AI-generated pull requests, reports, and supply-chain attacks are making the community side of open source harder to sustain.
2
Open weights models can win through lower cost and better verification systems even when closed models are more intelligent or faster.
3
American AI labs should release open weights before foreign models become the industry standard through lower prices and wider adoption.
Summary
Saoud Rizwan argues that open source has split into two parts. The contributor community is shrinking because AI-generated pull requests, issues, bug reports, and security submissions are expensive to review and difficult to trust. Projects such as Zig, curl, and tldraw have responded by banning AI contributions, considering an end to bug bounties, or closing pull requests. A compromised litellm release shows the security cost of relying on third-party packages. Rizwan believes the part of open source that allows people to use, modify, and build on software will grow through open weights models. He compares their economics with Facebook's Open Compute Project, where shared designs helped standardize hardware and reduce costs. A Cline test found GLM cheaper and better at verification than Opus on a real bug, while Coinbase reportedly cut AI spending nearly in half by routing to GLM and Kimi. Rizwan asks American labs to release open weights before foreign models win adoption through price and infrastructure scale.
Cline grew because open source let developers inspect and trust it
Saoud Rizwan says Cline began as an open source coding agent before cloud subscriptions and prompt caching made model access cheaper. Users once paid for every API request, and some spent hundreds of dollars a day. Open source helped them inspect the code, connect any API, and feel confident that they were not being overcharged. The community also shaped features such as custom rules and plan mode. Rizwan says Cline would not have become as successful without that ability to build trust and learn directly from users.
AI-generated contributions have turned open source communities toward distrust
Rizwan says GitHub has become an archive of low-quality pull requests, issues, and security reports. Projects that depend on third-party contributors now have to judge whether submissions reflect responsible use of AI. Zig bans AI use in pull requests, issues, and comments because its maintainers value developing trusted contributors over simply receiving more code. curl is considering ending its bug bounty after decades because AI-generated reports consume too much attention. tldraw closes pull requests automatically, and GitHub added a setting to disable third-party pull requests.
A compromised litellm release shows how third-party dependencies can amplify one attack
Rizwan describes litellm, a Python package with about three and a half million downloads a day. Attackers used a GitHub app to steal publishing tokens and release a compromised version. It installed software that harvested API keys, SSH keys, and crypto keys, along with a backdoor for remote command execution. The release remained live for three hours. A bug that crashed Cursor when running the litellm MCP server led a security researcher to investigate. Rizwan says the incident could have caused catastrophic damage, especially among enterprise users with internal gateways.
Open weights models can compete through price and system design
Rizwan says the advantage of the most intelligent closed model matters less once open weights models are capable enough for many tasks. The surrounding system can supply context, tools, project skills, rules, verification, and quality gates. A less capable model may produce similar results if it uses more tokens. In this view, reliability depends more on the system around the model and less on trusting the model or developer to use it responsibly. Businesses can also route work between models to find the lowest cost for each task.
GLM beat Opus on cost and verification in a Cline repository test
Cline tested GLM and Opus on a real bug from its own repository after questioning benchmark claims that GLM was better. Both models fixed the issue, but GLM used twice as many tokens and cost half as much. Opus finished faster and used half as many tool calls. GLM removed dead code and checked that the build compiled. Opus left type errors that broke the production build. Rizwan interprets this as evidence that GLM spent more tokens verifying its work, which is acceptable when those tokens are cheaper.
Companies are already routing work toward cheaper open models
Rizwan cites Brian Armstrong saying Coinbase has defaulted its internal language-model gateway to GLM and Kimi. He says this cut AI spending by nearly half while token use continued to grow. He expects companies to build internal routing and agent infrastructure around dollar-efficient choices, even if that means giving up access to the newest feature in a product such as Claude Code. The decision depends on the cost of completed work, rather than on which provider has the latest application feature.
Open Compute shows how shared designs can reduce the creator's own costs
Rizwan compares open weights with Facebook's Open Compute Project. In 2011, Facebook published designs for data centers, servers, networking, cooling, and related hardware after larger companies had already built similar infrastructure. The supply chain reorganized around the shared designs. Manufacturers moved from small runs of custom equipment to large standardized runs, which lowered prices across the industry, including for Facebook. Rizwan argues that an industry may standardize on something people can build on even when it is not the best available option.
Inference economics may make foreign open weights the default
Rizwan says large AI infrastructure investments will make open weights cheaper. He cites estimates of nearly three trillion dollars in spending and more than 100 gigawatts of new data center capacity by 2030. Hosting providers can reduce inference costs with dedicated hardware, caching, batching, and specialized chips. One estimate he cites puts the cost of inference on a one-trillion-parameter model 90 percent below today's cost by 2030. If foreign open weights become the standard, he worries that later quality improvements from GPT, Claude, or Gemini may not persuade companies to switch.
Rizwan asks American labs to release open weights without opening their research
Rizwan distinguishes open weights from publishing the research that produced a model. He asks American labs to release models that industry can use and build on, while keeping research advantages private. Open weights could increase competition, adoption, and price value for customers. He also connects the issue to safety research, arguing that losing adoption to foreign models could reduce the influence of American labs that invest heavily in safety. Cline has acted on this position by launching an open weights subscription with volume discounts and inference-provider partnerships.
"The intelligence is better placed in the system and guard rails around the model so that you don't have to be as reliant on the model or your end developer's responsible use of the model itself."09:18
Who should watch
You maintain an open source project and are spending more time filtering AI-generated contributions than reviewing useful changes.
You are choosing between closed and open weights models and need to compare total cost, verification, and infrastructure rather than model rankings alone.
You work at an AI lab or build internal model infrastructure and want to understand Rizwan's argument for releasing open weights early.
