owned this note
owned this note
Published
Linked with GitHub
# Examples of Open Source Business Models
[TOC]
## Book Dash Excerpt
Open source has been an important theme for AI communities in academia and industry. Popular open source software libraries like OpenCV and Scikit-Learn and open dataset and model providers like Hugging Face and Open AI have shaped how AI algorithms are discovered, improved, shared, and deployed. While each organisation pursued different paths towards sustainability, reflective of when they were founded, some themes emerge around important initial steps.
Of the four organisations of focus in this case study, OpenCV and Scikit-Learn started as community-driven projects. Both had a small group of motiviated individuals with a mission of creating "optimised" and "accesible" code within their domains and found other interested supporters within their organisation or networks to help continue building out the project. These grassroots efforts served to generate interest among many across academia and industry and foster a sense of community among these different user groups. The Scikit-Learn community has recently spun out a company called [Probabl](https://techcrunch.com/2024/02/01/probabl-is-a-new-ai-company-built-around-popular-library-scikit-learn/?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAAFCXO8XIZzp8FeD8sLQNQ70Zxza_SBRbDTbwcmZLpIqLQMuDvh7_fXvKsEAnVwaJ0P9RMvta-MYuE_Qi7-TJKoJ97Ab-nyHJHYncFVWTAdrdFc4Ao4qOTx6Y4PD73yJIi3-gFt6w2S3IVYcA4PnbNAg5xVbdrtXPVF-CsnFvhX9T), to offer services such as training and consulting relating to the scikit-learn codebase.
In contrast, OpenAI and Hugging Face started as investor-funded organisations to tackle specific topics such as AI Safety and Chatbots. While Open AI began as a self-defined open source non-for-profit and Hugging Face as an AI product company, both have taken different paths as they have grown. After a private investment from Microsoft, Open AI split into a not-for-profit and for-profit arm. Around the same time, they also began exploring new models of releasing code and models in a more closed or tiered-access way, different from the original "open appraoch." In contrast, Hugging Face has moved in a different direction, transitioning from an AI product buisness model to become an open source AI model and data platform, with over 100,000 pre-trained models and 10,000 datasets hosted on the platform. It has also built [BigScience](https://bigscience.huggingface.co/), a multi-language Large Language Model (LLM) openly with collaborators around the world.
From these examples open source AI sustainability models a few themes stand out:
- Identifying a strong organisational partner such as a national institute like INRIA or a known industry programme like Google Summer of Code can be a strong catalyst when starting out both for resources and community building
- Finding strong first supporter(s) is essential for sustaining development and to spread the word to others
- Identifying a value add to the ecosystem beyond the "open" aspect is helpful, for example OpenCV's focus on optimisation for computationally heavy image processing or Scikit-Learn's simplicity of API design for ease of use
- Projects that start with a closed focus, in the private sector, can become open and vice versa
- Starting with international teams can help the growth of an open source project, as demonstrated by OpenCV, Scikit-Learn, and Hugging Face
- Build for the existing ecosystem's needs - OpenCV and Scikit-Learn's compatibility with Python has encouraged growth with developers already using Python and SciPy for their work.
| Organisation | Founding Year | Founder & Founding Organisation | First Supporter(s) | Motivations |
| -------- | -------- | -------- | -------- | -------- |
| OpenCV | 2000 | Gray Bradsky, Intel Engineer | Intel researchers in Russia | "...not only open but also optimized code for basic vision infrastructure. No more reinventing the wheel."" |
| Scikit-Learn | 2007 | David Cournapeau, Google Summer of Code | Matthieu Brucher (thesis) & INRIA (French National Institute for Research in Digital Science and Technology) | "designed to be simple and efficient, accessible to non-experts, and reusable in various contexts" |
| OpenAI | 2015 | $1B investment from Sam Altman and Elon Mask to tackle challenges around AGI | $1B investment from Microsoft which led to capped-for-profit and not-for-profit division | "We will attempt to directly build safe and beneficial AGI, but will also consider our mission fulfilled if our work aids others to achieve this outcome." |
| Hugging Face | 2016 | Clement Delangue and Julien Chaumond, Hugging Face chatbot apps | VC funding | "We think the direction of responsible AI is through openly sharing models, datasets, training procedures, eval metrics, and working together to solve issues" |
## Anne Hari Meeting
- Framed around the initial steps/decisions to get started with governance/funding
- Raising potential issues
- Aim: open-source and for-profit are not binaries, but your journey can combine elements of both
- Open research vs. open source software
- Avoiding the Benevolent Dictator for Life model
- Invest in Open (investinopen.org/)
- What does open infrastructure look like in terms of funding & support?
