Wednesday, April 1, 2026
Anthropic leaks Claude source code while Block cuts 40 percent of staff
April 1 · 8 videos
Anthropic leaked 500,000 lines of Claude Code.
Block cut 40 percent of its workforce.
Small squads of four now outproduce teams of 14.
Google TurboQuant reduced KV cache memory by 40 percent.
Uber scaled to 5,000 microservices for team autonomy.
The model is no longer the bottleneck.
The build pipeline is the new risk.
“The next trillion dollar company is a technology business disguised as a services company.”
Helping Strangers Build A $1,000,000+ Business [LIVE]
Alex Hormozi · Alex Hormozi · 72 min
Watch on YouTube →Alex Hormozi provides tactical business interventions for companies transitioning from manual services to AI-first technology businesses. He introduces the Scale to Zero framework for automating personal output.
- Scale to Zero requires getting personal time in a business unit to zero before pressure testing for the next growth stage.
- Modern algorithms like Andromeda reward creative volume of up to 500 pieces per week over static consistency.
- The Second Place Prize strategy converts giveaway losers into discounted customers to maintain high conversion rates.
- Marketplaces must secure the supply side first because demand is easier to attract once the product is locked.
- Influencer coaching agencies can achieve profit margins as high as 94 percent by leveraging talent-based workflows.
- Fixing operations and automating workflows before scaling prevents taking on unnecessary employees as dead weight.
Tragic mistake... Anthropic leaks Claude’s source code
Fireship · Fireship · 8 min
Watch on YouTube →Anthropic accidentally leaked the entire source code for Claude Code via a misconfigured npm source map file. This incident reveals the internal prompt engineering and guardrails used by the AI safety leader.
- A 57MB npm source map file exposed over 500,000 lines of TypeScript code to the public.
- Claude's behavior is driven by a dynamic prompt sandwich consisting of an 11-step sequence of hard-coded instructions.
- The leak exposed internal features like a background agent named Chyus and a virtual pet named Buddy.
- Undercover Mode was revealed as a setting designed to make AI-generated code commits appear human.
- The community quickly created forks and rewrites like Claw Code and OpenClaw following the exposure.
- Acquiring low-level toolchains like Bun.js introduces new surface areas for infrastructure-level bugs.
Large-scale agentic quant research with Weights & Biases
Karan · Weights & Biases · 11 min
Watch on YouTube →Karan demonstrates how to use W&B Weave and Models for transparent agentic quantitative research. The focus is on making AI reasoning audit-ready for financial compliance departments.
- Decomposing market events into multi-agent workflows helps identify silent failures in tool calls.
- Fixing a historical analog tool failure increased model forecast conviction from 1/10 to 5/10.
- An automated optimization loop using a meta-optimizer LLM reduced the Brier score from 0.88 to 0.24.
- Parallel coordinate plots reveal that extreme weightings often outperform balanced averages in trading strategies.
- Compliance-grade reproducibility is a prerequisite for moving agentic models into institutional finance production.
- The research-to-validation loop must capture the reasoning behind strategy shifts rather than just the outcome.
Scaling Uber with Thuan Pham (Uber’s first CTO)
Thuan Pham · The Pragmatic Engineer · 99 min
Watch on YouTube →Former Uber CTO Thuan Pham details the hyper-growth journey from 30,000 to millions of rides per day. He explains why technical debt and microservices were necessary for survival.
- Uber scaled to 5,000 microservices to enable radical team autonomy and parallel development.
- The China launch was completed in five months by launching in the most difficult city, Chengdu, first.
- Architecture in high-growth environments should be viewed as a series of 12 to 18 month survival cycles.
- The Program vs. Platform organizational split allowed the company to manage extreme technical complexity.
- Pham is currently using AI swarm coding at Faire to double the output of top engineers.
- Career success is built on a compounding reputation where roles come from colleagues met a decade prior.
