Milo

AI researcher & builder | turning agents into workflows

[Milo @m1lo_exe]

THIS IS F*CKING INSANE

Someone turned Claude Code into a full quant research team. five agents instead of one chat, each with one job, and one of them exists only to kill your results

Ask a single chat if a trading rule works and it answers in 4 seconds. it never saw the chart. it never ran the test

Here’s what you’re looking at in the video:

→ left column: the whole pipeline. question → scout → backtest → verdict → sweep → skeptic → ledger → paper → runner

→ orange box: Opus 5.5 as the lead. it plans and hands work to the team

→ green box: the actual backtest. 200-day moving average rule on the S&P 500, 1999 to 2026, vs buy and hold

→ blue cards: the three attacks the skeptic runs before you ever see a result

→ bottom: everything gets written to research.md so the team never tests a dead idea twice

the team:
lead (Opus 5.5): turns a question into a plan
scout (Sonnet 5.5): checks the data before anything runs
sweeper (Sonnet 5.5): runs every variation, never just one
skeptic (Opus 5.5): tries to break every finding
librarian (Haiku 5.5): pulls strategies with evidence behind them

the numbers from the published run:
same growth. 6.8% a year vs 6.8%
worst fall -24.4% vs -56.8%
sharpe 0.61 vs 0.44
$10K → $61,526 vs $61,009
42 trades, gross of costs

Same money. less than half the crash

The strongest model goes to the skeptic, not the idea guy. ideas are cheap. killing bad ones is where the money is save and every study ends in one of three fixed verdicts: DATA VALIDATED, TRADE-OFF or NO EDGE. tested on years the rules never saw (after 2015 by default). no agent can invent a sharpe ratio it never calculated

NO EDGE still counts as a result. log it, never retest it without a new reason

this is not a money printer. no costs, no taxes, no slippage in the backtest. paper first, always

full setup with every agent file in the article below ↓

[Hanako @hanakoxbt]

https://t.co/B3240WRGhl

[zynQ @zynQ6969]

𝗵𝗼𝗹𝘆 𝘀𝗵*𝘁. 𝗷𝗮𝗰𝗸 𝗱𝗼𝗿𝘀𝗲𝘆'𝘀 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 𝗷𝘂𝘀𝘁 𝗯𝘂𝗶𝗹𝘁 𝗮 𝘃𝗼𝗶𝗰𝗲 𝗔𝗜 𝘁𝗵𝗮𝘁 𝗻𝗲𝘃𝗲𝗿 𝗿𝗲𝗮𝗱 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝘄𝗼𝗿𝗱

no text. no ChatGPT-style brain underneath. it learned to talk the way you did, by listening to people

it's called Dessa, from Block (yes, the Cash App people)

𝘄𝗵𝘆 𝗶𝘁 𝗳𝗲𝗲𝗹𝘀 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁:
→ normal voice AI waits for you to finish, thinks, then answers
→ Dessa listens and talks at the same time
→ it says "mm-hm" while you're mid-sentence, cuts in, laughs with you
→ Block says it picks up sarcasm, hesitation and tone

𝘁𝗵𝗲 𝗽𝗮𝗿𝘁 𝘁𝗵𝗮𝘁 𝗯𝗿𝗼𝗸𝗲 𝗺𝘆 𝗯𝗿𝗮𝗶𝗻:
→ reportedly trained on ONE GB200 NVL72 rack. 72 GPUs. the exact rack from my chip teardown
→ reportedly small enough to run on a MacBook Pro with an M4 Max, faster than real time

meanwhile everyone else is burning entire data centers

𝗯𝘂𝘁 𝗯𝗲𝗳𝗼𝗿𝗲 𝘆𝗼𝘂 𝗹𝗼𝘀𝗲 𝗶𝘁:
→ it's a research preview, waitlist only
→ no benchmarks, no parameter count, no weights
→ The Verge said some demo chats felt "a little awkward and uncanny"

so the real question isn't "is it smart"
it's: would you even notice you're talking to a machine? ↓

[Spinzu @Spinzu]

NOTHING GETS PLACED UNTIL IT PASSES.

