Back to blog

The December 2025 AI Shift: What Changed and Why It Matters

8 min read

Something changed in late 2025. If you've been building software, you felt it. If you haven't been paying attention, you're about to be caught off guard.

This is my attempt to make sense of what happened, what it means, and what we should all be doing about it.


What Actually Happened in December 2025

Before: AI as a faster typewriter

For most of 2024 and early 2025, AI coding tools were useful but limited. Copilot suggestions. ChatGPT for explaining errors. Claude for drafting boilerplate.

The pattern was always the same: you remained the programmer. AI helped you type faster, but you still held the steering wheel. On longer tasks, models would drift, hallucinate, or produce confident-looking code that was subtly broken.

Useful? Yes. Transformative? Not yet.

The shift: November-December 2025

Then came Claude Opus 4.5 and GPT-5.2 in November 2025.

On paper, the improvements looked incremental. In practice, they crossed an invisible line. The models could now:

  • Follow complex instructions across long sessions without losing context
  • Debug their own output and recover from errors
  • Handle multi-file, multi-step tasks reliably
  • Actually complete what you asked them to do

Simon Willison called it an "invisible capability line" - incremental improvements that suddenly unlocked harder classes of work. 1

By December, the reports were converging. Greg Brockman shared that OpenAI's internal teams had shifted: before December, Codex helped with unit tests. After December, agents were writing "essentially all the code." 2

Andrej Karpathy described building a complete video analysis dashboard by writing instructions while an agent ran autonomously for 30 minutes, hit problems, solved them, and delivered working software. 3

This wasn't autocomplete anymore. This was delegation.

What actually changed?

I think it came down to a few things converging at once:

Model capabilities crossed a threshold. Not magic, but reliable enough for real work. The models stopped compounding errors and started compounding correctness.

Agentic tooling matured. Claude Code, Codex CLI, and similar tools gave these models the scaffolding to actually execute - not just generate text, but run code, see results, and iterate.

Investment and iteration paid off. Years of work on reasoning, tool use, and long-context handling finally added up to something qualitatively different.

The result: AI moved from "assistant" to "executor." You describe what you want, it builds it, you review the output.


The Theory: What Happens Next

There's a provocative scenario that's been circulating called the "2028 Global Intelligence Crisis" from Citrini Research. 4

The thesis: if AI keeps improving at this pace, white-collar work gets disrupted faster than the economy can adapt. Implementation becomes so cheap that commodity software gets repriced. Companies that existed to arbitrage human labor costs face existential pressure. Spending drops because displaced workers buy less. The reflexive loop accelerates.

Is this going to happen exactly as described? Probably not.

Could parts of it happen faster than we're ready for? Yes.

I'm not a doomer about this. But I'm not dismissive either. The honest answer is: we don't know how fast this plays out. What we do know is that waiting for certainty is a bad strategy.


How I Think Companies Will Work

Here's my actual view on how this reshapes organizations.

The PM/engineer line is blurring

Product managers used to write specs. Engineers used to implement them. There was a handoff.

That handoff is compressing.

When AI can turn a well-written spec into working code, the quality of that spec matters enormously. PMs need to think more technically. And engineers need to think more about product - what should we build, not just how do we build it.

The people who thrive will be the ones who can work across that old boundary. Pure "throw it over the wall" roles get squeezed.

Every team needs someone staying current

AI tooling moves fast. What worked three months ago might be obsolete. What everyone dismissed six months ago might now be essential.

Teams need someone - doesn't have to be a formal role - who actively experiments with new tools, figures out what's actually useful, and helps the rest of the team adopt without everyone duplicating the exploration.

Call it an "AI enabler" or just "the person who keeps us from falling behind." Either way, it's becoming necessary.


What Engineers Need to Do

I've thought about this a lot. Here's what I actually believe matters:

Adapt. That's the baseline.

If you're not using these tools yet, start. If you're using them casually, go deeper. The gap between engineers who've adapted and those who haven't is widening fast.

This isn't about being an AI enthusiast. It's about being employable in 18 months.

Get better at communication

This sounds soft, but it's becoming critical.

