Apr 22, 2026 · 1h 14m · latent-space
AI-Native Engineering: 100% adoption, 5x search throughput, unlimited tokens — Mikhail Parakhin
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Shopify CTO Mikhail Parakhin joins host Shawn 'Swyx' Wang to discuss achieving 100% internal AI adoption, scaling agentic coding workflows, and deploying proprietary infrastructure like Tangle, Tangent, and Liquid Neural Networks. The conversation provides an extensive architectural deep dive into token economics, autonomous optimization loops, and the future of AI-driven commerce.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 4.3 Guest teaching 3.8 Guest disagreement 1.6 The hosts pushing back 1.9
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 10:44 Rejecting parallel agent swarms
Mikhail forcefully rejects the popular industry trend of uncoordinated parallel agent swarms, calling them almost useless compared to sequential critique loops.
Hardest push from the hosts ▶ 33:25 Challenging the AutoML resurgence narrativeThe host refuses to accept hype around auto research blindly, challenging why past AutoML failed and pressing for tangible production flaws.
Biggest teaching moment ▶ 40:14 Exposing the flaw in ungrounded customer simulationsMikhail educates the audience on why agentic user simulations produce meaningless prompt echoes unless calibrated against decades of historical A/B behavioral data.
The host holds their own ▶ 49:34 Host maps trajectory modeling and ergodicityThe host demonstrates strong statistical mastery by visually framing customer simulation as stochastic path modeling and ergodicity beyond static A/B test summaries.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Host Welcome and Channel Subscription Announcement | 2 | 1 | 1 | 0 | The host provides housekeeping and introduces the guest's extensive background at Microsoft and Shopify. The exchange is lighthearted and conversational, touching on internal tooling like QMD and SQLite. | |
| Tracking Internal AI Tool Adoption and Token Distribution | 3 | 3 | 1 | 2 | The host walks through internal adoption slides while the guest explains the phase transition in usage and CLI preferences. The host gently probes why a heavily skewed token distribution among engineers might be problematic. | |
| Agent Critique Loops and Rigorous PR Review Pipelines | 4 | 4 | 3 | 2 | Mikhail defends Jensen Huang's token budget claims and criticizes uncoordinated agent swarms as an anti-pattern. He explains why standard public PR review tools fail to meet Shopify's high-latency, frontier-model review requirements. | |
| Scaling CI/CD Pipelines and Reconsidering Microservices | 6 | 3 | 2 | 3 | The host introduces strong conceptual analogies including global mutexes and Netflix Chaos Monkey to analyze CI/CD bottlenecks. Mikhail is receptive, acknowledging that agentic code velocity might even force a reconsidering of microservices. | |
| Tangle: Collaborative ML Pipelines and Content-Addressed Caching | 5 | 4 | 1 | 1 | The host connects Tangle's design to his past quant finance experience with Airflow and Dagster. Mikhail elaborates on the system's lineage from Ether and Nirvana, emphasizing content-addressed caching across teams. | |
| Tangent and Autonomous Research Loops in Production | 4 | 4 | 2 | 3 | The host questions why historical AutoML efforts flopped and asks about the failure modes of current auto research loops. Mikhail openly shares practical limitations, recounting a 400-experiment run that produced only a single hit. | |
| SimGym: Grounding Customer Agent Simulations with Historical Data | 4 | 5 | 2 | 2 | Mikhail explains that simulating customer behavior without historical conversion data is futile since models merely mirror prompt instructions. He details the multimodal and browser infrastructure necessary to achieve high correlation with real shopper behavior. | |
| Counterfactual Modeling, Trajectories, and Advanced Statistics | 7 | 5 | 1 | 2 | The host actively illustrates trajectory modeling and ergodicity to demonstrate how multi-step simulation surpasses naive summary A/B testing. Mikhail enthusiastically builds on this with Shopify's HSTU counterfactual rollouts and Chinese Restaurant Processes. | |
| Universal Commerce Protocol and Global Catalog Search | 5 | 6 | 2 | 3 | Mikhail provides a deep technical breakdown of Liquid AI neural networks as non-transformer state space architectures used for sub-30ms search. The host probes why SSMs historically struggled and questions compute-scaling bottlenecks. | |
| Shopify's Engineering Moat and Specialized Technical Hiring | 3 | 3 | 1 | 1 | The host recaps Shopify's technical moat and asks for specific hiring targets, leading Mikhail to emphasize ML and distributed databases. The segment wraps with technical hiring requirements and closing reflections. |
Statements from this episode (28)
Parakhin: Shopify built agent memory systems before Karpathy proposed them
“Just even yesterday, Andrej Karpathy was famous in tweeting about, oh, there's some ways that you can organize your agents to store the data and then look up the data so that you don't have to research or lose context every time. And a little bit tongue-in-che…”
Parakhin: Daily AI tool adoption at Shopify approaches 100 percent
“So you could see that it approaches really a hundred percent by now. It's hard not to do your job now without interacting deeply, at least with one tool.”
Parakhin: CLI AI tools outpace IDEs like Cursor at Shopify
“The other thing I would claim you could see is that CLI-based tools and tools that don't require you to look at the code becoming more popular, and you could see, yeah, various versions of Cloud Code and Codex and Pi and internal development tools taking off e…”
Parakhin: Shopify funds unlimited tokens and discourages models below Opus
“And we effectively fund unlimited tokens for everybody. We do try to control the models that people use, but from the bottom, not from top. Like we basically say, hey, please don't use anything less than Opus.”