- Fragility of single contributors/maintainers (https://xkcd.com/2347/)
- Developing that research into OSS or feeding into upstream OSS environments like Jupyter
## Research Notes
1. OpenCV
- OpenCV was built to provide a common infrastructure for computer vision applications and to accelerate the use of machine perception in the commercial products. Being a BSD-licensed product, OpenCV makes it easy for businesses to utilize and modify the code.
- 47 thousand people of user community and estimated number of downloads exceeding 18 million.
- used extensively in companies, research groups and by governmental bodies
- OpenCV started as the brainchild of Gray Bradsky, once a computer vision engineer at Intel, around the early 2000s. Bradsky and a team of engineers, mostly from Russia, developed the first versions of OpenCV internally at Intel before making v0.9 of it open source software (OSS) in 2002. Bradsky then transitioned to Willow Garage, with the former founding members of OpenCV. Among them were Viktor Eurkhimov, Sergey Molinov, Alexander Shishkov, and Vadim Pisarevsky (who eventually started the company ItSeez, which was acquired in 2016 by Intel), who began supporting the young library as an open source project.
- 1999: Launced at Intel Research to advance CPU-intensive applications; main contributors were optimization experts in Intel Russia & Intel Performance Library Team
- Initial Goals included:
- advancing vision research with "not only open but also optimized code for basic vision infrastructure. No more reinventing the wheel"
- disseminate vision knowledge with common infra, advance commerical applications with portable
- performance-optimized code available for free
- 2000: Alpha verision released at IEEE CVPR
- 2006: first 1.0 version
- 2009: OpenCV2 - changes to C++ interface
- 2012: Support taken over by non-profit OpenCV.org which maintains the developer and user site
- 2016: Intel acquired Itseez, a leading OpenCV developer
- 2020: OpenCV AI Kit for hardware supporting Spatial AI
- Links
- https://en.wikipedia.org/wiki/OpenCV#cite_note-10
- https://opencv.org/about/
- https://www.oreilly.com/library/view/mastering-opencv-4/9781789533576/2de0893d-e450-417d-b336-a4143799b43d.xhtml
2. Sci-kit Learn: A Python Machine Learning Library
- 2007: Developed in Google summer of code; used in thesis work by a project contributor
- 2010: INRIA involved and launched first public release
- National Institute for Research in Digital Science and Technology (French national research institute for digital science and technology)
- Started as bridge between academia and industry in 1967
- 1998: I-Source France's first tech VC organisation
- 2020: 30+ active contributors and paid sponsorship from INRIA, Google, Tinyclues, and Python Software Foundation
- PSF sponsors: Bloomberg, Google, Meta, AWS Open Source, NVIDIA, Azure, QRT, CapitalOne, Salesforce, IBM, Red Hat, etc.
- Permissive simplified BSD license; Linux distributions; encourages academic and commerical use
- Built of SciPy, which includes NumPy, SciPy, Matplotlib, IPython, Sympy, Pandas
- Extensions or modules of SciPy are called SciKits
- Vision: level of robustness and support for use in production systems
- Ease of use, code quality, collaboration, documentation and performance
- Standard documentation and API structure
- Designed to be simple and efficient, accesible to non-experts, and reusable in various contexts
- Design choices for the API: simple, elegant for composition and reusability
- Links
- https://machinelearningmastery.com/a-gentle-introduction-to-scikit-learn-a-python-machine-learning-library/#:~:text=Scikit%2Dlearn%20was%20initially%20developed,published%20in%20late%20January%202010
- API design for machine learning software: experiences from the scikit-learn project: https://arxiv.org/abs/1309.0238
- Original Author (David Cournapeau) https://en.wikipedia.org/wiki/David_Cournapeau
- Matthieu Brucher
3. YOLO
- Real-time, multi-object detection and classification
- Before YOLO, object detection was slower (e.g. F-CNN, Fast R-CNN, Faster R-CNN); YOLO made it real-time without compromising accuracy
- 2015 YOLOv1: You Only Look Once: Unified, Real-Time Object Detection”
- 2016 YOLOv2: “YOLO9000:Better, Faster, Stronger”
- 2018 YOLOv3: “YOLOv3: An Incremental Improvement”
- 2020: YOLOv4: Alexey Bochkovskiy, et all in their 2020 paper “YOLOv4: Optimal Speed and Accuracy of Object Detection”.
