had, we patented it.
Packaged it.
Slapped it on a plastic API endpoint.
And now we're selling it.

The knowledge transfer happened at GPU speed. The wisdom transfer did not.

🧬 "Your Scientists Were So Preoccupied With Whether They Could"

Here's the thing about the AI labs that's easy to forget when you're in the day-to-day of shipping: the people building this stuff are genuinely brilliant. Hammond's scientists were too. That's what makes the critique land — it's not about competence.

It's about sequencing.

Could is an engineering question. It has tests, benchmarks, evals, a definition of done.

Should is a systems question. It has stakeholders you haven't identified yet, timelines you can't compress, second-order effects that don't show up until the T-Rex is already outside the paddock.

The industry is exceptional at could. It has built entire institutions, research agendas, and conference tracks around could. The should infrastructure is a few nonprofit orgs, some policy papers, and a lot of people saying the right things on panels while the deployment roadmap runs on a different schedule entirely.

Malcolm saw this in the genetics lab. The capability arrived before the governance did. Not because anyone was evil. Because could has a finish line and should doesn't.

🎟️ It Was the AI Labs' Hubris First

To be fair, the original sin belongs to the labs.

They built the capability. They ran the demos. They published the benchmarks showing GPT-whatever could pass the coding interview, ace the system design question, implement the algorithm faster and quieter than the senior engineer who'd been doing it for eight years. They built the story before the product. They sold the product before the story was true. They made "autonomous agentic engineer" sound like a solved problem while the thing still hallucinates package names and confidently writes tests that test nothing.

Malcolm called it in the genetics lab: "you read what others had done and took the next step". The labs read the transformer paper, took the next step, took the next step, took the next step — and by the time anyone asked whether the thing understood code or was doing something statistically adjacent to understanding code, the enterprise contracts were already signed.

The paddocks were always going to break.
They just need a Dennis to abuse the system and make Ray’s everywhere scream at the screen "PLEASE!"

🏢 Then Every Tech Company Picked Up the Gun

Hammond at least believed in what he built. Pathologically, delusionally, fatally — but genuinely. He thought the park was a miracle. He thought the fences would hold because he needed them to hold because the alternative was unthinkable.

Your VP of Engineering doesn't have that excuse.

They've read the same incident reports you have. They know the agentic system misses context that lives in Slack threads and in the heads of people who've been paged at 2am by that service three times. They know acceptance criteria in Jira is a lossy compression of what actually needs to happen. They know the thing confidently ships code that is locally correct and systemically wrong — because the system it lives in isn't in the ticket.

They know all of this.

And they're running the pilot anyway.

Because the math works at the spreadsheet level. Because the board asked about AI strategy. Because a competitor announced something. Because could is a question that got answered in the demo and should is a question that gets answered later, apparently, by whoever's left.

This isn't Hammond's hubris — visionary and blind.

This is Gennaro's hubris. Eyes open. Coupon day ready.

🌿 "Dinosaurs Had Their Shot"

Hammond's condor gambit is the most underrated moment in the scene.

He reaches for a sympathetic case — an animal on the verge of extinction — to reframe what he's doing as conservation, as stewardship, as good. And it's not wrong, exactly. It's just a frame swap. Replace the thing you're actually doing with an adjacent thing that polls better.

You see this move constantly in AI discourse.

We're not automating jobs, we're augmenting workers.
We're not replacing expertise, we're democratizing access.
We're not building a surveillance layer, we're improving safety.

These aren't necessarily lies. Some of them are partly true. But they are consistently deployed to bypass the harder version of the question — the one about what you're actually doing and what gets selected out in the process.

The condors aren't the thing. The condors are the thing that makes you stop asking about the thing.

🔩 What "Fully Autonomous" Actually Means in the Codebase

Let me be specific about what's being tested, because the abstraction lets people feel better about it than they should.

The loop looks like this:

1. Agent reads the Jira ticket — the title, the description, the acceptance criteria, the linked Confluence doc that hasn't been updated since 2022
2. Agent reads the relevant codebase — or a window of it, because context limits are real and your monorepo is not
3. Agent writes the implementation — correctly, often, for the happy path
4. Agent opens the PR — with a description that sounds like a human wrote it, because it was trained on humans writing things
5. Engineer reviews it — the engineer whose job this is, who now has a new job, which is validating the output of the thing that is here to replace them

Step 5 is doing a lot of weight-bearing work that nobody in the planning meeting is talking about.

Because reviewing code you didn't write, for a system you didn't design the implementation of, against requirements you know are incomplete, from a generator that is confidently wrong in ways that are hard to spot — that is a different and harder skill than writing the code yourself.

We're not removing the engineering judgment from the loop.
We're making it load-bearing while pretending we removed it.

☄️ The Hubris Isn't "AI Can't Do This"

I want to be careful here because the wrong version of this argument is easy to make and easy to dismiss.

The wrong version: AI can't actually write good code, engineers are irreplaceable, this is all hype.

Some people believe that. Some of those people are right about some things and wrong about others. But that's not the argument worth making in 2026, because the capability question is genuinely unsettled and moving fast and anyone who's certain in either direction is selling something.

The actual hubris isn't believing the AI can code.

It's believing that software engineering is the same thing as code generation.

A Jira ticket is not a spec. Acceptance criteria is not a contract. The codebase context window is not institutional memory. The PR review is not equivalent to having been in the room when the architecture decision was made three years ago for reasons that made sense at the time.

The thing being automated isn't the typing.
It's the judgment that decides what to type.

And judgment isn't in the ticket.
It's in the person who knows what the ticket isn't saying.

👾 Training Your Replacement Is the Generous Reading

Here's the part that should make everyone uncomfortable.

When companies run these pilots with senior engineers in the loop — reviewing, correcting, refining — they frame it as "human oversight." Safety. Responsibility. Malcolm-approved chaos theory mitigation.

But the correction data goes somewhere.

Every time an engineer catches a mistake, rewrites a section, adjusts the approach — that's a labeled training example. The gap between what the model produced and what the engineer accepted is signal. Valuable signal. Signal that gets used to close the gap.

The pilot isn't just testing whether the AI can do the job.

The pilot is the mechanism by which the AI learns to do the job.

You're not evaluating your replacement.
You're onboarding it.

And the thing is — most engineers in the loop know this. They can feel it. They're not naive. But the alternative is opting out of the process entirely, which doesn't stop the process, it just removes the one part that might catch the thing before it ships something catastrophic.

So they stay in the loop.
And the loop tightens around their necks.

Final Thought

Malcolm's line that everyone quotes is the could/should one. It's the bumper sticker. It's on the poster.

But the line I keep coming back to is quieter. Gennaro says things are different than we feared — meaning better, meaning the park looks fine, meaning maybe the pessimists were wrong.

And Malcolm doesn't argue the data. He just says: I know. They're a lot worse.

Not because the fences had already failed. Because he understood what the fences were for, and he understood the people who built them, and he understood that the park looking fine was exactly the condition under which nobody would fix the thing that was about to break.

The agentic engineer pilots are looking fine.
The metrics are promising.
The PRs are getting merged.
The tour cars are moving.

The question was never whether the fences would hold.

It was whether anyone would hear them failing before the first guest was eaten.