TL;DR. First-order learning is acquiring a skill. Second-order learning is using what you already know to make the next one cheaper, and it is where learning compounds. The research on transfer is sobering about the grand version of this idea: a set of meta-analyses on chess, music and working memory training concluded that far transfer “rarely occurs”. So we measured the ordinary version instead. Using the skill and knowledge ratings for 910 occupations in the ONET 31.0 database, we calculated how much of what one occupation requires another occupation already supplies. Four findings. First, moving to one of the five occupations ONET lists as most closely related to yours requires 46% less new learning than a random move (20.7 against 38.6 points of level shortfall). Second, the ten essential skills, such as reading comprehension, writing, speaking, mathematics and critical thinking, travel almost anywhere: even a random move keeps a median 93.3% of the required level. Third, domain knowledge is what does not travel: 32.9% of close moves arrive with at least 90% of the required knowledge already in place, against 10.5% of random moves. Fourth, knowledge makes up about 58% of the learning load at every distance. The rule that follows: choose the next thing by adjacency, carry your general skills across deliberately, and spend your learning budget on the knowledge gap.
Why “learning how to learn” is not enough
The appealing version of second-order learning says that some activity trains the mind in general, so that everything after it becomes easier. The evidence for that version is weak. In a review of three meta-analyses, Giovanni Sala and Fernand Gobet examined chess instruction, music instruction and working memory training and found small to moderate effects on children’s cognitive and academic skills, but effects that shrank as the quality of the experimental design rose, for example when active control groups were used. Their conclusion was that “far transfer of learning rarely occurs”.
That does not mean nothing transfers. Susan Barnett and Stephen Ceci argued in 2002 that the century-long argument about transfer had stalled because studies compared “apples and oranges”. They proposed nine dimensions along which transfer can differ, three describing what is transferred and six describing the context, among them the knowledge domain, the physical, temporal, functional and social context, and the modality. They also distinguished the specificity of what is learned: a specific procedure such as carrying a one when two digits sum past nine, a more general principle such as breaking a problem into subproblems, and meta-cognitive skills such as looking at a problem from multiple angles.
For anyone planning what to learn next, that framing raises a practical question the laboratory literature cannot answer: in real work, which parts of what you know already cover the next thing, and which parts must be learned from scratch? That is a measurement problem, and there is a public dataset built for it.
How we measured carry-over
ONET, the US Department of Labor’s occupational database, rates every occupation on how much of each skill and knowledge area it requires. Version 31.0 splits skills into two groups: 10 essential skills (reading comprehension, active listening, writing, speaking, mathematics, science, critical thinking, active learning, learning strategies and monitoring) and 25 transferable skills (such as coordination, persuasion, negotiation, complex problem solving, programming, systems analysis, time management and management of financial resources). It also rates 33 knowledge* areas, from economics and accounting to psychology, law and government, and computers and electronics. Each rating includes a level on a 0 to 7 scale.
O*NET also lists up to 20 related occupations for each occupation, in three tiers. Across the 910 occupations with complete ratings, we used 4,428 closest pairs (the top tier, about five per occupation), 4,472 next-closest pairs, 8,876 supplemental pairs, and, as a baseline for a far move, 19,980 random pairs.
For each pair we asked one question: of the level the target occupation requires in each area, how much does the source occupation already require? Summed across areas, that gives a carry-over percentage. The shortfall, the level points the target requires beyond the source, gives the learning load.
Finding 1: general skills travel, knowledge does not
Table 1. How much of the next occupation’s requirements you already carry (CEOtudent calculation from O*NET 31.0)
| Distance of the move | Pairs | Essential skills: median carry-over | Transferable skills: median carry-over | Knowledge: median carry-over | Moves with at least 90% of essential skills covered | Moves with at least 90% of transferable skills covered | Moves with at least 90% of knowledge covered |
|---|---|---|---|---|---|---|---|
| Closest related (top tier) | 4,428 | 95.6% | 90.3% | 85.3% | 75.3% | 51.3% | 32.9% |
| Next closest (second tier) | 4,472 | 95.2% | 89.5% | 83.6% | 72.6% | 47.9% | 27.3% |
| Supplemental related | 8,876 | 94.8% | 87.5% | 82.5% | 66.7% | 41.3% | 26.5% |
| Random occupation | 19,980 | 93.3% | 80.1% | 70.6% | 57.2% | 27.9% | 10.5% |
Carry-over for a pair = sum over all areas of the smaller of the two occupations’ required levels, divided by the sum of the target occupation’s required levels. Level ratings flagged “not relevant” are counted as zero. Random pairs drawn with a fixed seed; source and target always differ.
