The horrors persist, but so do we

Landing.Jobs and Damia published the Tech Talent Trends Report & Salary Benchmark 2026 in May, combining a survey of 1,115 tech professionals with salary expectations from 6,500-plus candidates gathered inside live recruitment processes.¹ It is the most complete public picture of the Portuguese tech market anyone has put out, and reading it end to end put a number on something I had only been feeling.

Average tech gross annual salary is 53.671€, up 0.9% year on year after 2025 delivered 14%. Seven years of growth flattening into a plateau. Product companies pay into the mid sixties for seniors while consulting sits at 51.985€ and public sector at 42.188€. 74% of respondents use AI coding tools. 54% are open to or actively looking for a new role, 40% would relocate for a foreign company, and only 13% currently work cross-border.

And the juniors are gone. The recruitment candidate pool went from 75% senior to 91% senior in a single year, with juniors at 2%. In the survey, 47% have more than nine years of experience and 13% have under three.

I wrote three parts on AI earlier this year, covering where the tooling came from, why the marketing around it is dishonest, and where the actual leverage sits for a working developer. This is the part I could not have written then, because the labour market had not finished moving yet. It has now moved enough to describe.

The funnel

Here is what my own job search actually looks like once you apply the constraints that matter.

             every opening on the board
                         |
                         v
   +-------------------------------------+
   |  1. reads like work I have done     |   full stack w/ CMS,
   |                                     |   DevOps-adjacent, MarTech
   +-------------------------------------+
                         |
                         v
   +-------------------------------------+
   |  2. pays above the Tuga rate        |   > 53.671 EUR
   +-------------------------------------+
                         |
                         v
   +-------------------------------------+
   |  3. remote-first, foreign payroll   |   72.012 vs 54.722 EUR
   +-------------------------------------+
                         |
                         v
   +-------------------------------------+
   |  4. survives the AI resume filter   |   before a human reads it
   +-------------------------------------+
                         |
                         v
                    what is left

Every stage of that funnel has become narrower in the last eighteen months, and stage four did not exist in a meaningful way three years ago.

Stage one: the lane

My lane is full stack engineering aimed at CMS work, DevOps-adjacent undertakings, and MarTech. That is a real specialisation with real depth, and it is also increasingly hard to name in a way that a search filter will match, because the titles have collapsed inward. More on that below.

Stage two: Tuga pay

"Tuga" is Portuguese slang for a Portuguese person or a Portuguese thing, used affectionately by locals and less affectionately when attached to a payslip. Tuga pay is the going domestic rate, and the going domestic rate is low for the skill on offer.

The report puts the average at 53.671€ gross. Against Portuguese cost of living that is a decent salary. Against the international market for the same skills, delivered remotely, at the same quality, it is a discount that companies have learned to depend on. Portugal has spent a decade marketing itself as an engineering hub, and a meaningful part of that pitch to foreign employers has been arbitrage.

The workforce knows. 40% say they would relocate for a foreign company while only 13% actually work cross-border, and that gap between wanting out and getting out is the most honest number in the entire report. It is not disloyalty. It is arithmetic.

Stage three: the remote chase

Which is why remote is not a lifestyle preference here. It is the pay rise.

The report's own gradient is unambiguous. Fully remote without office options averages 72.012€ for seniors. Fully remote with flexible office is 59.156€. Hybrid at two days or less is 57.227€, hybrid at three or more is 55.179€, and full office is 54.722€. A 31% spread across the same seniority band, sorted almost perfectly by how little the employer expects to see you.

Satisfaction tracks it exactly. On a ten point scale, fully remote scores means of 9.2 and 9.5. Full office scores 5.5 with a mode of 1, meaning the single most common response from people in full-office arrangements is the lowest score available.

So everyone chases the same thing: a remote seat on a foreign payroll. Which means the competition for those seats is global, and the person you are up against is not in Lisbon.

Stage four: what AI did to the front end, and what it did to titles

I went looking for a figure I half-remembered, a 37% collapse in frontend roles. It does not exist in that form, and the real number is more interesting.

