Model Brew

Fine tuning as a service with Continual Learning → Provable Knowledge Layer | |

[Model Brew @MBrew26730]

Our preprint, about continual learning to prevent catastrophic forgetting with our CRMA. https://t.co/uhCB0aTLgz

[Model Brew @MBrew26730]

One is about teaching a model what you do. The other is about the facts it works from, and only ever needs today's version.

That split is the whole design.

[Model Brew @MBrew26730]

Your enterprise has two kinds of data.

Static data: skills and subject matter that change slowly. ModelBrew's Add a Domain teaches a model new domains without breaking the old ones. 1/2

[Model Brew @MBrew26730]

Dynamic data - Fees, rates and policies change constantly, and a model that learned the old answer keeps giving it. Verifiable Knowledge Layer keeps those facts outside the model, so you update a fact once,answers use it the same day, and every answer shows where it came from

[Model Brew @MBrew26730]

Export it. Self-host it. Update it continually without catastrophic forgetting.
That’s exactly what we built at ModelBrew.
Own the model. Own the knowledge. Kill the learning loop.
Try it free → https://t.co/QYMW9Gy4Hm
#AI #OpenWeight #FineTuning #ContinualLearning

[Model Brew @MBrew26730]

Satya Nadella called it the Reverse Information Paradox.
Every time you use a closed AI model, you pay for intelligence twice:
Once in tokens.
Again in the proprietary knowledge you leak through prompts, corrections, and the learning loop.1/n

[Model Brew @MBrew26730]

Your “alpha” slowly compounds… for the provider.
There’s a better way.
Fine-tune open-weight models you actually own.
Keep the knowledge inside your model. 2/n

[Model Brew @MBrew26730]

Stop pretending RAG or fine-tuning solves knowledge control.
We built a verifiable knowledge layer:
Teach → Answer → Erase (with certificate).
Weights stay frozen. Facts stay under your control.

[Model Brew @MBrew26730]

New feature energy:
An editable knowledge layer that sits beside the model.
Add facts in seconds. Delete them with proof. Zero cross-contamination.
This is what governed AI should feel like.

[Model Brew @MBrew26730]

Most AI can’t actually delete a fact.
Ours can and gives you a verifiable certificate when it’s gone.
Live Knowledge: teach it, use it, erase it.

[Model Brew @MBrew26730]

CRMA on Mistral-7B (5 domains, 3 seeds):
Naive LoRA: +43% drift
CRMA: −0.17% drift
Base stays frozen. New domains stack cleanly.

[Model Brew @MBrew26730]

Catastrophic forgetting is still the default.
It doesn’t have to be.

[Model Brew @MBrew26730]

RAG breaks.
Normal fine-tuning forgets.
There’s a third option: keep adding knowledge without wrecking what the model already knows.

[Model Brew @MBrew26730]

If you can’t delete a fact from your AI and prove it’s gone, do you really own it?
We’re building the version that can.

[Model Brew @MBrew26730]

Most fine-tunes forget the old stuff when you teach them something new.
We fixed that.
One model, multiple domains, almost zero forgetting.

[Model Brew @MBrew26730]

@Real_Cambria How far r u in the process?

[Model Brew @MBrew26730]

@karpathy ‘Explain in plain english like you are explaining to a 5th grader with 2 examples’ is what i usually do

[Model Brew @MBrew26730]

@Ronanchamberss Done ✅ .

[Model Brew @MBrew26730]

@opey_aro Similar research at https://t.co/QYMW9Gy4Hm, we are working on continual learning as well.

[Model Brew @MBrew26730]

Battle is ON
Grok bot vs Meta Muse vs Chat GPT dots

Lets see who wins.