Read the table from left to right and the pattern is plain. Essential skills barely care about distance: from the closest move to a random one, median carry-over falls only 2.3 points, from 95.6% to 93.3%. Knowledge falls 14.7 points, from 85.3% to 70.6%. The share of moves where the knowledge is essentially already in place collapses by more than two-thirds, from 32.9% to 10.5%.
This is the Barnett and Ceci distinction showing up in labour-market data. The general layer, reading, writing, speaking, reasoning, learning, is portable almost everywhere. The specific layer, knowledge of a domain, is portable only nearby.
Finding 2: adjacency roughly halves the learning load
Table 2. New learning required by a move, in level points of shortfall (CEOtudent calculation from O*NET 31.0)
| Distance of the move | Essential skills | Transferable skills | Knowledge | Total | Share of the load that is knowledge |
|---|---|---|---|---|---|
| Closest related (top tier) | 2.10 | 6.51 | 12.07 | 20.68 | 58.4% |
| Next closest (second tier) | 2.29 | 7.03 | 13.27 | 22.58 | 58.8% |
| Supplemental related | 2.86 | 8.28 | 14.53 | 25.68 | 56.6% |
| Random occupation | 4.58 | 11.79 | 22.18 | 38.55 | 57.5% |
Shortfall = sum over all areas of the target’s required level minus the source’s, where positive. Means across pairs.
Two things follow. A move to one of your closest related occupations requires 46.4% less new learning than a random move, 20.68 points against 38.55. And the composition of what you must learn hardly changes with distance: about 58% of it is domain knowledge at every tier. Adjacency does not change what kind of learning the next thing demands. It changes how much of it there is.
Within the essential skills, the closest moves most often ask for more science and mathematics: 14.1% and 9.3% of closest-tier moves require at least one extra level point in them. Next come learning strategies and active learning, at 5.6% and 4.0%. Even the skill of learning itself sometimes needs an upgrade for the next role.
Finding 3: direction matters
Carry-over is not symmetric. What you already know can cover most of a target’s needs one way and much less the other way.
Table 3. Illustrative moves (CEOtudent calculation from ONET 31.0; pairs chosen as examples, whether or not ONET lists them as related)
| From | To | Essential skills carried | Transferable skills carried | Knowledge carried | Largest knowledge gaps to close |
|---|---|---|---|---|---|
| Software Developers | Data Scientists | 88.8% | 98.1% | 77.9% | Physics, Sociology and Anthropology, Geography |
| Data Scientists | Software Developers | 99.3% | 68.5% | 82.8% | Customer and Personal Service, Telecommunications, Transportation |
| Management Analysts | Data Scientists | 90.4% | 92.2% | 87.4% | Physics, Computers and Electronics, Engineering and Technology |
| Writers and Authors | Market Research Analysts and Marketing Specialists | 80.5% | 69.6% | 90.0% | Mathematics, Customer and Personal Service, Sales and Marketing |
| Graphic Designers | Web and Digital Interface Designers | 91.7% | 83.9% | 92.0% | English Language, Mathematics, Computers and Electronics |
| Secondary School Teachers | Training and Development Specialists | 96.3% | 82.4% | 77.2% | Sales and Marketing, Personnel and Human Resources, Production and Processing |
| Registered Nurses | Medical and Health Services Managers | 98.7% | 79.9% | 68.3% | Economics and Accounting, Administration and Management, Personnel and Human Resources |
| Customer Service Representatives | Sales Representatives of Services | 96.1% | 87.7% | 73.3% | Law and Government, English Language, Education and Training |
The software pair shows the asymmetry most clearly. A software developer already carries 98.1% of the transferable skills a data scientist needs; a data scientist carries only 68.5% of a software developer’s. The nurse-to-manager move shows the general pattern: almost all the essential skills carry over (98.7%), but less than 70% of the knowledge does, and the gaps are exactly the ones you would expect, economics and accounting, administration and management, personnel. That list is the learning plan.
Some gaps look odd at first sight, such as physics and geography for data science. They are what O*NET’s ratings say the target occupation requires as a whole, which is broader than any one person’s version of the job. Treat the gap list as a checklist to confirm against the actual roles you are targeting, not as a syllabus.
The Second-Order Learning Loop
The data point to a method. We call it the Second-Order Learning Loop.
