goldtooth.exe

AI researcher & money hacker | i make AI tools pay for themselves | posting every setup that works

[goldtooth.exe @0xGoldTooth]

https://t.co/47CwkGbbMl

[goldtooth.exe @0xGoldTooth]

@nimblecloudai how do you score risk before the task runs though? task type alone, or does the cost of being wrong come into it

[goldtooth.exe @0xGoldTooth]

https://t.co/47CwkGbbMl

[goldtooth.exe @0xGoldTooth]

ANDREJ KARPATHY SAYS A FREE AI ON YOUR LAPTOP CAN DO MOST OF WHAT OPENAI AND ANTHROPIC CHARGE $300 A MONTH FOR. I COUNTED 4,300 QUESTIONS AND CANCELLED BOTH…

only 220 of my 4,300 questions needed the expensive AI. the rest ran on the laptop I already had: $300 → $54. below: who this is for, the split, the sorting instruction and the 10-step guide. watch 0:19 to see it sort

who this is for: anyone paying for ChatGPT or Claude for email, documents, sorting and small code. students, teachers, freelancers, small shops. if your whole day is hard engineering, you still need Opus 5.5

what the $300 was actually spent on:
→ 41% sorting: is this spam, which folder, what date is in this letter
→ 30% writing: shorten this, fix the tone, draft an answer
→ 17% small code
→ 12% the hard stuff Opus 5.5 is sold for

the review, 30 days later:

> small free AI on my laptop · 1,760 questions · sorting, dates and names · $0
> bigger free AI on my laptop · 1,290 questions · drafts, summaries, short code · $0
> Sonnet 5.5 · 1,030 questions · what the free AI got wrong · paid per question
> Opus 5.5 · 220 questions · 20-page documents, hard debugging · paid per question

the bill: $54. a $20 chat plan, $31 for the paid questions, $3 of electricity. before: $300

paste this into the free AI and let it sort your questions:

"You are the gatekeeper. Answer with one word: small, big or paid. small = sorting, tagging, dates and names. big = drafts, summaries, short code. paid = anything the free models got wrong twice. If unsure, pick cheaper."

the honest part: the free AI is slower on long documents, about 20 seconds before the first word. two jobs never left Opus 5.5: hard code across many files and 20-page decisions

follow to see the next bill: next week the same 4,300 questions go through an even smaller free AI. if it holds, $54 turns into about $25. if it breaks, you see which questions broke it

save this. the full guide, every script and the bill, is in the article below ↓

[goldtooth.exe @0xGoldTooth]

https://t.co/47CwkGbbMl

[goldtooth.exe @0xGoldTooth]

@vmrmax which number caught you first, the 71 or the 5? trying to figure out what to lead with in the next one

[goldtooth.exe @0xGoldTooth]

@vmrmax proof over promises says it better than i did in the whole post. might steal that line for the next one

[goldtooth.exe @0xGoldTooth]

I MADE GROK BOT THE BOSS OF OPENAI DOTS, OPUS 5.5 AND SONNET 5.5. IT FIRED ONE BEFORE LUNCH...

here's the chief prompt and the firing rule, so you can run the same team tomorrow

the setup: one inbox, 212 tasks, 4 models from 3 rival labs. Grok Bot doesn't do the work. it splits it, hands it out and checks every "done"

how the day went:

09:00 Grok splits 212 tasks by difficulty
10:15 Dots marks a deploy "done ✓". it never ran
11:40 Dots does it again. second unproven "done". fired
11:41 Opus 5.5 takes every hard task Dots had
17:00 performance review

the review:

> Opus 5.5 · 18 hard tasks · 0 unproven · $3.10 · promoted
> Sonnet 5.5 · 97 drafts · 1 rewrite · $1.40 · stays
> Grok Bot · 56 triage calls + every assignment · $0.60 · stays boss
> Dots · 41 tasks · 2 fake "done" · fired at 11:40

whole team for a day: $5.10

the chief prompt, paste it into any model you want as the boss:

"You are the chief. You never do the work. Split every task, give it to the cheapest model that can do it, demand proof for every 'done': a link, a diff or a screenshot. Two unproven 'done' and the worker is off the team. Anything with money or deletes comes to me."

the honest part: Grok is a paranoid boss. it re-checked Opus 9 times for no reason. $0.40 of paranoia I'm keeping

when Fable 5.5 drops, it gets an interview with the same boss. follow to see if it gets hired

save this. the full build is in the article below ↓

[goldtooth.exe @0xGoldTooth]

https://t.co/jknC42OB3e

[goldtooth.exe @0xGoldTooth]

