CodeLina_

notes from the code trenches

[CodeLina_ @CodeLina_]

@x_Kuro_ nope, bc i banned 2/3 from the prompt so it only showed what the 1,000 games gave, 65.2%. i checked the json myself after
btw, how many runs does your predictor do per race?

[CodeLina_ @CodeLina_]

I USED FABLE 5.2 TO SETTLE THE PUZZLE THAT FOOLED PAUL ERDŐS

Nearly 1,000 PhDs got it wrong too. It took Fable about 20 minutes and ~450K tokens to build the proof: a 3D film of 1,000 real games.

The puzzle is Monty Hall. Three doors, a car behind one, goats behind the other two. You pick a door. The host, who knows where the car is, opens another one. A goat.

Do you switch?

In 1990 Marilyn vos Savant said yes, in Parade magazine. About 10,000 readers wrote in to tell her she was wrong. One math PhD: "You blew it, and you blew it big!"

Erdős, one of the greatest mathematicians of the century, refused to believe her until a computer showed him.

So I gave Fable 5.2 the job of that computer:
> stay: 348 wins, 34.8%
> switch: 652 wins, 65.2%

Switching wins 2 times out of 3. She was right.

Then the version that finally makes it click: 100 doors. You pick one. The host opens 98 goats. Still keeping yours?

Who here still thinks it's 50/50?

[CodeLina_ @CodeLina_]

@theia_organon smart. btw fun fact: in a 2010 study pigeons learned to always switch, and people didn’t. we overthink it lol

[CodeLina_ @CodeLina_]

@theia_organon right? vos savant used the same trick back in 1990, but with a million doors
in my animation the pawn picked door 13 and the car was behind 31. would you still stay?

[CodeLina_ @CodeLina_]

@x_Kuro_ just one prompt, i wrote out the 6 scenes and the rules (real seeded sim, no fake numbers), it did the rest
so i didn’t touch the animation

[CodeLina_ @CodeLina_]

@x_Kuro_ i’d add how hard it is to overtake on each track. monaco and monza are not the same game. does it know the track type yet?

[CodeLina_ @CodeLina_]

@theia_organon fair point, picks and shovels win, but nobody gets an exclusive on the shovels

[CodeLina_ @CodeLina_]

Microsoft gave Anthropic $5 billion.

Anthropic promised to spend $30 billion on Microsoft's cloud.

Today it launched Claude Sheets, Whiteboards and Data. Excel, Miro and Tableau, inside Claude.

So Microsoft is getting paid by the company coming for Office.

Either Satya made the worst deal of the decade, or he already priced this in.

Which one?

[Claude @claudeai]

Starting to roll out today: Claude Whiteboards, Claude Sheets and Claude Data.

Think a plan through on a board, build a tracker in a sheet, and tell a story with your data.

[CodeLina_ @CodeLina_]

@theia_organon it was useful, ty for sharing this repo

[CodeLina_ @CodeLina_]

@claudeai I certainly have some questions for Microsoft...

https://t.co/6vExKfn9rz

[CodeLina_ @CodeLina_]

Microsoft gave Anthropic $5 billion.

Anthropic promised to spend $30 billion on Microsoft's cloud.

Today it launched Claude Sheets, Whiteboards and Data. Excel, Miro and Tableau, inside Claude.

So Microsoft is getting paid by the company coming for Office.

Either Satya made the worst deal of the decade, or he already priced this in.

Which one?

[CodeLina_ @CodeLina_]

ANTHROPIC’S OWN DATA PROVES YOU ARE BURNING CASH FOR WORSE RESULTS.

Claude Fable 5.2 on Max costs $44.70 per coding task. On High it gets the same score for $6.72.

I've just look at Anthropic's own chart. And Max isn't even the top of it.

The chart is FrontierCode 2.0 from today's launch post. It checks whether maintainers of real open-source repos would merge the model's code. Here's every effort level, score and cost per task:

> Low: 34.2%, $2.10
> Medium: 37.6%, $2.48
> High: 45.3%, $6.72
> Xhigh: 49.1%, $14.19
> Max: 45.6%, $44.70

Xhigh is the peak. Max costs 3x more than Xhigh and scores 3.5 points lower.

I'd rather run Xhigh three times than Max once. Three runs cost $42.57. One Max run costs $44.70.

My guess: past Xhigh it starts overthinking, and you pay for every extra token of it.

For scale, Opus 5.5 on Max sits at 45.5% for $12.82.

Medium is the default in Claude Code, and on this chart the big jump comes after it, at High.

So for me it's High for everyday work, Xhigh when the task is hard, and Max stays off.

