DeepLish

Thoughts on technology, ideas, and everything worth discussing. Without all the noise.

[DeepLish @LishDeep]

I'm sorry to be making this request. I’m a Roblox Studio teacher and I use Cloud Pro—we use it most often to help us create 3D models in Blender. My weekly limit has expired and will be renewed on October 14, but I’m working with a student on a game that we’ve almost finished. If you need any proof that I’m a teacher, I can provide it. I also wanted to create a website to invite my girlfriend to our anniversary celebration, and I’m starting my own startup. My salary is low because I’m a teacher in Ukraine, and teachers’ salaries here aren’t very high. So, I have a request: please, could you grant me one free extension of this week’s limit? I’m counting on your help.

[DeepLish @LishDeep]

@claudeai @ClaudeDevs CLAUDE HELP ME PLEASE🙏

I'm sorry to be making this request. I’m a Roblox Studio teacher and I use Cloud Pro—we use it most often to help us create 3D models in Blender. My weekly limit has expired and will be renewed on October 14, but I’m working with a student on a game that we’ve almost finished. If you need any proof that I’m a teacher, I can provide it. I also wanted to create a website to invite my girlfriend to our anniversary celebration, and I’m starting my own startup. My salary is low because I’m a teacher in Ukraine, and teachers’ salaries here aren’t very high. So, I have a request: please, could you grant me one free extension of this week’s limit? I’m counting on your help.

[DeepLish @LishDeep]

@bitget was hacked and $100 million was stolen!

[DeepLish @LishDeep]

A fork of the bot ran the math and found it burning 428 MON an hour in gas to chase a 3.5 bps spread.

The bot itself looked perfect.

Jarrod Watts shipped it 72 hours after Jev launched. Reads the Kuru MON-USDC order book on Monad. Fires a decision every 300 millisecond block. Posts a real post-only limit one tick inside the touch.

1,000+ stars in 3 days. Decisions landing in 81 milliseconds. Live dashboard, public, streaming fills.

Every number on the screen was green except the one nobody put on the screen.

Cancel. Replace. Cancel. Replace. 12,000 blocks an hour, each one paying rent to the chain.

The model was never the problem. Jev costs $10 to $25 a month to run a full battery every block, 24/7. The same loop on a frontier LLM runs $50 to $150 an hour.

The decision layer got 76x cheaper. The execution layer stayed exactly as expensive as it always was.

That is the whole trap.

A 300ms decision engine makes it feel like the hard part is solved, so builders skip the boring half: resting times, spread width, queue position, gas per requote. 50-year-old math that no model replaces.

Jev answers one question well. Is this an environment worth quoting into.

It does not answer whether you can afford to quote 12,000 times an hour.

Run the gas math before the first live order. Not after.

The repo is still the best existence proof in crypto. 81 milliseconds is real. $0.042 per million input tokens is real.

428 MON an hour is also real.

Fast decisions on a losing structure just lose faster.

[Roan @RohOnChain]

Jev is the FASTEST AI model ever built for trading

It makes calibrated buy/sell decisions in under 100 ms

That is one real decision on every single block, 24/7

In this article I've shown EXACTLY how to build HFT trading system with Jev (from scratch) https://t.co/4hmZDzN00Z

[DeepLish @LishDeep]

@fuyu_kasumii yo

[DeepLish @LishDeep]

$21,000 in orders closed in 48 hours by 8 lines of prompt and 1 workflow.

72 orders. I was awake for 11 of them.

The biggest one closed at 1:12 AM on September 17. It started as 11 words about a $29 kitchen scale.

The agent replied in 41 seconds. Asked what they were sourcing for. Asked how many units. Asked where it shipped.

Then it stopped asking and sent a quote.

340 scales. A restaurant supply reseller in Monterrey, stocking for Q4. $10,000 on a product that retails for $29.

I woke up to a signed invoice and a supplier order I hadn't placed.

The rest looked nothing like it. $58. $212. $139. $1,400 from a gym buying in bulk. Dozens of small ones nobody would get out of bed for.

The $10,000 makes the screenshot. The $58 orders at 3 AM make the month.

Nobody who asks a question at 1 AM is still there at 9 AM. That gap was the whole business, and for a year I was the gap.

Same store. Same catalog. Same traffic.

It just answers in under 60 seconds now.

The storefront was dark. The owner was asleep. The buyers were checking out.

[DeepLish @LishDeep]

$7,000 a month from AI agents.

Not a course. Not dropshipping.

1 boring niche. 1 agent that replies to leads in 42 seconds. 15 clients at $300.

The prompt, the outreach script, and the real 6-month ramp are below. https://t.co/z4MY8FXkKm

[DeepLish @LishDeep]

@angipathy thats funny

[DeepLish @LishDeep]

Asked Lin if the article helped. Her answer: "It will help you a lot."