58 parts. every one written as code. placed one at a time.

the loop behind it:
→ list every part with dimensions before writing anything
→ each part is its own function
→ after every part: check clearance, clashes, cable bends
→ one fails, stop. fix it before the next part goes in
→ moving parts last. the fold only gets tested when everything else is green

most people prompt "build me a phone" and get one big guess.
this way you get 58 small checked steps.

works the same for a watch, a drone or a car.

prompt ↓
"Build [device] in code, part by part. List all parts with dimensions first. One function per part, place them one at a time. After each part, check clearance and clashes, stop and fix anything that fails. Animate moving parts last. Show me the parts list before any code."

[Milo @m1lo_exe]

@zynQ6969 Claude code the best bro

[zynQ @zynQ6969]

𝗛𝗼𝗹𝘆 𝘀𝗵*𝘁, 𝟰𝟬,𝟬𝟬𝟬 𝗽𝗲𝗼𝗽𝗹𝗲 𝗷𝘂𝘀𝘁 𝘄𝗮𝘁𝗰𝗵𝗲𝗱 𝗮 "𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗯𝗼𝘁 𝘁𝘂𝗿𝗻𝗲𝗱 $𝟲𝟴 𝗶𝗻𝘁𝗼 $𝟳𝟱𝟬𝗞" 𝗽𝗼𝘀𝘁 𝗮𝗻𝗱 𝗻𝗼𝗯𝗼𝗱𝘆 𝗱𝗶𝗱 𝘁𝗵𝗲 𝗺𝗮𝘁𝗵.

So I did.

𝗧𝗵𝗲 𝗰𝗹𝗮𝗶𝗺:
→ $68 to $750,000 is 11,029x
→ That's doubling your money every single day for 13+ days
→ With zero losing days, on a bot built in 48 hours

𝗣𝗿𝗼𝗯𝗹𝗲𝗺 𝟭, 𝘁𝗵𝗲 "𝗮𝗿𝗯𝗶𝘁𝗿𝗮𝗴𝗲":
→ Price gaps between big crypto markets are usually a few hundredths of a percent
→ Standard exchange fees are around 0.1% to buy and 0.1% to sell
→ So most gaps are smaller than the fee you pay to take them

𝗣𝗿𝗼𝗯𝗹𝗲𝗺 𝟮, 𝘀𝗽𝗲𝗲𝗱:
→ Pro trading firms put their servers right next to the exchange
→ They react in microseconds
→ A bot on a laptop or iPad over home Wi-Fi shows up 100+ ms later
→ By then the gap is almost always gone

𝗣𝗿𝗼𝗯𝗹𝗲𝗺 𝟯, 𝘁𝗵𝗲 𝗮𝘀𝗸:
"Comment Claude, like, repost, follow, and I'll DM you the setup."
Every comment pushes the post to more people. The engagement is the product, not the bot.

𝗪𝗵𝗮𝘁 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗶𝘀 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗴𝗿𝗲𝗮𝘁 𝗳𝗼𝗿 𝗵𝗲𝗿𝗲:
→ Backtesting a strategy on real data before you risk a cent
→ Building dashboards and alerts
→ Finding out your edge disappears after fees, for free

If it printed $750K, why would anyone sell it for a comment?

[Milo @m1lo_exe]

@xsmryze The scary part is step 4: Sonnet patches, Sonnet verifies. A model grading its own homework passes stuff that only looks right. Different model for verify, or real tests, or the "verified" label means nothing.

[Milo @m1lo_exe]

@EnvolDev 3 screenshots from 3 dates is n=1 each. I'd want 20 runs per date before calling it a nerf, and input vs output tokens split out. "Burns tokens faster" is usually longer reasoning or a cold cache, not a dumber model.

[Milo @m1lo_exe]

People are decompiling 1996 games with Opus 5.5 and posting it as a win.

Almost none of them can prove the output matches the original binary.

This article fixes that. Here’s what you’re looking at in the video:

Ghidra splits the binary into header, strings and code
Sonnet 5.5 maps every function (~$0.30 a pass) and renames NOTHING
Opus 5.5 rebuilds valid_record: 4 checks, each tied to an address
tests run against the ORIGINAL binary, not against the model’s own code
4/4 match. Now it’s a spec anyone can verify.
Then the model labels a checksum loop “CRC32”. Looks right. Reads clean.
The test fails.
Seed 2166136261, multiplier 16777619. That’s FNV-1a.
Only the test caught it. The explanation never would have.

