An underexplored alternative to donation is hiring people from low-income contexts to do paid work on meaningful problems.
Here’s a rough estimate of “happy” hourly rates for remote intellectual manual labor (data labeling, checking, summarization, interpretability grunt work), in USD:
Estimated Happy Rates ($/h)
Country
p25
p50
p75
Brazil
2.35
3.35
4.69
Argentina
2.11
3.02
4.23
Colombia
3.93
5.61
7.85
Peru
2.38
3.40
4.75
Chile
4.75
6.79
9.50
Bolivia
1.70
2.45
3.40
Paraguay
2.05
2.95
4.10
Ecuador
2.70
3.85
5.40
Mexico
2.90
4.10
5.80
Nigeria
0.70
0.99
1.39
Ghana
0.63
0.90
1.26
Kenya
1.24
1.77
2.48
Uganda
0.55
0.80
1.15
Tanzania
0.60
0.88
1.25
South Africa
2.07
2.96
4.14
Egypt
1.46
2.09
2.92
Morocco
1.85
2.65
3.70
Tunisia
1.95
2.80
3.90
India
0.95
1.40
2.10
Bangladesh
0.55
0.80
1.20
Pakistan
0.65
0.95
1.40
Sri Lanka
0.85
1.25
1.85
Vietnam
1.35
1.95
2.80
Philippines
1.60
2.30
3.30
Indonesia
1.10
1.60
2.40
Thailand
2.10
3.00
4.30
Malaysia
2.60
3.70
5.30
Nepal
0.60
0.88
1.30
Cambodia
0.75
1.10
1.60
Mongolia
1.10
1.60
2.30
Fiji
2.40
3.40
4.90
Samoa
2.10
3.00
4.30
Tonga
2.20
3.10
4.50
There exists a very large supply of people who are both willing and happy to do careful cognitive work at rates that are trivial by rich-country standards, if the work is structured and paid.
Some reasons this possibility can be quite good and interesting:
It allows money to be converted into actual work on impactful tasks, even if that work is initially “intellectual manual labor” (labeling, checking, summarizing, auditing, interpretability grunt work, etc.).
It treats people as participants rather than recipients. Receiving payment for work tends to be more humanizing than receiving aid, because it encodes agency, skill, reciprocity, and contribution.
It onboards people into the global intellectual labor market: deadlines, quality standards, tooling, communication norms. Those skills compound and transfer.
It can operate without heavy intermediary organizations, which reduces overhead and incentive distortion and keeps the causal chain legible: money → work → output → learning.
A lot of important research and analysis is not bottlenecked on genius so much as on coordination, paradigms, and tooling. Once those exist, large amounts of careful human attention can be usefully applied in parallel.
My usual joke is “GPT-2 has 12 attention heads per layer and 48 layers. If we had 50 interns and gave them each a different attention head every day, we’d have an intern-day of analysis of each attention head in 11 days.”
This is bottlenecked on various things:
having a good operationalization of what it means to interpret an attention head, and having some way to do quality analysis of explanations produced by the interns. This could also be phrased as “having more of a paradigm for interpretability work”.
having organizational structures that would make this work
building various interpretability tools to make it so that it’s relatively easy to do this work
I think there are real downsides of mixing unrelated goals (in this case: improving livelihoods/skills for educated people in LMICs, and getting work done).
remote work requires people who already have computer access, reliable internet, professional skills, and proficient English (or whatever language you need). So these are people who are already relatively well-off in their setting.
management capacity is often a bottleneck, so rather than onboarding people to things like deadlines and quality standards, for the sake of getting the work done efficiently you might rather pay a higher rate to get someone who doesn’t need as much hand-holding. (Maybe this isn’t relevant if the work you want done isn’t itself aiming at a positive impact, and you’re ok with your widget business running less efficiently in order to offer a jobs program.)
If you have needs that can be met just as well by remote workers in LMICs, seems great! But I wouldn’t start with the premise that this is your best option for improving the world.
interpretation: this is a very conservative lower bound on people who could plausibly do high-quality remote cognitive work using tools like chatgpt (incl. translation). this is not a hiring claim; it’s an order-of-magnitude sanity check.