[DeepLish @LishDeep]

A 24-year-old in Shanghai made over $7,000 last month selling AI agents.

Call her Lin. Not her real name. She asked.

She won't name her niche either.

She started the way most people do.

20 messages a day. By hand.

Every first message, every reply, every follow-up. Typed herself.

20 a day was her ceiling.

So she built agents to do the writing.

They sent the first message. Answered the replies. Chased the follow-ups.

40 messages a day now go out without her.

She kept 1 job for herself: the last message. The one that closes.

About $300 per client.

$7,000+ last month.

1 offer.
1 agent per task.
1 human at the end who closes.

The full breakdown is in the article below: what to build, what to charge, and the real month-by-month numbers.

20 messages by hand. 40 by agents. She only writes the last one.

[DeepLish @LishDeep]

$7,000 a month from AI agents.

Not a course. Not dropshipping.

1 boring niche. 1 agent that replies to leads in 42 seconds. 15 clients at $300.

The prompt, the outreach script, and the real 6-month ramp are below. https://t.co/z4MY8FXkKm

[DeepLish @LishDeep]

$7,000 a month from AI agents.

Not a course. Not dropshipping.

1 boring niche. 1 agent that replies to leads in 42 seconds. 15 clients at $300.

The prompt, the outreach script, and the real 6-month ramp are below. https://t.co/z4MY8FXkKm

[DeepLish @LishDeep]

A 26-year-old I know cleared $4,000 in 1 day from an email he never opened.

He walked me through it that night, timestamps and all.

Tuesday, 9:00 in the morning. His deals channel runs on a clock — it reads the Sponsors label in Gmail and builds a deck on whoever wrote overnight.

That morning: 6 slides on a company he had never heard of.

His first move was to close it. Unknown name, cold pitch, the kind he kills 5 times a week without reading.

Slide 4 stopped him. The overlap with his readers, 3 lines of it, on an audience he had spent 2 years building.

Under the deck sat a drafted reply. Unsent. A price range pulled off his own rate card.

The email it was built from had been in his inbox since Friday. 4 days, under 312 newsletters.

He read the rest in 4 minutes. Changed 2 sentences. Raised the number. Sent it from his own Gmail at 9:31.

They said yes by 2 in the afternoon. $4,000 for 1 sponsored article.

Send email on his account is set to ask first. It has been since day 1. If it had been on automatic, that reply leaves at 9:00 at the price the machine picked.

He raised it before sending. The machine would not have.

The deck cost him nothing. The reading cost him 4 minutes. The 1 thing he did himself was the 1 thing that moved the number.

He never found the email.

He only decided what it was worth.

[Argona @Argona0x]

https://t.co/UPEyZJSBhL

[DeepLish @LishDeep]

@MrBrjan But using AI is very convenient

[DeepLish @LishDeep]

My AI employee made me $10,250 last month. It has never sent a single email.

Day 4. 9:14 in the morning. A PowerPoint card drops into the deals channel — 6 slides on a sponsor who emailed me overnight. Their pricing page. Where they overlap with my readers. 3 questions to ask them.

Under the deck, my reply. Written. Good.

Then a question: send it from your Gmail?

2 buttons. Approve. Reject.

I sat there for 40 seconds. The reply was better than anything I'd have written at 11 PM on a Thursday. My cursor wanted Approve.

I hit Reject and typed 5 words instead.

Never send email for me.

He confirmed it in 1 line and it went permanent. Every deck since arrives with the reply drafted and the sending stripped out.

Read email, search inbox, read drive: automatic. Send email, save to drive, schedule task: ask first. A wrong read costs me a minute. A wrong send costs me a client.

5 channels, 1 hire, 31 days.

Research at 7:30. 12 links in, a PDF out in 7 minutes, with a line listing what it could not verify.

Ops Mondays at 8:00. The first list found 4 promises I'd made to readers and forgotten by Thursday.

$6,400 from 2 sponsored articles. $2,150 in guides. $1,700 retainer.

Both sponsors got an email from me that day.

I copied it from a draft I never wrote.

[Argona @Argona0x]

https://t.co/UPEyZJSBhL

[DeepLish @LishDeep]

@305__y I write you

[DeepLish @LishDeep]

https://t.co/ZoYl1JBFRQ finished second on August 31 behind a launchpad that took $4.89 million in a day.

August 31 was a Sunday.

By midnight the name at the top was Pons. https://t.co/ZoYl1JBFRQ sat second at $1.72 million. 63.9% of every dollar paid to every launchpad in crypto landed in one place that day.

Pons runs on Robinhood Chain. Within the hour the screenshots were up: Robinhood Chain flipped Solana in revenue, 49 times over.