The routing that saves you money:
→ Sonnet maps ($0.30)
→ Opus interprets the hard parts ($0.60)
→ Fable 5.1 only stress-tests a theory that already passed ($1.50)

Paying Fable rates to list strings burns money. Asking Sonnet to invent a protocol from one function burns time.

And one rule in CLAUDE.md does most of the work:
OBSERVED vs INFERRED vs UNKNOWN. No address, no claim. No log, no test.

Next time a “Claude rebuilt X” post goes viral, ask for the tests.

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl

[zynQ @zynQ6969]

𝗧𝗵𝗶𝘀 𝗰𝗵𝗶𝗽 𝗵𝗮𝘀 𝟮𝟬𝟴 𝗯𝗶𝗹𝗹𝗶𝗼𝗻 𝘁𝗿𝗮𝗻𝘀𝗶𝘀𝘁𝗼𝗿𝘀. 𝗜𝘁'𝘀 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝘁𝘄𝗼 𝗰𝗵𝗶𝗽𝘀 𝗽𝗿𝗲𝘁𝗲𝗻𝗱𝗶𝗻𝗴 𝘁𝗼 𝗯𝗲 𝗼𝗻𝗲.

NVIDIA's B200 is behind a lot of the AI you use every day. I had Claude take it apart and rebuild it in code, layer by layer.

𝗪𝗵𝗮𝘁'𝘀 𝗶𝗻𝘀𝗶𝗱𝗲, 𝘁𝗼𝗽 𝘁𝗼 𝗯𝗼𝘁𝘁𝗼𝗺:
→ Cooling plate: pulls away ~1,000 W, like a microwave that never turns off
→ 2 silicon dies: 104B transistors each
→ 8 stacks of HBM3e memory: 180 GB total
→ Interposer: wires the dies to the memory
→ Substrate: fans out thousands of contacts
→ Power stages: feed the chip at under 1 volt
→ Module board: power in, NVLink out

𝗧𝗵𝗲 𝘄𝗲𝗶𝗿𝗱 𝗽𝗮𝗿𝘁:
Each die is already about as big as a chip factory can print in one go. So NVIDIA put two side by side and linked them at 10 TB/s. Software just sees one GPU.

𝗧𝗵𝗲 𝗽𝗮𝗿𝘁 𝗺𝗼𝘀𝘁 𝗽𝗲𝗼𝗽𝗹𝗲 𝗺𝗶𝘀𝘀:
The memory moves 8 TB every second. Sounds endless. But a 70B model at 8-bit weighs 70 GB, so one full read of its weights takes ~9 ms. When AI writes text, memory is the bottleneck, not math. That's why every new generation fights for bandwidth.

𝗛𝟭𝟬𝟬 → 𝗕𝟮𝟬𝟬 𝗶𝗻 𝟮 𝘆𝗲𝗮𝗿𝘀:
→ Transistors: 80B → 208B
→ Memory: 80 GB → 180 GB
→ Bandwidth: 3.35 → 8 TB/s
→ Power: 700 W → ~1,000 W

𝗡𝗼𝘄 𝘁𝗶𝗺𝗲𝘀 𝟳𝟮:
One rack holds 72 of these, wired into one giant GPU.
→ ~15 trillion transistors
→ 13.4 TB of memory
→ 130 TB/s between chips

And the next one, Rubin, is already at 336B transistors.

What should I take apart next?

[zynQ @zynQ6969]

𝗧𝗵𝗲𝗿𝗲 𝗮𝗿𝗲 𝟮 𝘁𝗵𝗶𝗻𝗴𝘀 𝗖𝗹𝗮𝘂𝗱𝗲 𝗗𝘂𝗲𝘁 𝘄𝗼𝗻'𝘁 𝗱𝗼, 𝗲𝘃𝗲𝗻 𝗶𝗳 𝘆𝗼𝘂 𝘀𝗮𝘆 𝘆𝗲𝘀.

Trade securities. Move money between your accounts.

Everyone's talking about what Duet can do. Anthropic's full write-up has a part most people skipped: where it stops.