- 2020: YOLOv5 by the company Ultranytics
- No paper released - does it justify the YOLO branding as it's just a PyTorch implementation of YOLOv3
- Authenticity of performance cannot be guaranteed without the paper
- Feb 2020: Joseph Redmon response to NeurIPS debate about needing section considering broader impact of their work, including possible societal consequences
- "I stopped doing CV research because I saw the impact my work was having. I loved the work but the military applications and privacy concerns eventually became impossible to ignore"
- https://twitter.com/pjreddie/status/1230523827446091776?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1230524770350817280%7Ctwgr%5E%7Ctwcon%5Es2_&ref_url=https%3A%2F%2Fcdn.embedly.com%2Fwidgets%2Fmedia.html%3Ftype%3Dtext2Fhtmlkey%3Da19fcc184b9711e1b4764040d3dc5c07schema%3Dtwitterurl%3Dhttps3A%2F%2Ftwitter.com%2Fpjreddie%2Fstatus%2F1230524770350817280image%3D
- Links
- https://medium.com/analytics-vidhya/all-about-yolos-part1-a-bit-of-history-a995bad5ac57
- https://machinelearningknowledge.ai/a-brief-history-of-yolo-object-detection-models/
- https://medium.com/syncedreview/yolo-creator-says-he-stopped-cv-research-due-to-ethical-concerns-b55a291ebb29
4. Hugging Face
[Open AI notes](https://docs.google.com/document/d/1a8An8saSo0Chuhv4rr5xprhijb73PVK8XGQfvd6W9hU/edit)
- Hugging Face allows users to build, train, and deploy state of the art models using the reference open source in machine learning
- For-profit Unicorn; $161.3M
- From Series C announcement
- Hugging Face is the fastest growing community/platform for ML
- 100,000 pre-trained models and 10,000 datasets hosted on the platform
- We think the direction of responsible AI is through openly sharing models, datasets, training procedures, eval metrics, and working together to solve issues
- Open source and open science bring trust, robustness, reproducibility, and continuous innovation
- BigScience, collaborative workshop around study/creation of LLMs with 1,000 researchers, largest open source multilingual language model
- 2016: founded by Clement Delangue and Julien Chaumond for social AI-run chatbot iOS applications (Chatty, Talking Dog, Talking Egg, and Boloss)
- To accomplish this, developed own NLP model called Hierarchical Multi-Task Learning (HMTL) and managed a library of pre-trained NPL models under PyTorch-Transformers
- 2018: $4M seed round
- 2019: $15M Series A, including OpenAI CTO
- 2021: $40M series B
- 2022: $100M series C
- Links
- https://golden.com/wiki/Hugging_Face-39P6RJJ#:~:text=Hugging%20Face%20is%20an%20artificial,in%202016%20by%20Julien%20Chaumond.&text=Hugging%20Face%20is%20a%20company,based%20in%20Brooklyn%2C%20New%20York.
5. Open AI
- OpenAI has the for-profit Open AI LP and non-profit OpenAI Inc
- AGI as humanity’s “biggest existential threat”; focus on long-term human impact
- 2015: Founded with $1B from Elon Musk and Sam Altman
- Other investors include Reid Hoffman, Peter Thiel, Jessica Livingston
- 2019: $1B investment into for-profit arm by Microsoft; Transition from non-profit to capped for-profit
- 2020: GPT-3 LLM trained on trillions of words from the Internet; API as product
- OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity.
- We will attempt to directly build safe and beneficial AGI, but will also consider our mission fulfilled if our work aids others to achieve this outcome.
