Daniel Björkegren points out (h/t Deena Mousa's newsletter) that marginal returns to intelligence from advanced AI will be lower in LMICs due to scarcer AI complements, lower digital legibility, and smaller knowledge sectors, so AI that augments knowledge workers is likely to disproportionately benefit richer countries:

The economic implications of this transformation can be characterized by the marginal returns to intelligence (Amodei 2024): how much can we improve economic outcomes as we better generate ideas, process data, and apply knowledge? Intelligence allows us to solve scientific problems, design better products, better anticipate demand, and ensure the right quantities are stocked in the right places. Low-income countries will benefit from innovations developed in rich ones. But within many LMICs, the complements to advanced AI are scarcer, including data centers, reliable electricity, and digital records, as well as experienced knowledge workers. Data centers can be located in countries that already have good infrastructure (‘the cloud’) and accessed remotely. But LMICs are less digitally legible: AI will be less able to understand and act in markets, firms, homes, clinics, and schools that do not record data in structured forms. Overall, we would expect LMICs to be at a disadvantage in integrating advanced AI (Korinek and Stiglitz 2021).

A crucial distinction is that LMICs have much smaller knowledge sectors. LMICs employ fewer than 10% of workers in skilled knowledge work, like managers, technicians, and professionals, relative to 41% in high income countries (Silva 2026). Current AI tools require substantial human guidance. So, firms in rich economies are pursuing a grafting strategy: existing knowledge workers are being asked to integrate AI into their roles, starting from producing slides and emails, and scaling to more sophisticated tasks. In countries with smaller knowledge sectors, there are fewer workers and processes to graft AI onto. Thus a key question is whether advanced AI will mainly empower existing workers, or automate knowledge work completely. In wealthy countries, advocates concerned about jobs suggest that AI systems be designed to augment rather than automate (Acemoglu, Autor, Johnson 2026). But in low-income countries, the more urgent question may be how to provide knowledge services when few knowledge workers are available. Fully automating knowledge work could in fact augment less educated workers, who could ask AI to complete macro tasks like developing marketing strategies, rather than micro tasks like reformatting spreadsheets. However, even automated systems will likely require oversight from entrepreneurs and scientists with deep expertise, which may be sufficiently available only in wealthier countries like Brazil and India.

If AI allows LMICs to grow automated knowledge sectors, would the returns be high or low? One indicator is in wages paid to human workers. The wage returns to college education are slightly higher in lower income countries (Psacharopoulos and Patrinos 2018 and 2025), but educated people often earn higher wages abroad, and some domestic knowledge workers are working on rich countries’ knowledge problems in call centers and business process outsourcing. Lower income economies may not currently be structured to fully tap the decision making entailed in knowledge work (Engbom et al. 2025). If we tasked millions of data scientists with helping smallholder farms, the returns are unlikely to be large: agriculture is constrained elsewhere.

However, if the price of some forms of intelligence declines by orders of magnitude, it may become worth applying intelligence to problems that were never worth assigning a human to. Small manufacturers might generate nuanced designs that would have required a team of industrial engineers, and implement advertising campaigns that would have required large creative teams. Many regions have struggled to agglomerate sufficient human talent; since automated intelligence can be accessed anywhere, it could make businesses more mobile. These opportunities could more fundamentally change economic structure.

So what can LMICs do? Daniel suggests these:

The most capable AI systems currently require large-scale frontier models and large amounts of compute. Governments, firms, and NGOs will need to work with the frontier labs to ensure that the most advanced models speak local languages and understand local contexts. Ensuring that there are multiple suppliers for both models and data centers can reduce prices and risks of lock-in and geopolitical disruption (Athey and Scott Morton 2025).

Governments will also need to push to make economic activity digitally legible, from markets to clinics to schools.

It is also important to ensure that AI can be productively used. That may require training humans to be more productive users of AI, both in applying the tool and having the deeper world knowledge needed to direct it. Firms can also invest in developing AI tools that are complementary to the industrial structure of LMICs, including tools for small scale entrepreneurs who have less education, and for agriculture, like weather forecasting.