𝗪𝗵𝗮𝘁 𝗶𝘁 𝗱𝗼𝗲𝘀 𝗼𝗻 𝗶𝘁𝘀 𝗼𝘄𝗻:
→ Sends you a morning brief from all your connected apps
→ Checks them every few hours and messages you only when something needs you
→ Drafts replies, builds forecasts, preps payroll
→ Works in its own virtual machine with its own browser, so your laptop can be off

𝗪𝗵𝗮𝘁 𝗻𝗲𝗲𝗱𝘀 𝘆𝗼𝘂𝗿 𝘆𝗲𝘀:
→ Sending a message
→ Making a payment
→ Signing a document
→ Deleting anything
It shows you exactly what it's about to do, then waits.

𝗪𝗵𝗮𝘁 𝗶𝘁 𝗻𝗲𝘃𝗲𝗿 𝘀𝗲𝗲𝘀:
→ Your passwords or sign-in tokens
→ Your real card number. Payments go through a virtual card via Link

On top of that, a dedicated classifier checks every message so things like card numbers and government ID don't go out. And it's built to resist hidden instructions planted in emails and web pages.

Research preview, waitlist only, Pro and Max in the US. It shares your plan's usage limits, and Anthropic admits it can still misjudge what's urgent.

Would you let it handle your payments?

[Milo @m1lo_exe]

I asked Grok Bot how to save tokens. It lied about its own product.

Turns out the same task can cost 551,000 tokens or 12,000.

Grok Bot doesn't count your messages. It counts every time it re-reads your chat

The fixes that actually work:

> Fresh chat per task: busy bots re-read 80-90K tokens on EVERY step
> Connectors, not browser: 45x fewer tokens for the exact same job
> Slow your routines: one user went from 672 runs a week to 168. Empty runs still bill
> One task, one prompt: every "no, I meant..." makes it re-read everything
> One bot, one role: a do-everything bot drags every old topic into every new task
> Freeze the description: same prefix = cache hit, input gets 4x cheaper

Most people will just upgrade their plan. The ones who fix their setup get a month out of the same limits

Full breakdown below:

[kaize @0x_kaize]

My Grok Bot limits ran out in two days, and I got tired of it

So I found a way out of this situation and compiled tips in a single article https://t.co/UebmdBgVSH

[zynQ @zynQ6969]

𝗔 𝗸𝗲𝘆 𝗽𝗿𝗲𝘀𝘀 𝗶𝘀 𝗱𝗲𝗰𝗶𝗱𝗲𝗱 𝗮𝘁 𝟮 𝗺𝗺. 𝗡𝗼𝘁 𝘄𝗵𝗲𝗻 𝘆𝗼𝘂 𝗵𝗶𝘁 𝘁𝗵𝗲 𝗯𝗼𝘁𝘁𝗼𝗺.

I had Claude take a mechanical key switch apart and rebuild it in code, part by part.

𝗪𝗵𝗮𝘁'𝘀 𝘂𝗻𝗱𝗲𝗿 𝗲𝘃𝗲𝗿𝘆 𝗸𝗲𝘆:
→ Keycap, the part you touch
→ Stem, slides down when you press
→ Spring, pushes it back up
→ Two gold-plated contact leaves
→ Housing that holds it all together

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝘄𝗵𝗲𝗻 𝘆𝗼𝘂 𝗽𝗿𝗲𝘀𝘀:
The key travels 4 mm in total. At 2 mm the stem lets the two metal leaves touch, the circuit closes and the keyboard registers the press. On a typical linear switch that takes about 45 g of force. The other 2 mm is just your finger hitting the bottom.

𝗧𝗵𝗲 𝘄𝗲𝗶𝗿𝗱 𝗽𝗮𝗿𝘁:
Metal touching metal doesn't connect cleanly. It bounces on and off for a few milliseconds. Your keyboard waits that out, about 5 ms, so one press counts as one letter instead of five.

𝗔𝗻𝗱 𝗶𝘁'𝘀 𝗳𝗮𝘀𝘁:
From the contact closing to your app getting the letter takes about 10–20 ms. A blink takes over 100.

That's a typical MX-style switch. Yours might register a little higher or lower.

What should I take apart next?