Parakhin: AI token usage at Shopify skews heavily toward top percentiles
“And I would claim the super interesting part here is that you could see that the distribution becoming more and more skewed. The top percentiles grow faster. So that means the people in the top 10 percentile, they, their consumption grows faster than 75 and so…”
Parakhin: Jensen Huang is directionally correct on developer token budgets
“I do think Jensen gotten a lot of bad press saying, oh, of course you're, you know, this the cake seller says we don't need enough cakes, you know, like, of course. But I actually think that's undeserved. I think he, he's actually right. I do think. He's direc…”
Parakhin: Top AI models write code with fewer bugs than average humans
“I would claim by now, good model writes code on average with fewer bugs than average human.”
Parakhin: No existing third-party AI PR review tool meets his standards
“I haven't found a good PR review tool that, that does what I think should be done”
Parakhin: Effective PR review requires largest pro-level models, not fast tools
“At PR review time, you want to run the largest models. That means codex or cloud code is not going to cut it. You need to have pro-level models if you really want to stem the tide of bugs from going into production”
Parakhin: Slower AI PR reviews actually save total deployment time
“It actually, in terms of the overall time to deploy, it's total time savings if you spend more time on a longer model, like, thinking for an hour, because then you don't have to spend all that time During testing and rolling, you know, rolling back the deploym…”
Parakhin: Human-centric CI/CD systems must be redesigned for AI agents
“Clearly the old thing were designed for humans will need to be morphed into something new.”
Parakhin: Airflow is poorly suited for rapid ML experimentation compared to Tangle
“Airflow is great, but Airflow is more about you have something and you want to repeatedly run it in production on schedule. It's less about you as a team developing things and being able to share and you grabbing the standard pipeline and saying, hey, I want t…”
Parakhin: Content-addressed caching yields big compute savings by deduplicating cross-team work
“The main savings are coming from the fact that you ran it, you got your job done, and you moved on. Then somebody else in some department you don't know existed, runs the same task, but on the newer version. Like right now, you can't, in, in most of the organi…”
Parakhin: Shopify increased search throughput 5x to 4,200 QPS via auto-research
“Our search recently we moved from, it's hard to win, quote, from 800 QPS to 4200 QPS with the same quality just by pure optimizations and not a research loop that kept Running and changing code in our index surf on the same number of machines, just increasing …”
Parakhin: Auto-research excels at obvious optimizations but fails on out-of-distribution tasks
“Autoresearch is very good at doing kind of obvious things that you don't have bandwidth to do, or you didn't notice, or maybe you're not aware of like some standard practices. It is not good at doing something completely out of distribution, something that, yo…”
Parakhin: Simulating customers without historical data produces ungrounded agent responses
“If you don't have the historical data, All you can do is prompt agents in the vacuum, and they will do exactly what you prompt them to do.”
Parakhin: Shopify aimed for 0.7 correlation between SimGym and real add-to-carts
“Internal goal was to hit 0.7 correlation with add to cart events, for example, like that, that if we run real A, B test experiment, that it should go and replicate same sort of success that, that humans had or lack thereof.”
Parakhin: Increasing product image size usually tanks ecommerce sales
“Usually people's intuition here, by the way, is that I increase my images. I'll have more because they look nicer. You know, designers all look sparse and big images. Like, usually your sales tank, right?”
Parakhin: Shopify uses large HSTU architecture to model merchant trajectories
“Internally we have this system. We talked about it briefly once at NeurIPS. We have a huge HSTU-based system that models the whole companies.”
Parakhin: Counterfactual behavioral modeling was impossible before LAMPS and HSTUs
“Being able to model something complex as human beings or companies and model counterfactuals on it where you can have interventions in the future and optimize when to make intervention, what kind what kind of intervention to make. It's such an unlock that prev…”
Parakhin: Shopify brought back Chinese Restaurant Processes for buyer clustering
“Recently we're looking at it, and we had to bring back CRPs, you know, Chinese restaurant process. It's a, like, way of aggregating and like naturally grow clustering across specifically to answer questions that like you were just posing on how, how if buyers …”
Parakhin: Liquid is the only genuinely competitive non-transformer architecture Shopify found
“That's why we at Shopify, when we tried multiple, and we constantly try multiple models, multiple companies, we found that for small, particularly with low latency applications, when you have low latency and or if you need longer context lengths, Liquid was th…”
Parakhin: Shopify runs a 300M parameter Liquid model under 30ms for search
“We run it at 30 milliseconds, a tiny model, like three hundred million parameters, In, but we run it in 30 milliseconds end to end for search when you type a query, and then we produce all the possible things with what you can mean by that query and some, you …”
Parakhin: Liquid-transformer hybrids are probably the best neural network architecture available
“I think especially in their hybrid form, when combined with Transformer, like in Mamba fashion, they probably the best architecture I'm aware of, like, period.”
Parakhin: Liquid AI could beat frontier models with equal compute
“I think if they if they had similar level of compute, they would be very competitive and maybe even beat the largest models, at least from what I've seen.”
Parakhin: Bing Sydney first launched in India using Megatron, not OpenAI
“The funny thing, I mean, the most interesting anecdote is that Sydney was first shipped in India for and it was not noticed for a long time. And first implementation of Sydney didn't even have open AI model under it. It was during Megatron. Microsoft and the N…”
Parakhin: Bing Sydney's personality was deliberately engineered, not purely emergent
“What almost everybody doesn't fully realize is that it wasn't by accident that Sydney was Sydney. I mean, we spent a lot of effort on personality shaping. We, I mean, it was a bit of my Yandex legacy where previously we did this Alice digital assistant which w…”