BREAKING: GPT-6 Astra absolutely crushes Tiny Computer Bench™️, by finally hacking in to this dumb little toy computer I bought off Amazon

edits courtsey of codex x astra

it's the new Tamagotchi, except you feed it tokens 😂

claire you really made me buy this

Will send you my CLI and skills

When no other benches are trustworthy, TCB never fails hahaha

Can't wait to build something fun for my kiddos with this! @clairevo can you send me the CLI astra made for you to get into this thing?

This is the only “BREAKING” post worthy of the alert.

Hold up now. Codex hooks? Tell me more (I wonder if we can go the other way tooooo)

I can imagine having those around the office.

@whatRanjuSaid so cute na

FINE I bought one. Let’s do this! Send me things

this is soo sick. i want to get into hardware someday

Nice! But I bet it costs more in tokens than this PC could ever dream of being worth 😅

toy computer aside, benchmark contamination makes this result hard to trust without held-out tasks

#AIMadeForLosers

hacking a dumb amazon toy computer as a bench is so unserious and somehow the most convincing astra clip i saw today

First objection to make, and almost nobody makes it. Held-out has to mean held out by date though, or it does nothing: split a corpus the model already read and both halves are dirty, so the reserved part certifies nothing. Does that result quote a cutoff or only a split?

AOC is working with agents now? Looks like be a congresswoman was not a good job

I'm going to get this an sit it on my desk and just show episode after episode of Mad Men on a loop. Is there a retro TV version?

I appreciate the effort to make benchmarks more appealing 😂😭😭😭

@mweinbach How can I make this? lol I want one lol

Now this is a benchmark I care about

Don't make me buy this. This is too cool. 🫥

@mweinbach Were you able to get to deploy a custom firmware or are you just sending updates over Bluetooth? A few of us have been hacking away on these for a week or so in our spare time. Glad to see someone else thought the same thing after seeing it.

@mweinbach Right now it’s streaming over Bluetooth. My kids actually like the pre installed software but I’m planning to flash it next. Dm me if you’ve learned anything cool!

Model ≠ pipe. Truell put OpenAI at ~5%.

Haha love this

Astra "No problem"

This is great haha! Makes me want to try connecting my smart TV and turning the whole living room screen into a Codex workstation.

this is so cuteeee

I love that little device - I am thinking of doing something similar with a small e-ink screen

Omg. Told Fable 5.1 I wanted a tamagotchi style game and gave it a picture of my daughters hamster and grok imagine image gen pipeline. She’s in love. 10/10 would recommend

I neeeeed it

Have you tried getting it to load your agent fully into a phone ?! I’m going to doing an experiment with open claw as soon as I get access 🫣

Benchmarks are a psyop, this is the real eval.