You’re probably seeing more anti‑AI posts lately. This one captures the mood—engineers burning out, “vibe‑coded” demos getting applause, and scepticism that anyone is making real money from it. The author even suggests stopping AI use for a while to see what truly breaks. Interesting.
Problem
AI is in a weird spot.
Early on, we over‑hyped it, tried it, shrugged, moved on. Then GPT‑4 arrived and things felt legitimately useful. Look at that trajectory! Surely it will continue! CTOs and VPs started prioritising roll‑outs.
Fast‑forward to now: most developers in most organisations have LLM‑style tools in the IDE. Yet… productivity isn’t obviously up. The leap from GPT‑3 → GPT‑4 didn’t repeat with the next wave to GPT-5. We’ve also learned that vibe coding is mostly garbage, and that spec‑driven development outperforms seemingly magical prompting (shocking nobody: clearer instructions yield better results).
Meanwhile, adoption of next‑gen, spec‑first, agentic tooling is slow. Many devs are disillusioned; through failures of vibe coding aventures, they’ve stopped trying to tune the tools or lean into next‑generation tooling entirely. Maybe some tests were good to generate, but the real breakthroughts weren't coming. Leadership, however, hasn’t updated expectations. Exhibit A: in April 2025, Shopify CEO Tobi Lütke told staff that before asking for more headcount or resources, teams must show why they “cannot get what they want done using AI,” and said “reflexive AI usage” is now a baseline expectation — a stance that feeds into performance reviews (The Verge: https://www.theverge.com/news/644943/shopify-ceo-memo-ai-hires-job). Hopes stayed high while reality plateaued. That mismatch burns teams.
Why It Matters
- Misaligned expectations create pressure to ship low‑quality “AI‑assisted” code that’s costly to review and debug.
- Process debt: we skipped specs and context in the prompt‑engineering rush; now we’re paying interest.
- Tool fatigue: developers assume “AI doesn’t work” rather than “we used it badly.”
- Opportunity cost: we’re under‑investing in workflows that actually make prediction engines useful.
Reality Check
LLMs predict tokens. They are not magic. Prediction quality rises with better information and clearer constraints. Model upgrades help (longer context, cheaper inference, better efficiency), but the real gains come from how you use the models.
Where To From Here?
Option 1: Give Up
AI coding tools weren’t the saviour. Generated code is questionable by default, and we didn’t see time savings in backlog grooming, requirements capture, or customer‑feedback loops. You can choose this. It’s just not very… problem‑solvery.
Option 2: Lean Into The Next Wave
Same premise, better execution: feed the model richer, structured context.
What this looks like in practice (lightly):
- Rules & prompt modes: shared, versioned instructions per repo/team.
- Spec‑driven development: AI proposes diffs that trace back to explicit requirements and tests in files like
design.md,requirements.md,tasks.md. - Codebase understanding: auto‑generated (or human‑curated) architecture docs to ground changes.
- Persistent Context: A saved and used chat history.
- Server‑side agents: read access across repos, tickets, specs, and logs to gather what they need before they write anything.
- RAG on Internal Context: Find every line of code, chat message, document, Jira ticket you can and index it with RAG or similar techniques for deep research questions and other tasks later on.
(Anything else? Let me know!)
Opinion: I favour option two. The models will keep improving, but the step‑change comes from using them properly—with better inputs, clearer specs, and tighter feedback loops.
Key Takeaways
- Misaligned expectations are the real drag; reset them.
- Don’t give up; tooling + process is the lever (in moderation).
- Standardise artefacts enough to make AI useful; avoid cargo‑culting.
- The win isn’t “smarter models”; it’s smarter usage.
Closing
If LLMs are just predictors, your job is to feed them the right signals. Structure beats vibes. Specs beat spells. Context beats chat.
Good luck. Happy coding.