Indeed's Hiring Lab reported in July 2026 that software development postings have grown nearly 15% since Claude Code launched in late February 2025, while overall postings fell 7%.² Software development is recovering faster than the broader market. But it remains roughly 27.5% below February 2020 levels, and the composition of the recovery is the part that matters: 71% of the increase between May 2025 and May 2026 came from senior roles, and 37% came from jobs with AI in the title.

There is the 37%. It is not a collapse in frontend work. It is a description of who the recovery is for.

The same analysis found that between May 2022 and May 2026, the more exposed an occupation was to AI, the more its postings declined, and that this relationship flipped in the final year as AI-exposed roles rebounded hardest. Read those together and you get a market that punished AI-exposed work, then started paying a premium for people who could do that work with AI, and skipped the part where anybody junior got hired.

Frontend absorbed this first because frontend was the most legible target. Component scaffolding, CSS, form wiring, and the long tail of React boilerplate are exactly the well-specified, verifiable, low-ambiguity tasks where the tools genuinely deliver. That is the same boundary I drew in Part III, and it turns out that boundary maps almost perfectly onto a job description.

So the market went wise, or believes it did. The prevailing view is that you do not need a frontend developer, you need a full stack engineer with AI leverage. Whether that belief is correct is a separate question from whether it is operative, and it is operative.

The consequence is convergence. Everybody is running mid-lane now. Backend people who have not written CSS since flexbox landed are full stack engineers. Frontend people who have never configured a reverse proxy are full stack engineers. I am a full stack engineer, and I actually am one, which does me less good than it should, because the title has been diluted by everyone reaching for the same handhold at the same time.

Titles are how the market indexes people. When a title stops discriminating, the filtering moves somewhere else. It moved to the resume screen.

Recruitment got upended in both directions

Candidates started using LLMs to apply faster and tailor harder. Volume per opening went up. Employers responded by tightening automated filters. Candidates responded by optimising against the filters. That loop is now the dominant feature of technical hiring, and nobody involved is measuring quality at either end.

What comes out the other side is a process where an AI reads your CV and applies rules that the person who commissioned it could not articulate and cannot audit. Ten years of professional experience with checkable references can score below a bachelor's degree, because a degree is a clean boolean and a career is prose. The most common outcome is the "due to high demand" template, sent before a human ever opened the file.

The interview stage has picked up its own version. Recruiters increasingly run AI-generated technical questions against candidates, then evaluate the answers against an AI-generated rubric, frequently without the domain knowledge to judge either. You can give a correct, experienced, context-aware answer and lose the round because it did not match what the model expected. Being right is no longer sufficient. Being aligned with somebody else's model output is the new bar, and it is a bar nobody can see.

Vendors in this space advertise 85 to 95% accuracy at identifying qualified candidates. Those are vendor numbers, unaudited, on a task with no agreed ground truth. The confidence is the product, which is a sentence I have now written about three separate industries in the same series.

The juniors

The best evidence on this is not in the Portuguese report. It is the Stanford Digital Economy Lab work by Brynjolfsson, Chandar, and Chen, built on ADP payroll records rather than surveys.³ Workers aged 22 to 25 in the most AI-exposed occupations, software development explicitly among them, show a 13% relative employment decline since generative AI came into wide use. Employment for experienced workers in the same occupations held steady.

The mechanism they identify is the important part. This is not a layoff wave. It is a hiring collapse. The jobs are not being cut, they are simply never being posted, which makes the damage invisible in every metric that tracks separations.

Portugal's 2% junior candidate pool is the same phenomenon showing up in a much weaker dataset. The report attributes it to AI making seniors more productive, without measuring AI anywhere in the analysis. The attribution is unsupported in that document and correct anyway.

Where the AI market cannot see itself

Now the part the discourse keeps skipping.

Google's DORA programme is the most rigorous ongoing measurement of software delivery we have. Their 2025 report found that 90% of technology professionals now use AI at work, and that AI adoption is positively associated with delivery throughput. That is a reversal from 2024, when the same programme found AI associated with reduced throughput. Good news, honestly reported.

The finding that did not reverse is the one nobody quotes. AI adoption continues to show a negative relationship with software delivery stability.⁴ More throughput, less stability, in the same dataset, in the same year. DORA's own framing is that AI is an amplifier: it accelerates the work and exposes whatever was already weak downstream.

