Brief — April 1, 2026

Library Blog Radar Briefs Intelligence About

🤖 OpenAI Closes $122B Funding Round at $852B Valuation

OpenAI confirmed a record-breaking fundraise co-led by SoftBank and a16z, pushing valuation to $852B. ChatGPT now has **900 million weekly active users** and 50M+ subscribers. The company is simultaneously shuttering cost-heavy products (Sora short-form video app) and retreating from aggressive spending plans — a clear pivot toward IPO discipline. - **Source:** CNBC, March 31 - **Pillar:** Markets & Capital / AI Infrastructure - **Asymmetric angl

CNBC, March 31

🔍 Claude Code Source Leak — Fake Tools, Anti-Distillation & "Undercover Mode"

Anthropic accidentally shipped a `.map` file with the Claude Code npm package, exposing full source code. Key findings from analysis: - **Anti-distillation layer:** When enabled, Claude Code injects fake tool definitions into API calls to poison training data for competitors building copycats - **Frustration regexes:** Code monitors for user frustration signals and adjusts behavior accordingly - **Undercover mode:** The CLI can operate in a disgu

alex000kim.com / Hacker News, March 31

🧠 TinyLoRA — Learning to Reason in 13 Parameters

Arxiv paper showing that with the right LoRA architecture, models can develop basic reasoning capabilities at extreme compression. 136 points on HN today. Significant for edge AI and on-device intelligence. - **Source:** arxiv.org/abs/2602.04118 - **Pillar:** Frontiers / AI Education - **Take:** Aligns with a broader theme — intelligence is becoming cheaper, smaller, faster. The implications for education tools running on low-end devices are mass

arxiv.org/abs/2602.04118

💡 1-Bit Bonsai — First Commercially Viable 1-Bit LLMs

PrismML launched Bonsai 8B: 1.15GB RAM, 14× smaller than full-precision, 8× faster, 5× more energy efficient. Bonsai 1.7B runs at 130 tokens/sec on iPhone 17 Pro Max. This is a genuine milestone for on-device AI. - **Source:** prismml.com / HN Show HN - **Pillar:** AI Infrastructure / The New Playbook - **Take:** The moment LLMs run reliably on phones with <1.2GB footprint, the education platform equation changes. No cloud. No cost per query. Ful

prismml.com / HN Show HN

🧬 Neanderthals Survived on a Knife's Edge for 350,000 Years

New science.org research reveals Homo neanderthalensis navigated repeated near-extinction events over 350k years — population bottlenecks, climate stress, and interbreeding with Homo sapiens all played roles. 86 HN points, serious anthropological weight. - **Source:** science.org - **Pillar:** Frontiers - **Take:** Beautifully relevant to any content about resilience, adaptation, and what survival actually looks like over deep time. Strong Resear

science.org

Blog Pitches

The Intelligence Compression Thesis — Connect 1-bit Bonsai, TinyLoRA, and on-device AI into a coherent argument: the real disruption in AI isn't the frontier model race, it's the race to zero cost inference. Strong Research Lab fit.

What the Claude Code Leak Reveals About the AI Arms Race — Anti-distillation architecture, fake tools, undercover mode — unpack what this tells us about competitive dynamics in AI development. Who's winning, who's scared, and why.

What Neanderthal Survival Teaches Us About Building in Uncertainty — Long-form Frontiers piece. Connects 350k years of near-extinction to the resilience required for building companies and navigating the current AI moment.

LessWrong

Product Alignment is Not Superintelligence Alignment (and we need the latter to survive) — by plex — 78 karma. Sharp distinction between aligning products to be helpful vs. aligning superintelligent systems to human values. Relevant to anyone thinking about AI safety seriously. [Read →](https://www.lesswrong.com/posts/product-alignment-is-not-superintelligence-alignment-and-we

You can't imitation-learn how to continual-learn — by Steven Byrnes — 169 karma (Curated). Key insight into the limits of current training paradigms for AGI. LW curated. Worth a read if you're tracking AI capabilities research.