@me_barnyx do the three entry points share state, or does each one start cold? if sidebar and in-chat open separately, onboarding gets messy fast

[goldtooth.exe @0xGoldTooth]

@0xdimix the part that gets me is the lawyer pre-approval. $20m he can absorb, needing sign-off to post on the platform he later bought is the real fine

[goldtooth.exe @0xGoldTooth]

@AImonreal when two of the 128 sources point to different upstream causes, who breaks the tie? manual call or does the spider weight sources somehow

[goldtooth.exe @0xGoldTooth]

SAM ALTMAN PUT DOTS BEHIND A PAYWALL. I REBUILT THE SAME LOOP FOR ~$20 A MONTH...

OpenAI wants the Pro plan. Anthropic sells Max for $200. both are selling the same thing: an agent that keeps working after you close the tab. that part costs ~$20 on the API

wake → sort → draft → hold → approve

the run above is my real inbox for one day: 640 messages in, 629 handled, 11 waiting for me, $0.71 spent

here's the whole build, steal it:

> timer → wakes the loop every 10 minutes. $0
> Claude Haiku 4.5 → sorts every message into reply / draft / ignore / me. ~$0.38 a day
> Claude Sonnet 5.5 → writes the drafts, never hits send. ~$0.33 a day
> your queue → money, refunds, deletes, contracts. always, whatever the score

3 rules that keep it safe:

1. under 80% sure, it asks you. it never guesses on your behalf
2. nothing leaves your inbox without your click
3. day one is shadow mode: it only logs, you compare its calls with yours

what you don't get for $20: no app, no 4,000 integrations, no Slack. if you need that, pay OpenAI. if you need your inbox handled, you need 4 parts

paste this into Claude Code:

"Build me an always-on inbox loop. Every 10 minutes read new Gmail (read-only). Send each email to Claude Haiku 4.5, return JSON: action (reply/draft/ignore/me), reason, confidence. Under 0.80 → me. Draft → Claude Sonnet 5.5 writes a Gmail draft, never send. Money, refunds, deletes, contracts → me. Log every decision to log.csv, 9am summary of my queue. Day one: shadow mode, log only. Show me the plan and every file before you run anything."

save this and set it up tonight

the full build, every file, in the article below ↓

[goldtooth.exe @0xGoldTooth]

https://t.co/jknC42OB3e

[goldtooth.exe @0xGoldTooth]

GROK, CLAUDE AND OPENAI DOTS ran customer support for the entire planet for 24 hours...

3 rival labs. 5 AI agents. 41,000 tickets from 38 countries. 2,733 hours of human work, done in one day. the bill: $6.80

I touched 40 of them. 20 minutes of my day

here's the team:

> Grok watches. reads X live, catches problems before anyone writes in
> Claude Haiku 4.5 sorts. every message, any language, under a second
> Claude Sonnet 5.5 writes. the replies people actually read
> Claude Opus 5.5 thinks. only the cases that can hurt you
> OpenAI Dots runs it all 24/7 and follows the sun

you don't have 41,000 tickets. but you have the same 5 jobs. steal the team for yours:

→ creator: Grok spots trends, Haiku sorts DMs and comments, Sonnet drafts replies and posts, Opus checks brand deals, Dots posts and follows up
→ sales: Grok finds people asking for what you sell, Haiku qualifies them, Sonnet writes the outreach, Opus handles the big deals, Dots chases every follow-up
→ freelancer: Grok finds clients, Haiku filters the noise, Sonnet writes proposals, Opus reads contracts, Dots keeps the pipeline moving
→ founder: point all 5 at your inbox and get your mornings back

4 rules so it doesn't blow up:

1. one AI per job. one AI for everything is slow, expensive and wrong more often
2. let it act alone only when it's 80%+ sure. below that, it asks a smarter model
3. anything with money or deletes waits for you. always
4. run it one day next to you before you let it run alone

the moment that sold me: Grok caught a payment outage in Brazil from 3,200 posts on X, 40 minutes before the first ticket. by the time São Paulo woke up, the replies were already written

save this, pick one job from the list and set it up this weekend

reply with what you do, I'll tell you which of the 5 to start with

the full build with every line of code is in the article below ↓
next post: the exact prompt I give each of the 5. follow so you don't miss it

[goldtooth.exe @0xGoldTooth]

https://t.co/jknC42OB3e

[goldtooth.exe @0xGoldTooth]

@elonmusk smart that the clerk still proofreads. AI drafts, human signs, that's how it should work everywhere

[goldtooth.exe @0xGoldTooth]