Table 4. The Second-Order Learning Loop (CEOtudent editorial framework)
| Step | What you do | Why, from the evidence above |
|---|---|---|
| 1. Inventory what you carry | List your skills and knowledge areas and rate your level honestly, using the O*NET areas as a checklist; our skill audit gives a worksheet | You cannot reuse what you have not named |
| 2. Choose by adjacency | Shortlist next roles or skills from the occupations closest to yours, and check both directions | The closest tier requires 46% less new learning than a random move; direction can change carry-over by 30 points |
| 3. Diff, do not restart | For the target, list only the areas where its required level exceeds yours | About 58% of the gap is knowledge; the essential skills are usually already there |
| 4. Carry the general layer across on purpose | Before starting, write down how your existing reading, writing, reasoning and learning habits will be applied in the new domain | Far transfer rarely happens by itself; the general layer transfers when you apply it deliberately |
| 5. Learn the knowledge gap with proven methods | Use retrieval practice and spacing on the new domain’s core concepts; see our ranking of 12 study techniques by research strength | Knowledge is the bulk of the load and the part techniques most reliably help with |
| 6. Map new concepts onto old ones | For each new concept, write the nearest thing you already understand and where the analogy breaks | Principles transfer further than procedures; naming the principle is how you reuse it |
| 7. Log what transferred | After each project, record which old skills did the work and which gaps surprised you | This is your own transfer data; it makes the next choice of “next thing” better |
Steps 4 and 6 are where the “second-order” part happens. The laboratory evidence says a general capacity does not quietly upgrade everything else. The occupational data say the general layer is nonetheless present in almost every move. The gap between those two facts is closed by intention: you carry a principle into a new domain by naming it and testing it there, not by hoping it shows up.
What this does not tell you
Occupations are not people. O*NET rates what an occupation requires on average. Your own profile may be stronger or weaker than your job title’s, so the percentages are a starting point for your own inventory, not a verdict on it.
Levels are not hours. A shortfall of one level point in knowledge can mean a weekend or a degree. For time estimates, see our time-to-competence data for 20 skills.
Coverage is not competence in context. Barnett and Ceci’s context dimensions, the physical, social and functional setting, are not in these ratings. Knowing accounting as a nurse-turned-manager is not the same as having used it under a budget deadline.
The data are American. O*NET describes US occupations. The structure, general skills portable and domain knowledge local, is likely to hold elsewhere, but the specific percentages are US figures.
The CEO and the student
The CEO move is capital allocation: you already own a large asset, the general skills that carry to almost any next role, and a set of specific assets that carry only nearby. Choose the next investment where your existing assets cover most of the requirement, and put new capital only into the gap. The student move is to treat each transition as a learning experiment you log, so that your second transition is cheaper than your first and your third cheaper still.
For the general layer itself, our piece on the meta-skills that make every other skill easier covers what to build; for designing the sprint into a new domain, see fast upskilling protocols; and for the question of what actually predicts learning speed, see why learning styles are a myth.
FAQ
What is second-order learning?
First-order learning is acquiring a skill or body of knowledge. Second-order learning is using what you have already learned to make the next thing faster to learn: choosing targets that overlap with what you know, reusing general skills deliberately, and learning only the gap.
Does learning one skill make you better at everything?
Not in the broad sense. Meta-analyses of chess, music and working memory training found that apparent general benefits shrank in better-designed studies, and the authors concluded that far transfer rarely occurs. What does carry over widely are general skills such as reading, writing, speaking and reasoning, which O*NET shows are required at similar levels across most occupations.
How much faster is it to learn something adjacent?
In our O*NET analysis, moving to one of the closest related occupations requires 46.4% less new learning, measured in level points of shortfall, than moving to a random occupation. That is a measure of how much there is to learn, not a guarantee of time saved.
What usually has to be learned from scratch?
Domain knowledge. It makes up roughly 58% of the learning load for a move at every distance, and it is the part whose carry-over drops most as moves get further away.
How do I find the occupations closest to mine?
O*NET publishes related occupations for each occupation, in tiers from closest to supplemental. Use them as a shortlist, then check the carry-over in both directions, because it is often asymmetric.
Sources
- National Center for ONET Development, ONET 31.0 Database, US Department of Labor. Essential Skills, Transferable Skills, Knowledge (importance and level ratings), Related Occupations (relatedness tiers) and Occupation Data files for 910 occupations; all carry-over, shortfall and example figures in this article computed from these files.
- Sala, G. and Gobet, F. (2017). Does Far Transfer Exist? Negative Evidence From Chess, Music, and Working Memory Training. Current Directions in Psychological Science, 26(6), 515-520. Three meta-analyses; effect sizes inversely related to design quality; conclusion that far transfer of learning rarely occurs.
- Barnett, S. M. and Ceci, S. J. (2002). When and Where Do We Apply What We Learn? A Taxonomy for Far Transfer. Psychological Bulletin, 128(4), 612-637. The nine-dimension framework of transfer content and context; the distinction between specific procedures, general principles and meta-cognitive skills.
Tables 1 to 3 are CEOtudent calculations from ONET 31.0; the carry-over and shortfall measures and the random baseline are our analytical choices. Table 4 is a CEOtudent editorial framework.*
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