The diversity of institutional conditions in low- and middle-income countries may be a comparative advantage. Wealthy countries have evolved similar institutions around human knowledge work; tweaks may lead to local optima. In contrast, systems in low-income countries can differ greatly. Tailoring to different constraints can generate opportunities: for example, Kenyan entrepreneurs coping with unreliable network connections developed techniques to create on-device AI models that are seeing demand around the world (Fastagger). Or, also in Kenya, 90% of people resolve disputes outside the formal justice system (Kenya 2020), and just two doctors serve every 10,000 people, compared to 37 in the United States (WHO, 2022). Firms and NGOs may find creative new solutions, such as offering more efficient ways to settle disputes outside of court, or dynamic medical advice. Governments can take advantage of opportunities to design new regulation for AI, rather than retrofit regulation designed for humans. A lack of established institutions around human knowledge work could also allow harm: what happens when medical AI makes mistakes and there are limited mechanisms to address malpractice? It will take care to develop appropriate new institutions.

Relatedly, Daniel also has a great post on how the poorest use AI. A quote:

AI usage can provide a new window into the needs of the poor, analogous to Google Trends. This can help AI labs and a variety of organizations better serve these populations. We saw early examples of this among teachers in Sierra Leone, who submitted requests for not only facts and lesson plans, but also on handling reports for insurance claims and navigating interpersonal situations with students and supervisors. Another study found that one of the top uses of ChatGPT among gig workers in India and Brazil was for health queries.

What could be computed: An easy start would be to take the standard categorization of requests already reported by the labs (such as writing, technical help, or mapping to industries) and report them specifically for the subset of users in marginalized groups (defined by having cheaper devices, speaking local languages, or using from remote areas). However, these taxonomies are built around knowledge work, and may systematically undercount the ways poor people find the technology useful. Thus, it would be helpful to develop new taxonomies to understand poverty-specific needs, including particular uses within agriculture, health, navigating government bureaucracy, and business advice.

What we might find: The poor are likely to use AI differently from the wealthy: almost no software development, some use for navigating bureaucracy and social problems, more for help with homework, and less for writing assistance. Anthropic has already reported that Claude users in lower-income countries are more likely to request help with coursework. A further breakdown will help us understand if AI is being used only in the wealthiest schools or broadly, and help school systems ensure it is used in ways that support learning.

This topic became more salient to me after attending Deena's EAG talk on how LMICs should respond to AI, which feels like it should be a much bigger topic than it currently is.

Would you be comfortable sharing what opportunities you gave to, broadly speaking?

(I'm not at all an expert on any of this, please discount appropriately)

1. Agree with reasoning for directional adjustment and bounds, magnitude-wise seems a bit overcorrected? SemiAnalysis' figures roughly suggest 15M center. But you're on track to becoming correct given token efficiency trends anyhow
   1. I wish I had a more empirically-grounded sense of how token usage varies by type of task, fixing task duration at 8 hours for a human professional (that you'd pay $400/day for, say). My guess from comparing model vs human jaggedness (e.g. this) is that leadership-level / early-employee / entrepreneurial / high-context / taste-heavy work would require way more tokens to get 8 hours of work done than the routine analyst-type / junior SWE etc tasks typical of benchmarks
2. My sense is global average cost per token will go down a lot due to the following, but very unclear as to the mix
   1. a key driver of inference demand going forward being very cache tokens-heavy agentic workflows
   2. a rising share of demand being satisficing not maximising w.r.t output quality for ever-growing task share (e.g. plan with Opus -> code with Sonnet or even DeepSeek models at 1-2 OOM cheaper price point)
   3. race to the bottom pricing wars (DeepSeek again)

SemiAnalysis' recent newsletter provides some data points on token spend vs labor cost ROIs for actual 1-20 hour tasks. 

SemiAnalysis has written and talked extensively about our Claude Code usage, but it is important to emphasize that agentic AI is no longer limited to just coding. Our analysts are using agents every day to convert excel models into dashboards, create charts for all our notes, build financial models and analyze company earnings, and much more. These are all tasks that either 1) we simply wouldn’t have been able to do before or 2) would’ve previously taken our junior analysts many hours, taking them away from far more value added tasks.

The table below shows a handful of real examples from our own workflows, comparing token spend against what the equivalent human labor would have cost:

... We estimate that the true blended price per million tokens for running Opus 4.7 on agentic tasks at $0.99 despite the sticker price being $5/$25 per MTok. Agentic workloads have extremely high input-to-output ratios (our Claude Code usage has a ratio of about 300:1) and high cache hit rates (90%+). Because cached input tokens only cost $0.50/MTok, most of the tokens end up in the cheapest tier. We walk through the full methodology here.

Eyeballing, it looks like 8 hours of analyst-type work costs them $7-30 in Opus 4.7 token spend, so (very roughly) 7-30M tokens at their true blended price of ~$1 per M tokens, in contrast with the post's 40-1,300M token estimate, and already squarely here. I expect token usage to drop further for a given task with more advanced models, and also to vary a lot depending on (essentially) how much the big AI companies prioritise RLVR-ing them and on model jaggedness, but also for doable tasks to get much more complicated, like this and more.