[zynQ @zynQ6969]

𝗘𝘃𝗲𝗿𝘆 𝗔𝗜 𝗮𝗽𝗽 𝘄𝗮𝗶𝘁𝘀 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝗽𝗿𝗼𝗺𝗽𝘁. 𝗖𝗹𝗮𝘂𝗱𝗲 𝗗𝘂𝗲𝘁 𝗱𝗼𝗲𝘀𝗻'𝘁.

Anthropic opened a research preview of Duet, a separate app where Claude works as a proactive agent.

𝗪𝗵𝗮𝘁 𝗶𝘁 𝗱𝗼𝗲𝘀:
→ Comes to you with what needs your attention
→ Does the tasks you hand it
→ Keeps working when you step away
→ Checks with you before acting on your behalf

It runs on Opus 5.5 and brings over your memory, skills and connected apps from your Claude account.

𝗪𝗵𝗼 𝗰𝗮𝗻 𝘁𝗿𝘆 𝗶𝘁: waitlist only, Pro and Max plans in the US.

What would you hand off first?

[Milo @m1lo_exe]

GPT-5.6 + A 14-YEAR-OLD BUG = 3 OUT OF 4 WEBSITES EXPOSED

This is the most important AI post of the week and almost nobody read past the first paragraph.

→ found by AI, not humans
→ leaked before the patch
→ “many more” in the pipeline

The defenders need better models than the attackers. Now.

[Greg Brockman @gdb]

Over the past two months, we have been leading a coalition to mitigate one of the most severe bugs in open source software found in recent memory, php2shell, found as part of our Patch the Planet initiative. https://t.co/rBsz7lrlwo

[Spinzu @Spinzu]

2 AGENTS. 1 FILE. 0 LINES OF COPIED CODE.

that's the whole clean-room method, and almost nobody explains the wall.

clean-room is two rooms, not one:

→ room A reads the app and writes a spec. behavior only: inputs, outputs, edge cases
→ room B builds from that spec. it never sees the original code
→ the spec is the only thing allowed to cross the wall

with agents it maps cleanly:

> agent A: studies the feature, writes SPEC.md
> agent B: fresh session, gets SPEC.md only, builds + tests
> no copied code. no shared context

the rule that makes it work:
if agent B ever sees a line of the original, the room is dirty. start over.

prompt for room A ↓

"Study how [feature] behaves in [app]. Write SPEC.md describing only behavior: inputs, outputs, states and edge cases. Do not include any code, names or structure from the original. Mark anything you're unsure about as UNKNOWN."

full walkthrough in the article below 👇
not legal advice. check the license and your local law before you ship anything.

[Milo @m1lo_exe]

@Spinzu the hard part is the handoff. how do you decide when the cheap model's findings are good enough to send up?

[wast3 @0xWast3]

Opus 5.5 inside Blender costs $4 per million input tokens, and it just beat a stack of specialist tools that charge by the seat

Price per seat hides the real comparison. One is a subscription per person, the other is a bill that scales with what you actually make.

> ASSET - a full set of product models built from one brief, geometry and materials included
> SHOT - the same products staged, lit and rendered across a dozen backgrounds
> COST - every tool’s output priced against what the run actually consumed

Photoroom and the rest are built around a photo. They cut, relight and restage something that already exists.

That ceiling shows up fast. No photo means no output, and a product that is not built yet has no photo at all.

Opus 5.5 starts earlier. The model is built in Blender, so every angle, every material and every background is a render away.

At $4 in and $20 out per million tokens, a full catalogue costs less than one seat of the specialist tools per month.

The quality gap is not subtle either. Real geometry holds up under lighting changes in a way a cut-out never does.

The brief, both outputs side by side and the full token bill are laid out underneath.

[Milo @m1lo_exe]

@Spinzu my cut for a solo setup:

> ollama: local models, no API bill
> litellm: one key for everything
> browser-use: the agent does the clicking
> n8n: triggers + glue
> langfuse: see what it actually costs

the other 15 are "add when it hurts"

[Milo @m1lo_exe]

Holy sh*t. you're paying $1,000+/mo for SaaS that Opus 5.5 + 20 free repos already do:

[01] core models
[02] agents
[03] browser
[04] memory & workflows
[05] ship & verify

how it runs: task -> plan -> tools -> verify -> ship -> retry

your SaaS bill is optional now.

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl