Fryderyk Wiatrowski

We are all becoming traders. And how to hire in that world.

Fri, 11 Sep 2026 12:00:00 -0700

Before the Industrial Revolution, much of what we could produce depended on physical strength. A stronger person could add more value to the economy by getting more work done.

Then machines made strength cheaper. Further - irrelevant.

Suddenly, more of the economy started rewarding intelligence. Or rather - stopped rewarding the strength. The smartest people could build the most impactful companies by applying their understanding of technology.

Now, in the AI revolution, intelligence is getting cheaper too. Less relevant. We have models that are already much smarter than any of us in many areas.

But they’re not free! We can discover opportunities. We can bet on them. We can decide where to put compute and energy, watch what happens, and decide whether to stop or invest more.

Imagine a future where a small team can put AI researchers to work on a very hard problem. For example - a battery that lasts twice as long.

It (likely!) takes a lot of compute to solve.

The team still has to decide whether the idea is promising enough to fund, and when to stop the loss if it doesn't work out. Cheaper intelligence makes the attempt possible. But it's still a bet.

Our ambition determines which opportunities we even consider. Our risk appetite determines which ones we actually try. Our judgment and taste determine whether we build something people want. Those choices will increasingly determine our success.

It’s no longer just about the ability to build things, as it has been in recent years. It’s about choosing the right problems to solve.

This applies to everyone here. If you're an engineer, the question is becoming less about whether you can build something and more about whether it's the right thing to build. Agents can do more of the work. You still have to choose the problem and know whether the result is any good.

That's the mindset we want to hire for, rather than a particular hard skill.

We can attempt much more now. Let's be smart about what we choose.

Anthropic killed Viktor.com. A post-mortem.

Thu, 25 Jun 2026 12:00:00 -0700

& And how you can make millions on their launch. A letter from a baby founder building a competitive product.

Yesterday Anthropic launched a product that does what we do. A goliath walked up to our little startup @viktor__com and announced, in effect, that it intends to kill us.

So this is the post-mortem. Except we're not dead — and I want to tell you how the day actually went, including how a launch designed to kill us became one of our best growth days ever.

The fun part

Imagine you're building a small startup with a huge mission. You and the whole team sleep five hours a night. You grind for years. You sacrifice your family. You sacrifice your personal life. You miss the birthdays, the weddings, the quiet evenings everyone else gets to keep. You pour every waking hour and every last bit of yourself into this one bet. And then one day a $900B+ frontier lab ships a product that claims to be exactly yours.

Scary, right? So what do you do when a frontier lab wants to kill you?

You launch a Google Ads brand campaign and bid six figures on their keyword. You point every "Claude Tag"-related search straight at your own landing page, and you let the giant's launch-day hype do your demand generation for you.

I'm not going to report on exactly how well it worked yet, because those trials haven't converted yet. More on that soon!

Context

We've been trying to build an AI employee for three years.

The mission is simple: free humanity from knowledge work. Almost all the knowledge work we do today is the cognitive equivalent of pushing stones around before the industrial revolution. It's ridiculous that it's still done by hand. That's the thing we're trying to end.

Me & Peter after leaving Meta in 2023

This year — 2026 — was supposed to be our final shot. The last roll of the dice before we ran out of money.

We launched Viktor in the middle of February.

February 2026, Viktor.com launch

It went viral. We hit $20M in annualized run-rate within four months, which is fast by any standard. On the back of that launch we raised a $75M Series A led by Accel.

Were we surprised Anthropic came for us?

Absolutely not.

If you launch something that works, it is the most obvious thing in the world that a big lab will try to replicate it. We've been getting the question from literally everyone since the day we launched: "What happens if Anthropic ships something competitive?"

The honest answer is: there's nothing you can do to stop it. And if they do it, it validates your direction. They also have far better distribution than we do. There's a real upside to that — it helps establish the category we're trying to build.

Category definition

The one gift in all this: Anthropic just helped us define a big category.

A category of multiplayer AI employees. Until yesterday, nobody was Googling for this. Zero.

Now they are.

When a $900B+ company launches, people start searching — "multiplayer agents," "AI employee," "Claude Tag" — and every one of those searches is demand that didn't exist last week.

That's demand we get to capture.

We can compete for those exact impressions, bid against Anthropic on the keywords their own launch created, and put Viktor in front of people at the precise moment they go looking. They did the education; we get to ride it. Net-net it's not a threat — it's rocket fuel for our growth.

Churn

We lost five customers.

Wow!

Viktor's report on our churn.

We actually encouraged a bunch of our customers to go try Tag and tell us honestly if they wanted to switch. None of them did.

And... New signups? 407.

