Why OpenAI Should Be Scared of Open Source

The episode’s central argument is that open source could challenge centralized AI companies not just through model quality, but by moving value toward the software that manages users’ context and work. Hermes is designed to learn preferences, create skills from use, retain context, and route tasks across models. If users keep their knowledge and workflows in that harness as models change, they may be less tied to any one provider. Hermes can use models from Claude and OpenAI through API keys, and the discussion notes that OpenAI also supports open models in this context-sharing setup. The speakers argue that controlling an AI stack can produce different outcomes from relying on the same models as everyone else.

The speakers see product design as a key distinction between usable open source and components that still require technical assembly. OpenClaw helped popularize agents, but they describe its setup as clunky and say projects built with it can be difficult to demonstrate or transfer. Optimized model endpoints from Together, Modal, Base 10, and Fireworks likewise require technical know-how. Hermes aims to handle memory, preferences, model routing, and infrastructure with better defaults. Drew compares its packaging and updates to Ubuntu’s role in making Linux more accessible. Pi is lightweight and easy to fork, he says, but users have to build their own adapters.

Hermes’s adoption beyond conventional developers is presented as evidence of product-market fit, though not proof of monetization. Drew points to meaningful GitHub forks, including adaptations by small Chinese robotics companies that connect Hermes to their hardware and tools. A robotics manufacturer’s embedded-hardware interface is one example. The discussion suggests that such systems may benefit from routing different tasks to different models rather than relying on one model. The speakers say Hermes may appeal first to prosumers and later to less technical users; OpenClaw’s setup burden, by contrast, is likened to capable but cumbersome crypto wallets. They also cite Nous’s tens of millions of model downloads, while acknowledging uncertainty about the exact count.

Nous’s open-source work is linked to technical research rather than token-driven promotion. Tarun describes work on low-bandwidth training, synthetic model evaluation, YARN for longer context, and a distributed training run of roughly 40 billion parameters. The speakers say crypto participants contributed to the rise of open-source and open-weight models over the prior six months by applying lessons from crypto to specific technical problems. They contrast this with weak attempts to repurpose blockchain infrastructure. Early decentralized training faced a shortage of strong open models and difficulty combining optimizations; one speaker estimated in 2023 that it had about a 1% chance of overcoming its technical problems and was less optimistic about decentralized inference. Prime Intellect is cited as another team pursuing serious decentralized-AI research.

Nous is also pursuing enterprise customers, where Hermes can be cloned and customized to connect with company data and workflows while tailoring permissions, privacy, and compliance. Robot Ventures led Nous’s $90 million round, with Microsoft and Nvidia participating; Tarun said the investors may help bring Hermes to businesses. Hermes is installed by default on NVIDIA DGX-BARC machines. The speakers describe Portal subscriptions and OpenRouter usage as growing, but caution that enterprise interest is not secured until contracts are signed. They also emphasize community sentiment, retention, and stickiness as potentially more important than near-term monetization, particularly in a downturn.

The discussion sees possible business value in controlling user-facing harnesses and the flow of model requests. OpenRouter routes requests to providers that quote separate input-, output-, and cache-token prices, with cache tokens potentially hundreds or thousands of times cheaper because they require less computation. Provider selection reflects price and quality measures such as uptime and tokens per second. One speaker reports experiments suggesting providers may advertise low prices to win requests without passing caching savings on to users, while warning that the comparison to MEV is limited. The speakers argue that routers face margin pressure, while harnesses and inference providers may capture more value. They predict companies could monetize order flow through provider rebates or guaranteed-volume payments, but present this as a possible direction, not an established outcome.