Thursday 2 April 2026

Models

Caltech Researchers Claim Radical Compression of High-Fidelity AI Models [TLDR AI]

PrismML has developed a proprietary 1-bit technology that allows AI models to run locally on edge devices without compromised performance. Caltech owns the IP with PrismML as the sole exclusive licensee.

MiniMax Releases M2.7 Agentic Model [TLDR]

MiniMax released M2.7, a model that participates in its own training and handles complex multi-step agentic tasks at levels comparable to Sonnet 4.6 on OpenClaw benchmarks. Also open-sourced MIT-licensed Office Skills library for Word, Excel, PDF & PPT generation.

Quantum Computing Advances Threaten Encryption [TLDR]

Two independently written whitepapers conclude that building a utility-scale quantum computer capable of cracking elliptic curve encryption requires far fewer resources than previously thought. Neither paper has been peer reviewed yet.

Agents & Tools

Claude Code Source Code Leak — Major Story [ALL SOURCES]

Anthropic accidentally shipped source maps in an npm package, exposing Claude Code’s entire 500,000-line codebase. The leak was rapidly mirrored and forked tens of thousands of times. Key architectural discoveries:

• Three-layer memory system: MEMORY.md index, on-demand topic files, and searchable session transcripts with “autoDream” sleep mode for memory consolidation
• KV cache fork-join model for subagents: parallelism is essentially free since subagents inherit full context
• Custom tool suite: Grep, Glob, LSP implementations for superior repository navigation
• Aggressive prompt caching and file read deduplication
• Less than 20 default tools (up to 60+ total) — restraint over tool bloat
• Internal unreleased features: ULTRAPLAN, KAIROS, MAGIC DOCS Anthropic acknowledged the mistake and issued DMCA takedowns. Attacker npm packages emerged to target developers downloading the leaked code.

Claude Code Computer Use Capabilities Added [TLDR AI]

Anthropic introduced computer use in Claude Code, enabling agents to interact with applications, navigate interfaces, and iteratively test and fix code through a closed-loop workflow.

JustPaid Replaces Human Developers with AI Agents [TLDR]

A California startup used OpenClaw and Claude Code to create 7 AI agents that built 10 major features in a month (each would have taken human developers a month or more). Plans to replace all 9 remaining employees with AI.

Agent Lightning (Microsoft) [TLDR AI]

Microsoft released Agent Lightning on GitHub — a trainer that can optimize any agent system with zero code changes.

Gemini API Docs MCP and Agent Skills [TLDR AI]

Google released Gemini API documentation MCP server and developer skills to prevent coding agents from generating outdated API code. Achieves 96.3% pass rate on Google’s evaluation set.

Research & Engineering

Aurora: RL-Based Real-Time Speculation [TLDR AI]

Together AI released Aurora, an RL framework that learns directly from live inference traces and continuously updates the speculator. Delivers 1.25x additional speedup over well-trained static speculators, with real-time adaptation across shifting traffic.

What is Inference Engineering? [TLDR]

Deep-dive into inference engineering: the emerging discipline around how models generate outputs token by token, and techniques developers use to optimize open models’ inference behavior.

Economics of Generative AI: Two Years Later [TLDR AI]

Semiconductors still capture ~70% of all AI revenues. Infrastructure remains the only truly competitive layer. “Selling the shovels” remains the most profitable AI strategy.

Compute Wars: OpenAI vs Anthropic [TLDR AI]

Anthropic more than doubled its compute capacity to reach close to OpenAI’s levels (enabling Opus 4.5). OpenAI will pull away in H2 2026 but 2027 should be close. Both companies planning massive capacity increases.

Subprime Technical Debt Crisis [TLDR]

Businesses are using AI under the assumption it will eventually solve current technical debt, but this may prove incorrect — accruing debt faster than AI can address it.

Oracle Layoffs Amid Heavy AI Investment [TLDR]

Oracle laying off workers in what appears to be thousands of reductions as the company pivots toward AI.

Industry & Security

OpenAI Raises $122B at $852B Valuation [AINews, TLDR]

OpenAI closed the largest funding round in Silicon Valley history at $122B. Revenue of $24B ARR (growing 4x faster than Google/Meta at their peak). “Soft IPO” with $3B from private investors and ARK Invest ETF inclusion. ChatGPT hasn’t crossed the 1B WAU target for end of 2025.

LiteLLM Supply Chain Attack / Mercor Breach [TLDR AI]

A compromise of LiteLLM’s opensource project, linked to TeamPCP hacking group, affected thousands of companies. Mercor (AI recruiting startup) confirmed a security incident. Lapsus$ extortion group claims access to stolen data.

Axios Open Source Project Hijacked by Hackers [TLDR]

Malicious versions of the popular Axios HTTP library were published, putting millions of developers at risk. North Korean hackers blamed.

Other Notable Items

Apple Testing Multi-Command Siri Feature [TLDR]

Apple is testing a Siri feature to process multiple requests in a single query, expected to debut at WWDC on June 8.

Google Veo 3.1 Lite [TLDR AI]

Google released a lower-cost video generation model via Gemini API, offering same speed as Veo 3.1 Fast at under half the cost.

Claude Dispatch and the Power of Interfaces [TLDR AI]

Essay on how AI capability has outpaced accessibility — most disappointment stems from wrong interfaces rather than wrong models.

Sources Processed

• TLDR (general tech, Apr 1 edition)
• TLDR AI (Apr 1 edition)
• AINews (Latent Space, Apr 1 edition)

Total stories distilled: 20