hacky fermi table
country
population
internet users
final pool (÷8000)
brazil
203,000,000
170,520,000
21,315
argentina
46,000,000
41,400,000
5,175
colombia
52,000,000
40,040,000
5,005
peru
34,000,000
24,480,000
3,060
chile
19,500,000
17,940,000
2,243
bolivia
12,400,000
7,440,000
930
paraguay
7,500,000
5,850,000
731
ecuador
18,300,000
13,725,000
1,716
mexico
129,000,000
96,750,000
12,094
nigeria
227,000,000
88,530,000
11,066
ghana
34,000,000
18,020,000
2,253
kenya
55,000,000
23,650,000
2,956
uganda
49,000,000
14,210,000
1,776
tanzania
67,000,000
20,100,000
2,513
south africa
62,000,000
44,640,000
5,580
egypt
112,000,000
80,640,000
10,080
morocco
37,000,000
31,080,000
3,885
tunisia
12,300,000
8,733,000
1,092
india
1,430,000,000
800,800,000
100,100
bangladesh
173,000,000
70,930,000
8,866
pakistan
241,000,000
86,760,000
10,845
sri lanka
22,000,000
11,880,000
1,485
vietnam
101,000,000
75,750,000
9,469
philippines
114,000,000
83,220,000
10,403
indonesia
277,000,000
182,820,000
22,853
thailand
71,000,000
60,350,000
7,544
malaysia
34,000,000
32,980,000
4,123
nepal
30,500,000
13,420,000
1,678
cambodia
17,000,000
9,520,000
1,190
mongolia
3,500,000
2,905,000
363
fiji
930,000
697,500
87
samoa
225,000
157,500
20
tonga
107,000
74,900
9
key takeaway:
even after filtering to internet users only and then applying an extremely harsh 95%×95%×95% filter, many countries still have thousands to tens of thousands of plausible high-quality contributors. at global scale, talent supply is not the bottleneck; coordination, tooling, and trust are.
“”″
(I know this estimation relies on some independence assumptions. Regardless, it is meant to be illustrative, not authoritative.)
An underexplored alternative to donation is hiring people from low-income contexts to do paid work on meaningful problems.
Here’s a rough estimate of “happy” hourly rates for remote intellectual manual labor (data labeling, checking, summarization, interpretability grunt work), in USD:
Estimated Happy Rates ($/h)
There exists a very large supply of people who are both willing and happy to do careful cognitive work at rates that are trivial by rich-country standards, if the work is structured and paid.
Some reasons this possibility can be quite good and interesting:
It allows money to be converted into actual work on impactful tasks, even if that work is initially “intellectual manual labor” (labeling, checking, summarizing, auditing, interpretability grunt work, etc.).
It treats people as participants rather than recipients. Receiving payment for work tends to be more humanizing than receiving aid, because it encodes agency, skill, reciprocity, and contribution.
It onboards people into the global intellectual labor market: deadlines, quality standards, tooling, communication norms. Those skills compound and transfer.
It can operate without heavy intermediary organizations, which reduces overhead and incentive distortion and keeps the causal chain legible: money → work → output → learning.
A lot of important research and analysis is not bottlenecked on genius so much as on coordination, paradigms, and tooling. Once those exist, large amounts of careful human attention can be usefully applied in parallel.
— Buck’s comment on “How might a herd of interns help with AI or biosecurity research tasks/questions?”, EA Forum
I think there are real downsides of mixing unrelated goals (in this case: improving livelihoods/skills for educated people in LMICs, and getting work done).
remote work requires people who already have computer access, reliable internet, professional skills, and proficient English (or whatever language you need). So these are people who are already relatively well-off in their setting.
management capacity is often a bottleneck, so rather than onboarding people to things like deadlines and quality standards, for the sake of getting the work done efficiently you might rather pay a higher rate to get someone who doesn’t need as much hand-holding. (Maybe this isn’t relevant if the work you want done isn’t itself aiming at a positive impact, and you’re ok with your widget business running less efficiently in order to offer a jobs program.)
If you have needs that can be met just as well by remote workers in LMICs, seems great! But I wouldn’t start with the premise that this is your best option for improving the world.
I agree this cannot replace donation-based interventions! It is still feels potentially underrated and underconsidered.
I do agree that management and structure are the hardest parts. I do imagine many EA orgs have solved harder problems in the past.
I think automatic dubbing services have become good enough to make English fluency not be a hard requirement anymore for many potential jobs.
Here is a super hacky fermi-gpt estimate of a headcount of potentially hireable global workers:
“”″
hacky fermi estimate — internet users → elite tail
definitions (clean + explicit):
population: total population (≈2024–2025)
internet users: people using the internet (any device)
final pool (÷8000): internet users filtered by three independent 95th-percentile criteria
high cognitive ability (≈95th percentile)
hardworking (≈95th percentile)
ethical / trustworthy (≈95th percentile)
combined ⇒ (1 / (20×20×20) ≈ 1 / 8000)
interpretation: this is a very conservative lower bound on people who could plausibly do high-quality remote cognitive work using tools like chatgpt (incl. translation). this is not a hiring claim; it’s an order-of-magnitude sanity check.
hacky fermi table
key takeaway:
even after filtering to internet users only and then applying an extremely harsh 95%×95%×95% filter, many countries still have thousands to tens of thousands of plausible high-quality contributors. at global scale, talent supply is not the bottleneck; coordination, tooling, and trust are.
“”″
(I know this estimation relies on some independence assumptions. Regardless, it is meant to be illustrative, not authoritative.)
Somewhat related: https://x.com/i/status/2021218105154756893