Arpit (@arpittalksstuff) header
Arpit (@arpittalksstuff) profile picture

building @just_unmute · talk to claude code & codex by voice · shipping in public from bengaluru · @iitgn

Bengaluru
Joined August 2022
Photo by Arpit (@arpittalksstuff)Photo by Arpit (@arpittalksstuff)Photo by Arpit (@arpittalksstuff)Photo by Arpit (@arpittalksstuff)Photo by Arpit (@arpittalksstuff)Photo by Arpit (@arpittalksstuff)Photo by Arpit (@arpittalksstuff)Photo by Arpit (@arpittalksstuff)Photo by Arpit (@arpittalksstuff)Photo by Arpit (@arpittalksstuff)
i gave fable 5.2 one prompt and went to sleep 11 hours later it had built the entire history of earth as a fully interactive 3d globe. big bang to today drag the timeline to any year, spin to any place, click anything to go deeper. this is just ww2 the last thing on the timeline is the ai that built it
I wanted to test Fable 5.2's quality with one-shot prompts without human feedback. Everything in this game is from scratch: models, sounds, textures. ~$550 at API pricing, ~2.7M output and 445M input tokens, heavily cached. Took 8 hours! Play it here: claude.ai/artifact/BGY1SztrA…
13
8
2
45
9,490
the laziest possible approach already beat reducto reducto turns messy pdfs (bank statements, contracts, scanned forms) into clean data an ai can actually use. i'm trying to build an open source version of it with fable first test: no pipeline, no tricks. just show claude haiku the page and ask it to read it it got 80% of a public benchmark's checks right. reducto's own model got 76% catch: reducto sells a version for $0.001 a page. haiku cant go that cheap
1
2
34
the laziest possible approach already beat reducto reducto turns messy pdfs (bank statements, contracts, scanned forms) into clean data an ai can actually use. i'm trying to build an open source version of it with fable first test: no pipeline, no tricks. just show claude haiku the page and ask it to read it it got 80% of a public benchmark's checks right. reducto's own model got 76% catch: reducto sells a version for $0.001 a page. haiku cant go that cheap
trying to beat a yc company with fable 5.2 reducto turns messy pdfs into clean data for ai. on their own benchmark they beat azure, aws and gpt-4o i'm giving fable that exact benchmark and their official grader, and asking it to build an open source version that wins every number goes up here, even the ugly ones. the first one already surprised me
2
3
4
7
497
this is a whole journey, not a one off post. follow along if you want the next parts, i'll keep posting every number as fable builds it, good or bad
26
next experiment: trying to beat a yc company with fable. first result already surprised me xitter.twitterapi.xyz/arpittalksstuff/status…
the laziest possible approach already beat reducto reducto turns messy pdfs (bank statements, contracts, scanned forms) into clean data an ai can actually use. i'm trying to build an open source version of it with fable first test: no pipeline, no tricks. just show claude haiku the page and ask it to read it it got 80% of a public benchmark's checks right. reducto's own model got 76% catch: reducto sells a version for $0.001 a page. haiku cant go that cheap
8
trying to beat a yc company with fable 5.2 reducto turns messy pdfs into clean data for ai. on their own benchmark they beat azure, aws and gpt-4o i'm giving fable that exact benchmark and their official grader, and asking it to build an open source version that wins every number goes up here, even the ugly ones. the first one already surprised me
1
195
testing a theory right now opus 5.5 plans and orchestrates. fable 5.2 only executes, one big job per session, no back and forth long chatty fable sessions burn tokens and waste what its actually good at will share numbers once its done
1
78
what i noticed after letting fable 5.2 run a whole project alone long running doesnt really mean long. its more like wide and deep at the same time. research in one corner, deep code in another, testing somewhere else, and every branch has its own mess inside it. fable just holds all of it, in one shot. havent seen that from any model before stuff thats working for me - write a proper spec. honestly thats the only bottleneck now, not the model - let it research as it goes and dont let it assume anything. its research is good enough to trust - make it check its own work. opus 5.5 did this too but missed a lot. fable actually catches its own mistakes - then just step back and let it steer this is what one prompt and 11 hours got me
i gave fable 5.2 one prompt and went to sleep 11 hours later it had built the entire history of earth as a fully interactive 3d globe. big bang to today drag the timeline to any year, spin to any place, click anything to go deeper. this is just ww2 the last thing on the timeline is the ai that built it
1
2
119
wanted to post a fable 5.2 example last night. its still running. almost 6 hours in lol
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.
2
1
131
i gave fable 5.2 one prompt and went to sleep 11 hours later it had built the entire history of earth as a fully interactive 3d globe. big bang to today drag the timeline to any year, spin to any place, click anything to go deeper. this is just ww2 the last thing on the timeline is the ai that built it
38
try it: earth-whole-story.pages.dev what's inside: - 13.8 billion years on one scrollable timeline - continents drifting from real plate data - 67,000+ events, every country - 16 scroll stories (ww2, black death, spice routes, silk roads...) - 12 trails (follow the buddha, the story of salt...) - kid mode, sources on everything
4
3
20
839
Arpit retweeted
Put it right here” is now a 3D workflow. Draw on your scene. Tell Mixar what you want. Watch your agent build it. Draw to Life mixar.app
1
3
3
8
510
i said friday night. anthropic said monday.
gpt 6.2, fable 5.5 and qwen 4 all supposedly this week. one of them is shipping friday night and ruining everyones weekend
1
103
The cheap model is the one your agent actually runs all day.
Releasing Muse Ember today -- our smallest model that delivers higher intelligence per dollar than others in its class. Built for token-heavy work that agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. Available today on the Meta Model API and in Muse Code.
2
133
gpt 6.2, fable 5.5 and qwen 4 all supposedly this week. one of them is shipping friday night and ruining everyones weekend
1
2
248
this bridge only exists where my lantern shines 🏮 first time ever making a game: a papercraft adventure about carrying the last flame home. built with Opus 5.5. the courier, lanterns and ruins were all generated in Mixar. would you play this? #indiegame #gamedev
10
6
2
22
456
half my job now is saying "yes continue" out loud to a laptop
34
i want two numbers on every ai demo now. what it cost and how many times u got mad at it
40