Building Affitor. AI-native affiliate marketing platform.
San Francisco, CA
Joined July 2016
- Tweets2.1K
- Following118
- Followers1.7K
- Likes2.1K
HUGE.
Full HUGE* If True episode on Neuralink and the future of brain computer interfaces
I toured @neuralink and got a close up view of their new R1 surgical robot, to see the future they and @elonmusk are building and how it is already changing lives...
00:00 What if you put a chip in your brain?
01:02 What does a human brain feel like?
02:17 How does your brain work?
03:58 Can a brain computer interface read your thoughts?
05:19 How do you get a chip into your brain?
08:07 What’s inside Neuralink’s implant?
10:06 How do you build a brain implant?
11:07 How does robot do brain surgery?
12:20 What could you do with a brain chip?
15:05 How are brain computer interfaces already being used today?
16:07 How are brain chips huge if true?
Past 30, I know which one title I would keep.
Not investor, not economist, not researcher, not marketer. None of those. I pick Engineer.
I like hard problems. I have never enjoyed easy work. I can work through the night and get up early the next morning, seven days a week, and I have done exactly that for a decade and a half. Not because engineering pays well or sounds impressive.
What I want to leave behind is not a number in an account. It is things that still run after I walk away.
If I could only pick one, I would still pick Engineer.
Affiliate programs never had a public quality score.
OpenAffiliate now scores them, and the formula deliberately ignores how much a program pays. 70 points go to publishing the terms an affiliate needs before committing: commission stated, how long it keeps paying, cookie window, minimum payout, payout methods and schedule, how joining works, promotion rules. The other 30 go to three terms that protect the partner.
A program paying 50 percent still fails if nobody can find the conditions attached to it.
1,258 programs are live, 566 verified against the brand's own terms and scored A+ down to F. Listing is free and position cannot be bought.
I canceled two of my four Claude subscriptions last month.
Opus 5.5 just made me renew, and I am adding three more. Same tasks, fewer tokens burned, longer useful sessions before the context falls apart. The math flipped in a single model release.
These accounts run production work that ships to real users, not experiments. When the cost per useful output drops this fast, you do not scale down. You scale up and ship more.
Everyone says agents will kill affiliate marketing.
I spent 10 years in this industry, founded 3 teams and ran a fourth, and I keep seeing the opposite. Every wave of automation stripped out the grunt work and multiplied returns for the people who adapted to it.
Agents finding winning products. Agents scanning competitor ads. Agents localizing content across markets. Affitor is the piece I am building to put that layer in the hands of creators and brands at once.
This is not the end of the affiliate model. It is the affiliate model running while you sleep.
Your last customer left no signal before buying.
They saw your post, scrolled past, came back weeks later, watched one video without liking it, then opened ChatGPT and asked about your product. Then bought. No click, no comment, no form fill. Nothing your dashboard would count.
The person who paid you never entered your funnel. They entered a conversation with a model you cannot track. Your attribution report shows a direct visit from nowhere, and your content report says engagement was flat.
The last mile of buying now runs through an LLM you do not own.
My long posts with media had a higher ceiling, not a higher middle.
This is for founders who write long posts on a Facebook Page and wonder whether attaching media is worth the effort. I pulled my own Page archive and kept only posts of 1,500 characters or more. 142 had media. 11 had none.
The typical post, the median, was identical in both groups. The gap showed up at the top: the best tenth of media posts reached more than twice the reactions of the best tenth of text-only posts.
Now the honest part. Eleven text-only posts is a tiny group, and one or two outliers could swing its top end a lot. This is one Page, observed after the fact, no control. It points a direction. It settles nothing.
What I do with it: I stopped expecting media to rescue an average post, and I still attach it when a post has a real shot at travelling. The media posts had more upside, not a better floor.
On my Page, long posts with media topped out higher, and that is all the data says.
Threads does not predict shares. It has no head for them.
This is for founders who cross-post to Threads and wonder why share bait that works elsewhere falls flat there. Meta's own system card for the Threads feed lists the prediction heads its ranker uses. There are 10. Seven are about tapping, clicking through, or time spent viewing. One is a like, one is a reply, and one is negative: scrolling past instead of engaging. None predicts a repost, a quote or a share.
That changes what a good Threads post is. The first line has to beat the scroll-past head, the only negative signal on the list. After that, the post has to hold attention and pull a reply. A post built to be reposted is aiming at a target the model does not score.
My own account matches this loosely: views track likes more closely than reposts and quotes. That is one account, observational, no control, so I hold it as a hint.
