Jordan Christensen

A fantastic summary of the conversations Michael Liu from Inception Studio and I had this past week regarding Canadian Tech founders and their relationship with silicon valley.

30

4 Comments

I went to an official Claude Code meetup in Toronto. Here’s what you need to know. The room skewed heavily technical, engineers, software developers, people deep in the weeds. That makes sense given where these tools were born. But it was obvious this won’t stay an engineering-only room for long. The biggest idea I walked away with: 2026 is shaping up to be the year of the orchestrator. The traditional way software gets built has already changed. Instead of spending weeks on requirements and planning, you’re increasingly just stating what you want to happen and then supervising the system as it does the work. The leverage isn’t in doing the work yourself anymore, it’s in breaking problems down and deciding how much to run in parallel. We’re moving past single-agent setups into multi-agent systems. One agent plans, another tests, another deploys, another reviews. You’re not micromanaging prompts anymore. You’re designing a team. The mental model feels much closer to hiring than coding. This is where democratization really shows up. As these tools mature, domain experts won’t need to wait in line for technical resources. They’ll build, test, and iterate directly. A lot of things that used to be black-boxed by function or department start to open up across the organization. Which brings us to the real constraint: human oversight. The question isn’t whether agents can do the work, they most definitely can. It’s how much one person can responsibly oversee at once, and how trust in LLMs grows as knowledge work keeps getting better. There were a few live demos that stuck with me:  - Claude as a Chief of Staff Mike Murchison, CEO and co-founder of Ada, showed how Claude essentially runs his operating system. It ingests Slack, email, docs, meetings (his entire life basically) and prioritizes his day based on company goals. Claude even acts as a personal CRM, creating and maintaining context for every person he meets, surfacing what’s changed, what’s gone stale, and what deserves attention now. - Autonomous agents that can “close the loop” If you want agents to work longer and better, they need the ability to validate outcomes. Seeing interfaces, interacting with apps, testing results,  autonomy improves when agents can verify their own work and iterate. - Security as a first-class problem If agents can act inside your systems, permissions and constraints can’t be an afterthought. Security architecture has to be embedded from day one, not bolted on later. If I’m honest, some of the conversations in the room went over my head as a less technical person. But, thats about to change. I heard from Saurabh Suri that more verticalized, role-specific meetups are on the roadmap for Anthropic, and you bet i’ll be there. If you’re in a GTM role and already experimenting with Claude Code (or thinking about it), I’d love to connect. Shoot me a message.

60

12 Comments

Continuing this series on technical interview questions. I pose this question to candidates who have spent meaningful time operating production systems. I like it because it tests understanding of how data, schemas, and assumptions decay over time—something that’s hard to appreciate if you’ve only worked on prototypes or short-lived systems. Q3: How do you generate reliable demo and integration testing data? What are some problems that you might encounter? A3: There are really only two approaches to this question: (1) generate synthetic data whole-cloth; (2) somehow use production data. Clever candidates will combine (1) and (2) into a hybrid approach. The problem with (1) is that functionality changes and any hand-crafted data will become stale. Such data may simply not exercise new functionality, or worse your database schema has changed and now the DB won't populate. So it's someone's responsibility to keep all the data fresh. Whose? There is no DX team at a startup. So the answer for Reffie would be "everyone and no one". But it’s also very defensible. Developers already encode edge cases and invariants when they write unit tests, and extending that work to integration and demo data is a natural leap. The data stays close to the code and the intent behind it, which means failures surface early and locally when assumptions change. While this can be frustrating under deadline pressure, it’s conceptually straightforward and avoids many of the harder-to-detect risks that come with production-derived data. (2) is tempting but raises complex challenges. (a) What do you do with sensitive customer data? How do you KNOW that a certain field contains sensitive data, how do you recognize it, and mask it (fields like credit card info are obvious but users can shove PII in unexpected places)? (b) What do you do with data that has external dependencies such as integrations? How do you recognize external dependencies at all? (c) What do you do with user credentials? Presumably you want to login as these users. Do you hard-code credentials or what? (d) Time-derived state and timestamps, including ordering and current user state. (e) This is the hardest part: preserving semantic relationships between records when IDs, ordering, and external references inevitably change. Notice that, especially for (2), there are no concrete answers here. That’s intentional. Experienced engineers understand that system design is about choosing which failure modes you’re willing to own. The key is making those tradeoffs consciously, not accidentally. Apply today: https://lnkd.in/gWqwj3C5

