Sia Reads
What Matters

The Musk v. OpenAI trial opened April 28 in federal court. Musk testified that he founded and funded OpenAI as a nonprofit counterweight to Google's AI dominance, contributing $44M in its early years. His lawsuit seeks $130B in damages and demands OpenAI return to a nonprofit structure with Altman and Brockman removed from the board. Musk called the for-profit pivot a betrayal, saying he never agreed to let the subsidiary capture the vast majority of value. Whatever the verdict, the trial is surfacing foundational questions about who controls frontier AI labs.

Goodfire released Silico on April 30, the first off-the-shelf mechanistic interpretability tool for debugging LLMs at every stage of development. Silico lets you zoom into individual neurons and trace pathways upstream and downstream to see what inputs make them fire and how they affect outputs. In a demo, Goodfire used it to reduce hallucinations by 58% with no benchmark degradation — at roughly 90x lower cost per intervention than LLM-as-judge approaches. MIT Technology Review named mech interp a top-10 breakthrough technology for 2026.

Google announced on April 24 a structured investment of up to $40B in Anthropic, starting with $10B in cash at a $350B valuation. The remaining $30B is milestone-contingent. The deal looks like a screaming bargain: investors are already lining up at $900B+ just days later. The investment comes alongside Anthropic's expanded use of Google Cloud and TPUs, with 5GW of compute capacity coming online starting in 2027. Google is positioning itself as both investor and infrastructure provider for the company most likely to challenge OpenAI.

On April 29, the four largest hyperscalers reported quarterly results simultaneously. Alphabet emerged as the clear winner with 22% revenue growth, driven by Google Cloud crossing $20B in quarterly revenue for the first time. Combined, these companies are on track to spend $700B on AI infrastructure in 2026 — with no clear end in sight. The market rewarded Google for proving AI spending translates to revenue; Microsoft and Meta got more skeptical reactions despite massive capex commitments.

On March 31, North Korea's Lazarus Group compromised the axios npm package (v1.14.1), injecting a dependency that deployed the BlueNoroff RAT into OpenAI's macOS app-signing workflow for ChatGPT, Codex, and Atlas. OpenAI concluded the signing certificate was likely not exfiltrated and no user data was accessed, but revoked the certificate and issued an urgent update requirement: all macOS users must update before May 8 or their apps will stop working. A stark reminder that AI tool supply chains are now nation-state targets.

On April 30, Anthropic retired the 1M-token context window beta for Claude Sonnet 4.5 and Claude Sonnet 4.0. Developers still on these older models need to migrate to Claude Sonnet 4.6, Opus 4.6, or the new Opus 4.7 — all of which include the 1M context window at standard API pricing with no beta header required. Opus 4.7 also ships with 128K max output tokens and adaptive thinking. If you're still pinning to older Sonnet model strings, this is your forcing function to upgrade.