AVODA Group

Work It and Keep It: Open-Source AI as Digital Commons Stewardship

Genesis 2:15 gives the first human job two verbs: work the garden and keep it. The Hebrew verb for work there is ‘abad, the root of avodah, and it describes what a founder owes the open-source AI she builds on. Open-source AI is a digital commons, made of models, code and training information released so that anyone may use, study, modify and share them. Stewarding that commons means working what you received and keeping it in better shape than you found it, and the simplest way to do both is a ledger of what you took and what you gave back.

Key Takeaways

  • The Open Source Initiative released its Open Source AI Definition on October 28, 2024. It grants four freedoms (use, study, modify, share) and requires enough training-data information for a skilled person to build a substantially equivalent system.
  • A 2024 Harvard Business School working paper estimates that firms would spend 3.5 times more on software if open source vanished, and that 96% of its demand-side value comes from 5% of developers.
  • Scripture treats received goods as a trust. Leviticus 19 reserves the edges of the field for the poor and the foreigner, and John 4:38 tells the disciples they are reaping where others labored.
  • The strongest critique, from David Gray Widder, Meredith Whittaker and Sarah Myers West in Nature, holds that openness alone leaves AI power concentrated. We concede it, and place East African stewardship in the layer we can own: language data, evaluation and documentation.
  • The Gleaning Ledger has four columns (Received, Permitted, Returned, Guarded). One sheet per product is enough to start this week.

What counts as open-source AI, and what is only open weights?

Start with the document. The Open Source Initiative, founded in 1998 and steward of the Open Source Definition, published version 1.0 of its Open Source AI Definition at a conference in Raleigh in October 2024. It came out of a co-design process that InfoWorld reported involved more than 25 organizations, Microsoft, Google, Amazon and Meta among them. An open-source AI system, in the definition’s words, must let you “Use the system for any purpose” and also study, modify and share it.

The definition then asks for more than a downloadable file. It requires the code used to train and run the system, the parameters, and “Data Information” detailed enough “so that a skilled person can build a substantially equivalent system.” Many models marketed as open publish the weights and stop there, often under a custom license with its own use restrictions. A February 2026 study of AI model-release data, published on a Duke University site, called openness in AI “nascent and inconsistently understood” and reported friction around licensing practices.

So a founder in Kampala who says “we use an open model” may mean three different things. She may be running a system that meets the full definition. She may be running open weights under a permissive license. Or she may be running weights under a license that limits who can use them and for what. Each carries different obligations, and the first stewardship act is simply knowing which one sits in your stack.

Mozilla’s Ayah Bdeir, who leads AI strategy there, defended the definition’s compromise on training data: she warned that an ideologically pristine standard that model builders fail to meet “could end up backfiring.” That realism is useful for us. A Luganda customer-service bot built on open weights is legitimate work. It is also work on borrowed ground, and borrowed ground has terms.

The commons rests on very few hands

The economics are lopsided in a way every builder should feel. Manuel Hoffmann, Frank Nagle and Yanuo Zhou, in a Harvard Business School working paper dated January 2024, estimated the cost of recreating the most widely used open-source software once at $4.15 billion. They put the demand-side value, the cost to every firm that would have to build it internally, at $8.8 trillion. If open source disappeared, they found, firms would spend 3.5 times more on software than they do now.

A caution belongs here. The $8.8 trillion is a replacement-cost estimate, the Linux Foundation sponsored the research, and one critic on the Open Path blog pointed out that the paper’s goods-market approach produces a figure “four orders of magnitude lower, at $177 million.” Read the trillions as an order of dependence, and hold the exact figure loosely. That paper also measures software in general. A comparable estimate for open AI models, or a reliable count of East African firms building on them, is missing from everything we have read, and AVODA holds zero internal data on it either. We prefer to say so plainly.

One finding survives every method argument: “96% of the demand-side value is created by only 5% of OSS developers.” A very large field is being kept by a very small crew. When one of those maintainers burns out or walks away, everyone downstream inherits the damage, often before they have learned the maintainer’s name.