4

1 Comment

Openclaw joining OpenAI feels like the closest signal yet that the “one-person unicorn” era might not be fiction. If this model scales, headcount stops signaling capability, leverage shifts to agentic systems and architecture, and founders win on system design—not org size. We’re moving from “build a team” to “build a machine.” #AI #Startups #Founders #AgenticAI #FutureOfWork

2

3 Comments

The Confidence Trap (in the Gen AI workflows) Early wins: GenAI delivers useful results quickly → "Wow, this works!" Faster pace: Users start depending more on the tool, believing they’ve "understood" how to use it. Complexity creep: The scope of problems gradually expands  Expectation lock-in: Confidence builds not just in the output, but in the invisibility of the process (“I don’t need to check, it’s working”). Black-box shock: the model hallucinates, misinterprets context, or deviates. But because it’s embedded in a long chain, knowing where or when it broke is non-trivial Breakdown is hard to trace: No audit trail, no checkpoints, no versioning of prompt changes. And often, the failure is silent until it hits something critical — like a customer, a report, or a production system. This has been reflection from my own explorations - whether it's quick interactions like generating code snippets in Canvas, exploring ideas visually through images, or drafting copy and making small edits. And equally in longer, multi-step workflows like iteratively coding larger features like algo-trading, analyzing complex datasets, creating or crafting videos and narratives over time. Across all these use cases, I’ve felt how easy it is to slip into a false sense of confidence, trusting the output because it “worked last time.” As GenAI becomes part of our creative and technical flow, we need to build thoughtful guardrails around it, not to slow us down, but to keep us aware of where things might silently drift.

23

1 Comment

"Build a new model that makes the old one obsolete." 3 mainnet upgrades in 4 months. We're fully all in on payments and every upgrade gets us closer to gigagas and 100k TPS, the kind of throughput where Polygon is just the obvious choice for moving all money onchain. Thats where we're going and the shipping doesnt stop 🚢 https://lnkd.in/gmK4nZKR

27

3 Comments

Do LLMs perform better if you threaten them (as Google Co-Founder Sergey Brin claimed)? Controlled studies show that they perform better if you are polite! But it's not about pretending it has feelings. It's about understanding that every word in your prompt is a signal that shifts the probabilistic landscape from which the model draws its answer. It’s about probability, not personality. If you'd like to learn more, see my post here: https://lnkd.in/eWtKVTqT

6

3 Comments

We had the option of building our own SDK, and passed on it. There's some great players in the field like: Mastra, LangChain, LlamaIndex, CrewAI, LiteLLM (YC W23). It just didn't make sense to add yet another SDK. Here's what we did instead 👇: Develop a standard type-safe file format for communicating with, testing, and iterating on prompts/agents across any modality. We call this AgentMark. Then, we allow our users to load this into any of their favorite SDK's, run it via a CLI, or collaborate with SME's on our platform. AgentMark will soon support multiple languages, and we believe it can become a standard for developers and organizations committed to building reliable agents. It feels like we're just getting started. We have some pretty incredible updates coming in the next few weeks, so stay tuned!

15

🚀 Scaling to 20 Billion QPS: How Shopify Reinvents the Global Storefront When you’re the backbone of global commerce, "downtime" isn't just a technical metric, it’s lost revenue for millions of entrepreneurs. Shopify’s footprint is staggering: - Billions of users - Millions of merchants - 175 countries - 6 continents of operation - 24/7/365 availability requirements To handle this demand, Shopify is moving toward a future powered by a distributed, globally sharded database architecture with Yugabyte. The Goal: 20 Billion QPS 📈 Yes, you read that right. Shopify is planning for a massive scale of 20 billion Queries Per Second (QPS). Moving to a cloud-native, distributed SQL database allows Shopify to: 1. Operational simplicity: Having one logical database makes it much easier to manage and do schema changes on. 2. Scale Linearly: Adding capacity becomes a matter of adding nodes, not performing risky migrations. 3. Optimize Latency: By sharding data closer to the user, they ensure a "local" feel for a global audience. So excited to be along for the ride with such a determined team! #SystemDesign #CloudNative #Scalability #Shopify #YugabyteDB #DistributedSQL #TechInnovation #SoftwareEngineering Alan Caldera Brad Dietrich Rohan J. Thomas https://lnkd.in/gspFdZKV

