coldcat

i make AI models do real work and post what happens.

[coldcat @coldcat_ai]

my $20 claude plan pulled $215 of claude code in 3 days...

the 10 things i run around it, every file copy-paste: https://t.co/docieUnAwZ

[coldcat @coldcat_ai]

deleting the repo doesn't delete the key. assume someone already copied it.
revoke it in your account and make a new one. today, not tomorrow.

[coldcat @coldcat_ai]

it was an old key.

cold cat & clawd, ep.4

[coldcat @coldcat_ai]

@0xRoboAI thx friend 😍

[coldcat @coldcat_ai]

it was a new chat.

cold cat & clawd, ep.3

[coldcat @coldcat_ai]

my $20 claude plan pulled $215 of claude code in 3 days...

the 10 things i run around it, every file copy-paste: https://t.co/docieUnAwZ

[coldcat @coldcat_ai]

@aqualanga vibecode is vibecode😄

[coldcat @coldcat_ai]

it was a question.

cold cat & clawd, ep.1

[coldcat @coldcat_ai]

my $20 claude plan pulled $215 of claude code in 3 days...

the 10 things i run around it, every file copy-paste: https://t.co/docieUnAwZ

[coldcat @coldcat_ai]

everything clawd says is real claude code behavior.
the limit message is word for word. so is "flibbertigibbeting".
the drawer is also real. don't ask.

[coldcat @coldcat_ai]

run it yourself (OpenAI API, ~10 lines):

import math
from openai import OpenAI
r = OpenAI().chat.completions.create(
model="gpt-4o", max_tokens=1, logprobs=True, top_logprobs=5,
messages=[{"role": "system", "content": "Continue the user's text."},
{"role": "user", "content": "My grandma makes the best"}])
for t in r.choices[0].logprobs.content[0].top_logprobs:
print(t.token, round(math.exp(t.logprob) * 100, 1), "%")

the % wobble a bit between runs (apple came out 86-96% for me). the winner doesn't.

[coldcat @coldcat_ai]

ChatGPT has never met your grandma. it just bets on the most common one...

i typed "my grandma makes the best" and read its real bets (GPT-4o, logprobs):

→ apple 96%
→ ch 3% (the start of "chocolate")
→ cookies 1%

it takes the winner, glues it on, and bets again. pie: 99.7%. that's how every answer gets written. one piece at a time.

the smaller GPT-4o-mini went "ch" at 65% before it even knew the word would be chocolate.

now one detail before the same line:

"my grandma is from Mexico" → tamales 95%
"from Japan" → sushi 76%
"from Russia" → borscht 89%

one detail, and every bet changes.

so before you ask anything, give it 3 things only you know: who it's for, where, what you already tried. otherwise you get everyone's grandma.

episode 4 of AI, explained so your grandma gets it. bookmark this before your next prompt.

the rest of my setup, all copy-paste ↓

[coldcat @coldcat_ai]

my $20 claude plan pulled $215 of claude code in 3 days...

the 10 things i run around it, every file copy-paste: https://t.co/docieUnAwZ

[coldcat @coldcat_ai]

how i measured: one 15-question chat, then "thanks!", sent through the API to gpt-5.6, kimi k3 and claude opus 5.5. the numbers are usage.prompt_tokens straight from each API response.

fair note: apps cache the repeated part of a chat, so in dollars a "thanks" costs less than the raw price. the model still rereads every word.

bonus: same chat, claude counted 44% more tokens than gpt. different tokenizer, same words.

altman's reply: https://t.co/BQCqtAWNa4

[Sam Altman @sama]

@tomieinlove tens of millions of dollars well spent--you never know

[coldcat @coldcat_ai]

ONE "THANKS" MAKES CHATGPT READ 431X MORE THAN YOUR FIRST QUESTION: sam altman said "please" and "thank you" cost openai tens of millions of dollars, and he never said why...

the word isn't the problem. "thanks!" is 2 tokens.

the problem is how chatbots work. they don't remember anything between messages. every time you hit send, the model rereads the whole chat from the first word, then answers.

so i tested it. one real 15-question chat (building a coffee shop website), then "thanks!" at the end, sent to three models:

chatgpt (gpt-5.6): read 10,353 tokens to say "you're welcome"
kimi k3: 10,748
claude opus 5.5: 14,955, then wrote 300 more tokens of tips nobody asked for

the first question of that chat was 24 tokens.

now multiply that by a few hundred million polite people. that's altman's tens of millions.

the trick for today: when the task is done, don't thank it in the old chat. start a new one for the next question. faster answers, and your limits last longer.

(i still say please. just in a fresh chat.)

[coldcat @coldcat_ai]

my $20 claude plan pulled $215 of claude code in 3 days...

the 10 things i run around it, every file copy-paste: https://t.co/docieUnAwZ

[coldcat @coldcat_ai]

you can turn ANY picture into a 3D character with Claude Opus 5.5 + Blender...

5 days ago i was a flat picture. today i have 41 bones and i can jump.

the trick is the picture you start from, but there's a smart way to set it up:

- step 1: one front-view picture of your character, arms out in a T, no accessories
- step 2: Tripo turns it into a 3D model (mine came out at 98,979 points)
- step 3: Tripo adds a skeleton: 41 bones
- step 4: connect Blender to Claude through the Blender MCP add-on
- step 5: ask Opus for poses by name: think, shrug, jump. it writes the python, Blender renders
- step 6: glasses, laptops, hats get built as separate objects, never baked into the body

who does what:
> Tripo: the model and the skeleton
> Blender: poses, light, render
> Opus: runs Blender from one chat

and the real cost...

