Products & Tools

AI Reverse Engineering Tool REA Hit 234,477 GitHub Stars in 8 Days

REA's viral growth signals that AI can collapse the engineer-hours behind technology imitation — and The Economy's analysis asks whether Europe's comparative advantage returns as a result.

When AI Makes Copying Cheap: Can Europe Rewrite Its Comparative Advantage?
When AI Makes Copying Cheap: Can Europe Rewrite Its Comparative Advantage?Jakub Żerdzicki / unsplash
By Sophie Lindqvist8 min read

Why it matters

  • REA (Reverse Engineer Anything) grew from under 10,000 to 234,477 GitHub stars between October 6 and October 14, 2026.
  • A 2025 METR study found skilled developers were 19% less efficient with LLM assistance on 246 coding tasks, despite expecting a 24% gain.
  • German engineering/IT job openings fell ~28%, from an average of 148,230 to 106,310 between 2024 Q1 and 2025 Q2.
  • China's 2024 R&D spend was 3.63 trillion RMB, up 8.9% year on year, with ~81% going to experimental development of existing products.
  • World Bank data show ~20% of the European and Central Asian workforce is exposed to AI, but fewer than 10% of firms have adopted it.

The AI reverse-engineering tool REA jumped from under 10,000 GitHub stars on October 6, 2026 to 234,477 stars by October 14 — an eight-day surge that, according to a new analysis by The Economy's Editorial Board, could mark the beginning of a structural shift in who wins engineering work worldwide.

The piece, published November 15, 2026 and titled "When AI Makes Copying Cheap: Can Europe Rewrite Its Comparative Advantage?", makes a specific claim: LLM-based agents that reverse-engineer compiled code are collapsing the engineer-hours needed to replicate technology. That undercuts a comparative advantage East Asia has held for roughly 70 years — and it could hand some of that work back to Europe.

The stakes are straightforward. If the cost of copying software-heavy technology falls far enough, the geography of engineering-intensive production changes. Whether Europe actually captures that work depends on adoption speed, and the transition will produce losers as well as winners.

How expensive was copying before AI?

The analysis opens with Sony. In 1953, the company paid Western Electric $25,000 for a transistor licence, contingent on Japanese government clearance of foreign currency allocation. The licence came with Bell Labs documentation. That documentation was not enough to build a commercially viable product — Sony needed roughly two more years of engineering work before the TR-55 shipped in 1955.

The pattern persisted. When the Walkman reached the West in 1979, the sequence was the same: licence or purchase the technology, then spend engineer-hours to develop, replicate or reproduce it. Comparative advantage belonged to whichever economy could convert acquired knowledge into mass-producible products at the lowest cost.

Samsung followed the same playbook in memory chips. After licensing Micron's 64K DRAM in June 1983, Samsung produced a prototype by November 1983. For the next generation, the 256K chip, Samsung skipped licensing and reverse-engineered competitors' technology, delivering a prototype by July 1985. By 1993, Samsung was the world's largest DRAM manufacturer.

China has run a similar approach at unprecedented scale, backed by national policies, supplier localisation and steady expansion of its domestic engineer workforce. The article cautions against conflating these economies — Japan built its position on consumer electronics, Korea on memory chip production, China on a vast domestic manufacturing footprint. But the common thread held: all three could buy engineer-hours for learning and replication more cheaply than European and U.S. economies.

What is REA, and why did it explode?

REA — Reverse Engineer Anything — is the clearest evidence that this cost is now falling fast. Developed by GitHub user 'Morluto', REA uses LLM-based agents to drive existing reverse-engineering tools such as Ghidra, IDA and Hopper to analyse binary code.

In an interview with Scientific American, Morluto said he built REA because people have grown used to coexisting with black-box software.

REA is not an isolated case. The article notes that an AI agent, Claude Opus 5.5, recently helped a developer produce open-source replicas of key Adobe products. Open-weight models such as GLM-5.3 from Z.ai, it adds, remove concerns about vendor lock-in.

The Swiss Institute of Artificial Intelligence (SIAI), in a November 3 memo, argues that reverse engineering is primarily pattern recognition — a task type for which LLMs are well suited.

Why AI helps least on familiar code — and most on foreign code

Here the analysis leans on an uncomfortable finding. In a 2025 study, METR measured skilled software developers on 246 coding tasks. With LLM assistance, they were 19% less efficient on average — even though they expected a 24% gain beforehand and believed they had achieved a 20% gain afterwards.

SIAI reads that result as defining a lower bound: LLM augmentation delivers least where developers already know the codebase. Reverse engineering sits at the opposite extreme. By definition, the codebase is foreign — which is exactly where pattern-recognition agents should add the most value.

There is a hard limit, though. Hardware cannot be duplicated the way software can. A photolithography system would have to be reverse-engineered physically, with considerable risk of irreversibly damaging it, and it depends heavily on engineers' tacit knowledge and trade secrets. TrendForce, citing The National Interest, reported that China attempted to reverse-engineer ASML equipment with limited success — though the article flags this as not independently verified.