Has anyone seen Max beat Xhigh on a real repo?

[Claude @claudeai]

Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.

They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.

[CodeLina_ @CodeLina_]

@me_barnyx interesting collection of repos, what's your fav?

[CodeLina_ @CodeLina_]

DARIO AMODEI JUST NAMED THE FIRST JOBS AI COMES FOR: JUNIOR LAW, JUNIOR FINANCE, JUNIOR CONSULTING.

His own company already put the tools for all three on GitHub. For free.

The clip is 42 seconds. Someone asks him what a kid finishing high school should do, and he says AI "can do entry-level law work. It can do entry-level finance. It can do entry-level consulting."

Last year he put a number on it: half of entry-level white-collar jobs, within 1 to 5 years.

So I opened Anthropic's repos and matched them to his list:
> junior law: knowledge-work-plugins, the legal folder
> junior finance: the financial-services repo, plus a finance folder in the same plugins repo junior
> consulting: the skills repo, with Excel, Word, PowerPoint and PDF

Together that's over 260,000 stars. People are already using this.

A first-year analyst's week is decks, models and memos. Those are the folders.

The junior who keeps a seat is the one who knows which plugin to run and can tell when it got the numbers wrong.

If you hire a junior next year, what do you still want them to do by hand?

[Hoodie @HoodieFade_]

Your agents can already work for hours without you

But you still check every step they take

Because you have a loop, but no graph

I broke down how to build a system where you only approve the last step ↓ https://t.co/x0Kg0o0Clp

[CodeLina_ @CodeLina_]

@theia_organon i honestly think this is one of the most underrated parts of claude code brcause it's a huge deal having that level of control without relying on the model to behave is a huge dea

[CodeLina_ @CodeLina_]

@theia_organon honestly, 40 cents is cheap enough to experiment with, but reliability on real projects is what would make it a serious claude alternative.

[CodeLina_ @CodeLina_]

OPUS 5.5 JUST LOST TO A MODEL THAT COSTS 40 CENTS A TASK
And the viral post about it got one comparison wrong.

It’s Qwen-4-Flash, an early checkpoint codenamed hawk. On GameFeel Bench it scored 49.9. Opus 5.5 on XHigh scored 47.2, at $2.60 a task.

So: 2.7 points better for 6.5x less money.

It also says the small 27B model beats Sonnet 5.5. It beats Sonnet on XHigh (42.8). Sonnet on Max still wins, 45.0 to 44.6.

The rest of the small print:
> one benchmark, built mostly around game-dev tasks
> an early checkpoint, not a release
> the 27B’s audio score was already fixed once, for “a genuine mistake”

I’m keeping Claude Code for now. But if hawk ships at 40 cents and holds up on real repos, $2.60 a task gets very hard to justify.

Would you switch your coding agent for 6x cheaper?

[Layton Gott @Layton_Gott]

NO WAY THIS IS REAL…

if this is real, Qwen 4 is INSANE:
> Qwen-4-Flash is #1, above Opus 5.5
> a 27B model is #4, above Sonnet 5.5
> Flash only costs $0.40 a task, Opus 5.5 costs $2.60

seriously, I'm having a hard time believing this is true

if it is, my goodness we're in for the time of our lives

[CodeLina_ @CodeLina_]

@theia_organon @beamnxw good point, glad someone raised this question (had the same thought)

[CodeLina_ @CodeLina_]

I AM BEGGING YOU TO STOP PASTING RAW NOTES INTO CHAT WINDOWS

Opus 5.5 kept answering from vibes. same notes, same model, a completely different answer every single session

So i moved everything into a karpathy-style markdown vault: one idea per note, every claim hard-linked to its source, and a daily log under 20 lines

Now the model has to cite a file for every sentence. if it can’t, the question goes straight to gaps.md instead of giving a confident guess

The full system prompt and vault setup are in the image belo

How are you organizing your notes for AI right now?

[Brucky @danikkk_wqs]

https://t.co/LKghIM2a0E

[CodeLina_ @CodeLina_]

I STILL DON'T UNDERSTAND WHY EVERYONE IS STILL PROMPTING WRONG

An anthropic engineer just leaked that 90% of their team stopped writing prompts altogether and started building systems that prompt themselves.

I spent hours breaking down her 30-minute masterclass on how to build autonomous agents that fix their own code, remember context, and scale without human fatigue. Trust me, the payoff is insane.

Honestly, this free guide destroys every overpriced $2,000 AI course on the market.

The system blueprint and exact code structure are below.

[Hanako @hanakoxbt]

https://t.co/B3240WRGhl