Epoch BOTEC-ed a related question last year, prior to Claude Code: How many digital workers could OpenAI deploy? My main takeaway was "worker equivalents is probably more misleading than helpful if people just skim headline numbers" (which everyone does, speaking as someone who sometimes needs to produce headline numbers). 

On the tasks that AIs are able to perform today, how many “human-equivalent digital workers” could frontier AI labs deploy to work on them?

Based on a speculative back-of-the-envelope calculation, we estimate that companies like OpenAI have the hardware to deploy on the order of 7 million digital workers, with a wide 90% confidence interval of 400,000 to around 300 million.2 This doesn’t mean that OpenAI could do the jobs of 7 million human employees today, because AIs can’t fully substitute for humans. But as AI progress continues, AIs will be able to perform an increasing fraction of the tasks that humans currently do.

Thanks, that's useful to know and positively optimism-inducing. 

The Frontier in 2025 (data), by Gavin Leech, Lauren Gilbert, and Ulkar Aghayeva, rated 202 of the biggest breakthroughs of last year. Some favorites, mainly public health- and society-related:

1. Diagnostics on a phone with no doctor needed (source)
2. Murder rates worldwide have fallen 25% since 2000 (source) ("On average! Potentially some confounding from improved trauma emergency care converting murders into attempted murders")
3. 5 factors explain most of the genetic variance in common mental illnesses (source)
4. Large effect for 5-MeO-DMT for treatment-resistant depression (source) ("Recall that major depression is maybe 2% of the total global burden of disease")
5. For the first time in recent history, China’s emissions might be falling (source)
6. The first evidence of a solar take-off in Africa (source)
7. A tiny number of people are functionally cured of HIV. The antibodies responsible may have been identified (source)
8. An E. coli vaccine is currently undergoing Phase III human testing (source): "E. coli is the second-most lethal bacterium in the world, with about a million deaths a year. There are currently no effective vaccines for it"
9. Extreme poverty drops from 27% of India to 5% in one decade (source)
10. Observational follow-up on the Covid vaccines shows a large decrease in all-cause mortality (source)
11. First-in-human 'prime editing' gene therapy. Cured an inherited immune disease (source)
12. Last year's biannual HIV shot available in low-income countries, $40/year (source)
13. AI designs antibodies that can turn on or off membrane signaling proteins implicated in many diseases (source)
14. AI for antibiotic design (source): "7 of 24 AI-designed and custom-synthesized compounds show selective antibacterial activity, including against N. gonorrhoeae and S. aureus"
15. AI generator for antibodies against specific protein targets (source)
16. Two promising drugs to prevent secondary and post-surgical stroke (source)
17. Challenge trial on a salmonella vaccine showed roughly 70% effectiveness (source)
18. Tiny demo of a 90% effective malaria vaccine which only takes one dose (source) ("However, the Leiden study was n=15. The followup PfSPZ-LARC2 study won't be finished until 2027 and is also n=22(!). Last year's R21 vaccine was 75% effective but takes 4 doses")
19. A candidate "gene drive" for eliminating malaria reduced parasite hosting from 80% to 30% (source)
20. A new class of treatment for malaria: 97% cured and it shouldn't suffer existing drug resistances (source): "the first new class of malaria drug approved in more than 25 years"
21. Three new countries certified malaria-free (source): Georgia, Suriname, and Timor-Leste
22. Four countries eliminated trachoma, a disease that causes blindness (source): Senegal, Egypt, Mauritania and Fiji
23. The cost to treat drug-resistant TB drops below $300 (source): "bedaquiline now available at $63 per treatment course, bringing the price of the complete BPaLM treatment to $284"
24. First successful transplant of a non-human lung into a human (source)
25. The oldest baby in history: an embryo frozen in 1994 was brought to term and resulted in a healthy baby boy (source): "The biological mother of the baby was 62 years old at the time of his birth. While incremental, this points towards eventually allowing for delayed IVF, which would be socially transformative"
26. First human infant cured of a lethal genetic disease with a personalized gene therapy (source)
27. Trevogrumab could potentially prevent muscle loss in the sedentary (source)
28. Approval of a strong non-opioid painkiller targeting a pathway specific to pain neurons (source): "It's strong, as strong as hydrocodone or low-dose morphine. It is claimed to be nonaddictive, which is the somewhat unlikely part. It does hit the brain much less, which might work"
29. Tooth regrowing procedure enters human trials (source)
30. Rising Internet access reduces prevalence of female genital mutilation (source)

Why they did this:

A couple of years ago, Gavin became frustrated with science journalism. No one was pulling together results across fields; the articles usually didn’t link to the original source; they didn't use probabilities (or even report the sample size); they were usually credulous about preliminary findings (“...which species was it tested on?”); and they essentially never gave any sense of the magnitude or the baselines (“how much better is this treatment than the previous best?”). Speculative results were covered with the same credence as solid proofs. And highly technical fields like mathematics were rarely covered at all, regardless of their practical or intellectual importance. So he had a go at doing it himself.