The card dates from March 2025, so recheck it before building on it.
On Threads, write for the person who stays and answers, not the one who passes it on.
LinkedIn dwell is not 15 seconds. It is a percentile.
This is for founders who post on LinkedIn to reach buyers or hires outside their network. LinkedIn published the paper behind its current feed ranker, a transformer called Feed SR. One of its two main objectives is Long Dwell: a classifier for whether time on a post clears a percentile threshold, for example the 90th, set per content type, ranking position and platform, and recomputed every day.
The 15 seconds people quote is in the paper, as an illustrative example. It was never a published threshold. And nothing published backs "a comment is worth 5x a like" either.
So the real question is whether your post sits near the top of the dwell distribution for its own format: a document against documents, a text post against text posts.
The second part matters more for reach. Out-of-network candidates come from a fine-tuned LLM dual encoder that matches post text to members. Retrieval reads your words first. Write "KV cache" and "inference latency", not "the AI stuff".
On LinkedIn, precise nouns are distribution.
The model writes the code. A different vendor signs it off.
My GitHub shows 1,974 commits across 22 repositories this month, and the busiest single day went past 700. That pace is survivable because of one rule I do not bend: whatever writes the code is never what approves it.
Claude does the building, on four accounts running in parallel. Codex, Grok and Cursor run review and QA before anything reaches production. Different vendor, different blind spots. A model is most generous about its own output in exactly the places it was already wrong.
Throughput stopped being the hard part. An honest second opinion is.
X publishes its ranking weights. The famous ones are not in it.
This is for founders who use X to grow a product. For years people have repeated that a reply is worth 13.5x or 75x a like. Those figures trace back to raw constants in a 2023 code release. X never published them as multipliers, and the signal behind the 75 does not exist in the current system.
The current production weights are public, and they reward different things. The biggest positive weight in the whole system is someone copying your post's link to send it elsewhere: 40x a like. A reply counts far more than a like, and it gets a further 4x boost when you and the replier follow each other both ways, on your original posts.
So the useful moves are plain. Write things people want to pass on outside the app, not things that farm a like. Reply to people in your niche, and follow back the ones who are actually your peers, because a mutual follow is where replies count most.
It is a snapshot of one commit, and the repo changes almost daily.
Stop optimizing for a multiplier X never published.
Son Piaz retweeted
Why the fear ? Why would Aliens ever want to invade Earth for our resources ? They are far more powerful and have technology far more advanced. They choose to study us , soul advancement 🩷
🤖 Made with AI
Most affiliate directories show you the rate. Not the takers.
We shipped the Affitor program marketplace: 57 affiliate programs from SaaS brands, and every card carries three numbers instead of one. The commission. How long the cookie lasts. How many partners actually joined.
That third number is the one nobody publishes. 50% with 6 partners and 12% with 80 partners are not the same offer, and without that column you are reading a pitch, not a market.
Sort by most partners and the ranking stops being marketing.
Start01 shows the work, not just the result.
This is for founders who want to build real products with AI and want to see what that looks like before they try it. Start01 is a series I recorded and streamed, and it covers the whole arc, from why to build to what you do once you can.
The why: how I market an AI product, how vibe coding turns into income, why AI agents let one person run a whole business, and where your edge sits once anyone can build software.
The build, from zero: three apps in one live session, building an app with AI instead of learning to code first, shipping an AI product fast with Claude Code, and why the fastest way to learn AI coding is to build something real.
What comes after: taking an AI product to international markets, the self-running system I built with AI, and one agent that manages other agents. That last one is the shortest video in the set and the place it took me longest to reach.
All 11 videos are free on YouTube. They are in Vietnamese: I am serving the Vietnamese market first, and English comes next.
Watching a product get built teaches more than another finished demo.
The link is in the reply.
I wrote a posting rule from 30 posts. 295 posts reversed it.
This is for founders who post to grow a product and keep notes on what works. In late September I compared 30 of my own Threads posts and concluded that adding an image lifted median views about 1.7x. It looked clean, so it went straight into my checklist.
A few days later I pulled every post on the account, 295 of them, and ran the same comparison. The direction flipped. Text-only posts had the higher median, by a small margin. My best post ever on that account is a short text post with no media at all.
Even the bigger set is observational: topics drifted, followers grew, and older posts had more time to collect views. A small sample amplifies all of that, and one lucky image post can carry it.
The fix is boring. One variable at a time, posts in pairs, and 20 to 30 pairs before anything becomes a rule. Until then it is a hunch, labeled as one.
A format rule built on 30 of your own posts is a mood, not a finding.