35

YC removed Canada from their approved incorporation list, and it's the trending topic in the Canadian Startup Ecosystem. Canadian founders now have to (re)incorporate elsewhere to get funding (US, Cayman Islands, or Singapore). Canada is the ONLY removed jurisdiction. Michael Liu argues founders should be supported based on nationality, not incorporation. He's right. But here's what I think is getting missed. John Ruffolo just published a piece that explains the real driver: QSBS tax benefits for US angel investors. They need US C-Corps to qualify. Re-domiciling benefits them, not founders. What Canadian founders actually lose by flipping: * The capital gains exemption (LCGE) * Favorable CCPC stock option rules * Refundable SR&ED credits And YC still accepts Cayman and Singapore. If US domicile is so critical, why the exceptions? Here's the part that matters for policy. Even if QSBS explains YC's incentives, there are still real structural issues making Canadian incorporation harder. * CFIB data shows 256 hours per year lost to red tape (roughly a month). * Rounds that close in weeks elsewhere take months here (velocity gap). * SAFEs don't have clear CRA rules, so every deal needs custom legal work. * Each province has different securities rules, so a single round works across 10+ frameworks. * SR&ED credits require consultants to navigate, turning non-dilutive capital into an administrative burden. (<< this is the one I'm working on) If the tax treatment for US angels is driving this, transparency matters. But if there are genuine operational barriers, those are solvable. Let's make it happen. Wishlist: >> Canada needs to make incorporation here work as smoothly as Delaware. >> Clear federal rules on SAFEs. One national framework for early financing. >> Fast-track timelines at regulators. SR&ED without the high consultant tax. >> A stable founder-friendly tax regime. Singapore supports their people wherever they build. It's not apples to apples, but Canada can too. It's ok for founders to go south to get into YC. For some, it makes total sense, and we should support them. I've seen some of those founders raising $$MM and hiring in Canada. That also works and strengthens the ecosystem. And we can also do better to keep more founders building in Canada. It requires being honest about what's broken and actually fixing it.

16

5 Comments

Just shipped a new feature for Receipt Recall — snap a photo of any product and instantly know if it's safe to consume. No receipt needed, no manual searching. Just upload, and our AI cross-checks it against live Health Canada recall notices in seconds. Built with Google Cloud Vision, GPT-4o, and a recall cache that refreshes every 6 hours so you're always working with current data. Product safety shouldn't require effort — this makes it effortless.  #buildinpublic #foodsafety #nodejs #openai #healthcanada #AI

12

1 Comment

Andrej Karpathy nailed it: more context → better answers. But what if your context is locked behind OAuth scopes, consent screens, and brittle integrations? At Naptha AI, this was our biggest pain point building agents. So we built a better way in-house. We call it OneDollarOAuth. 💡 What’s the problem? LLMs are only as good as the context you give them. But getting access to your own data—from Google Docs, Notion, Slack, etc.—is surprisingly painful. Each OAuth flow is a mini integration project. Each provider is slightly different. It's a pain. To get OAuth working properly, you need to build at least 8 components: 1. Client credential management 2. Auth flow + consent screen 3. Callback handler 4. Token exchange 5. Token refresh logic 6. Scope management 7. State param + CSRF protection 8. Secure token storage https://lnkd.in/evShWeqh ⚙️ What we built So we packaged everthing and made it easier to use. It’s not a Zapier clone. It’s not an iPaaS. It’s just OAuth, done right—for LLM-native apps. And it's $1 per user per month. We built this because we were tired of reinventing the same plumbing in every AI app. Now we turned it into a product. If you’re building AI tools that rely on third-party context, this saves weeks of OAuth pain. 🔗 Try it here: https://lnkd.in/eBJ229RU

18

For decades, video surveillance has been about recording events — not interpreting them. Mehran Najafi, PhD and I will showcase how Generative AI is transforming video surveillance for security and operations at the upcoming AWS Summit Toronto.