This year, in partnership with Renaissance Philanthropy, we took a more systematic approach. So, how did the world change this year? What happened in each science? Which results are speculative and which are solid? Which are the biggest, if true?

I liked GiveDirectly's recent update via GWWC's email newsletter:

Donations from Giving What We Can community members were delivered to Masauli, Chirtera, and Mtembo villages in Chiradzulu district in Malawi. Together, we funded transfers for 954 Malawians in poverty across all three villages. How did families spend their transfers? Here’s what follow-up surveys show:

Hear directly from Emily and David, who are just a few of the people in Masauli village who received transfers from you and other GWWC members:

Emily and her husband, Evance

“My husband and I relied on farming and casual labor to survive,” said Emily. “We grow maize, tomatoes, and cabbage, but without fertilizer, our harvests were always small. Our house also had no windows, and we used sacks instead of a proper door. I always felt unsafe because I worried people could steal from us.” “When I received my first cash payment, I used it to buy fertilizer. This changed everything. Our harvest increased from six bags to ten, and now we have enough food to last until the next harvest. This has brought a sense of peace since we know we have food. With the second payment, I focused on improving my home. I bought 11 bags of cement and used them to put in a proper floor, plaster the house, and add windows and doors. Now my house is safer, and I feel proud of it. I also started thinking about the future. I bought a goat as an investment and started a small business selling beans. I’m making sure I don’t just spend the money, but use it to build something for my family. My husband and I had been sleeping on a mat since we got married. Now, we sleep on a mattress. No more body pains when we wake up. We can sleep comfortably, something we only heard about from other people before. I also bought a bicycle, so I no longer have to borrow from my neighbors when I want to go to the hospital, market, or maize mill. When other women talk about how nice my house looks, it makes me feel proud. I can see the difference in my life now. I have more dignity, security, and hope for the future.”

David and his wife, Lucy, with their harvest

“Every year, we harvested between 15 to 20 bags of maize, but even with all that, we could not develop our home,” said David. “We had three children in secondary school, so all our money went to their education. As parents, we chose to sacrifice for them, but we always wanted to do more.” “We thought about starting a business, because farming is no longer reliable: the weather changes, and farm inputs like fertilizer are expensive. But we never had enough capital to start. Then GiveDirectly came. With our transfer, we opened a small shop using K200,000 (~$115). At first, we sold simple things, like eggs, drinks, and bread. These were the items that people bought most. The business started well, and we were making K40,000-50,000 (~$23-29) a day. We followed one important rule we were told by a friend: never keep money without restocking. Every day, we used the money to buy more goods: salt, biscuits, sweets, and more. The shop kept growing, and now we’re planning ahead. We want to build a proper shop structure. We’ve already bought 22 iron sheets and 3 bags of cement. We are waiting for the rains to pass so we can mold bricks. We’e also planning to build two rental houses for students from a nearby secondary school. We also want to buy livestock, goats, pigs, and cows, so that we can continue supporting our son who is still in school. For us, this money did not just start a business. It gave us a new path.”

All this for slightly over $800 per beneficiary. Hell of a benchmark, cash transfers.

My impression (not the author): 

• area is from slide 8 of SpaceX's FCC Starship Gen2 filing (says V2 in the table but the 2000kg mass is V3-scale)
• wattage is Forethought's guess (257 m² area x 1361 W/m² solar constant x 20% AM0 cell efficiency x 0.721 system derating = 50,400 W), not including 95% capacity factor and 8% annual degradation. It's 2.5x what Musk claimed FWIW
• not sure where array mass comes from 

Seems to be a formatting error and it's supposed to be in the main text, referencing the table.

I resonated with a lot of this, especially prior to 2022. Speaking only for myself, I think a lot of it was downstream of what Ozy Brennan wrote in The Life Goals of Dead People, but I was (unbeknownst to myself) much better at rationalisation than introspection, so it took a long time for me to realise this.