277 karma · joined November 5, 2018
I suppose that is what will keep happening. At some point, someone will decide to re-write everything from scratch and re-discover lower layers of abstraction or challenge them.
1. https://github.com/geniusrise/geniusrise - core framework
2. https://github.com/geniusrise/geniusrise-huggingface - hf modules
3. https://github.com/geniusrise/geniusrise-openai - openai modules
4. https://github.com/geniusrise/geniusrise-listeners - streaming data input
5. https://github.com/geniusrise/geniusrise-databases - database input
6. https://github.com/geniusrise/geniusrise-prompt-actions - functional integrations (RAG-able and GPT function call-able, WIP)
7. https://github.com/geniusrise/geniusrise-indexing - vectorizing for RAG usecases (WIP)
8. https://github.com/geniusrise/geniusrise-exit-proxy - cached LLM interface with MITM-auditing (WIP)
## Asides
I think the core framework can be AGPL but the modules must be MIT / Apachev2.
I really wanted to create an elaborate example in the guides but could not find time, - something like load and vectorize SNOMED-CT or UMLS and use it to NER / RAG EHR docs. Or maybe a usecase of doctor communicating to patient in another language (a major problem in India), with reverse translation verifying translated output using the KG. These kinds of stuff are soon to come. Or discourse segmentation for better chunking for RAG usecases.
I'm not sure if I should add cyberpunk-ed scientists as banner images. I tried with mathematicians like Voevodsky to Andre Joyal to John Baez, but couldn't. Actual geniuses tend to not be famous, hence SDXL fails I guess.
I plan to also write this framework in scala. The category-theorizing of neural networks is amazing!!! https://github.com/bgavran/Category_Theory_Machine_Learning. I hope Bartosz Milewski approves.
I love Alan Turing, but cuz of "The Chemical Basis of Morphogenesis". It introduced me to the wonderful world of complex systems. Hence, his image as banner.
I'm also working on a cli library called "isomorphic", wraps over argparse and provides cli, api, yaml, json interfaces.
Yes, gradio integration is also underway.
Finally, to huggingface.
1. https://github.com/geniusrise/geniusrise - core framework
2. https://github.com/geniusrise/geniusrise-huggingface - hf modules
3. https://github.com/geniusrise/geniusrise-openai - openai modules
4. https://github.com/geniusrise/geniusrise-listeners - streaming data input
5. https://github.com/geniusrise/geniusrise-databases - database input
6. https://github.com/geniusrise/geniusrise-prompt-actions - functional integrations (RAG-able and GPT function call-able, WIP)
7. https://github.com/geniusrise/geniusrise-indexing - vectorizing for RAG usecases (WIP)
8. https://github.com/geniusrise/geniusrise-exit-proxy - cached LLM interface with MITM-auditing (WIP)
## Asides
I think the core framework can be AGPL but the modules must be MIT / Apachev2.
I really wanted to create an elaborate example in the guides but could not find time, - something like load and vectorize SNOMED-CT or UMLS and use it to NER / RAG EHR docs. Or maybe a usecase of doctor communicating to patient in another language (a major problem in India), with reverse translation verifying translated output using the KG. These kinds of stuff are soon to come. Or discourse segmentation for better chunking for RAG usecases.
I'm not sure if I should add cyberpunk-ed scientists as banner images. I tried with mathematicians like Voevodsky to Andre Joyal to John Baez, but couldn't. Actual geniuses tend to not be famous, hence SDXL fails I guess.
I plan to also write this framework in scala. The category-theorizing of neural networks is amazing!!! https://github.com/bgavran/Category_Theory_Machine_Learning. I hope Bartosz Milewski approves.
I love Alan Turing, but cuz of "The Chemical Basis of Morphogenesis". It introduced me to the wonderful world of complex systems. Hence, his image as banner.
I'm also working on a cli library called "isomorphic", wraps over argparse and provides cli, api, yaml, json interfaces.
Yes, gradio integration is also underway.
Finally, to huggingface.
