Feed

Great presentation from Nathan Lambert at "The Curve" on the state of open source models.

Link: https://docs.google.com/presentation/d/1f1Et0Mz8zb1yVCnCgdYSy4tAa0Kv_gKT4wPEg1XPdUA/edit?slide=id.g38edb366806_0_6#slide=id.g38edb366806_0_6

The rise of the Chinese ecosystem has been blistering, starting from negligible downloads at the start of 2025 to well over 50% share in September of this same years, compounding across ~20 different organizations.

Qwen's ecosystem in particular has been on a tear, emerging as the starting point of choice for AI developers. It's interesting to think through the incentives of Chinese Labs and whether they will continue to open source.

• Continuing to lag the US closed source models, open source is a great way to close the gap
• Leading in open source is fantastic marketing to continue to attract top AI talent
• The willingness to pay for software / AI in China is de minimis, leading large ecosystems to monetize gains AI capabilities through adjacent products
• The Chinese government is highly focused on AI diffusion through the rest of society with its "AI+" initiative; ubiquitous, free intelligence is aligned with that vision
• Culturally, DeepSeek has set the tone for open sourcing SOTA level models which will be a difficult precedent to challenge and remain competitive.

Overall, my read is that China seems likely to continue to open source based on the above incentives. If the US does not want Chinese models to become the starting point for every developer looking to tune or post-train an AI for their product, they have some work to do...

‍

‍

PSA: Delphi is co-hosting a summit on Open Source AI in San Francisco on Oct. 24

This is a one-day, single-track event for founders and engineers focused on the frontier of open source and distributed AI, part of the Linux Foundation's Open Source AI Week.

We are bringing together senior researchers and leaders from OpenAI, Google DeepMind, Meta, NVIDIA, alongside the founders and maintainers of top open-source AI projects like Letta, Cline, and Qdrant. They will be joined by leading research labs in open and distributed AI.

The agenda is technical, with no marketing pitches. You will hear directly from the people building foundational models and core infrastructure. The goal is to move beyond hype and focus on making AI universally accessible, provably fair, impartial and beneficial for all.

This is an opportunity to connect with the maintainers of leading open-source projects and the founders of emerging AI-native companies. Should be an awesome group.

***Tickets can be brought here***

If keen, please feel free to dm me on twitter @ponderingdurian with a bit of your background / experience in AI and may be able to provide a discount to listed pricing :)

‍

‍

Really enjoyed this interview by Dwarkesh Patel with Casey Handmer on the future of energy. Casey makes a compelling case for solar as the torch-bearer for the coming energy transition.

For the first time in my life, I started contemplating buying rural desert land in Nevada and West Texas.

Highlights:

• China currently has 20x more annual solar manufacturing capacity than the US, which could give it a significant edge in the AI race due to massive energy needs, but the US can catch up within five years through automation, cheap natural gas, and financial advantages. (PD: Seems like a very aggressive timeline and unlikely)

• China's energy security is vulnerable due to reliance on imported oil and geopolitical risks, while the US benefits from isolation and stable neighbors.

• AI will drive hundreds of gigawatts of new energy demand; solar is positioned as the scalable solution, potentially powering off-grid data centers with batteries by 2027, as natural gas turbines face supply limits.

• Solar costs are dropping rapidly (43% reduction per production doubling), making it cheaper than running existing coal plants; by 2040, new data centers could be 100% solar-powered.

• US environmental regulations (e.g., NEPA) slow solar deployment—Texas installs 10x more than California due to lighter rules—highlighting regulatory reform as key to competing.

• Large-scale solar for 5 GW data centers requires ~50,000 acres, feasible in places like Texas, but permitting and non-contiguous land use are challenges.

• Batteries enable temporal arbitrage, storing solar for peak times and reducing grid dependency, with per capita battery capacity surging (e.g., from 10g to 100kg via EVs).

• As batteries proliferate, grid utilization will decline, electron travel distances will shorten, and mobile battery markets could emerge for flexibility.

• GDP underestimates AI's impact due to deflation; energy consumption (e.g., solar-powered AI) is a better metric, as AGI could automate $60T in global labor value.

‍