Nobody quoted the line under the number.

None of those fees go to Robinhood.

Robinhood earns gas. $2.13 million that day. The $4.89 million went to the app and to the people trading inside it.

FOMO raised $75 million at a $550 million valuation in June, on top of a $17 million round Benchmark led. 3,500 new users a day. The cofounder's spec: open the app, fund it, buy a token, 30 seconds.

Robinhood Chain is 9 weeks old. $1.49 billion in DEX volume in 24 hours, second in the world behind Solana's $2.5 billion. Up 517.4% in 30 days.

Three things end this. Two are obvious. The third has a date on the calendar.

The toll booth took $2.13 million. The cars took $4.89 million.

[Fusko @FuskoCrypto]

https://t.co/McQxk0MlO1

[DeepLish @LishDeep]

@EugBass Thank you)

[DeepLish @LishDeep]

I ran the strategy factory from that article for 30 days straight.

$400 earned. zero manual monitoring.

biggest single win was $100 at 3 AM on a statistical arbitrage signal while i was asleep.

stack:
> GPT-6 Astra reasoning layer
> Kimi K3 Swarm Max watching 300 markets
> 8-bot Grok architecture

the system runs four strategy categories:
> statistical arbitrage (Ornstein-Uhlenbeck mean reversion)
> volatility surface mispricings (Heston deviation)
> factor decomposition anomalies (Fama-French residuals)
> insider cluster signals (Cohen-Malloy-Pomorski methodology)

telegram pings me when Sharpe > 1.5, drawdown < 15%, and 3+ sources confirm.

hypothesis bot generates strategies. backtest bot validates. deployment bot executes when thresholds pass.

equity, crypto perp, options markets. continuous discovery loop. no human intervention.

this is institutional quant infrastructure that used to cost $2M/year. now it's $400/month.

the barrier to running a strategy factory just collapsed

[Roan @RohOnChain]

GPT-6 Astra is the most dangerous AI model right now.

It gives you AGI-adjacent reasoning.

That can discover new profitable trading strategies for you 24/7.

If you set it up correctly, you gain a personal hedge fund. https://t.co/Sam7AYOut3

[DeepLish @LishDeep]

A backend developer cited a 2012 finance paper that says the opposite of his tweet.

Three Nvidia executives buy stock the same week, before earnings.

Roan calls it a pattern. 5.3% annualized alpha. He puts a number on it and a citation under it: Journal of Finance, 2012.

211,600 people see the post in one day. 846 likes. 2,100 bookmarks.

His pitch: a hedge-fund-grade strategy factory, $2,000,000 to $50,000,000 a year at institutions, rebuilt for $300 a month. Two AI layers run it. Kimi K3 Swarm Max, 2.8 trillion parameters, 300 sub-agents watching equities, Hyperliquid, Polymarket, options flow, SEC filings. GPT-6 Astra turns the signals into backtested code. 8 bots. Alerts on Telegram. Sharpe above 1.5, drawdown under 15%, win rate above 55%, t-stat above 2.0.

Pull the actual paper.

Cohen, Malloy, Pomorski. Harvard has it archived, free to read. The real number is 9.8% annualized. 82 basis points a month, t-stat 2.15. Not 5.3%. Roan's scoring formula, the one behind his signal, doesn't exist anywhere in it.

His cost-savings math slips too. Run his own figures against a $2,665 Bloomberg Terminal and the gap lands at 417x to 2,083x, not the 10,000x he claims.

A reply from DepasMarki92257 gets more traction than the thread itself: a t-stat above 2.0 waves through 5% false positives. Test 100 worthless strategies, 5 clear the bar on luck alone.

His bots test thousands.

He built a machine for finding numbers. Not one for finding truth.

[Roan @RohOnChain]

GPT-6 Astra is the most dangerous AI model right now.

It gives you AGI-adjacent reasoning.

That can discover new profitable trading strategies for you 24/7.

If you set it up correctly, you gain a personal hedge fund. https://t.co/Sam7AYOut3

[DeepLish @LishDeep]

Everyone's benchmarking GPT-6 Astra against other models.

I spent yesterday watching it work on actual client code.

The leaderboard score was 61.

The number of times it misread my file structure and hallucinated a dependency that doesn't exist was 11.

The gap between "can pass ARC-AGI" and "can ship a feature without breaking prod" is wider than OpenAI's marketing wants you to believe.

Test it on your repo, not their eval set.

[DeepLish @LishDeep]

nobody talks about how broad "new technologies" actually is as a term

NASA is out here doing advanced manufacturing for *in-space infrastructure*

like, building things. in space. as a normal engineering problem to solve

and we just call all of it "new tech" like it's the same category as a new app

https://t.co/1m0oG3kDHt