John 4:38 names our position with uncomfortable precision. Jesus tells his disciples he sent them to reap a harvest others worked for: others did the hard labor, and they have entered into it. Paul presses the same point as a question in 1 Corinthians 4:7: what do you have that you did not receive? Every model running on a Kampala server answers that question with a long list of strangers.

Gleaning, Boaz and the edges of the field

The Torah had a structure for a commons long before anyone wrote a license. Leviticus 19:9 and 10 tell the landowner to leave the edges of the field unharvested and to leave the fallen grain and the dropped grapes for the poor and the foreigner. Leviticus 25:23 grounds it: the land belongs to the Lord, and Israel lives on it as resident foreigners. Psalm 24:1 widens the claim to the whole earth and everything in it.

Ruth is the case study. She was a Moabite widow, vulnerable twice over, and she went out to glean in the field of a man called Boaz. Gleaning was hard physical work from morning onward. Boaz responded by telling his workers to let her glean among the sheaves and to pull out stalks for her on purpose (Ruth 2:15 and 16). The owner kept the law and then went beyond it, and the gleaner worked the edges with dignity. Both of them were being faithful.

Here the analogy turns uncomfortable for East African builders. In open AI, most of us are Ruth. The large fields were planted by laboratories in California, Paris, Beijing and elsewhere, and we gather at the edges with fine-tuning scripts and borrowed GPUs. Gleaning is honorable work, and building a business on open weights is the gleaner’s labor dignified by law, which is why we affirm it plainly. The idol appears when the gleaner starts treating the field as a mine: extract everything, return zero, and call the extraction efficiency.

Paul supplies the corrective in 2 Thessalonians 3:8. He reminds the church that he paid for the bread he ate and worked night and day so that he carried his own costs. That is the posture the commons asks of us. Take freely, because the field was opened for exactly that, and then pay for your bread in the currency the field runs on.

Does openness move power? The Nature critique

The strongest case against everything above appeared in Nature in November 2024. David Gray Widder, Meredith Whittaker and Sarah Myers West examined what open AI actually provides across models, data, labor, frameworks and computing power. Their conclusion is blunt: “openness alone does not perturb the concentration of power in AI.” They go further and argue that open AI systems are “highly dependent on the resources of a few large corporate actors,” and that talk of openness is frequently used in ways that deepen that concentration.

We concede the core of it, and the concession is large. Downloading a model from a frontier lab leaves the compute, the capital and the research talent exactly where they were, far from Kampala, Addis Ababa and Dar es Salaam. A Ugandan firm running open weights remains downstream of decisions made in boardrooms far from its own. If someone sells you open-source AI as liberation, the three authors are right to be suspicious.

Their own paper still names three things openness does provide: “transparency, reusability, and extensibility.” Those three are exactly the tools a gleaner needs. Transparency lets a Ugandan developer inspect what a model does with Acholi or Luganda. Reusability lets a ten-person firm start from a capable base on day one. Extensibility lets local builders add the layer the big labs underweight, and that layer is where our stewardship lives. The Sunflower local-language work in Uganda and the InkubaLM small-model approach both point the same way: the defensible African contribution is the language data, the evaluation sets and the documentation that make a general model useful here.

The Gleaning Ledger, column by column

The Gleaning Ledger is a single sheet per product with four columns. Take the illustrative case of a CTO at a nine-person Kampala fintech who has fine-tuned an open-weights model to answer SACCO members’ questions in Luganda. She is invented for illustration; the sheet she would fill in is real enough to copy.

Received. List every open model, dataset, tokenizer and library the product depends on, with the version and the maintainer. Expect to find dependencies you had forgotten. This column is the 1 Corinthians 4:7 answer written out.

Permitted. For each line, record the license and what it allows. Does it meet the OSI definition, use a standard permissive license, or carry custom restrictions on users and purposes? Does the release publish its Data Information? A founder who skips this column is signing a lease she has yet to read.

Returned. Record one contribution per quarter, sized to the firm. A bug report with a reproducible case counts. So does a Luganda evaluation set released under an open license, a correction to a model card, a translated README, or a monthly payment to a maintainer whose library carries your product. The return is the gleaner paying for her bread.