That is the whole argument in one sentence from a source nobody can dismiss as a hater.

MIT's Project NANDA reported in 2025 that 95% of enterprise generative AI pilots produced no measurable P&L impact, across 52 executive interviews, 153 leader surveys, and 300 public deployments.⁵ The methodology has been fairly criticised and the number gets quoted more confidently than it deserves. Even discounted heavily, it points at a gap between adoption and demonstrated return that eighteen months of enthusiasm has not closed.

And METR's 2025 study remains the most uncomfortable result in the field: experienced developers working in repositories they already knew were 19% slower with AI assistance, while reporting that they felt roughly 20% faster. The perception gap ran to nearly 40 percentage points against the people best positioned to know.

Underneath all of it sits the thing that has not changed since I wrote Part II. The system is stochastic. Its memory is bounded. Apple's GSM-Symbolic work showed performance collapsing by up to 65% on the addition of a single irrelevant clause to a maths problem, which is not a bug to be patched but a description of what the artefact is.

You can harness it well. You can write beautiful tests. You can loop cleanly, keep your context lean, and understand system design and software architecture at a level most of the field does not. You will still catch it producing things you did not ask for, and the probability of that rises with the complexity of the request, because complexity is exactly what forces you to depend on context the model cannot reliably hold.

There is a genre of post right now asserting that engineers who have not internalised the tooling are NGMI. Some of that is fair. Most of it is peacocking, and the people posting it are hitting the same error rates as everyone else. They just do not post those.

The honest question is not whether LLMs can speed up production. They can, and mine do. The question is how many organisations can demonstrate sustained, measurable improvement since adoption, on numbers they would let an auditor see. That list is much shorter than the adoption figures imply, and the gap between those two lists is where the next few years of pain lives.

Meanwhile the failures keep surfacing. Reputable organisations shipping AI-authored bugs to production, published work with fabricated citations, applications falling over under conditions nobody modelled because nobody with context reviewed the change. The volume of quality quietly traded for speed in this period is going to keep researchers busy for a very long time.

My actual position

I am sceptical, and the scepticism is earned rather than performed.

Has AI increased my productivity? Yes. Do I use it every day? Yes. Are my numbers good? Yes.

I also see what happens when this scales. I see commitments made in production environments that should not have been made, on timelines that only work if nothing needs a second look. Slow used to mean something. It meant incrementally reviewed by someone with authority and context. AI review can happen without sufficient context, and the context deficit grows with the project, which means the review gets less trustworthy exactly as the stakes get higher.

The practical consequence for anyone in my position is that an AI-powered full stack engineer in 2026 has to become more DevOps, not less, purely to stay credible. If you are going to move that fast, you own the pipeline, the observability, the rollback, and the blast radius. You have to be able to see what you broke before someone else does.

And you cannot be a yes-man. The cost of saying yes has never been more chaotic, because yes now means yes at machine speed with a review layer that may not have understood the change.

But so do we

None of this is a reason to stop.

Keep building. Keep shipping, because shipped work is now the most legible proof that you can provide value, and it routes around a hiring process that has stopped being able to read a CV. Keep strengthening whatever your actual goals are, separate from what the market panic says they should be.

Keep using the tooling. You will be required to, unless you are in a genuinely niche or genuinely senior position, and the people who use it with discipline are still eating better than the people who refuse it on principle.

If you have any say in hiring, take the junior. It is a worse deal on a one-year horizon and a better one on a ten-year horizon, because the scarce input in 2036 will be people who can tell when the machine is wrong, and that faculty is only built by shipping something bad and paying for it under your own name. Whoever trains them owns that supply. Everyone else rents it.

Reality will come knocking eventually. Inference will get deeper, the resource spillage behind an inflated belief in these tools will show up on somebody's balance sheet, and companies will grow wise the way companies always do, which is slowly and expensively and only after it is embarrassing.

Until then, hunker in. Provide value. Learn. Persist.

The horrors persist. So do we.

Over and out.

Previously in this series:

1. Part I: Genealogy

2. Part II: The Confidence Game

3. Part III: Where the Leverage Is Now