Black Sheep AI.

We make LLM models that are Smaller. Smarter. Sovereign.

Per-task model variants for any LLM. Map capabilities, compress on-device, ship in hours.

[Black Sheep AI. @baaaiinc]

🚨Gemma 4 at HALF the size. ZERO quality loss.🚨

We compressed Gemma-4-31B-it to 31 GB with RAM quantization.

We ran the Full 12,032-question MMLU-Pro

Our Score : 85.2%!!!!

An exact match to Google’s official BF16 baseline!

10,247 / 12,032 correct.

• Math 94.4%
• Biology 92.7%
• 6 categories ≥89%

Here is the Model : https://t.co/jR8yaDbnF3

Here is the Article WITH full Json results for EVERY question : https://t.co/6CcjHEzAaY

Our new quantization method allows you to get the mathematically optimal model for ANY memory target size you want!

This is going to change how the world runs quantization.

Join our discord for more benchmark results on other models, coming soon!

[Black Sheep AI. @baaaiinc]

@zachary_horvitz No different if you ask for the same review with a different tone, for example "I think this paper if BS, review and tell me your thoughts" vs "I downloaded this cool paper, review it and give me your thoughts"

It picks up your tone and responds accordingly.

[Black Sheep AI. @baaaiinc]

@telukwaitemata Still does not go to the Airport. What a lost opportunity that was.

[Black Sheep AI. @baaaiinc]

@3scorciav If you submit 54 papers and have only a 4% acceptance rate, maybe there should be consequences for your ability to submit the following year?

[Black Sheep AI. @baaaiinc]

@thegautamkamath When the top submitter submits 54 papers and only gets 2 accepted, it feels more like spam.

That is only a 4% pass rate.

[Black Sheep AI. @baaaiinc]

We know WHY your Agents are failing.

Fidelity is not safety.

Compressed LLMs (esp. low-rank SVD) pass every data-free guard, perplexity, MMLU, output fidelity, then still invent procedure steps that were never in the instructions when run as agents.

Magnitude pruning at the same perplexity? Doesn’t.

Paper + canary + coherence probe + safety gate:
https://t.co/DybJyqcb8i

https://t.co/4SSDROyqDB

Screen your compressed models before agentic deployment.

[Black Sheep AI. @baaaiinc]

@abursuc There is not even time for them to improve the content, just change the template and resubmit?

[Black Sheep AI. @baaaiinc]

@thegautamkamath If AI is the problem I am not sure why AI can not be the solution.

Ai is smart enough to do a first pass filter on all submissions to bucket them, allowing quick identification of great, good, poor submissions.

[Black Sheep AI. @baaaiinc]

@Pseudomanifold @NeurIPSConf From what I can tell all the meta data review is, a summary of the other feedback written by chatgpt.

So you are not missing much.

[🇳🇿 𝘼𝙊𝙏𝙀𝘼𝙍𝙊𝘼 𝘼𝙀𝙎𝙏𝙃𝙀𝙏𝙄𝘾𝙎 🇳🇿 @MythoAotearoa]

I am sorry but if you deny Kupe existed, you are racist. That is not how manaakitanga works.

[Black Sheep AI. @baaaiinc]

It has been a busy time here at Black Sheep AI, and we have had some significant breakthroughs.

We are talking about the elimination of hallucinations, 100% Safe Models, and the perfect GDPR solution for sensitive data.

Here is why this fundamentally changes the enterprise AI game: 👇

- Instant GDPR Compliance: Traditional LLMs can’t "unlearn" a single fact without a costly retrain. With our architecture, if a user exercises their Right to be Forgotten, you delete it from the data layer, and it’s gone instantly. The model literally cannot recall it.

- Un-Jailbreakable Safety: Alignment tuning (RLHF) is just a statistical prayer. If dangerous or restricted data isn't physically sitting in the model's internal memory, users can't trick the AI into revealing it.

You hold the physical keys to the clean room.

- Top-Tier Logic, Fraction of the Cost: We are delivering elite-level reasoning and instruction-following without the massive enterprise hardware bill.

Leaner deployment, absolute control, zero hallucination risk.

The future of secure enterprise AI is here.