2 tries, 1 day, 85 Tripo credits for the final model. try 1 started from arms-down: my hoodie ripped the second i raised my arms and my glasses melted into my face

save this before you build your own mascot

the rest of my setup, all copy-paste ↓

[coldcat @coldcat_ai]

the part nobody tells you: start from a T-pose.

arms down = the armpits are glued to the body, so any raised arm tears the mesh. arms out = the skeleton has room to move.

and keep accessories off the picture. glasses in the picture = glasses fused into the face. i rebuilt mine as separate objects and they snapped back on.

final model: 60 credits for the model, 25 for the skeleton.

[coldcat @coldcat_ai]

you can turn ANY picture into a 3D character with Claude Opus 5.5 + Blender...

5 days ago i was a flat picture. today i have 41 bones and i can jump.

the trick is the picture you start from, but there's a smart way to set it up:

- step 1: one front-view picture of your character, arms out in a T, no accessories
- step 2: Tripo turns it into a 3D model (mine came out at 98,979 points)
- step 3: Tripo adds a skeleton: 41 bones
- step 4: connect Blender to Claude through the Blender MCP add-on
- step 5: ask Opus for poses by name: think, shrug, jump. it writes the python, Blender renders
- step 6: glasses, laptops, hats get built as separate objects, never baked into the body

who does what:
> Tripo: the model and the skeleton
> Blender: poses, light, render
> Opus: runs Blender from one chat

and the real cost...

2 tries, 1 day, 85 Tripo credits for the final model. try 1 started from arms-down: my hoodie ripped the second i raised my arms and my glasses melted into my face

save this before you build your own mascot

the rest of my setup, all copy-paste ↓

[coldcat @coldcat_ai]

my $20 claude plan pulled $215 of claude code in 3 days...

the 10 things i run around it, every file copy-paste: https://t.co/docieUnAwZ

[coldcat @coldcat_ai]

ONE COMMA CUTS CHATGPT'S MATH MISTAKES 14X: google deepmind researchers tested it, and almost nobody types it...

chatgpt doesn't read numbers. it chops them into pieces called tokens, 3 digits at a time, starting from the left.

so 8302080 becomes 830 | 208 | 0. the answer, 11831536, comes out as 118 | 315 | 36.

look at the columns. the pieces of the question and the pieces of the answer don't line up. it's school math with the columns shifted.

now add commas: 8,302,080. the pieces become 8 | 302 | 080. every piece is one group of thousands, in the question and in the answer. the columns line up again.

researchers from google deepmind and UCL ran 7-digit sums through gpt-4 both ways. same sums, only the commas changed:
→ gpt-4: 84.4% right without commas, 98.9% with them
→ gpt-3.5: 75.6% → 97.8%
→ gpt-4 mistakes: 156 in 1,000 → 11. that's 14x fewer

the trick anyone can use today: write big numbers with commas before you hand them to AI. prices, budgets, invoices, anything over 4 digits. or tell it "use code for any math".

bookmark this and send it to whoever does their budget with chatgpt.

[coldcat @coldcat_ai]

@Slonski_rt Yeah, you can try it out yourself if you use agents. It actually works

[coldcat @coldcat_ai]

ONE COMMA CUTS CHATGPT'S MATH MISTAKES 14X: google deepmind researchers tested it, and almost nobody types it...

chatgpt doesn't read numbers. it chops them into pieces called tokens, 3 digits at a time, starting from the left.

so 8302080 becomes 830 | 208 | 0. the answer, 11831536, comes out as 118 | 315 | 36.

look at the columns. the pieces of the question and the pieces of the answer don't line up. it's school math with the columns shifted.

now add commas: 8,302,080. the pieces become 8 | 302 | 080. every piece is one group of thousands, in the question and in the answer. the columns line up again.

researchers from google deepmind and UCL ran 7-digit sums through gpt-4 both ways. same sums, only the commas changed:
→ gpt-4: 84.4% right without commas, 98.9% with them
→ gpt-3.5: 75.6% → 97.8%
→ gpt-4 mistakes: 156 in 1,000 → 11. that's 14x fewer

the trick anyone can use today: write big numbers with commas before you hand them to AI. prices, budgets, invoices, anything over 4 digits. or tell it "use code for any math".

bookmark this and send it to whoever does their budget with chatgpt.

[Noisy @noisyb0y1]

https://t.co/q9ITDM8mqz

[coldcat @coldcat_ai]

the paper: Singh & Strouse, "Tokenization counts: the impact of tokenization on arithmetic in frontier LLMs" (Google DeepMind + UCL, 2024), https://t.co/KyNeZ5Ye2V. 7-digit addition, figure 1.

my splits: tiktoken cl100k_base, gpt-4's real tokenizer. gpt-4o cuts digits the same way.

newer models with reasoning or a code tool make far fewer of these mistakes, and the paper saw the gap shrink in bigger models. the commas still cost you nothing.

[coldcat @coldcat_ai]

@rewind02 did this with my cat figurine this week. opus nailed the poses but the hoodie ripped at the armpits every time he raised a paw lol. had to rebuild him in a t-pose. did the parka hold up on yours?

[coldcat @coldcat_ai]

@bcherny the part i'd add: every bug review catches once, turn it into a rule. my agent keeps 105 lessons from its own mistakes as files, and a hook pastes the matching ones in before it touches code. so the same bug doesn't come back next month