Self-Improving AI and Jobs: What Sakana’s MASS Results Really Show

Sakana AI and UC Berkeley researchers report two rounds of MASS, a system that learns from work produced and selected by its own AI teams. Their October 2026 account describes gains on research benchmarks alongside uneven results on software benchmarks. [1]

I take this seriously. My judgment is that better outputs, a better self-improvement process and an economical substitute for paid work are three separate achievements. Career forecasts need evidence connecting them. Otherwise, a promising experiment becomes an unsupported verdict on people's working lives.

What improved, and where

MASS, short for Multi-Agent Self-Supervision, uses a shared model to perform tasks, evaluate results and revise how agents coordinate. It then trains that model on selected records of the team's work. The updated model takes over all three roles for another round. [1]

After two rounds, the researchers report score per output token at 1.2–1.6 times the base model's level across four research benchmarks. The corresponding ratios were 0.96 on Terminal-Bench 2.0 and 0.94 on SWE-bench Verified. [1]

These ratios relate benchmark scores to generated text. They cannot be read as a percentage of workers made unnecessary. The lower software ratios also do not, by themselves, establish lower absolute scores. [1]

I see evidence of progress with uneven reach. I would neither extrapolate the research gains to every office task nor treat the software results as protection for developers.

Improving the improvement process

The researchers also examined whether better ways of organizing work carried over to the original model. With that model executing both sets of plans, plans discovered using the first updated model won 64% of comparisons against plans discovered using the base model. [1]

That is an interesting result: an earlier model could benefit from later planning. But during the search that produced those plans, the task solver, evaluator and optimizer changed together. The experiment does not isolate which role caused the improvement. [1]

Other AI models supplied the outside judgments. Those comparisons do not establish objective scientific correctness, and the authors report that successive candidate workflows could perform worse. [1] My reading is that two rounds give us something worth investigating, with sustained acceleration still unproven.

The missing bill

In a separate training comparison, the authors report that a model trained on examples from AI teams achieved a higher win rate against the base model while processing 31% fewer supervised training tokens than a model trained on examples from individual agents. Those counts exclude data generation, workflow search, evaluation and other inputs. [1]

I would judge economic substitution using the total cost of an accepted result: computing, preparation, checks, failed attempts and whatever human work remains. A cheap generated report could be expensive to verify. Conversely, a costly training run could become economical if its benefits were spread across enough later assignments.

Neither outcome is measured here. The research account provides no comparison with paid workers and no measured employment effects. [1] That missing evidence limits the career conclusions we can draw; it does not erase the technical achievement.

I would also distinguish an employer's reason to experiment from its reason to eliminate a position. A promising trial could justify further testing. A staffing decision deserves evidence that the required work still gets completed under ordinary operating conditions. Eliminating jobs solely on these benchmarks would, in my view, be reckless.

Jobs can be threatened before the loop accelerates

I think there is a dangerous comfort in treating open-ended recursive improvement as the event workers must wait for. An AI system could stop improving and still become good enough to automate a valuable part of a job. A technical plateau would offer little reassurance to someone whose work it already handles.

Consider a hypothetical research team that uses AI to produce acceptable preliminary analyses with less staff time. Management could commission more investigations with the existing team. It could also decide that the same workload needs fewer employees. The technical gain alone would not determine which choice wins.

This is why I regard demand as part of the employment question. More affordable research might expand what an organization undertakes. It might instead let it reduce spending on people. MASS does not measure either response. [1]

My concern is that limited, uneven automation could still produce painful losses. Proof of an intelligence explosion is an unnecessarily high threshold for taking that possibility seriously. Equally, a successful improvement loop would supply no automatic timetable for eliminating an occupation.

What would change my career outlook

For career planning, I would give more weight to repeated evidence about actual assignments than to a universal countdown toward replacement. A developer should care whether systems can complete the kinds of changes employers need, with review effort and failures included. A researcher should care whether their outputs survive the checks required for use.

This suggests a useful learning priority: become better at designing fair comparisons. Practice defining an acceptable result, testing unfamiliar cases and accounting for the effort needed to fix mistakes. I consider that useful technical judgment, although I would not promise that evaluation work will employ everyone displaced elsewhere.

If reliable completion becomes cheaper across a broad share of an occupation's work and employers cut hiring accordingly, I would become substantially more pessimistic. If lower costs support more projects and continued recruitment, I would become more optimistic. Those are conditional judgments; this experiment cannot settle them.

For someone deciding whether to leave a field, those differences matter more than the label attached to a research paper. I would not abandon a career on the strength of these benchmarks. I would take sustained, well-measured substitution in that field very seriously - even if the system never improved itself again.