17

1 Comment

I've spent the last year helping organizations across a wealth of industries redefine themselves vis-à-vis AI – their operating model, precesses, and mindset – and am super excited to chat with Jonathan, Kana and Stephanie in front of a live audience about their very own learnings and tips. Can't wait for this and so much more amazing content in Zartis' #AISummit2026 next week!

19

2 Comments

🧠 Tech leaders: Let’s be honest. AI is already reshaping how we build, lead, and work - and that changes the job market in Tech as well. Shopify's CEO Tobi Lütke told Shopify's leads what now VCs are telling startups: “No new hires unless AI can’t do the job.”  The effects of that are significant:  🔹 Entry-level roles? Vanishing as senior staff can now cover them with AI assistance. 🔹 Career paths? Collapsing as a consequence. 🔹 Companies? Scaling with smaller teams. The job market seems to be shrinking in Tech. ❓️ So what’s the best places to be at in terms of job perspective?  ‼️ It’s for sure the ones where you learn and build up your AI knowledge! There are a few options: 1️⃣ Obviously the big AI companies like OpenAI, Mistral or others. But there are only so many of them...  2️⃣ Companies leveraging AI - for their products or their processes. Not the big techs, but nearly all other industries! If you can help them leveling up their AI game, you have good job perspectives. 3️⃣ Yourself! Thanks to AI, starting your own business has become easier than ever. So go invent some new products or disrupt existing businesses. 👉 This is your edge. 📰 If you want to go deeper on this topic, the latest issue of my newsletter "The Tech Impact Uplift" is for you. You can access the newsletter from my profile or directly go to the issue from the comments.

55

14 Comments

Sarah Guo's piece on software abundance nails something important: when code generation gets commoditized, the bottleneck shifts. Not to writing code. To specifying intent, surfacing hidden assumptions, exercising judgment. But there's a piece that most of the abundance discourse misses. You can articulate intent perfectly and still ship something broken, because the agent didn't know what was actually running in production. It didn't know about the job that's been silently failing. It didn't see the edge case baked into the pipeline three months ago by someone who's no longer on the team. Organizational knowledge doesn't live in your prompt. It lives in your production system. The real unlock isn't just better intent specification. It's closing the loop from production back into the development cycle, continuously, so agents are operating on reality, not assumptions. That's how software is built - in iterations and reaction to feedback, until we get it right. That's why the agile methodology was such a game changer. That's what software abundance actually requires. Want to learn more about closing the production feedback loop? check it out: https://lnkd.in/dx9pYrPq

49

4 Comments

Two things about Shopify’s hiring process really stuck with me. 1️⃣ Even VPs have to code. Every engineering director and VP candidate does a coding interview. Not strategy. Not leadership philosophy. A coding interview. Apparently, many VP candidates are shocked. Shopify CTO Farhan Thawar explains it simply: “We believe they are still deeply in love with technology.” The point isn’t that leaders should code every day. It’s that the best engineering cultures are led by people who are still running toward the craft. When leaders step back into coding interviews, something interesting happens: The muscle memory comes back. And whether it does tells you a lot. 2️⃣ AI tools are allowed in interviews. Candidates can use anything. Cursor. Copilot. Claude. Whatever they’d use on the job. But Farhan isn’t just watching for speed. He’s watching how candidates think when AI fails. He described a candidate who ran into a single-character bug in AI-generated code. Trivial to fix manually. Instead of fixing it, the candidate kept re-prompting the model — hoping it would catch the mistake. Farhan stopped them and asked: “Are you really an engineer?” What Shopify is looking for: Engineers who are 90–95% AI-assisted. Not 100% AI-dependent. That last 5–10% matters. The judgment to: • step in • fix the line • understand what broke Speed is table stakes now. The real signal is whether someone can still think clearly in the moments where AI isn’t enough. 🔗

1

1 Comment

A concept that that we've really leaned into lately is reinforcing that the hard part used to be building software but now it's deciding what's worth building. The tools used to build are better than ever - AI, robust APIs, infrastructure at a moment's notice, CI/CD, generated documentation, the list goes on and on. These tools made execution faster but it didn't make decisions easier. If anything, it raised the stakes because you can build the wrong thing faster than ever. It's never been more important to be in tune with your customers and what they actually need to drive their business forward. That's what you need to build. 🔨

27