The direction still matters. An increasing share of modern technology is software — machine control units, automotive systems — and every reduction in replication engineer-hours erodes East Asia's advantage in copying others' products.

How does Heckscher-Ohlin explain the shift?

The framework here is the Heckscher-Ohlin theorem: countries export goods that intensively use their abundant factors. For about 40 years, lower engineer-hour wages pushed engineering-intensive replication toward East Asia.

The article's worked example makes the mechanism concrete. Economy A (Europe) pays engineers €80 per hour and has abundant capital, automation and sophisticated local demand. Economy B (East Asia) pays €25 per hour. A task needing 10,000 engineer-hours costs €800,000 in A versus €250,000 in B — so the work goes to B.

Now let AI cut the task to 500 engineer-hours. The gap shrinks to €27,500, small enough to be offset by compute and energy prices, IP accessibility, lead times, and the value of engineers sitting next to the end user.

The key insight is what did not change. Relative factor abundance stays the same; the activity itself flips from labour-intensive to capital-intensive — AI tools, computing, physical equipment — and capital is what Europe has. A change in an activity's factor intensity can switch comparative advantage.

The article also connects this to an earlier Edgeworth-box analysis of U.S.-China trade. AI does not change labour endowments, but as a labour-augmenting technology it changes effective endowments and therefore the efficient allocation of tasks. Formalising this, the authors note, would require an extended H-O model that drops the assumption of a common production function.

One caveat follows immediately: this need not mean reshoring to Europe. East Asian economies get the same engineer-hour savings and keep supply-chain advantages — scale, supplier networks — that may compound their cost position. H-O predicts the advantage can change; it does not say in which direction.

What would Europe actually gain?

Europe brings capital, advanced manufacturing infrastructure, research institutions and sophisticated end users. If AI cuts the engineer-hours for adapting products, some engineering tasks could move back toward local markets. The article's example: a German machine manufacturer might stop outsourcing development of control units once local AI-assisted development is cheap enough.

The effect is bounded. It covers engineering stages only, and it will not close the wage differentials that keep physical production offshore.

What do the labour-market numbers show?

Two data points anchor the labour story. Advertised engineering and IT job openings in Germany fell about 28%, from an average of 148,230 to 106,310 between 2024 Q1 and 2025 Q2. The article calls this suggestive of falling engineer-hour demand, while acknowledging macroeconomic conditions could explain it too.

The second point is the adoption gap. World Bank data show roughly 20% of the European and Central Asian workforce is exposed to AI, yet fewer than 10% of firms have adopted it, and mostly for basic applications. The World Bank identifies this slow uptake as a key regional challenge — and the article notes that Europe's highest-wage economies are also its fastest AI adopters.

Who wins and who loses?

Drawing on CEPR research by Antràs, Kulesza and Presbitero (Discussion Paper No. 21948) on the U.S.-China tariff war, the article sketches the distributional picture. Since 2017, China's share of U.S. imports has fallen from 22% to roughly 8% in 2026. CEPR found that export similarity is a poor predictor of how individual economies fare, decomposing effects into export and import competition plus consumer and input-cost savings.

Applied here: Asian economies would face higher export competition, European consumers and producers gain from lower costs, and Asian producers gain from faster, cheaper technology development. The effects are mixed across economies.

The transition itself will hurt some places and firms. Reallocating manufacturing locations, engineers and capital flows means the gains to some economies will not fully offset losses elsewhere — the new equilibrium will not be Pareto-improving.

Can East Asia pivot instead of fading?

The open question is whether East Asian economies reshape their R&D. China spent 3.63 trillion RMB on R&D in 2024, up 8.9% year on year. Korea spent roughly 5% of GDP — the second-highest in the OECD behind Israel — with 79.2% undertaken by companies.

But the composition matters more than the total. Around 81% of China's R&D goes to experimental development of existing products, with under 7% on fundamental research; Korea's fundamental-research share is about 15%. Whether these economies can shift toward fundamental research, the article argues, will determine if they keep their lead in developing genuinely new technologies.

The bottom line: if Europe adopts these tools fast enough and pairs them with its capital and local sophisticated demand, it can compete across a wider range of activities with fewer engineer-hours. The race, as the piece closes, is over how quickly each economy integrates AI into production — a contest that will unfold over the coming years.

Source: economy.ac

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

630 articles

Related articles

  1. Eurozone AI investment surges 15% as adoption doubles
  2. 96% of Developers Don't Trust AI Code. Teams Are Rewriting Review.
  3. US Data Brokers Sell $500 Million in RL Environments to Chinese AI Labs
  4. Santander Counts €194 Million From AI as Earnings Season Puts Numbers on the Hype
  5. Eric Schmidt likens OpenAI and Anthropic to colonial trading firms

« Previous article