Guarded. Write down what you will keep out of the commons, and why. Customer transaction records, SACCO member identities and a congregation’s prayer requests belong to the people who shared them, and church data carries a particular trust. Boaz left the edges of his own field; the grain he gave was his to give. Consent governs this column, and generosity with other people’s data is theft with good intentions.

The impulse behind the ledger is right, so follow it. Then expect it to fade. The first sprint deadline will push the Returned column to “next quarter,” and the quarter after that will look the same. Good intentions about the commons last about as long as the first outage. What has to be true instead is structural: a recurring calendar line for the quarterly return, a small fixed budget line for maintainers, and one named person who owns the sheet. Keep the action; give it a skeleton.

Frequently Asked Questions

What is the difference between open-source AI and open weights?

Open weights means the trained parameters are downloadable. Open-source AI, under the Open Source Initiative’s 2024 definition, also requires the training and inference code, plus training-data information detailed enough for a skilled person to build a substantially equivalent system. Many models called open meet the first standard and miss the second, so read the license.

Is it wrong for a Christian founder to build a business on free AI models?

Building on open models is honorable gleaning, and the Torah protected the gleaner’s right to gather. Leviticus 19 and the book of Ruth treat gathering at the field’s edge as legitimate work. The obligation runs the other way too: receive freely, then return something, as Paul paid for his bread in 2 Thessalonians 3:8.

Does open-source AI reduce Big Tech’s power in Africa?

Very little, on the evidence so far. Widder, Whittaker and West argued in Nature in 2024 that openness alone leaves power concentrated, because compute and capital stay with a few large firms. Openness does provide transparency, reusability and extensibility, and East African builders can use those to own the local language, evaluation and documentation layer.

What can a small firm contribute back to open-source AI?

Small firms can contribute bug reports with reproducible cases, evaluation sets in local languages, corrections to model cards, translated documentation, and regular payments to the maintainers they depend on. A 2024 Harvard Business School paper found that 5% of developers create 96% of open source’s demand-side value, so small, steady support reaches a very thin crew.

What should stay out of the commons?

Data that belongs to someone else. Customer records, member identities, health details and a congregation’s prayer requests were shared in trust, and releasing them apart from their consent breaks that trust. The Gleaning Ledger’s Guarded column exists so a team settles this in advance, in writing, before a deadline makes the decision for them.

Next Step

Fill in the first two columns of a Gleaning Ledger by Friday: everything you received, and what each license permits. Then take the AVODA Business Capability Assessment, which shows where your operations can carry a recurring commitment like a quarterly return, and saves you promising the commons something your calendar cannot keep.

Related Reading

Sources and Evidence

  1. Open Source AI Definition 1.0, Open Source Initiative: the four freedoms, the Data Information requirement, and the code and parameters requirements.
  2. The Open Source Initiative Announces the Release of the Industry’s First Open Source AI Definition: release date of October 28, 2024, OSI founded in 1998, Ayah Bdeir’s comment on the training-data compromise.
  3. OSI unveils Open Source AI Definition 1.0, InfoWorld: co-design process involving more than 25 organizations including Microsoft, Google, Amazon and Meta.
  4. The Value of Open Source Software, Hoffmann, Nagle and Zhou, HBS Working Paper 24-038 (SSRN): $4.15 billion supply-side value, $8.8 trillion demand-side value, 3.5 times software spending, 96% of value from 5% of developers.
  5. Questioning “The Value of Open Source Software,” Open Path: Linux Foundation sponsorship and the $177 million goods-market estimate.
  6. Why ‘open’ AI systems are actually closed, and why this matters, Widder, Whittaker and West, Nature 635 (2024): the hostile witness on concentration of power, and the three affordances of open AI.
  7. The State of “Open” Source AI: Exploring Data on AI Model Releases: openness in AI described as nascent and inconsistently understood, with friction around licensing.

Leave a Comment

Your email address will not be published. Required fields are marked *