@PatrickToulme

I’m seeing reports today $META is launching a cloud business to sell AI chips. Looks like my below prediction came true.

Some thoughts on $QCOM acquisition of Modular. Many companies can make AI silicon; very few companies can make useable AI silicon that customers actually want to use. That is the key point. 1. The most important part of AI silicon is the software stack. This means compiler, runtime, kernel, drivers etc. This is the main reason $NVDA CUDA has been so successful. Modular is a world-class compiler team. Arguably one of the best in the world. They will supercharge $QCOM software stack. 2. Compiler team acquisitions are difficult for mainstream analysts to understand. Compilers are very complex software and not near as buzzy as a consumer product or new AI agent. It is hard for analysts to understand there is this special complex software that actually makes other software run on your chip. And without this software your chip is virtually useless. 3. @clattner_llvm is world-class at 1. Building very complex software systems. 2. Building world class software teams. Qualcomm acquiring him and his team is a massive immediate value add to their organization. In general, I wish more large scale organizations did acquisitions at the infrastructure layer like this. It is an acquisition that is harder for Wall St to understand than a pure consumer software acquisition, but arguably much more valuable.

Open-sourcing HarnessGym. Most coding-agent evals ask: “Did the agent solve the task?” HarnessGym asks something different: Can the agent learn what harness was missing, build it, prove it works, and make the next fresh run stronger? agent attempts task → reflects → builds tooling → qualifies it → next run starts stronger 🧵 https://t.co/ZjmQ0CNGHP

Human code review cannot scale with agentic coding. IMO the only way to review vast amounts of agentic code is with vast amounts of verifier code review agents. Agents reviewing agents

Some thoughts on $META and why it is undervalued Meta is spending hundreds of billions on AI compute. Critics say this is wasteful etc as Meta does not have a leading frontier LLM right now. In my view, this capex spending gives Meta massive leverage. Compute is as good as cash in this market, as there is virtually infinite demand for compute right now. 1. Meta could tomorrow decide to rent all of its compute to Anthropic or OpenAI. Now they have a leading AI cloud and are printing money. We just saw @xai do this with Colossus 1. 2. They could lower their own AI costs by hosting free open source models (GLM 5.2) on their compute instead of paying for Claude and Codex. 3. They could also continue to train LLMs on their compute and obtain a frontier model. Compute gives you optionality, and optionality is underpriced when nobody knows who wins