Untitled topic

Every important system you use today - LinkedIn job matching, Spotify's recommendations, CapitalOne's fraud detection, Youtube's video ranking, DoorDash's food search - was made good by a very small number of very expensive humans.

Staff ML engineers. Applied scientists. Senior data engineers. Group product managers. Whatever the title, we have all been doing the same job: sit inside a system, understand it deeply, run hundreds of experiments, and compound its performance over years. Tiktok's, Spotify's & Instagram's feeds did not become good on its own. It took a team of specialists iterating on ranking models, exploration strategies, and multi-objective optimization across years of A/B tests. his. Making the systems you engage with better.

These people are the reason your feed is good, your fraud gets caught, your search results make sense, and your notifications are relevant.

Here is the problem: there are maybe a few thousand of such people on the planet. But every company in the world needs one. The subset of humans who can actually make production systems measurably better is far smaller that what we want it to be to power the industry.

This means most systems in the world are not getting better. Not as much as they should.

Your bank fraud system. Your insurance company risk model. The food delivery app search ranking. The e-commerce recommendation engine. These are all systems that would benefit enormously from the same kind of iterative, experiment-driven science that made Tiktok, Youtube, Spotify or Netflix work. But the companies that own them cannot hire the people who know how to do it. Not at the salaries required. Not in the timelines required. Often not at all.

This has been true for 20 years. It is the reason the gap between the top companies who. have this talent and the others who don't keeps widening. The constraint was never compute or data. It was the humans.

That constraint is breaking now. Almost.

The shift from human-led to platform-led applied science means the operational overhead that used to consume 80% of a specialist time - data cleaning, feature engineering, pipeline debugging, experiment infrastructure - is collapsing. Agentic tooling is compressing the iteration loop from quarters to days. The experimentation cycle that used to require a dedicated team of engineers and scientists supporting one system can now be run by a single person across multiple systems in parallel.

Internally, we have started calling this person the Octopus Scientist.

Not because they are smarter than the specialists who came before. But because the tooling now lets one person touch five systems instead of one. Run A/B tests across three companies in a week instead of one company in a quarter. Do twelve months of applied ML in thirty days. Ouch.

This is not about the role. It is about what happens to all the systems that were previously starved of talent.

The food delivery company that could never hire a Spotify-caliber search ranking team can now have an octopus scientist improve their system in 30 days and move on. The insurance company whose fraud model has not been meaningfully updated in three years can get the same quality of experimentation that a FAANG company runs internally. The marketplace whose recommendation engine was built once and never iterated on can finally start compounding.

I gave a keynote talk on this at the Agent4IR Workshop at KDD last year. The core argument: the best ML talent of the last decade was locked inside a handful of companies, improving a handful of systems. The next decade is about that same caliber of science reaching the thousands of systems that never had access to it.

The Octopus Scientist is what happens when the tools finally catch up to the scale of the problem.

There are millions of data driven systems in the world that should be getting better and are not. That is about to change.