A real-ish AI labEst. before the funding round

The AI lab that accidentally published its business plan.

OpenishAI is a public-benefit AI company making capable models in public, with the receipts attached and the investor update accidentally CC’d.

Openish-0.1 / live-ish Training in public

0.1

small enough to understand

8B parameters

Nothing to hide
except the latency

weights: public
ego: pending

Live from the transparency departmentWeights outReceipts attachedNo tasteful fog machines

Our public position

We believe the future should come with a changelog.

The industry has perfected the soft-focus launch film. We are trying something radical: saying what we built, how we built it, and which part we are still figuring out.

The weights are public.

Download them. Fine-tune them. Discover that our benchmark score was rounded up in a meeting.

The data has footnotes.

Every source, filter, and questionable PDF gets a line in the manifest. We even label the vibes.

The roadmap has feelings.

When a launch slips, the changelog says “we were wrong” instead of “we are thoughtfully iterating.”

The model, in public

Small enough to inspect. Weird enough to remember.

Openish-0.1 is our first general-purpose model. It writes code, summarizes meetings, and has a suspiciously nuanced view on oat milk.

parameters

42GB

download

openish playground

ready

Choose a prompt

youExplain OpenishAI in one sentence.

aiPress run. It is not a trap.

Apache 2.0 licensed No gated waitlist Eval suite included

Read the model card

Subject to change in public

The roadmap has no offsite.

A living document, updated when we learn something or when a founder says “quick question” in Slack.

Bring your best prompt injection. Please be gentle with the interns.

The awkward investor update

Receipts, not vibes.

A complete public record of what we shipped, what it cost, and which metric we quietly removed from the deck.

Get the monthly ledger

public / ledgerv. 06.11

Training data manifest

1,204 sources · 18 questionable PDFs

published06.11.24

Safety evaluation suite

Public red-team set · 3,800 prompts

published05.28.24

Series A investor letter

Revenue: spiritually positive

published05.03.24

Compute emissions report

A little embarrassing, actually

in review04.19.24

Showing all, regrettablyArchive

Come for the weights

Stay for the footnotes.

A monthly note from the lab. New releases, clean receipts, and one paragraph that should have stayed in the draft.