Claude Code Leak: What 512K Lines of Source Code Reveal About Anthropic’s Agent Strategy
Anthropic accidentally published Claude Code’s full TypeScript source code in a source map file bundled with the 2.1.88 npm update — 512,000+ lines of internal code. The leak, quickly forked 50,000+ times on GitHub, reveals the architecture behind one of the most widely used AI coding tools.
Key findings from the leak:
- KAIROS: An always-on agent that runs persistently in the background, watching files and executing tasks without manual triggering. This is the “agent mode” Anthropic has been hinting at — now confirmed in production.
- Ultraplan: A parallel subagent orchestration system that breaks complex tasks into hundreds of concurrent subprocesses, each verifying its own outputs before merging. This is the engine behind the 750K-line migration feat.
- The Tamagotchi pet: 18 species of AI creatures with a rarity system — a gamification layer that responds to your coding activity. It’s real, it’s shipping, and it’s surprisingly detailed.
- Memory architecture: A multi-tier system that stores session history, user preferences, and long-term context across sessions.
- Internal developer comments: “the memoization here increases complexity by a lot, and I’m not sure it really improves performance” — raw developer honesty, now public.
Anthropic called it “a release packaging issue caused by human error, not a security breach” and fixed it within hours.
Why it matters: The leak is the most transparent view yet of how a frontier AI lab builds agentic software. KAIROS and Ultraplan are the features that matter — they show Anthropic is building toward persistent, autonomous agents that operate alongside humans rather than waiting for prompts. The Tamagotchi is a reminder that even serious AI companies experiment with playfulness in their products.
Meta’s $145B Cloud Pivot: From Social Media to AWS Competitor
Mark Zuckerberg told shareholders that building a cloud business is “definitely on the table” after Meta’s $145 billion AI infrastructure buildout. The company has been laying off headcount (8,000 workers cut) while pouring capital into GPU clusters — a pattern that mirrors Amazon’s early days when AWS emerged from excess compute capacity.
Meta already operates one of the world’s largest private cloud fleets. If it goes public with cloud services, it would compete directly with AWS (34% market share), Azure (23%), and GCP (11%). The difference: Meta has zero enterprise cloud experience but has the advantage of a greenfield architecture unburdened by legacy enterprise customers.
The $145B figure includes both CapEx and operating costs across Meta’s GPU clusters, data centres, and fibre networks.
Why it matters: The cloud market hasn’t had a serious new entrant in a decade. Meta’s entrance would reshape pricing, availability, and competition. The question isn’t whether Meta has the hardware — it’s whether it can build the enterprise sales, support, and compliance infrastructure that AWS and Azure spent 15 years perfecting.
Waymo’s Ojai Robotaxi: 42% Fewer Sensors, $75K Cheaper, Built in China
Waymo opened its 6th-generation Ojai robotaxi to riders in San Francisco, Los Angeles, and Phoenix. Key specs:
- Sensor reduction: 42% fewer sensors than the 5th-gen system — removing lidars and reducing camera count while maintaining Level 4 safety
- Cost: ~$75,000 less per vehicle
- Manufacturing: Built by Geely’s Zeekr in China
- Design: Purpose-built for driverless operation with removable steering wheel and roomier cabin
- Expansion: Planned fleet expansion through 2026
Why it matters: The cost reduction is the headline. The $75K saving per vehicle means lower per-mile operating costs, which changes the economic model for robotaxi fleets. A fleet that needs fewer lidars and cameras can scale faster and cheaper. The Geely connection also raises questions about US-China tech dependencies in critical transportation infrastructure.
ADL Study: Major AI Models Show Systemic Anti-Israel and Antisemitic Bias
The Anti-Defamation League’s comprehensive evaluation of GPT-5.5, Claude Opus 4.8, Gemini 3.5, and Meta’s Llama 4 found that all four models produce anti-Israel and antisemitic outputs at significantly higher rates than for comparable geopolitical topics. Models denied Jewish ties to Jerusalem, made historically inaccurate claims about Israel’s founding, and framed Israeli actions without necessary historical context.
A companion ADL AI Index found no model could consistently detect antisemitic content across all tested categories, with a 40+ point performance gap between best and worst.
Why it matters: This is the strongest evidence yet that political bias in training data is not a problem any single lab has solved. The finding has implications for content moderation, education, journalism, and any field where AI-generated information about contested histories is deployed at scale.
Liquid AI LFM2.5-8B-A1B: On-Device MoE with 128K Context
Liquid AI released LFM2.5-8B-A1B — 8B total parameters but only 1B active per token, making it viable on entry-level laptops. Trained on 38 trillion tokens with a 128K context window, it achieves competitive benchmarks against 7B-8B dense models while using a fraction of the compute.
The architecture combines Mixture of Experts (MoE), Grouped Query Attention (GQA), and gated short convolutions — a genuinely different design from standard transformers. Multilingual performance shows massive gains: Thai +238%, Vietnamese +118%, Hindi +120%.
Why it matters: If a model with 1B active parameters can compete with 8B dense models, the efficiency gains reshape the on-device AI landscape. This affects everything from student laptops to IoT devices to privacy-sensitive applications that can’t use cloud APIs.
🔍 THE BOTTOM LINE
The Claude Code leak and Meta’s cloud pivot are both signals of the same underlying trend: AI infrastructure is scaling faster than anyone’s public roadmap. Anthropic ships half-finished features because the release velocity demands it. Meta builds so much compute it has to find new markets to monetise it. Waymo proves robotaxis don’t need $150K sensor suites. The technology is outrunning its own packaging — and that’s both the opportunity and the risk.
SOURCES
- The Verge — Claude Code source code leak
- CNBC — Anthropic tops OpenAI as most valuable AI startup
- Electrek — Waymo Ojai robotaxi 6th-gen Driver
- 247 Wall St / FrontierNews.ai — Meta $145B AI bet
- ADL — Anti-Israel bias in LLMs report
- Liquid AI Blog — LFM2.5-8B-A1B release
- The Decoder — MCP backlash
- VentureBeat — Claude Code leak analysis