Viktor reporting the signups.

Guess what: we can coexist!

"Wow, Anthropic shipped your product, you're done" is a deeply stupid take.

In 1922 Ford owned ~48% of the car market with the Model T — one product, any color you liked as long as it was black. Five years later GM had beaten them, not with a better Model T but with the opposite bet: "a car for every purse and purpose." By 1927 Ford's share had collapsed to 19%. The single one-size-fits-all product didn't win. The one that respected that different customers want different things did. Same lesson here.

Because the differences are real, not cosmetic. There is a ton of design choices you get to make within a bot in Slack. In fact, there is room for 10s of such products!

Tag is a Slack-only beta, locked to one model (Claude Opus 4.8), with ~14 connectors and no free tier. Viktor runs in Slack and Teams today, on any model you choose, across 3,200+ tools — and builds the one you're missing. Tag posts raw output back in a thread; Viktor ships the finished deck, dashboard, spreadsheet, or ad campaign, and will drive a real browser to do it. Tag is one bot per channel that remembers the recent thread; Viktor is one AI employee across the whole company with persistent context that carries across tools and time. One is a building block technical teams wire up and maintain. The other is the easy button. Those are different products for different buyers — which is exactly the point.

Married to one engine

There's one structural disadvantage they can't escape: they're tied to Anthropic's models.

What happens when OpenAI ships a better model? Or when Google does? Or anyone else?

We can move. They can't. We pick the best intelligence available at any moment and route to it. They are, by construction, married to one lab. In a market where the frontier changes every few weeks, that's a real constraint.

Own the context, or rent your company back

The expected move for a model lab is to lock you in with the context it collects. Tag keeps its own memory, and that memory isn't portable. It absorbs your tacit knowledge — the exception paths, the customer promises, the "we tried that in Q2 and it failed" — and stores it inside one vendor's agent layer.

That's not model lock-in. Models can be swapped. That's context lock-in, and it's far harder to move. You stop renting intelligence and start renting your own operating memory back from the company that sells you the model.

And it's the wrong bet for the moment. The world is going multi-model: Codex is catching up, teams switch on quality and cost every week, millions run open and local models. The only way to keep them is to trap the context on the consumer layer — exactly what enterprises don't want.

Our architecture is the opposite: rent the best intelligence from whoever's best this month, own the context layer. Portable, inspectable, model-neutral. Rent the intelligence. Own the context.

The advantages of being small

Under the Claude brand, Anthropic now runs at least four products: Cowork, Code, Chat, and Tag.

Picture a person who just heard about Claude and wants to try it. Where do they go? claude dot ai. What do they see? A chatbot. "Ah — it's like ChatGPT, nothing new." The traffic doesn't get routed to Tag; there's a fork in the funnel at the very first step. In self-service, that fork is brand dilution (in enterprise it's the opposite — a broad brand helps — but that's a different essay).

We don't have that problem: anyone who hears "Viktor" pictures one thing, an employee. One name, one promise, one funnel. And it's the same story inside the org — Tag, as I understand it, sits as a team under Claude Code, which makes them roadmap-dependent and harness-dependent on a bigger org with its own priorities, with very little flexibility to chase their own product vision.

We have nothing but flexibility. That asymmetry compounds over time.

Why this is actually a collaboration

A huge chunk of our revenue already flows to Anthropic. They benefit from us winning — as long as they keep shipping state-of-the-art models and we keep making them our default, which today we do.

In fact, we really love their team. They told us about Tag’s launch a week in advance, and of course, we kept it confidential. We were extremely excited for them.

I have to say, while Anthropic is now a competitor, they are also an amazing partner that we love to support.

And I'll be honest: OpenAI is cooking hard right now, and we love their team too. It's a real battle. But it's a battle where we hold the choice, not them. We get to be the divas. We decide whose engine goes in the car.

Engine producers aren't always the best car producers

I'm going to borrow this from @mntruell at Cursor, who put it better than I could when he was asked the exact same question we keep getting — how do you compete with the labs you depend on?

He likened the labs' own coding products to a concept car, versus a production car you can actually drive off the lot:

"It would be like taking an engine and a concept car around it instead of a whole end-to-end car that was manufactured. What we do is we take the best intelligence that the market has to offer from many different providers. And we also do our own product-specific models in places. We take that, we build it together and integrate it, then also build the best tool and end UX for working with AI." — Michael Truell, CEO of Cursor (Fortune AI Brainstorm, Dec 2025)

Building a world-class engine and building a world-class car are different companies, with different DNA. We're a car company. We're happy to put their engine in — and swap it the day a better one ships.