- OpenAI LP is governed by the board of the OpenAI nonprofit, comprised of OpenAI LP employees Greg Brockman (Chairman & CTO), Ilya Sutskever (Chief Scientist), and Sam Altman (CEO), and non-employees Adam D’Angelo, Reid Hoffman, Will Hurd, Tasha McCauley, Helen Toner, and Shivon Zilis.
- Our investors include Microsoft, Reid Hoffman’s charitable foundation, and Khosla Ventures.
- Links
- https://openai.com/about/
- https://en.wikipedia.org/wiki/OpenAI
6. Eleuther AI
- Open-source collective artificial intelligence project that aims to create a fully decentralized singleton artificial intelligence with an associated autonomous decentralized civilization
- Inspired by Vernor Vinge, Nick Bostrom, and Eliezer Yudkowsky
- EleutherAI was proposed under the premise that the creation of a singleton (a single instance of a governing artificial intelligence) in a decentralized manner is in the interests of the human race.
- The project's authors believe that the creation of a singleton would enable the human race to overcome the challenges posed by artificial intelligence
- 2011: group began, open to participation from anyone with Internet/Github (Gennady Korotkevich and Maxim Goncha)
- POC to explore feasibility of creating an intelligent digital organism capable of evolution and self-improvement
- Participate in the project by submitting proposals for the development of EleutherAI’s capabilities as well as in the project’s governance. EleutherAI’s developers conduct weekly updates on the project’s progress, which are broadcasted live on YouTube.
- 2015: Raise funding to build an autonomous civilization
- July 2020: build open source LLMs through self-organizing group of volunteers
- AI transcends nation states to control, contain, or regulate it
- Humble Beginnings
- The year is 2020, OpenAI has unveiled GPT-3, and the entire ML world is unimpressed. Well, not entirely… One small group of indomitable nerds still holds out against the scaling-deniers. And life is not easy for the symbolicist who garrison the fortified echochambers of Twitter…
- “It really was at first just a fun hobby project during lockdown times when we didn’t have anything better to do, but it quickly gained quite a bit of traction.” —Connor Leahy, EleutherAI
- GPT-NeoX-20B, a 20-billion-parameter, pretrained, general-purpose, autoregressive dense language model
- Stated mission to make LLMs accesible; downloadable for free under a permissive Apache 2.0 License
- “It’s actually very difficult to get the right hardware to train a large language model,” he said, adding that it requires very high capital investment and there are only a few hundred companies that have that hardware today. “Right now, it’s not something you could do anonymously.”
- Links
- https://www.eleuther.ai/
- https://blog.eleuther.ai/year-one/
- https://generative.ink/alternet/eleuther.html#:~:text=History%5Bedit%5D,of%20evolution%20and%20self%2Dimprovement
- https://spectrum.ieee.org/eleutherai-openai-not-open-enough
7. Meta’s Open source AI
- https://www.technologyreview.com/2022/05/03/1051691/meta-ai-large-language-model-gpt3-ethics-huggingface-transparency/
- Meta is making its model, called Open Pretrained Transformer (OPT), available for non-commercial use. It is also releasing its code and a logbook that documents the training process from October 2021 to January 2022
- “We strongly believe that the ability for others to scrutinize your work is an important part of research. We really invite that collaboration,” says Joelle Pineau, managing director at Meta AI.
The Rise of Hugging Face: Live Action Roleplaying for Wealth & Community
https://marksaroufim.substack.com/p/huggingface?s=r
- People use software not (only) because it solves a business problem but because it solves a core emotional need.
- https://twitter.com/marksaroufim/status/1378556625812746240?s=20
- Solving a technical problem means your users will exchange money for a solution to that problem. Solving an emotional problem means your users are instead community members, who will contribute code, ideas, blog posts, videos and whole bunch of externalities that come with attention to make your success assured.
- AI Business Models
- Service - charge user per inference (Clarifa.ai)
- Consultancy - Help businesses without ML use ML
- Platform - layer upon which others can build services (Weights and Biases)
- Media - publish groundbreaking papers and get attention (Open AI)
- Community - platform + strong group identity (Hugging Face)
- Hugging Face interacts with their users personally, frequently, and publicly
- Not just a "content strategy" going outwards towards users