Tokens & Signals · Tuesday, March 31, 2026

Claude Code Leaked: Anthropic’s Architecture Exposed

claude-codeveo-3.1-litebonsai-8bkat-coder-v2claude-opus-4-6-thinking-autogemini-3-1-pro-preview-highlfm2.5-350manthropicopenaigithubgoogleollamaapplealibabamercororacleliquid-aisupply-chain-attackagentic-architecturefundingcoding-agentslocal-llmon-device-aidata-breachai-infrastructurereasoning-benchmarkskarpathyadrgrondinarankomatsuzakishiri_shhkimmonismus
Tokens & Signals for 3/31/2026. We scanned ~1,200 Twitter accounts (1253 tweets), 13 subreddits (49 posts), Hacker News (11 stories), 8 newsletter posts, 5 podcast episodes, 209 Discord messages, and leaderboard data for you. Estimated reading time saved: ~12 hours.

TLDR & AI Twitter Recap

* Anthropic's Claude Code source code leaked: A packaging error in an NPM update accidentally shipped a 59.8 MB .map file — and with it, 512,000 lines of TypeScript source code. Their entire agentic architecture is now out in the open. news.ycombinator.com/item?id=47584540

* OpenAI closes $122B funding round: Their new $852B valuation makes them the most capital-intensive force in the industry — and it's not particularly close. x.com/OpenAI/status/2039085161971896807

* @karpathy on the Axios supply chain attack: "Supply chain attacks on AI coding tools are particularly nasty because agents auto-install dependencies with elevated permissions." x.com/karpathy/status/2038849654423798197

* GitHub kills Copilot PR ads: Developers pushed back hard after promotional "tips" started showing up inside human-authored code reviews, and GitHub quietly killed the pilot. news.ycombinator.com/item?id=47582984

* Google's Veo 3.1 Lite is live: A lower-cost video generation tier — half the price of the "Fast" version — now available through the Gemini API. x.com/opencode/status/2038743310559244582

* Ollama + Apple MLX: A new preview adds support for Apple's MLX framework, giving M-series Macs a serious performance boost for local LLM inference. ollama.com/blog/mlx

* 1-bit Bonsai 8B demo: 40+ tokens/sec in real time on an iPhone 17 Pro. Local edge AI is getting uncomfortably fast. x.com/adrgrondin/status/2039066539022778613

* Alibaba releases KAT-Coder-V2: A new agentic coding model that's putting up numbers that rival the top US frontier models. x.com/arankomatsuzaki/status/2038844459857985954

* Mercor AI breach: The recruitment platform had a misconfigured database exposed, leaking 4TB of data — resumes, transcripts, the works. x.com/shiri_shh/status/2038961296847376713

* Oracle workforce cuts: A 20% staff reduction, with the savings going straight into AI infrastructure. x.com/kimmonismus/status/2039059029381574936


Go deeper on what matters to you

Tap to expand

Best to Build With Today

* Codingclaude-opus-4-6-thinking-auto (Top-tier reasoning and agentic performance).

* Reasoninggemini-3-1-pro-preview-high (Leading the current reasoning benchmarks).

* Chatgemini-3-1-pro-preview-high (Consistent top ELO performer).

* Video generationVeo 3.1 Lite (High quality, 50% cheaper than the previous standard).

* Open-source/EdgeLFM2.5-350M (Best choice for resource-constrained, local tool-use agents).


Deeper Dives

💼 Industry & Business

* OpenAI's Massive Valuation

OpenAI just closed a record $122 billion funding round, pushing its valuation to $852 billion. The money is earmarked for compute scaling and AGI R&D.

Why it matters: The cost of staying at the frontier of AI is now comparable to the GDP of a small country. This is what the race looks like.

� Twitter

* Mercor AI Data Breach

Recruitment platform Mercor left a database misconfigured and paid for it — 4TB of data exposed, including resumes, transcripts, and KYC documents.

Why it matters: AI hiring platforms are becoming identity honey pots. One bad config and millions of users' personal data is out the door.

� Twitter

* Oracle's AI Pivot

Oracle is cutting 20% of its workforce to funnel resources into AI infrastructure.

Why it matters: It's pivot-or-die for legacy tech right now. Expect more rounds of mass restructuring as incumbents scramble to keep up with AI-native companies.

� Twitter

* GitHub Reverses Ad Policy

GitHub pulled the plug on Copilot ads in pull requests after developers made their feelings very clear.

Why it matters: Developers treat their code review tools like sacred ground. Try to monetize the "human" parts of the stack and you'll lose their trust — fast.

� Hacker News

🧠 Models & Research

* Anthropic Claude Code Leak

An NPM build error accidentally shipped a source map with 512k lines of TypeScript. System prompts, memory logic, API hooks — all of it is now public.

Why it matters: Anthropic's closed-source moat for their CLI agent just evaporated. Competitors now have a free, detailed roadmap of their agentic architecture.

� Twitter� Hacker News� Reddit

* Alibaba's KAT-Coder-V2

Alibaba's new model uses hierarchical task planning to tackle multi-file software engineering problems, and it's trading blows with Claude Opus 4.6.

Why it matters: The performance gap between US labs and international developers is closing fast — faster than most people expected.

� Twitter

* Liquid AI's LFM2.5-350M

A 350M-parameter model built specifically for edge tool use and data extraction. Tiny footprint, focused purpose.

Why it matters: Not every agent needs the cloud. This is what offline, on-device intelligence starts to look like.

� Reddit

🚀 Products & Launches

* Google Veo 3.1 Lite

A cost-optimized video generation tier delivering 1080p outputs at half the price of the "Fast" model.

Why it matters: The AI video tax keeps dropping. High-volume production workflows that couldn't justify the cost before are about to reconsider.

� Twitter� Reddit

* Ollama + MLX

Native MLX support means serious inference gains for anyone running local models on Apple Silicon.

Why it matters: Local LLMs on MacBooks just got meaningfully faster. This one's worth updating for.

� Hacker News� Twitter


Funding & Deals

* OpenAI: Secured $122B in funding at an $852B valuation.

* Mercor: Currently dealing with the fallout of a massive 4TB data leak.


Launches

* Veo 3.1 Lite: Google's efficient, low-cost video model.

* LFM2.5-350M: Liquid AI's new ultra-compact edge model.

* Ollama MLX Preview: New performance-focused Apple Silicon backend.

Closing thought: Between the Anthropic code leak and the Axios supply chain hack, this week is a brutal reminder that even the biggest names in AI are still tripping over basic security hygiene. Lock your dependencies, audit your build pipelines, and maybe double-check what's actually inside your next NPM package.