Siddhant

more intelligence less compute

[Siddhant @skcache]

@cdngdev Wait that was you??

[Siddhant @skcache]

@0xSigil Wow this is just amazing engineering!

[Siddhant @skcache]

@aaayandev @typesafeai @ycombinator This is awesome!!

[Siddhant @skcache]

@imnotchalk You’re right — and honestly, it’s not that AI replies are increasing, it’s that authentic human interaction is becoming harder to distinguish.

[Siddhant @skcache]

@iHarnoorSingh Yea remember they burned their entire years worth of AI budget in 4 months

[Siddhant @skcache]

@0xZergs Crine man, hopefully the airline found a solution tho

[Siddhant @skcache]

@danialhasan How about now?

[Siddhant @skcache]

Wow it only took one ChatGPT account for them to get access to their GitHub, Slack, and email.

And they opened a PR in the internal repo

[s1r1us @S1r1u5_]

On July 25, we hacked OpenAI.

Two bugs let us take over ChatGPT/Codex accounts of OpenAI employees (+some unaffiliated users) and reach connected services: Outlook, Slack, GitHub, etc.

We proved it with a PR in OpenAI’s internal codebase . It took us <72h. 🧵

[Siddhant @skcache]

@thdxr Absolutely, the traffic you can catch from this will be absurd. Also I need it

[Siddhant @skcache]

@danialhasan Would love to join

[Siddhant @skcache]

@notkevinzhang Thanks king

[Diogo Almeida @CompleteSkeptic]

you are not too stupid at all!

jev is meant as a tool for developers to put intelligence into software, so in an ideal world, average users would never know there was AI in their software, it would just silently be better

(kinda like not knowing what database your favorite software runs on)

[Siddhant @skcache]

@anshgrows Ansh! I’m more of a 2 guy myself

[Siddhant @skcache]

@rezoundous Hoping so, Astra is basically this for $20 users

[Siddhant @skcache]

@AdamHoltererer @Muse did this first

[Siddhant @skcache]

@CompleteSkeptic @justKDeng Can’t wait!

[Siddhant @skcache]

Source: https://t.co/oyFhFwGe4k

[Siddhant @skcache]

Jev is a new System One model from TypeSafe built for structured decisions instead of text generation.

New architecture, parallel sampling, Reinforcement Learning from Calibrated Decisions (RLCD), typed probabilistic outputs in 70–500ms.

I made a quick thread, take a look ↓

[Siddhant @skcache]

TypeSafe’s Doom demo runs ~10 Jev calls/sec for about $7/hour.

Jev is $0.042/MTok input, output is free, and latency is 70–500ms.

You can put model decisions inside tight software loops without cost dominating the design.

Excited to see where this goes, thanks for reading!