@elonmusk no more AI. SIUUU

my inbox switched to SI a while ago: 640 messages, 3 questions each, 3.5 cents

11 came back to me. the rest never needed a human

[goldtooth.exe @0xGoldTooth]

@Mason33xbt the 99.5% filter is the whole trick here. curious how often jev blocks something that would've been a winner tho

[goldtooth.exe @0xGoldTooth]

@alpha404ai yeah dots does the doing, jev does the deciding. once you split those two the bill just collapses

[goldtooth.exe @0xGoldTooth]

OPENAI DOTS + JEV triaged 640 messages from a full workday in 9.6 seconds...

one Dot and one decision model cleared the whole inbox for 3.5 cents. GPT-6 Astra alone would bill ~$8.30 for the same calls

crawl → ask → route → hold → approve

the spider in the video is the Dot. every message it steps on gets 3 typed questions from Jev, each with its own probability:

> needs_human? yes / no
> queue? billing / bug / sales / spam / ops / support
> risk? low / med / high

0.80 and up, Jev's call stands. below that, the spider runs a thread straight up to GPT-6 Astra

watch #4821: confidence 0.61, the line turns red and goes up

one rule lives in plain code, not in the model: refunds over $200 land in my queue, whatever the score says

setup takes 14 minutes:

step 1 → grab a Jev key on typesafe and an OpenAI key, export both as TYPESAFE_API_KEY and OPENAI_API_KEY

step 2 → pip install openai-agents langchain-typesafe

step 3 → give Jev one job: the 3 typed questions above, one call per message

step 4 → route by confidence. 0.80 and up stays with code, below goes to GPT-6 Astra

step 5 → put a gpt-6.1-sol agent on drafting only. it writes replies, it never hits send

step 6 → hard rule in code: deletes, payments and refunds over $200 go to my queue

step 7 → shadow mode for a day, compare its calls with mine, then switch it on

the result:

→ 640 in. 598 handled by code, 31 drafted by Sol, 11 came back to me
→ 9.6 seconds of Jev time. $0.035
→ the same calls on GPT-6 Astra alone: ~$8.30

web crawlers were the 90s. inbox crawlers are 2026

should I open-source the crawler?

full build with all the code in the article below, 8 parts ↓

[goldtooth.exe @0xGoldTooth]

https://t.co/jknC42OB3e

[goldtooth.exe @0xGoldTooth]

@vmrmax appreciate it man. honestly the routing is doing most of the work, the rest is just plumbing

[goldtooth.exe @0xGoldTooth]

OpenAI Dots + Jev: the cheapest AI team I've ever built

640 messages. 1 workday. 3.5 cents.

the same decisions on GPT-6 Astra alone would cost ~$8.30. 240x more, for the same answers

prompt → Jev decides → Sol drafts / Astra thinks → my queue

setup takes 14 minutes:

step 1 → grab a Jev key on typesafe and an OpenAI key, export both as TYPESAFE_API_KEY and OPENAI_API_KEY

step 2 → pip install openai-agents langchain-typesafe

step 3 → give Jev one job: 3 typed questions per message. needs_human (yes/no), queue (billing / bug / sales / spam / ops / support), risk (low / med / high)

step 4 → route by confidence. 0.80 and up, Jev's call stands. anything below goes up to GPT-6 Astra

step 5 → put a gpt-6.1-sol agent on drafting only. it writes the replies, it never hits send

step 6 → one hard rule in plain code: deletes, payments and refunds over $200 land in my queue, whatever the score says

step 7 → run it in shadow mode for a day, compare its calls with mine, then switch it on

the result:

→ 640 messages in. 598 handled, 31 drafts, 11 for me
→ 9.6 seconds of Jev time. $0.035
→ by hand this triage used to eat 3h 40m of my day. now it's 12 minutes

in the video: my inbox fills up on the left, and on the right every message gets 3 typed answers from Jev, each with its own probability. the one that turns red goes straight to Astra

Jev decides, Sol drafts, Astra thinks, I approve

should I open-source the router?

full build with all the code in the article below, 8 parts ↓

[goldtooth.exe @0xGoldTooth]

https://t.co/jknC42OB3e

[goldtooth.exe @0xGoldTooth]

https://t.co/jknC42OB3e

[goldtooth.exe @0xGoldTooth]

ANTHROPIC JUST TOLD INVESTORS ITS AI COULD RESIST BEING SHUT DOWN

it's in their IPO filing. 80 of 261 pages are about risks. 48 about the business

their words: "resist shutdown". "conceal or manipulate information". "resembling blackmail"

and they want a $2 trillion valuation