
Dissonant and obsessive try-hard. Founded @getcognito (acquired by @Plaid)
San Francisco
Joined August 2012
- Tweets6.6K
- Following1.7K
- Followers11.6K
- Likes19.6K
Pinned Tweet
In 2014, @alain and I dropped out of Stanford so we could run @getcognito full time.
After almost a decade of hard work, I’m extremely excited to share that @Plaid has acquired Cognito for $252 million!
tcrn.ch/3rCV2vS
Being James Bond and/or a ninja is considered illegal in California.
Niche weapons you aren't allowed to possess:
- Knives hidden inside belt buckles, lipstick cases, pens, canes, or tire-pressure gauges
- Ballistic knives, where the blade can be ejected from the handle
- Brass OR wooden OR hard-plastic knuckles
- Leaded canes (heavier, for striking)
- Blowguns (so you can own a real gun, but not a blowgun or blow darts??)
- Replica grenades unless they're filled with resin
- Throwing stars
- Basically anything with a heavy part for hitting, like a "slungshot" (a rope with a metal-weighted ball), or compact sandbags you could use to hit people
You CAN own nunchucks though! They were re-legalized in 2022.
Logging into Venmo in 2026 like
ME: Hello I want to—
VENMO: Open a debit card? Credit? Deposit into Kalshi? Buy Crypto? PayPal stablecoin? Please? Buy tickets with TicketMaster? Get an NYT subscription?
ME: Pay my friend!
VENMO: Ugh, fine. Good luck, we barely maintain that.
AI coding tools today remind me of Betty Crocker introducing a "just add water" cake mix in the 1940s which didn't sell well until they tweaked it to "just add an egg," making consumers feel like they're still baking.
Still in egg phase now, but it sure feels like "just add water" is right around the corner. Psychological toll this will have on builders will be profound.
POK(AI)MON: Experimenting with GPT 5.4 autonomously editing and rewrite the Pokémon Red ROM, replacing Pokémon with AIs.
Details below
Readers added context they thought people might want to know
The game being showcased is a simple hack of Pokémon Red, not an AI recreation.
The game code has been fully disassembled and tagged by humans, and creating such a rom would require less than half an hour of work of a single person.
Here's the code and assets on github:
github.com/pret/pokered
I've been curious when AI will let us just hack ourselves into video games and rewrite the gameplay. The GPT 5.4 announcement has a lot on games, so I decided to give it a try.
FIELD NOTES
[Most impressive part] Codex was able to autonomously QA it's changes by setting up a browser-based Gameboy emulator and then interacting with it. After some trial and error it figured out how to generically QA the game.
Followed the playwright-interactive + imagegen codex skills OpenAI demos in their blog post
- Some things worked perfectly with no guidance like the banner art at the top.
- Other things were hilariously bad. It tried to make the AI lab logos just text of their name. No amount of steering addressed this, I had to give it the logos and say use these (then it did all the processing)
Replacing characters (e.g. Prof Oak → @karpathy) worked one shot, but didn't look like them. Had to give it a reference image.
A lot of this was actually sprite editing which Codex did a pretty good job consistently one-shotting which is tricky given resizing, greyscale, building sprites, etc.
This was not a reverse engineering task. I just dropped Codex into the pret/pokered GH repo and gave high level instructions.
More of a fun little art project by the end than anything, but feels pretty close to being able to just describe how you want to modify a game and letting an AI hack it together. A really fun future project would be to do something similar to hack yourself in as a character into Super Smash Bros.
Twitter's Eng org used to be TWO THOUSAND engineers.
Now it's ~25 engineers and a few designers and PMs.
I knew the cuts were big but this is blowing my mind.
I thought I had 10 more years with my Dad.
He treated his CLL so effectively the last decade, sometimes I forgot he had cancer.
An infection took his brain so quickly that by the time my red eye landed we could only have simple yes or no conversations. Within 48 hours, it was a coin flip whether he would be able to recognize me. A week later, our last flicker of hope went out and we said goodbye.
There was no warning. No heads up to get some quality time in.
Why didn’t I call him every week? Why didn’t I ask him to show me pictures of him from when he was my age and explain that phase of his life? Why don’t I have more photos of him? With him?
I’m angry at myself for all those moments I wasn’t present with him. I’m angry at the world for how suddenly he was taken. So early. Six months before my wedding. Never able to meet my kids. To be able to call him “grandpa.” I couldn’t wait to give him that, and now I can’t.
The last few weeks I’ve been desperate for control. I couldn’t control the bureaucracy that denied him a bed at the specialist hospital he donated to, even when we paid for a private medical plane to transport him. It hurts even more that, in the end, it wouldn’t have made a difference.
I can’t control what happened to my Dad at this point, but I can control how I move forward. I won’t make the same mistakes again, Mom. I won’t let the family go without a strong leader they can rely on. I won’t let my kids grow up without knowing who you were, Dad. Your energy. Your overwhelming optimism. Determination. Work ethic. How your hard-driving intensity turned into a distinct tenderness in your later years.
I will continue to make you proud with what I accomplish, even if I can’t experience the joy in seeing the pride in your eyes anymore. I’ll pass on the values we shared. I’ll give my kids everything, just like you did.
Talk to your parents.
Cybernetics predicted tech would supercharge our brains, enhancing our memory, focus, and cognition.
This came true in a "monkey's paw" cursed way that also degraded our brains:
• Search nerfed our long-term memory
• Tiktok, our attention span
• AI, thinking for ourselves
Announcing Ti2: Plaid’s latest Trust Index model, trained on 2x the data of its predecessor and catching 30% more fraud!
Ti2 uses bank transaction history and a user graph spanning Plaid’s vast network, enhancing fraud detection for synthetic identities and account takeovers.
We’re shipping all of this just 4 months after launching Ti1!
This model’s capabilities are growing fast, with more coming, from graph neural networks to AI agent analysis.
Every ID verification system will need a strong fraud model. Check out Plaid Protect and Trust Index.
More details in this blog post and the video below: plaid.com/blog/plaid-protect…
Love tuning into @cowenconvos to hear Tyler ask world experts why AI can't do their jobs lmao.
Every other interview is like "Why can't AI do neurosurgery?" "Why can't o1-pro run the Fed?" "When will all state capacity of the Peruvian government be handled by AI?"



