Recent Reflections in AI

It’s been a while since I had a brain dump in AI. So without further ado:

Infrastructure:

• Chip-level bottleneck will give away to system-level bottleneck. Looking ahead, the constraint will shift from chips & compute to power generation, distribution, and grid resilience. US in particular faces a significant issue here - power supply has been essentially flat for the past 15 years and the grid is a mess.

• China has largely caught up on the model architecture side. They're behind on compute but making good progress. Huawei’s latest 7nm chip isn’t on par with Nvidia yet, but it fits what China needs (read more: https://lnkd.in/gNcySW7w.) Design won’t be the limiting factor — fab is. Even Jensen Huang recently acknowledged that China is not behind anymore in AI.

• Capex spend is still accelerating. Microsoft and Meta both reaffirmed their infrastructure investments for 2025. We remain in the early innings.

Model and Tooling:

• The gap in model performance is practically indistinguishable by now. Chasing SOTA is getting very expensive these days. Almost better to be a close second, and then learn and distill instead (see: OpenAI vs. DeepSeek).

• OpenAI's lead is shrinking. OpenAI continues to have a lead on data and mindshare thanks to it's head start, but Meta and Google still dominate distribution. Anthropic winning more mindshare on B2B and xAI is very compute-rich.

• As Yann LeCun recently said: “I’m not so interested in LLMs now.” I agree. Scientific AI is a much larger opportunity — but far harder. It demands more compute and doing labeling/RLHF is much harder (e.g., generating proteins with AI is one thing, verifying them at-scale in a wet lab is another).

• RL will be the future in AI. AI will eventually be able to learn from their own experiences and actions. You can already see hints of this in the latest LLMs. Rich Sutton & David Silver wrote a great piece about this: https://lnkd.in/gtYydAYY

• MCP is a gamechanger, but agent infra is still early — we need standards for communication (Google A2A is a start), permissioning, memory, payments, orchestration, etc.. Still lots to build.

Application:

• Robotics foundational training is improving quickly, but the field still face bottleneck on the data and manipulation/actuation side. China will likely emerge as the global leader in this space, though U.S. companies will still contend due to sovereign AI.

• Edge AI is becoming more viable, thanks to SLMs and improved hardware. Today’s 3B parameter models outperform original 175B ChatGPT. However, I'm not sure if edge has found that “killer” breakout use case yet.

• We’re going to see some highly successful consumer AI companies in the next few years. Previously, token costs were still too expensive to subsidize wide-scale usage by consumers. But they’ve come down enough now.

• 12 months ago, no GenAI application company had over $100M in ARR. Today, there are at least six — Cursor, ElevenLabs, Glean, Midjourney, Perplexity, Synthesia — and I expect a dozen more by next year

• The next big AI application? forecasting. Still an underexplored space from a fundamental research perspective, but the impact potential is massive.