Wookyoung Kim, Jisoo · July 20, 2025
Shopify, Muon Optimizer, GLP-1, OpenAI, Sam Altman
Weekly Commits collects the articles and information I read during the week in one place and introduces them with my perspective on why they were interesting.
A View of the Valley - Asher Brown Durand (1) How Shopify introduced AI internally From Tobi Lutke on x.com:
Using AI effectively is now a fundamental expectation of everyone at Shopify. It's a tool of all trades today, and will only grow in importance. Frankly, I don't think it's feasible to opt out of learning the skill of applying AI in your craft; you are welcome to try, but I want to be honest I cannot see this working out today, and definitely not tomorrow. Stagnation is almost certain, and stagnation is slow-motion failure. If you're not climbing, you're sliding.
In April 2025, Shopify CEO Tobi Lutke caused a major stir by declaring AI use mandatory. Farhan Thawar, Shopify's VP of Engineering, later summarized what had happened inside Shopify.
Using AI costs money. Using better AI costs even more. Within Shopify, there were concerns that adopting AI tools like Cursor would become too expensive, because giving everyone unrestricted access to the best AI could add $1,000-$10,000 per person.
At that point, Thawar decided that if AI could create value, people should use it. He created an AI token usage leaderboard inside Shopify and made it something people were proud to appear on.
Shopify also used MCP to connect all of the company's internal tools and built an internal LLM proxy so employees could easily interact with and switch between different models. (An effort to use the latest AI models faster depending on the situation.)
Using MCP and the LLM proxy, Shopify accelerated its workflows.
Sales engineers connected frequently used tools such as Google Drive, Slack, and Salesforce through MCP and built a workflow that could answer the question, "What should I do today?" with suggested actions.
When external questions came in for deals, such as RFPs, the Revenue Tooling team built an AI chatbot that could generate answers based on internal documents. Its biggest advantage was that it learned from answers that had led to won deals, making future answers better.
They had AI roast existing code, helping people write better code.
Shopify's adoption of AI feels a bit like Jeff Bezos's 2002 API mandate at Amazon.
Jeff Bezos's API mandate:
1/ From now on, all teams must expose their data and functionality through service interfaces. 2/ Teams must communicate with each other only through these interfaces. 3/ No other form of interprocess communication is allowed. Direct connections, directly reading another team's data store, shared-memory models, back doors, and everything else are prohibited. The only allowed communication method is calling service interfaces over the network. 4/ It does not matter what technology you use. HTTP, CORBA, Pub/Sub, custom protocols, and so on are all fine. Bezos does not care about the technology choice. 5/ Without exception, every service interface must be designed from the beginning to be externally exposable. In other words, teams must plan and design their interfaces so they can expose them to external developers. No exceptions. 6/ Anyone who does not comply will be fired.
Shopify's next step will be interesting. How much competitive advantage can expanding internal AI adoption create?
(2) Proving that the Muon Optimizer can scale to large-scale training
In deep learning models, optimizers have not advanced much since AdamW. Changing optimizers imposes a huge validation cost, and AdamW satisfies most requirements for deep learning training. That is why most recent optimizers are merely variations of AdamW.
However, in December 2024, OpenAI's Keller Jordan released Muon, a new optimizer, while trying to break the three-second barrier in the CIFAR-10 computer vision algorithm dataset speedrun. Using the Muon Optimizer, Keller Jordan achieved 2.59 seconds in the CIFAR-10 speedrun.
Still, Muon had mainly been used to reduce NanoGPT training time or to chase CIFAR-10 training speedruns, so questions remained about whether it could be applied to large models like AdamW.
In February 2025, China's Moonshot AI released Moonlight, an open-source model demonstrating the performance of the Muon Optimizer, and found that Weight Decay is needed for large-scale training with Muon.
However, when training a model with more than 100 billion parameters with Muon, they ran into a problem where attention logits became extremely large even with Weight Decay. As a result, it seemed impossible to build AI models as large as GPT-4.1 or Claude 4 Opus using the Muon Optimizer.
In July 2025, Moonshot AI eventually proposed QK-Clip, a method for controlling attention logits, and solved the problem of attention logits growing enormously. At the same time, it released Kimi K2, an open-source AI model trained with the Muon Optimizer; Kimi K2 was trained with 32 billion activated parameters and 1 trillion total parameters.
Kimi K2 outperformed GPT-4.1 and Claude 4 Opus on coding benchmarks. Kimi K2 and QK-Clip are significant because they proved that the Muon Optimizer can scale to large-scale training and showed that an AI model trained with Muon can reach frontier-model performance. This may be a moment when the paradigm for deep learning optimizers begins to shift.
(3) GLP-1 (Wegovy) is disrupting life insurance
U.S. life insurers can predict when a person will die with 98% accuracy.
Insurers use key health indicators such as HbA1c, cholesterol, blood pressure, and BMI to price insurance.
The recent issue is that HbA1c, cholesterol, blood pressure, and BMI can be significantly improved through GLP-1 (Wegovy) treatment.
GLP-1 treatment is becoming a problem for life insurers because two years after stopping GLP-1, key metrics such as BMI, blood pressure, and cholesterol return to their original levels. People can undergo life-insurance underwriting while taking GLP-1, then stop taking it; people who should look healthy in insurers' models get sick and die, creating headaches for insurers.
In fact, Miss Declines (dark green bars), cases where insurers failed to detect an applicant's health risk and therefore missed a chance to decline the contract or raise the premium, have continued to increase.
To solve this, the idea emerged for insurers to subsidize GLP-1 costs. If continuous GLP-1 use lowers BMI, blood pressure, and cholesterol and helps policyholders live longer, it could generate more revenue for insurers.
(4) Notes from leaving OpenAI
Former OpenAI employee Calvin's public departure retrospective has drawn attention. He compares OpenAI's organizational culture to a combination of "Los Alamos (the nuclear research lab) + a massive consumer service."
OpenAI's official roadmap is effectively blank; when an idea emerges, people attach themselves to it and a team forms.
Parallel experiments, where multiple teams attack the same problem in different ways at the same time, are treated not as "waste" but as a cost of faster learning.
Researchers choose their own topics and push them forward like "mini CEOs."
Unexpectedly, internal communication is not tightly aligned. Most communication is scattered across Slack, and some new products or changes in direction move forward with only a few people aware until the very end.
Some news circulates first through external articles or social media rather than internal announcements.
People even say "this company runs on Twitter vibes," because it takes social media so seriously.
OpenAI releases technology and models quickly, but core numbers such as revenue, performance, and unit economics are visible only to a very small group: an "open technology vs. closed numbers" structure.
Engineering is in a transitional state, with Rust and Go layered on top of a huge Python monorepo, and the company is said to be cleaning up the "scaling side effects" such as inconsistent style, duplicate libraries, and fragile CI.
It is a little worrying, but also interesting, because this is the first time this level of detail has come out of OpenAI, a company known for tight security.
(5) Sam Altman, The Vision 2030
Sam Altman's blueprint for the future is intricate and ambitious. Even the outline alone gives us expectations and imagination for a new era.
This book brings together Altman's publicly known plans, including OpenAI, UBD, basic income, and Worldcoin, and explains why he is pursuing these projects through his past and ideas. In particular, by using companies he invested in since his YC days as examples, it presents a convincing scenario for why he bet on those companies, even with his own money.
However, Altman's vision is so vast and complex that it cannot avoid criticism of technological solutionism. He still does not provide clear answers to fundamental questions about AGI that remains difficult to control, justice and fairness, and the traps of technological solutionism.