Open Source Is Dead. Long Live Open Source. — Saoud Rizwan, Cline
Saoud Rizwan, founder of Cline, discusses how AI has eroded trust in open-source communities, leading projects like Zig and curl to restrict AI-generated contributions and exposing severe supply chain risks, as seen with the litellm compromise. He argues that while community open source is struggling, the economic imperative of "open weights" models is rising. Citing the high costs of closed LLMs and the success of models like GLM at Coinbase, Rizwan draws parallels to the Open Compute project, predicting a commoditization of AI inference. He urges American labs to release open-weights models to maintain technological leadership against foreign competitors and prevent lock-in.
The speaker, Saoud Rizwan, founder of Cline, argues that while the community aspect of open source is “dead” due to the pervasive influence of AI, the principle of “open weights” models is emerging as an economic imperative that will shape the future of AI.
The Erosion of Trust in Open Source
Rizwan, whose project Cline started as an open-source coding agent fostering community trust and collaboration, laments the decline of the broader open-source community. AI has fundamentally altered software development, turning platforms like GitHub into “an archive of SLOP PRs and issues and security reports.” The former sense of community has been replaced by deep skepticism and distrust, as AI-generated code and reports introduce significant risks.
Examples of this shift include:
- Zig: Its code of conduct now explicitly bans all AI use in pull requests, issues, and comments, prioritizing the growth of human contributors over contributions, which AI assistance undermines.
- curl: The project is overwhelmed by AI-generated bug reports and is considering ending its decades-old bug bounty program.
- tldraw: Automatically closes all pull requests, regardless of origin.
- GitHub: Has introduced a feature to disable third-party pull requests altogether, a core functionality that initially defined the platform.
This erosion of trust extends to critical security concerns. Rizwan highlights the increasing danger of supply chain attacks, citing the litellm compromise. A Python package with 3.5 million daily downloads, litellm was compromised for three hours. Attackers exploited a GitHub app to steal PyPI publishing tokens, deploying a version that installed a credential harvester (API keys, SSH keys, crypto keys) and a backdoor for remote command execution. The compromise was discovered by pure luck when the malware’s bug crashed a user’s Cursor IDE.
The Economic Case for Open Weights Models
Despite the decline of community-driven open source, Rizwan asserts that open-weights models are becoming more critical than ever due to their economic impact. He presents compelling evidence of the unsustainability of current closed-source LLM spending:
- An unnamed company accidentally spent $500 million on Claude in one month due to lack of usage limits.
- Uber’s CTO reported that 95% of engineers used Claude, and 70% of committed code came from it, leading to a monthly spend of $2,000 per user and exhausting their 2026 budget in four months.
Rizwan notes that even AI labs are losing money, subsidizing their services (e.g., a $200 Claude plan yields $8,000 in API usage, Codec $200 for $14,000). Their strategy appears to be to lock in users by making engineers dependent on their tooling, pushing new features as workflows that standardize token usage, ultimately leading to “price gouging.” The shift of Anthropic and OpenAI from API businesses to application layers is seen as an attempt to build “moats” against future commoditization.
However, Rizwan argues this strategy is shortsighted. Developers and businesses prioritize value, and open-weights models, particularly those from China, are reaching an “inflection point.” While previously lagging, their “raw intelligence” lead is becoming less critical as cost takes precedence.
Crucially, Rizwan emphasizes that intelligence is better placed in the system and guardrails around the model, rather than solely relying on the model’s raw intelligence. With robust AI-native development infrastructure, including project skills, rules, verification systems, and quality gates, “even a mediocre model can produce similar results as a more intelligent model; it just might take more tokens.”
Cline’s own testing corroborates this:
- When comparing GLM and Opus on a real bug, GLM used twice as many tokens but cost half as much. It cleaned dead code and verified the build.
- Opus finished faster with fewer tool calls but left type errors, breaking the production build.
- This suggests GLM was “trained to spend more tokens verifying its output,” a valuable trade-off given cheaper tokens and the importance of the end result.
The industry is already standardizing on these open-weights models. Coinbase CEO Brian Armstrong reported that defaulting their internal LLM gateway to GLM and Kimi cut their AI spend by nearly half while token usage continued to grow.
The Open Compute Precedent and Future Commoditization
Rizwan draws a parallel to the Open Compute Project of 2011. Facebook, having been outpaced in data center infrastructure, open-sourced its data center designs. This led to the entire supply chain reorganizing around these shared standards, allowing manufacturers to move from small, custom production runs to massive, standardized ones. This commoditized components, drove down costs for the whole industry (saving Facebook billions), and established a market even if the initial designs weren’t “the best.”
The same commoditization is expected in AI inference. Estimates project nearly $3 trillion in spending and over 100 gigawatts of new data center capacity by 2030. Hosting providers are aggressively competing on efficiency, using dedicated hardware, caching, batching, and specialized silicon. By 2030, inference on a 1 trillion-parameter LLM is projected to cost 90% less than today. This mirrors the cloud commoditization of 2014-2015, where hyperscalers shifted from competing on raw compute to higher-level services.
As open weights allow aggressive cost optimization, mass adoption of these models is inevitable. The “absurd API costs” of closed labs will become unsustainable for most knowledge work.
A Call to Action for American Labs
Rizwan issues a fervent request to American labs: take “open weights more seriously.” He warns that if foreign open-weights models become the industry standard, US models like GPT and Claude could become irrelevant, and “marginal quality gains may not be enough to bring anyone back.” This would lead to a loss of control over the technology’s development, as the values embedded in AI are deeply tied to its creators’ nations.
He clarifies that this is not about open-sourcing research (which provides a competitive lead) but about releasing open-weights models. While open weights might make it easier for “copycat models,” they don’t allow “leapfrogging.” Instead, they foster industry usability, competition, and better value for customers, ultimately helping the US maintain its leadership.
Cline’s Commitment to the Open Weights Future
Cline is fully embracing this future. They recently launched an “open weights subscription plan,” offering significant discounts through volume-based pricing and partnerships with inference host providers. Their goal is to demonstrate that expensive closed-frontier models are not always necessary to get work done effectively, highlighting the advancements of models like GLM and DeepSeek. Cline remains open source, allowing users to bring their own API keys and integrate with various platforms, continuously adding the newest open-weights models for evaluation.