Wookyoung Kim, Jisoo · July 29, 2025
Alchemy, AI performance, economic control, KV-cache, repetition
Weekly Commits gathers articles and information seen over the week in one place and introduces them with a perspective on why they are interesting.
The Gold Bracelet - Amedeo Simonetti (1) A Mass-Production Process for Turning Mercury into Gold, from Marathonfusion.com:
Marathon Fusion has announced a solution to the transmutation of gold, one of alchemy's great challenges for thousands of years. Unlike previous attempts, our method is scalable to large volumes, practically achievable, and economically attractive. In deuterium-tritium fusion, high-energy neutrons induce "breeding" reactions that complete the fuel cycle by producing the tritium needed to sustain operation. Our approach uses those neutrons to induce a breeding reaction in mercury-198, producing mercury-197, which decays a few days later into gold's only stable isotope. With our approach, a power plant could produce 5,000 kilograms of gold per year per gigawatt of electric output (~2.5 GWth), without compromising either fuel self-sufficiency or power output. This marks the beginning of a new golden age: not only in critical mineral production, but also in energy, prosperity, and scientific discovery.
Fusion energy startup Marathon Fusion said it can use neutrons released during fusion reactions to mass-produce gold. What makes the research interesting is that it can be combined with the existing fusion power process, creating additional revenue through gold production.
Among fusion power methods, deuterium-tritium (D-T) fusion combines deuterium (D) and tritium (T) at temperatures above 100 million degrees Celsius. Because tritium is almost nonexistent in nature, it is produced through a breeding blanket. The main reaction is between neutrons and lithium.
This creates an important problem. In a D-T fusion reaction, one neutron is produced per reaction. To sustain the fusion process, the neutrons produced by D-T fusion must be used entirely by the breeding blanket to generate tritium. In actual fusion, however, some neutrons are lost to the outside, so additional neutrons must be produced for sustainable fusion power. This is called neutron multiplication.
Marathon Fusion proposed using mercury-198 to produce neutrons during neutron multiplication. Mercury-198 produces mercury-197 as a byproduct of the neutron multiplication process, and mercury-197 decays into gold-197 with a half-life of 64 hours. In other words, the process produces high-value gold while carrying out tritium production, an essential step in D-T fusion power. Marathon Fusion said that using this method, a D-T fusion power plant could produce 5,000 kg of gold per year per gigawatt of power generation without any loss.
One complicated issue is that if mercury isotopes other than mercury-198 are present during neutron multiplication, unstable isotopes of gold can be produced. In that case, the gold may be radioactive, and it is estimated that it may need to be stored for 14 to 18 years to eliminate the radioactivity.
This research makes alchemy, one of humanity's great dreams, real. If artificial gold continues to be produced this way, can gold continue to serve as a store of value? Even just improving the profitability of fusion power suggests that this research could have a massive impact on the world.
(2) AI Is Effective at Improving the Performance of Less-Skilled People
From OpenAI:
One study found that call center agents working with OpenAI's model were an average 14% more productive in responding to customer queries. In particular, less-skilled workers saw the largest productivity gains. There is also evidence that AI could help close skills gaps, as it augmented lower-performing consultants by 43% and higher-performing consultants by 17%.
OpenAI released Unlocking Economic Opportunity, a report on how ChatGPT affects productivity.
According to the report:
In call centers: among call center employees using OpenAI models, less-skilled workers saw large productivity gains.
In consulting: using AI improved task completion speed by 43% for lower-performing consultants and by 17% for higher-performing consultants.
In other words, AI provided greater utility to less-skilled people than to highly skilled people, and to lower-performing people than to higher-performing people. This report is interesting because it suggests that AI models may bring greater opportunity to everyone, rather than concentrating wealth and power in the hands of a few.
(3) AI Will Not Be Able to Control the Economy Perfectly
From Quora:
I assert that AI cannot be intelligent enough to run the modern economy perfectly. Why? The main reason is that AI itself will become part of the economy. Businesses and individuals will use AI to make decisions. Therefore, every AI would have to take into account the decisions of other AIs. But no AI will be so much more advanced than the others that this becomes possible. In other words, as AI grows more powerful, the complexity of the economy will also grow.
In 2018, Alex Tabarrok answered the question of whether AI could advance enough to manipulate the economy perfectly. His reason was that because businesses and individuals would use AI to make decisions, no single AI could have enough computing power to account for the behavior of every individual AI.
Reading this, I thought that as long as humans use AI to make decisions, it will be impossible for one AI to perfectly control every AI, or every human.
(4) The Single Most Important Metric for an AI Agent: KV-Cache Hit Rate
From manus.im:
I'd argue that the KV-cache hit rate is the single most important metric for a production-stage AI agent. It directly affects both latency and cost.
Manus, which builds AI agents, said it would focus on context engineering. Each time the team discovered a new way to form context, it rebuilt the agent framework, going through four rebuilds in total.
One thing Manus learned through this process is that the most important metric for an AI agent is KV-cache. KV-cache is an optimization technique that lets an LLM avoid repeated computation during inference. It stores keys and values from the self-attention calculation to avoid recomputing them.
KV-cache greatly reduces both time to first token and inference cost. For Claude Sonnet, the cost differs by nearly 10x depending on whether KV-cache is applied. Manus introduced ways to preserve context while maintaining the KV-cache.
Do not change the beginning of the prompt - especially avoid inserting changing values such as the current time every time, since that breaks the cache.
Do not edit previous conversation content - leave what was said as it is, and append only new content afterward.
Organize data in the same way - data such as JSON must be arranged in the same order every time for the cache to remain valid.
Clearly separate cache segments - decide in advance where caching should begin.
Send requests to the same server - when using multiple servers, keep sending the same conversation to the same server.
As someone who has not built an AI agent directly, Manus's view that KV-cache is important felt fresh because it differs from AI wrapper companies' usual advice to use the best possible AI API. It is also impressive to see context engineering being applied one step after another in real work.
(5) Repetition Reveals Difference
French philosopher Gilles Deleuze argues that true difference is revealed through repetition, and that this difference shows the uniqueness of each being. It is a complex idea, but we already experience it.
Music: when we listen to the same song repeatedly, we discover bass lines or subtle shifts in vocal emotion that we missed at first.
Food: when we eat a regular menu item again and again, we notice that today's touch in the cooking is different from yesterday's.
Art: when we look repeatedly at the same print, we begin to see tiny differences in the spread and density of the ink.
Reading: when we read a book repeatedly, we notice the author's intention or hidden meanings that we did not see at first.
Modern technology operates on the same principle. Just as humans discover difference through repetition, algorithms repeatedly learn from your behavior to identify what is unique about you.
Netflix and YouTube continually analyze your clicks, skips, watch time, and other behavior to understand your individual tastes.
This piece looks at how the philosophical concept of "discovering difference through repetition" became a core mechanism of modern digital personalization algorithms, and how it connects to our everyday experience.
(6) Recruiting Strategy Like Toss
I went to the Toss Makers Conference. After sitting there for six hours, the thought in my head was, "If I got a job at Toss, it would be fun and I would grow incredibly fast." Most attendees probably thought something similar. That is because the conference itself was designed as a recruiting event.