ALL INPredictions
Last updated Nov 29, 2025
Episodes / GPT-4o launches, Glue demo, Ohalo breakthrough, Druck's Argentina bet, did Google kill Perplexity?
Back
Chamath @ 00:28:35Wrong
aitech
In the AI era, companies that build internally useful but non‑core AI systems (such as customer support agents like Klarna’s) will increasingly choose to open-source those systems rather than maintain them privately, leading to a stronger push toward open-source infrastructure in AI than in prior software markets.
That's why I think the incentive for these folks is going to be to push this stuff into the open source... Release it in the open source, guys. Let the rest of the community take it over so that it's available to everybody else. Otherwise you're going to be stuck supporting it... So I also think the incentive to just push towards open source in this market, if you will, is so much more meaningful than any other market.View on YouTube
Explanation

Available evidence up to November 2025 does not show the pattern Chamath described.

  1. Companies like Klarna have not open‑sourced their internal customer‑support agents. Klarna’s widely covered AI assistant (used for two‑thirds of its customer‑service chats and doing the work of ~700 FTEs) is an internal, OpenAI‑powered deployment with no indication that the system or its orchestration code has been released as open source. It remains a proprietary capability used to drive profit improvement for Klarna, not a community project.

    • Evidence: Klarna’s own press and OpenAI’s case study describe the AI assistant as a Klarna product powered by OpenAI, with no mention of open‑sourcing the system or its components. (klarna.com)

  2. The dominant trend for customer‑service / internal “agent” systems is proprietary SaaS, not enterprises open‑sourcing their own agents.

    • Major CX and contact‑center vendors — Intercom (Fin 2), NICE CXone Mpower, Cisco’s Webex AI Agent, Zendesk’s Resolution Platform, Oracle’s AI agents for sales, and others — are all sold as proprietary products, not open‑sourced internal tools handed to the community.(en.wikipedia.org)
    • Telecom and large‑enterprise use cases similarly do not match the prediction. AT&T, for example, switched from ChatGPT to a hybrid system built on open‑source models (H2O.ai + Meta’s Llama 70B) to analyze customer‑service calls, but this is just internal use of open‑source components; AT&T has not open‑sourced its own end‑to‑end system.(businessinsider.com)

  3. Where open source is strong, it is mostly from AI vendors and research groups, not ordinary enterprises pushing out their non‑core agents.

    • Open‑weight / open‑source models from Meta (Llama 3.x), DeepSeek, AI2’s OLMo, etc., plus tooling like Langtrace, form a vibrant open ecosystem, but these are purpose‑built open projects by AI labs and dev‑tool startups, not companies dumping internal customer‑support agents onto GitHub.(en.wikipedia.org)
    • There are open‑source customer‑support platforms with AI agents (e.g., Chatwoot with its “Captain” AI agent), but these were conceived as open‑source products from the outset. They are not examples of mainstream enterprises open‑sourcing their previously internal, non‑core AI systems to offload maintenance.(github.com)

  4. The broader structure of the AI market contradicts the claim that incentives here push toward open source more than in prior software markets.

    • Analyses of open‑source AI note a mixed landscape: some major models and tools are open or open‑weight, but many key components (frontier models, data, and cloud platforms) remain closed and are used to reinforce proprietary ecosystems.(en.wikipedia.org)
    • Historically, previous software eras (e.g., operating systems, web servers, databases) also had massive open‑source pushes (Linux, Apache, MySQL, PostgreSQL). Current AI looks at best comparable in openness and, in core model access, often more closed, not clearly “more meaningfully” open‑source than any other market.

Given these points, the specific prediction — that companies building internal, non‑core AI systems like customer‑support agents will naturally choose to open‑source them, driving a uniquely strong open‑source infrastructure wave in AI — has not materialized by late 2025 and is contradicted by how most enterprises are actually behaving. Therefore, it is best classified as wrong.