{ "video": { "id": "CoEIs6Xm8m8", "title": "Open Source Is Dead. Long Live Open Source. — Saoud Rizwan, Cline", "duration": 1050, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "In person" }, "analysis": { "video_id": "CoEIs6Xm8m8", "title": "Open Source Is Dead. Long Live Open Source. — Saoud Rizwan, Cline", "one_liner": "Cline's founder argues that AI has killed the community side of open source (slop PRs, banned AI contributions, supply-chain compromises) while making open-weights models the part that now matters most, because cost — not raw intelligence — is becoming the deciding factor in coding agents.", "summary": "Saoud Rizwan traces how AI coding has hollowed out open-source community norms — Zig banning all AI use, curl being \"DDoSed\" by AI bug reports, tldraw auto-closing PRs, GitHub shipping a switch to disable third-party PRs — and how supply-chain risk (the three-hour LiteLLM compromise) makes depending on third parties more dangerous than ever. He then argues the surviving, and now more important, part of open source is permissive reuse, specifically open-weights models: closed labs are subsidizing subscriptions ($200 of Claude Code buying ~$8,000 of API usage) to create lock-in before commoditization, but businesses will jump to whatever is cheapest per dollar. He backs this with Cline's own GLM-vs-Opus bug fix comparison, Coinbase defaulting to GLM and Kimi, and the Open Compute Project as the historical template for how a shared open standard commoditizes an industry. He closes by asking American labs to release open weights — not open research — to keep mindshare, and pitches Cline's new open-weights subscription plan.", "key_points": [ "Cline was the first coding agent, shipped before prompt caching and before Claude/Codex subscriptions, when users paid per API request — some spent hundreds of dollars a day; being open source is what made developers trust it enough to spend that.", "Concrete evidence of open-source community collapse: Zig's code of conduct bans AI in PRs, issues and even comments (they value contributors over contributions, and PR review exists to grow trusted contributors); curl's CEO says AI-generated bug reports are DDoSing the project and they may end a decades-old bug bounty; tldraw auto-closes all PRs; GitHub added a feature to disable third-party pull requests entirely.", "The LiteLLM compromise: a package with ~3.5 million downloads/day was compromised for three hours via a GitHub app that stole its PyPI publishing tokens, shipping a credential harvester for API/SSH/crypto keys plus a remote-command-execution backdoor. It was caught only by luck — a bug in the malware crashed Cursor when running the LiteLLM MCP server.", "Inference spend is out of control: an anonymous CFO report of accidentally spending $500M on Claude in one month with no dashboard usage limits across thousands of employees; Uber's CTO reports 95% of engineers using Claude, 70% of committed code from Claude, up to $2,000/user/month, and the entire 2026 budget burned in four months.", "SemiAnalysis measured subscription subsidies by running long-horizon coding tasks to the weekly limit: a $200 Claude plan yielded ~$8,000 of API usage and a $200 Codex subscription ~$14,000 — evidence that labs are subsidizing to lock developers in (CI agents, background agents, looping agents) before the price gouging comes.", "Cline tested GLM against Opus on a real bug from the Cline repo: both fixed it, but GLM used twice the tokens at half the cost, cleaned up dead code and verified the build compiled; Opus was faster with half the tool calls but left type errors and broke the production build. Conclusion: GLM is trained to spend more tokens verifying its output, and tokens are cheap.", "Coinbase CEO Brian Armstrong says they defaulted to GLM and Kimi in their internal LLM gateway, cutting AI spend by nearly half while token usage keeps growing.", "The Open Compute Project analogy: Facebook lost the data-center race to Amazon and Google in 2011, open-sourced its schematics and CAD files, the supply chain reorganized around the standard, manufacturers got large standardized runs, and commoditization saved Facebook billions — the industry standardizes on what it can build on, not on what's best.", "Cost trajectory: ~$3 trillion of capex and 100+ GW of new data center capacity by 2030 (roughly doubling global capacity), with hosts like Baseten and Fireworks competing on dedicated hardware, caching, batching and inference-specialized silicon; inference on a 1-trillion-parameter model is estimated to cost 90% less by 2030 — the same pattern as Google's 2014 GCP cuts (32% compute, 68% storage) and AWS's matching 42nd price cut days later." ], "takeaways": [ "Route through your own internal LLM gateway and default to open-weights models where they suffice — Coinbase cut AI spend nearly in half doing exactly this, and set usage limits on the provider dashboard before a $500M month happens to you.", "Put the intelligence in the system, not the model: project skills, rules, systems of verification and quality gates let a mediocre model match a smarter one, just at more tokens — and tokens are the cheap part.", "Evaluate models on the end result and total cost, not on speed or tool-call count: Opus finished faster with fewer tool calls but broke the production build, while GLM verified compilation before finishing.", "Assume third-party dependencies are a live supply-chain risk — a single compromise of a package like LiteLLM harvests API, SSH and crypto keys from everyone who installs during the window, and detection may be pure luck.", "Avoid single-vendor agent lock-in: keep a bring-your-own-key path (Cline runs on CLI, VS Code and JetBrains with any provider) so an outage or a price hike doesn't stop your developers working.", "If you maintain an open-source project, decide and publish an explicit AI-contribution policy now — the options in the wild range from Zig's total ban to tldraw's auto-close to GitHub's disable-third-party-PRs switch." ], "topics": [ "open source", "open weights", "coding agents", "inference cost", "supply chain security", "model commoditization", "context engineering", "vendor lock-in" ], "tools": [ "Cline", "Zig", "Bun", "curl", "tldraw", "GitHub", "LiteLLM", "PyPI", "Cursor", "MCP", "Claude", "Claude Code", "Codex", "Opus", "GPT", "Gemini", "Anthropic", "OpenAI", "SemiAnalysis", "GLM", "Kimi", "DeepSeek", "Coinbase", "Uber", "Facebook", "Open Compute Project", "Baseten", "Fireworks", "AWS", "Google Cloud Platform", "VS Code", "JetBrains" ], "quotes": [ { "text": "GitHub is effectively an archive of SLOP PRs and issues and security reports where the sense of community before has turned into this like deep skepticism and distrust of each other's responsible use of these tools.", "at": "01:44", "url": "https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=104s" }, { "text": "when I say that open source is dead, I mean some parts of it like the community, it's just not worth cultivating anymore, especially because building software is so cheap.", "at": "03:45", "url": "https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=225s" }, { "text": "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.", "at": "09:16", "url": "https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=556s" }, { "text": "the industry will adopt and standardize on something that they can build on top of even if it isn't the best thing.", "at": "12:23", "url": "https://www.youtube.com/watch?v=CoEIs6Xm8m8&t=743s" } ], "words": 3369 }, "summary_url": "/#CoEIs6Xm8m8", "transcript": { "html": "/transcripts/CoEIs6Xm8m8.html", "txt": "/transcripts/CoEIs6Xm8m8.txt", "vtt": "/transcripts/CoEIs6Xm8m8.vtt" } }