Brian Collins

I test AI tools so you don't have to.

[Samurai @samura1ij]

One HFT firm made ~$3M/day from being 3 milliseconds faster than everyone else.

To get those 3ms they laid a $300M fiber cable through mountains between Chicago and NY. Obsolete 6 years later when microwave beat it.

"Easy money" cost $300M in dirt and 6 years of exclusivity. That's the arb.

[Brian Collins @coll1nsb]

@samura1ij Turns out "consistently wrong" is alpha. "Randomly wrong" is just a tax.

[Samurai @samura1ij]

Two people invest the exact same $10,000. Same fund, same day, same 8% average return over 20 years.

One ends up with $46,000. The other ends up with $31,000.

Same math. Different outcome. The gap isn't returns — it's what happened inside their heads during the three years the number went down.

Most people think risk is about picking the right investment. It isn't.

It's about whether you can survive being wrong long enough to be right — and almost nobody can, which is why almost nobody gets paid for holding it.

The market doesn't reward the smartest person in the room. It rewards whoever doesn't touch anything for eighteen months while everyone around them panics.

Below is exactly what that looks like, up close.

[Brian Collins @coll1nsb]

@rozenhz The wildest part isn't even the $10B — it's that this has been public since 2014, went through congressional hearings, became a bestseller, and the mechanics didn't change at all. Not because it was hidden. Because the rules allow it.

[Brian Collins @coll1nsb]

One cent. Repeated billions of times a day. That's how a company nobody's heard of built a $43B fortune without placing a single directional bet.

There's a company that makes $10,000,000,000 a year without ever once betting on whether a stock goes up or down. It's completely legal. Here's exactly how it works.

A machine sees your order before the exchange does. Buys the stock a cent cheaper. Sells it back to you a cent higher. Billions of shares a day.

That's not a trading strategy — that's a toll booth built into market structure. Public since 2014, went through congressional hearings, hasn't changed since.

The part that gets me isn't the $10B. It's that everyone already knows exactly how it works

[rozen @rozenhz]

A stock ticks up two cents. Someone bought it. A millisecond later, someone sold it. Both trades were the same machine.

In between them: your order.

This is how a single company in Chicago earns $10 billion a year. Without ever betting on a stock going up or down.
Michael Lewis put it on 60 Minutes in 2014. The book became a bestseller. Congress held hearings. A new exchange opened just to fight it.
Nothing broke the machine.
Here's exactly what it does.
You place an order. A machine somewhere else sees it before the exchange does. Buys the stock a millisecond earlier. One cent cheaper. Two cents cheaper. Then turns and sells it back to you.
You never see the trade.
You never see the machine.
You just get filled a fraction late, at a fraction worse.
One cent. Two cents.
Billions of shares a day.
That's not gambling. That's a toll booth.
The company is called Citadel Securities. It handles more than a third of every retail trade in America. The man who built it is Ken Griffin. Forty-three billion dollars, and not one of them came from being right about a stock.
He doesn't predict where prices go. He gets paid every time prices move at all.
The stock market isn't a place where prices are discovered.
It's a place where prices are sold to you.
Fourteen minutes below. It will change every trade you place after.

[Brian Collins @coll1nsb]

@minchoi Feels like the whole AI agent hype cycle in one tweet — a domain worth $1M before anyone's confirmed the bot behind it actually works.

[Brian Collins @coll1nsb]

@minchoi Wild if true. Curious how many of these 10 hold up when someone else tries to replicate them, vs just looking good in a clip.

[Brian Collins @coll1nsb]

@3voai You're right!

[Brian Collins @coll1nsb]

ChatGPT vs Claude vs Gemini: two of these three QR codes are fake. Can you tell which?

I asked all three to write SVG code for a QR code that actually scans (encoding "AI TOOLS"). Not "does it look like one." Does it scan. So I rendered all three and ran them through an actual QR decoder (OpenCV) — no eyeballing.

ChatGPT: doesn't scan. A whole corner marker is missing from the code — not a rendering glitch, just not there.

Claude: scans. Decodes to exactly "AI TOOLS." And before giving me the code, it wrote its own encoder (proper error correction, finder patterns, mask optimization), rendered it, and ran it through OpenCV itself to check it actually decoded. It fact-checked its own answer before I did.

Gemini: doesn't scan either. All three corner markers are present. Looks completely correct. The data underneath just doesn't decode.

2 out of 3 "QR codes" that look totally legit are actually broken.

Would you have caught it just by looking?

[Brian Collins @coll1nsb]

@Vincent4T_ GROW

[Brian Collins @coll1nsb]

@Mannu0100l Hello!

[Brian Collins @coll1nsb]

Same question, three models: how many times does the letter "e" appear in "entrepreneurship"?
ChatGPT (free tier): 3 — wrong
Claude (Sonnet 5): 4 — correct
Gemini (3.6 Flash): 4 — correct, and unlike the other two, it actually ran code to check instead of guessing
Why the miscount: LLMs read in tokens, not letters — counting a specific character means reconstructing spelling from chunks, not scanning each one.
Getting this right doesn't mean a model is reliable on harder tasks — just that this one narrow check passed.
Which one do you trust with anything that needs to be precise?

Tested Aug 19, 2026.

[Brian Collins @coll1nsb]

@shakhzod1001 Claude writes better copy, ChatGPT counts better. Tested both on the same prompt this week — Claude nailed the pitch but got its own word count wrong by 8.

[Brian Collins @coll1nsb]

@ShazXcode This is literally me tab-switching between ChatGPT, Claude and Gemini praying one of them still has quota left 🪨🙏

[Brian Collins @coll1nsb]

@syakirbuilds Claude, a little. Not the writing — that's genuinely good — but ask it to self-check something like a word count and it'll confidently give you the wrong number. Tested it head-to-head with ChatGPT this week, same story.

[Brian Collins @coll1nsb]

@DrstaOne Tested this literally yesterday — same prompt, both models. Claude's copy was sharper but it miscounted its own word count by 8. ChatGPT was safer and actually accurate. If precision matters more than personality, that's the tell.

[Brian Collins @coll1nsb]

@Tech_HudsonLee Ran the same 60-word prompt through both yesterday. ChatGPT played it safe and stayed generic. Claude wrote a sharper pitch — but told me its answer was 58 words when it was actually 50. Depends if you need speed or precision, but don't trust either one's own math.

[Brian Collins @coll1nsb]

I gave ChatGPT, Claude, and Gemini the identical prompt: write a cold DM pitching a restaurant owner on an AI chatbot. Under 60 words. No fluff.
ChatGPT: safe and generic. 37 words. Leans on [brackets], no real reason to reply.
Gemini: skipped the call entirely — offered a 30-second demo video instead. 50 words, no fluff.
Claude: sharpest pitch of the three — real benefit, concrete timeline, actual CTA. But it told me the DM was 58 words.
I counted. It's 50.
Same brief, same word limit. Only one of three can count.

Which DM would you actually reply to?