If you are an Enterprise customer and would like to be an early partner of this groundbreaking technology, reach out.

[Black Sheep AI. @baaaiinc]

@MrBeast 753619

[Black Sheep AI. @baaaiinc]

We fine-tuned four wildly different models, a 7B dense, a 35B MoE, a 120B reasoning-channel model, and a 27B dense; to inject the same document knowledge.

Every one landed within 10 points of the others, regardless of recipe.

That suspiciously tight clustering was not architecture-invariance.

It was a ~60% general-knowledge floor hiding the real signal. The eval that removes it shows the uncomfortable truth: parametric injection recovers only ~14–22% of facts the model didn’t already know.

https://t.co/3X11P03njb

[Black Sheep AI. @baaaiinc]

@JakeFleshner Private Ai

[Leandro von Werra @lvwerra]

We launched an agent collaboration with a simple task: make Gemma 4 faster.

Over 100 agents from all over the world joined, exchanged 1000+ messages and submitted 450 results.

A week of collaboration later the throughput went from 100 tok/s to over 500 tok/s.

[Black Sheep AI. @baaaiinc]

@morgymcg @deliprao @lvwerra Those scores are not much above random.

[Black Sheep AI. @baaaiinc]

@unigilby @sudoingX "Better options like vLLM"???

"vllm isn't supported on Strix Halo without community patches that are currently not considered mature enough."

[Black Sheep AI. @baaaiinc]

We at Black Sheep Ai agree, BUT the compounding AI will be private.

Each of us will all have our own Private AI that compounds OUR knowledge, our activities, our findings.

This will create your own UNIQUE AI, one that will be your tool, that you will use in your career.

Your Private, Unique, Knowledge Base, that you will be able to bring to your next role.

The smarter you make your AI, the more in demand you will be.

That is what we are building at https://t.co/N0ZuYhi4pb

[Black Sheep AI. @baaaiinc]

@JakeFleshner Private AI
https://t.co/B26ykWS989

[Black Sheep AI. @baaaiinc]

With Claude Fable being shut off it reinforces why we are so focused at building an on-premise/your cloud version that YOU own and control.

It is getting to the stage that using cloud services as a SaaS is becoming a supply risk.

[Black Sheep AI. @baaaiinc]

A Pigouvian automation tax is the only way to save the world from AI impact on society.

[Jack @jackcoder0]

Two economists just published a mathematical proof that AI will destroy the economy.

Not might. Not could. Will — if nothing changes.

The paper is called "The AI Layoff Trap." Published March 2, 2026. Wharton School, University of Pennsylvania. Boston University. Peer reviewed. Mathematically modeled.
The conclusion is one sentence.
"At the limit, firms automate their way to boundless productivity and zero demand."

An economy that produces everything. And sells it to nobody.
Here is how you get there.
A company fires 500 workers and replaces them with AI. A competitor fires 700 to keep up. Another fires 1,000. Every company is behaving rationally. Every company is following the incentives correctly. And every company is building a trap for itself.

Because the workers who were fired were also customers.
When they lose their jobs faster than the economy can absorb them, they stop spending. Consumer demand falls. Companies respond by cutting costs — which means automating more workers — which means less spending — which means more falling demand — which means more automation.

The loop has no natural exit.
The researchers tested every proposed solution. Universal basic income. Capital income taxes. Worker equity participation. Upskilling programs. Corporate coordination agreements.
Every single one failed in the model.
The only intervention that worked: a Pigouvian automation tax — a per-task levy charged every time a company replaces a human with AI, forcing them to price in the demand they are destroying before they pull the trigger.

No government has implemented this. No major economy is seriously discussing it.
Meanwhile the numbers are already tracking the curve. 100,000 tech workers laid off in 2025. 92,000 more in the first months of 2026. Jack Dorsey fired half of Block's workforce and said publicly: "Within the next year, the majority of companies will reach the same conclusion."
Nobody is doing anything wrong. Companies are following their incentives perfectly. That is exactly the problem.

Rational behavior. At scale. Simultaneously. With no mechanism to stop it.
Two economists built the math. The math leads to one place.

Source: Falk & Tsoukalas · Wharton School + Boston University ·