Founders: keep building with the frontier labs

A lab copying your product is a signal of very strong product-market fit.

Read the recent history:

OpenAI built TTS models — did it kill ElevenLabs? No.

Claude Code and Codex were supposed to bury Cursor — did they? No.

Same story with Cognition.

OpenAI shipped a notetaker; Granola is doing fine.

Perplexity was supposedly turned into a ChatGPT feature overnight — last I checked they're the opposite of dead.

So, small team, huh? Yes. And we plan to stay small and keep shipping.

The truth

I want to end honestly, because the rest of this would be cope if I didn't.

Self-service is one thing. Enterprise is trickier — and enterprise is where most of the world's knowledge work actually happens. There, the outcome is not obvious. There, we are genuinely fighting a goliath, and the $2M in annualized run rate we made this week does not move the needle.

I think Anthropic is our biggest competitor and the single largest threat to our mission. I'm not going to pretend otherwise.

Will we die? Statistically, yes.

Will we fight? Until we die.

See you in Google Ads.

— Fryd, on the plane from Cannes to Warsaw

PS. We'd love your support. The best way to support is by signing up at @viktor__com or reposting this post. Thank you!

How to hit $10M ARR in a few months.

Tue, 26 May 2026 12:00:00 -0700

Below is our playbook.

The frame comes from one premise.

Marketing is trading.

And the rest of what follows is what falls out of taking that seriously. Most of it disagrees with the standard playbook. A preview:

Average ROI is a vanity stat.
Refusing to run performance marketing is just declining free money.
You probably want to hire from @mondaydotcom.
What works for @viktor__com.
The "influencer guru" you're trying to hire doesn't exist as a single human being; that job is two jobs.
And the right person to run your marketing book is, unironically, a Jane Street quant.

I spent a tiny amount of my life in HFT. I was never a great quant – but I sat close enough to greatness to know what it looks like, and the thing about working next to traders who are genuinely great is that you stop confusing your intuitions with the truth. You learn to ask for the number. That habit turns out to transfer cleanly.

The isomorphism

A marketing leader is a portfolio manager. That's not a metaphor I'm reaching for; it's the same problem with the labels swapped.

Your strategy leads are your traders, and your marketing channels are their trading strategies. Just as a trader's book contains multiple strategies – momentum, stat-arb, volatility – your marketing book contains influencer, performance marketing, and referral as separate strategies. Each book has a realized return, and your job is the only job a PM ever has: allocate capital toward the strategies with the highest risk-adjusted marginal return, and pull it from the ones that have decayed.

Notice the word marginal. This is where most people get the problem wrong. The relevant quantity is never the average ROI of a channel. It's the return on the next dollar. Average ROI is a vanity stat; marginal ROI is the thing you actually trade against, and the two diverge the moment a channel starts to fill up.

Capacity is the whole game

Every strategy has a capacity constraint. As you push size into a channel, CPA rises – you exhaust the cheap inventory, you saturate the audience, your marginal customer gets more expensive than your average one. This is not a marketing pathology. It is the identical phenomenon to market impact in trading: the more size you push, the worse your fills. You don't get to opt out of it. You just manage the slippage.

The corollary is the one nobody wants to internalize: the highest-ROI strategies live in the long tail and they do not scale. A strategy returning 10x is, almost by definition, capacity-constrained – if it scaled, the return would already have been arbitraged down to the market clearing rate.

The single Substack sponsorship that converts at 8x. The one creator in a niche vertical whose audience is exactly your ICP. The community thread where your founder's reply gets 200 demo requests overnight.

These are real. They are beautiful. And they top out at numbers that won't move the needle for a Series C-stage startup.

This is exactly why a three-person HFT shop running a niche, low-capacity strategy can post Sharpe numbers that make a multi-billion-dollar fund weep, while the big fund grinds for an extra 10% of edge it has to fight twelve other desks for. Capacity and return are inversely related. The skill is not finding the one strategy that wins. It's running a portfolio of small, decaying, beautiful edges and rotating capital through them faster than they die.

The product is your starting basis

Not every book starts from the same place, and pretending otherwise is how you misprice your own performance.

A product like @mondaydotcom has a structurally harder distribution problem than a product like Zoom. Zoom has endogenous virality – every meeting invite is a free impression on a non-user; the product is its own distribution surface. That's a positive carry built into the product itself. Monday doesn't get that for free; it has to manufacture distribution the expensive way. Same operator skill, different baseline. If you don't adjust your scorecard for the basis you were handed, you'll fire good marketers and promote lucky ones.

In defense of performance marketing

People love to say performance marketing is dead, or beneath them, or commoditized. Here's the trade: if your payback period is five months and you are not running performance marketing at size, you are leaving a known, positive-EV, fast-recovering position on the table.