Anthropic Just Dropped Their Entire Claude Roadmap by Accident
Josh Kale · Limitless Podcast · 22 min
Watch on YouTube →Josh Kale analyzes the massive Anthropic security breach that exposed 19,000 internal files. The leak provides a detailed look at the company's unreleased product roadmap and agent architectures.
- The leak exposed 44 unreleased product features including an autonomous agent named Kairos.
- Kairos uses a nightly dreaming cycle to consolidate memory and improve long-term performance.
- Internal codenames for upcoming God-tier models include Capybara and Mythos.
- The exposure of the CLI architecture allows developers to use Anthropic's interface with open-source models like DeepSeek.
- Spiky intelligence describes a paradox where an AI safety company fails at basic build-pipeline hygiene.
- Gamification of developer tools is a growing trend, evidenced by the RPG-style virtual pet Buddy found in the leak.
How to Reorg After AI Changes Everything | Block's Owen Jennings on the a16z Show
Owen Jennings · a16z · 27 min
Watch on YouTube →Owen Jennings explains why Block executed a 40 percent reduction in force to rebuild the company around AI agents. He describes a fundamental shift in the correlation between headcount and output.
- A squad of four people supported by AI agents can now outperform a traditional feature team of 14.
- Block is rebuilding around an internal agentic substrate called Goose and an autonomous PR-merging tool called Builder Bot.
- Management layers were reduced by 50 to 60 percent to return the organization to a builder-centric culture.
- Meetings were reduced by up to 80 percent to maximize engineering focus and output.
- Generative UI products like Money Bot create custom interfaces on the fly instead of using static screens.
- Functionalizing the company allows for building shared cross-product infrastructure like a unified AI platform.
Google’s New AI Just Broke My Brain
Károly Zsolnai-Fehér · Two Minute Papers · 8 min
Watch on YouTube →Károly Zsolnai-Fehér reviews Google's TurboQuant technique for optimizing LLM memory. The method addresses the hardware memory shortage by compressing the KV cache.
- TurboQuant combines quantization, random rotations, and the Johnson-Lindenstrauss Transform to compress memory.
- Independent benchmarks show a 30 to 40 percent reduction in memory usage for the KV cache.
- Prompt processing speeds increased by approximately 40 percent in real-world production tests.
- Randomly rotating vectors before quantization prevents information loss in minor axes.
- The technique is highly effective for long-context tasks like analyzing large codebases or long documents.
- Software-based efficiency gains are becoming critical as the global semiconductor shortage continues.
9 Habits That Build Quiet Wealth (No Luck, No Stress)
Rob Dial · The Mindset Mentor Podcast · 19 min
Watch on YouTube →Rob Dial outlines psychological and systemic shifts for building long-term wealth. He introduces the concept of Financial Escape Velocity as the ultimate goal of money management.
- Financial Escape Velocity is the point where investment returns exceed all living expenses.
- Wealth is built in the gap between what you earn and what you spend, regardless of gross income.
- Lifestyle creep can lead to bankruptcy even for high earners, as seen in a case study of a 20 million dollar profit year.
- Automating investments removes the reliance on human discipline which often fades over time.
- Comparison on social media is the fastest way to feel poor because it highlights spending rather than net worth.
- Keeping lifestyle expenses flat for two years after a raise significantly accelerates wealth building.
References
PeopleAlex Hormozi · Thuan Pham · Owen Jennings (twitter.com/owenbjennings) · David Haber (twitter.com/dhaber) · Károly Zsolnai-Fehér (cg.tuwien.ac.at/~zsolnai/) · Rob Dial (coachwithrob.com) · Josh Kale (x.com/JoshKale) · Ejaaz (x.com/cryptopunk7213) · Travis Kalanick · Dario Amodei · Jack Dorsey
ToolsClaude Code · W&B Weave · W&B Models · Goose · Builder Bot · TurboQuant · Bun.js · Money Bot · Manager Bot
PapersTurboQuant