Introducing geniusrise, an agent framework and component ecosystem for building AI agent networks that are as flexible as your team.
landing page: https://geniusrise.ai (fancy but useless) docs: https://docs.geniusrise.ai (please check this out) github: https://github.com/geniusrise (for dear devs)
## Thought process
Since the ChatGPT disruption, I've been pondering on what the tooling layer is going to look like for building LLM-interfacing agents. Saw a plethora of tools coming out as we witness here every week. I'd broadly categorize them into the GUI drag-and-drop variety, and the airflow and MLflow competitors pivoting. Not counting experiment dumps.
My opinions kind of differed from these. The thesis that I chose to work with, is based on Conway’s law - the tooling for building agents must accommodate various team, communication structures and expertise available in an org. This scheme could only work if the tooling itself was more of a thin infrastructure glue layer supporting loosely coupled components such that each component could be built, tested, deployed and re-used independently while leveraging the context and expertise of the ones building them. Also these components need to be "Write once, run anywhere".
## Scheme
1. Developers build components relevant to their infrastructure and services.
2. ML engineers and data scientists get to work on data massaging, prompt engineering and model training without having to build data ingestion pipelines from whatever database engineering teams or whatever service product may be using.
3. Devops could deploy these components in their choice of runner (e.g. k8s) and orchestrate with their choice of orchestrator (e.g. airflow).
4. Product, engineers or whoever could take these components and compose them into workflows, experiment, test, then hand over to devops to deploy.
5. Some of these components end up as open source, and benefit the entire community.
Whichever of the above layer is not available, well, that’s where our open source modules kick in For example, no data science team? Use our standard huggingface components.
Thus, the builders build, the users use, and open source plugs in the holes.
The framework enables EVERYONE, without getting in the way.
## Community and Future
Guys, this is for you, and I want to make it useful. Some feedback would be nice, and it would help me create a better roadmap. I plan to make this framework production ready by December '23.
Having worked alone, and after countless sleepless nights and social sacrifices, I wanted to share this project before I start working on it full time from next week.
Meanwhile I forgot to activate github sponsorships Anyway, I am excited!
There is more in the pipeline - integrations (For example, model management, data governance and quality, runners, etc etc), modularization of the core framework, MOAR components, unit tests for components, etc
Hell, the batch data modules do not even have partitioning schemes yet
As the project grows bigger, I will be open to co-maintainers.
The north star goal would be to build something worthwhile that can be donated to Apache / CNCF.
Code is meant to be mostly read not written. And no, LLM explanations are shit useless, a language model simply cannot handhold a human to make them understand logic line by line, you NEED to read it, and hence code has to be "readable", which to an engineer means "precise".
What do you mean? Android is Linux!
- should I reset it every iteration? resulting in it behaving more like working memory maybe
- should I reset it every epoch? would that pollute the memory from previous iterations and what would happen if it got full?
- finally, why not maybe delete say 0.01% of the memory or maybe 1 cell per iteration randomly, this would imply the memory would not be that reliable as biological memory behaves and the neural net has to build resilience to use it effectively (by storing multiple copies of what is really useful (hyppocampal novelty detection type?)).
If you see toward the end where it generates Einstein's field equations, I had to start from real numbers (dedekind cuts) to real manifolds to pseudo-Reimannian manifolds to the curvature metric to the final thing.
A good example looks like:
``` I am trying to model associative memory that i may attach to a gpt model
here is the code for the memory:
....
can we keep the input vectors in another array so we can fetch items from it directly instead of having to reconstruct? ```
"I have soft power". Suspicious indeed.
I wrote this list for EMs in my org so they can have a handy list of tools and a good point to start digging deeper.
Of course, sharing such a long list will inevitably cause the TLDR effect. https://www.psychologytoday.com/us/blog/boundless/201903/the...
[Technically I imply a heavily skewed exponential distribution, maybe not exact numbers].
In the last decade, I might have very rarely seen a resume with more than say 10 years of experience at a single place. The average would be somewhere in the range of 1-4 years.