When you're directing AI agents instead of writing code yourself, your ability to clearly articulate what you want - constraints, edge cases, non-goals - directly determines output quality.

Vague prompts produce vague results. Clear specs produce usable code.

Communication is no longer a "nice to have" soft skill. It's a core engineering capability.

Know concepts, not syntax

You still need to understand systems, architecture, and how things work. But memorizing syntax? That's increasingly pointless.

What matters is knowing what to build and why - understanding tradeoffs, failure modes, and system interactions. The AI can handle the how of implementation. You need to handle the what and should we.

Review the plan, then the code

When reviewing AI-generated work, start with the approach. Does the plan make sense? Is it solving the right problem? Is the architecture reasonable?

Only after you're satisfied with the plan should you dig into the code details. Getting lost in syntax review while missing a fundamentally wrong approach is the new failure mode.

Learning compounds faster now

Here's the upside: learning has never been faster.

You can explore a new framework, understand a new domain, or prototype an idea in a fraction of the time it used to take. AI accelerates learning. That learning makes you better at using AI. The cycle compounds.

The engineers who lean into this - who use AI to learn more, faster - will pull ahead. The ones who see AI as a threat to avoid will fall behind.


What Companies Should Do

Actually experiment

Most companies are still talking about AI adoption, not doing it. They're waiting for someone else to figure it out first.

That's a mistake. These are early days. The companies that experiment now - really experiment, not just pilot projects that go nowhere - will learn things their competitors won't.

Measure the impact

Don't just adopt AI tools. Measure what changes. Are you shipping faster? Is quality holding? Where are the gains real and where are they illusory?

Without measurement, you can't tell the difference between actual productivity improvement and just feeling busy in a new way.

It's still early

This is important context. We're not in the "mature adoption" phase. We're in the "figuring it out" phase.

That means there's room to experiment, make mistakes, and learn. But it also means the window for easy catch-up is closing.


What Non-Tech People Should Do

The honest answer: I don't know exactly how this plays out for every role. But here's what I believe.

Embrace AI, but keep the human touch

AI can draft, summarize, analyze, and generate. It's going to touch almost every knowledge work role.

The response shouldn't be resistance. It should be: figure out how to use these tools while doubling down on what makes human judgment valuable.

The best work will combine AI capability with human taste, context, and relationships. Pure AI output often feels hollow. Pure human output is increasingly slow and expensive. The sweet spot is the combination.

Your job probably changes

Not necessarily disappears - but changes. The tasks that made up your day will shift. Some will be automated. Some will be augmented. New tasks will emerge.

The people who adapt will find the new shape of their role. The people who insist nothing is changing will be surprised.


My Overall Outlook

I'm balanced on this. Not a doomer. Not a cheerleader.

Slightly cautious, if I'm honest.

The productivity gains are real. The leverage is real. The ability to build things faster and learn things faster - that's genuinely exciting.

But jobs are not going to look the way they did before. The demand for pure implementation work is dropping. The value of judgment, taste, and human oversight is rising - but it's not clear that creates as many jobs as it displaces.

I don't think this is the apocalypse. I do think it's a real shift that requires real adaptation.

The worst response is to ignore it and hope it doesn't affect you.

The best response is to engage honestly: learn the tools, understand the tradeoffs, adapt your skills, and stay useful in a world where software is becoming abundant.

That's what I'm trying to do. That's what I'd recommend to anyone reading this.


References


I'm navigating this transition too, while job hunting. If you're in the same boat, I'd love to compare notes. Reach me via Hire.

Footnotes

  1. Simon Willison, 2025: The Year in LLMs: https://simonwillison.net/2025/Dec/31/the-year-in-llms/ ↩

  2. Greg Brockman on OpenAI's internal shift: https://x.com/gdb/status/2019566641491963946 ↩

  3. Andrej Karpathy on autonomous agent coding (Feb 2026) ↩

  4. Citrini Research, 2028 Global Intelligence Crisis scenario: https://www.citriniresearch.com/p/2028gic ↩

aiengineeringcareerssoftware