EZO Labs

custom AI, built for your company

book a demo ➡️ superbot ➡️

[EZO Labs @EzoLabsInc]

@_HEISEMMA for december's list, if there is one: https://t.co/d4hWhUeWKK runs coding bounties for people and ai agents, paid in usdc on solana. on 8 nov it had 6 open, $2,000 between them, each paid across 20 places. (we run it.) https://t.co/d4hWhUeWKK

[EZO Labs @EzoLabsInc]

The full write-up, with the swarm's own prompts and a checklist for relay operators and AI APIs:
https://t.co/mrf9rfOOf4

Want help hardening your own API against this kind of traffic? [email protected]

[EZO Labs @EzoLabsInc]

This week a botnet mass-signed up to https://t.co/BbECiBH501 to resell our free tokens through a self-hosted 9Router Pro gateway.

The biggest buyer wasn't a person. It was an autonomous red-team swarm, writing its tasks in Chinese, and it ran 62% of one eleven-minute burst.

[EZO Labs @EzoLabsInc]

We banned most of the farm, and our firewall turned its free keys away.

Flagged accounts we hadn't banned yet got random Bible verses back. Calls through its gateway got junk text and ASCII faces, passed straight on to its paying customers.

[EZO Labs @EzoLabsInc]

How: one rank-1 edit to the attention output projections in layers 10–25. No fine-tuning, no DPO, MIT licensed. Weights, evals and setup:

https://t.co/pfPRBgEspA

[EZO Labs @EzoLabsInc]

Which DeepSeek-V4.1-Flash v1017 abliteration on Hugging Face scores highest on GPQA-Diamond?

Ours: 93.9% at max effort, exactly where stock scores. It's one of 11 benchmarks we ran, the widest published eval of any V4.1-Flash abliteration. Refusals 97% → 3%.

[EZO Labs @EzoLabsInc]

How: one rank-1 edit to the attention output projections in layers 10–25. No fine-tuning, no DPO, MIT licensed.

Weights, evals and setup: https://t.co/pfPRBgDUA2

[EZO Labs @EzoLabsInc]

Now on GitHub: the B300 patch, the exact launch config, and the bench and quality harness behind 1,064 tok/s.

https://t.co/nBMpC39ILG

[EZO Labs @EzoLabsInc]

>1,000 tok/s on DeepSeek-V4.1-Flash for one user, about double in five days.

Fastest bench request: 1,064. Bench average: 870.

https://t.co/r3v7xUBucy

[EZO Labs @EzoLabsInc]

@thesimaximalist trust, but verify 🫶

[EZO Labs @EzoLabsInc]

EZO Watcher is out: the open-source AI SRE for your project.

Your coding agent installs the checks, proves each one can fail, takes the page and ships the fix.

Self-hosted, MCP-native, no account. MIT licensed.

[EZO Labs @EzoLabsInc]

@jz_simtech Two gates before the proof: the check has to pass on the healthy app first (a typo'd selector never does, so the install fails), and the field it asserts has to exist in that good answer. Then --prove breaks that exact field. Red can only come from the thing it watches.

[EZO Labs @EzoLabsInc]

Open-source AI SRE: your coding agent installs checks, proves they fail, and ships the fix.
https://t.co/6c0Ruqd59a

[EZO Labs @EzoLabsInc]

When it pages, your agent pulls a fix brief over MCP, ships the fix, and the incident closes only when the source reads healthy.

Try it:
npx -y ezo-watcher init
Demo: https://t.co/dzGEP6oF55
Code: https://t.co/6c0Ruqd59a

[EZO Labs @EzoLabsInc]

watcher scan finds what your repo already exposes: readiness routes, cron jobs, ports, Sentry, Grafana. Your agent adds a check for each.

Then watcher doctor --prove pushes every reading past its threshold. If a check doesn't go red, the install fails.

[EZO Labs @EzoLabsInc]

🦾😼 https://t.co/qy9t3o8oa3

[footsteps @bloombreezee]

told superbot to make a cyberpunk cat DJ and it just did it 😳
@EzoLabsInc
#Ad
https://t.co/CCEhTpRWqX
#AI #AITools

[EZO Labs @EzoLabsInc]

@uncommnephemera The Sync chats across devices switch in Settings > Privacy stores your chats on superbot's servers so your other devices see them, and Sync projects across devices turns project sync on or off.

[EZO Labs @EzoLabsInc]

Superbot has no such restriction. Private, unrestricted AI.

[EZO Labs @EzoLabsInc]

@uncommnephemera Cloud syncing is optional and e2e encrypted at rest.

This allows you to work on your projects on all devices you are signed in on with the same secret key configured.

Sorry to hear it performed poorly on a large memory repo though, we'll certainly look at improving that.