Good morning, founders 🕶️

Two rules run through this edition: whatever you reward gets farmed, and whatever you automate, you still own. On Polymarket, a rumored token may have turned two finished election markets into $600M of volume. In New York, CarGurus is finding out that its chatbot's "deals" count as its own promises. And OpenAI just reminded everyone building on GPT-4o that a model is a lease, not a purchase.

Three stories, one blind spot: an incentive or a dependency nobody was watching. Which number on your dashboard would survive two accounts trading with each other?

In today's newsletter, we'll get into:

Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.

Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.

Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.

Try Notetaker Free

This summer, Ethiopia's prime minister Abiy Ahmed won in a landslide: 438 parliamentary seats to the runner-up's 13. On Polymarket, traders put $173,000 on him becoming the next leader, and more than $41 million on his chief of staff, Shimelis Abdisa. That's 200-plus times more money on the man who didn't get the job.

The receipt: two election markets on Polymarket's offshore platform, Ethiopia and Romania, did $600M+ in volume, mostly on long shots who lost. Of Ethiopia's $285M+, about $284M came from pairs of accounts trading with each other, says Chris Park, an ex-Polymarket trader who now runs forensics firm MSR Decode.

Not quants chasing pennies. MSR Decode traced 38,000+ "trade loops" after the vote, and over 98% ended with neither side up or down. Median hold: nine seconds. The market only resolved after the October swearing-in, so it traded for 123 days after the election, a "zombie market" in economist Rajiv Sethi's words. Users asked: "THIS NOT RESOLVED?" In Romania, Sethi's algorithm flagged 99.6% of some weeks' volume as wash trading.

Probably a token that doesn't exist yet. In 2025, Polymarket's CMO told a crypto podcast there "would be a token." If allocations reward volume, a zero-profit, nine-second round trip isn't bad trading. It's a job application. Wash trading breaks Polymarket's terms, and the company says it flagged the activity and blocked the accounts. Sethi's caveat: "What appears to be wash trading isn't always wash trading." Nobody knows who the traders are.

Whatever you reward with points, tokens, leaderboards or referral credits will get farmed, and the metric dies with it. Pick the number you'll pitch investors on and make sure two accounts can't fake it by trading with each other. Build self-dealing detection before launch, and shut down stale campaigns, promos and markets on schedule.

If a Polymarket token ever ships, its allocation rules will tell us what those nine seconds were worth.

On Wednesday, October 14, OpenAI notified developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave the API on January 14, 2027. Every fine-tuned model built on them goes the same day. The recommended replacement for all four: gpt-6-luna.

That's 92 days. OpenAI's own deprecations page promises generally available models at least six months, unless safety or compliance forces a faster timeline. The entry cites neither. Two weeks earlier, on the same page, GPT-5.1 got its full six months.

Because the clock is shorter than it looks. Subtract the holidays and you have roughly ten working weeks. And swapping the model ID is the easy part: per The New Stack's read of the migration guide, Luna only supports function calling in Chat Completions with reasoning switched off. Reasoning plus tool calls means moving to the Responses API.

Treat any model like an apartment on a month-to-month lease. Put a routing layer between your product and the model, so the next deprecation email is a config change, not a fire drill. OpenAI says some customers can buy dedicated capacity to keep a model past shutdown. That's a sales call, not a plan.

Watch for pushback. 4o's fans got it back in ChatGPT once before, a day after GPT-5 replaced it.

Scale your growth: Join names like HubSpot, Ahrefs, and Deel. Put your product in front of 85,000+ startup founders. Let's talk.