Source: raw/Inside_Shopify_s_AI_strategy_with_CTO_Mikhail_Parakhin.md — official Anthropic YouTube channel (@claude), interview with Mikhail Parakhin, CTO of Shopify, youtube.com/watch?v=9GtcbnmldcY (fetched 2026-10-30). The interviewer is not named. The auto-captions are rough in places; unclear phrases are flagged below.
Parakhin’s argument is that AI’s bigger payoff is “raising the ceiling” (doing things that were impossible before), not “raising the floor” (doing existing work with fewer people). His working policy at Shopify follows from that: unlimited tokens for everyone, the largest model for all development work, and cheaper fine-tuned models only in production, which he calls a “barbell strategy.” He gives two examples: a “digital twin” that models each merchant as a sequence of actions, and a team-monitoring system he built after deciding he had been “wrong” that LLMs could not help him manage.
Key Takeaways
- Raising the ceiling beats raising the floor. Raising the floor is easy to measure (work engineers no longer do), but “you could have just been more diligent or throw more people at it.” Raising the ceiling has “no easy way to extend.” His definition: “you start doing something that previously was not even in consideration set.”
- His best problems need a human and the model together. He says he solved a problem he had worked on for seven years, involving the Wasserstein loss function (“making everything Gaussian”), which neither he nor the LLMs could solve alone. He calls the pairing a “centaur”, after the chess term: “I cannot solve it without the model, but the model cannot solve it without me either.”
- Measure productivity normalized for project complexity. Parakhin claims Shopify is “probably the best in the world right now at measuring how high is the floor.” Beyond counting PRs and PR complexity, Shopify estimates the complexity of whole projects, which lets it see “teams of people are able to execute on more projects faster… normalized for complexity.”
- Always use the largest model, and don’t cap tokens. “You don’t know unless you use the large model and then compare the results. You don’t know what you’re missing.” Shopify gives “unlimited tokens for everybody” and runs “the heaviest models we can find.” His one caveat is that you still need circuit breakers for “runaway processes that suddenly start burning tokens.”
- The barbell strategy. For development, testing and “auto research”, use the largest model. For production, use the smallest model that works, “often fine tuned”, and use auto research to tune it so that cost, throughput and quality balance out. He says he has repeatedly seen companies do the opposite (save on coding tokens, overspend on production models) and that he “will outcompete anybody who is not on that strategy.”
- Make sure the decision-makers use the models. His advice to other CTOs on talking to their CEO: everyone making decisions should be “actually using the model and using the product,” then keep pushing the ceiling until results show “if it works or it doesn’t work.”
- The flagship project is a digital twin of each merchant. Text is a sequence of letters and LLMs predict the next one; “companies’ actions are similar” (open a credit line, ship a product, start accepting cards, take a loan). Modelling a company as an action sequence gives “a digital twin of the company” that Shopify can run counterfactuals on (“what if we added an advertising campaign… what if we give them a loan”) and then act on in production with offers or messages. He says it has “very direct impact on our bottom line” and that it “was not humanly feasible previously.” The captions call it the “HST prediction system”, which may be mis-transcribed.
- LLMs as a management tool. “I thought that the LLMs cannot really help me manage the team better. I could not have been more wrong.” He built a system that analyses what is happening across the team and tells him proactively that a project looks likely to slip its deadline, and also detects “people unhappiness.”
- Technical judgement matters more, not less. “The LLM is very good at picking up signals and it tries to please… the answer is very often in the question.” Because building is now instant, a badly framed request skips the discussion that used to catch it, so “your consequences could be actually worse than before.”
- Engineers become second-level managers. They now supervise LLMs rather than people typing. The questions become how to dispatch work across agents, how to keep track of what each is doing, and how to avoid conflicts between them.
- On Fable 5. Parakhin says he went on record that Fable 5 “was such a huge step, like much, much bigger than all the previous steps.”
Try It
- Run a barbell audit. List where your tokens go. If development and testing use a cheap model while a production service runs a frontier model, invert it: frontier models for building and research, the smallest tuned model that passes your evals for production.
- Remove token caps for builders, but add a circuit breaker. Set a per-process spend or rate alarm that kills runaway loops instead of a per-person budget.
- A/B the largest model before deciding it is not needed. Run the same task on your current model and the largest one available and compare outputs before you settle on the cheaper one.
- Write neutral prompts for decisions. Because the model “tries to please”, state what you need and leave out the answer you expect, especially before asking it to build something.
- Prototype an action-sequence model of your customers. If you have event logs (signups, purchases, upgrades), treat each customer as a sequence of actions and test whether a model can predict the next one. That is the minimum version of Parakhin’s digital twin.
Related
- OpenAI Enterprise Signals — Frontier Firms vs Typical Firms — usage data on firms that spend heavily on tokens
- Coordinator-Pattern Cost (Big Models Plan, Small Models Execute) — a measured version of the big-model/small-model split
- Picking the Right Model — Building Evals for Model Selection — how to test whether a smaller model is good enough
- Claude Fable 5 and Claude Mythos 5 — the model Parakhin calls the biggest step yet
- 2026 — How AI Is Restructuring Organizations and Knowledge Work — where Shopify’s company agent (River) already appears
- Factory — Building Autonomous Software With Droids — a founder’s view of the same management shift
Open Questions
- What the “HST prediction system” is: the captions may be garbling a model name. Its architecture, data and measured revenue impact are not given.
- The name of Shopify’s productivity-measurement system is unclear in the captions (“the system get stuff done”).
- No productivity percentages are stated, despite the claim that Shopify can measure them “very precisely.”
- What the Wasserstein-loss problem was, and what Parakhin solved, is not explained.
- The Linear A anecdote (Parakhin says Boris tweeted that someone deciphered the Minoan script with Claude) is a secondhand claim. This transcript does not name the tweet or the person.
- Which Claude models Shopify runs, and how “unlimited tokens” is governed in practice beyond circuit breakers, is not stated.