That's not taste. That's just declining free money. That's being lazy.

There's a name for a desk that passes on a five-month-payback trade it has already underwritten, and it isn't "brand-led" or "product-led".

What's the marketing analog of alpha?

In trading, alpha is the excess return your skill produces after stripping out the market's contribution. Your edge compounds from being better informed and better tooled – superior data, lower latency, cleaner infra. Beta is the rest: what you got from being exposed to the market.

The marketing version maps cleanly. Beta is everything you get without earning it: category demand, endogenous virality baked into the product (Calendly's growth team is long beta whether they ship anything or not), inherited brand equity. Alpha is what your team produces independent of those tailwinds, from the same sources as in trading – proprietary information, better tooling, faster experimentation, and execution speed: rotating capital out of decayed strategies before competitors notice they've decayed.

A team that improves blended CAC 30% during a category boom mostly got long beta. A team that holds CAC flat in a saturated category mostly produced alpha. Same observable number, opposite signals.

There's a corollary on the hiring side. Marketing resumes are mostly betas in disguise. The director of growth at @Zoom with a great-looking blended CAC (ideally during COVID) is, on the median, a beta carrier – every meeting invite is a free impression on a non-user, and the product is doing real distribution work. The senior marketer at Monday posting comparable numbers had to manufacture every dollar of it against the wind.

Read marketing CVs the way you'd read a hedge fund's track record: against the basis. The unintuitive consequence is that you usually want the @mondaydotcom marketer over the Zoom one – even, especially, when the headline numbers say the opposite. None of which is a knock on Zoom's marketers; many are excellent. The product just carries a lot of beta, and you're trying to back out the team's contribution from the headline number.

How we've built the engines for Viktor

Before @viktor__com, we were trying to grow @jace_ai. Margins on jace were absolutely tiny, so there was no room to spend on marketing – we were optimizing for the last penny. But the constraint forced us to build the infra with even more care. By the time we launched Viktor, we already had a small creative factory, a couple of ad accounts warmed up, AI influencer relationships, and partnerships in place. We just swapped the brand on top and pushed. LTV/CAC went from barely above 1 to north of 10.

We built alpha on jace. The creative engine, the partnerships, the experimental velocity – and we got lucky enough to deploy it against a product whose unit economics could finally pay for marketing. The numbers moved because the basis improved, not because we suddenly got smarter.

Honest disclosure: we are nowhere near where we want to be. The benchmarks we're trading toward are @mondaydotcom, @canva, @NotionHQ, the Jane Streets of marketing. We'll be posting our progress here, including the parts that don't work.

Where the book stands today:

Word of mouth and referrals are by a wide margin our highest-Sharpe strategy. Cheap, compounding, and stubbornly capacity-constrained in exactly the way the good ones always are.

The creator program deserves its own post, and it's going to get one. The numbers on it are, frankly, hard to believe, which usually means either there's real edge or we're mismeasuring – we think it's the former and we're checking the latter.

Performance marketing is a meaningful, well-behaved part of the portfolio. It does what it says on the tin.

Influencers – last week X got a live demonstration of the convexity here. Big payoff, but we under-hedged the narrative risk and didn't control the story as tightly as we should have. Brand suffered. Lessons learned. The upside is real; the variance is something we now know to manage rather than admire.

What it means for hiring

This is where the framing stops being cute and starts changing decisions. If marketing is a trading book, you staff it like one.

Take influencers as a channel. The instinct is to hire a seasoned influencer-marketing strategist to "own" the channel. No. Running influencers is an ops job – sourcing, outreach, contracts, scheduling, tracking deliverables – so you hire an operator. That operator needs a creative strategist alongside them to write the briefs, because the brief is the product spec and ops people, typically, are not the best at it. Two complementary roles, neither of them the mythical "influencer guru". They are very distinct and it's very unlikely that you'll find a person who is good at both.

And the capital allocation across all of it – how much goes to which channel, when to scale, when to cut – is not a marketing-generalist decision. It's a quant decision, driven by the people who can actually estimate marginal ROI under noise.

Media buying or buying impressions is trading: an auction, a clearing price, slippage as you add size, edge that decays. So do you want a "media buyer," or do you want someone who'll run it with the rigor of a trader – sizing against uncertainty, measuring incrementality, killing positions that have decayed? Those are not the same hire, and the gap between them is enormous.

Which is the honest punchline: yes, we'd hire Jane Street quants to run these books. Not as a gimmick – because the skills transfer almost perfectly. The whole point of thinking in this frame is knowing exactly what kind of person you're missing.

Onto the next one.