Gianfranco's Best of June 2025 Reading List

Welcome to the June 2025 edition of my monthly reading list.  

This month, I've curated my favorite essays offering insights into the latest developments in AI, societal impacts of technology, strategic business thinking, financial trends, and personal growth.

If you only have a few minutes, these three posts were my favorite:

• Scaling Reinforcement Learning: Environments, Reward Hacking, Agents, Scaling Data ($) - Dylan Patel and AJ Kourabi
  • Reinforcement learning inverted AI economics overnight. Models burn 100x more inference compute than training. OpenAI's response? Abandon GPT-4o for inference-optimized base models like GPT-4.1. The constraint becomes the moat: bankrolling inference-heavy workloads while preventing reward hacking owns the future. China can't play—export controls killed their inference capacity. Data quality trumps quantity. Engineering environments that don't get gamed requires STEM PhDs writing rubrics. The bottleneck shifted from training clusters to distributed inference farms.
  • Read More
• Making Uncommon Knowledge Common - Kevin Kwok
  • Rich Barton built three billion-dollar companies with the same trick: weaponize hidden data. Before Barton, airline prices lived with travel agents. Home values belonged to brokers. Salaries stayed behind closed doors. He dragged them all into daylight—then owned every search result. The "Data Content Loop": surface proprietary information, index it at scale, dominate SEO forever. Glassdoor made companies take anonymous reviews seriously. Zillow's Zestimate gave buyers ammunition against brokers. Transparency wasn't charity. It was conquest.
  • Read More
• What Happened to Working Your Way Up from the Mailroom? - The Diff
  • Sidney Weinberg cleaned toilets at Goldman Sachs. Sixty years later, he ran it. That path died when we got too good at sorting talent. Harvard sorts perfectly—which is why the Thiel Fellowship produces unicorn founders at 70x Harvard's rate (13.8% vs 0.2%). Every status ladder that gets too precise creates outcasts who build their own. Jensen Huang's résumé would get auto-rejected at modern NVIDIA. The mailroom-to-CEO story died, but something better replaced it: founder-to-CEO. The market's efficiency at crushing dreams young is precisely what motivates the next generation to build alternative hierarchies.
  • Read More

Artificial Intelligence

• Software Is Changing (Again) - Andrej Karpathy
  • Programming is dead. Karpathy maps three paradigms: traditional code (1.0), neural networks (2.0), and natural language prompts (3.0). LLMs aren't tools—they're operating systems. The revolution: anyone who can write a sentence can now program. Partial autonomy beats full automation. Human oversight remains the killer feature. Software development just became a conversation.
  • Read More
• Scaling Reinforcement Learning: Environments, Reward Hacking, Agents, Scaling Data ($) - Dylan Patel and AJ Kourabi
  • Reinforcement learning inverted AI economics overnight. Models burn 100x more inference compute than training. OpenAI's scramble: abandon GPT-4o for inference-optimized base models like GPT-4.1. The constraint becomes the moat—only players bankrolling inference-heavy workloads while preventing reward hacking win. China can't play; export controls killed their inference capacity. STEM PhDs now write rubrics to stop models gaming their own rewards. The bottleneck shifted from training clusters to distributed inference farms.
  • Read More
• Why I don't think AGI is right around the corner - Dwarkesh Patel
  • LLMs can't learn on the job. Patel identifies the fatal flaw: models lack continual learning—they can't internalize feedback like employees. Without live adaptation, only 25% of white-collar work gets automated. His timeline: competent computer-use agents by 2028, human-level online learning by 2032. Then productivity explodes. Self-teaching, not scaling, decides who wins AGI.
  • Read More
• Claude 4, Anthropic Agents, Human-AI Agents - Ben Thompson
  • Anthropic built replacement workers, not assistants. Claude Opus 4 and Sonnet 4 operate autonomously—actual agents doing jobs. Thompson's twist: "human-AI agents" where humans and AI share the same ChatGPT session beat both solo humans and solo AI. The productivity unlock isn't automation; it's auditable AI collaboration. Hybrid beats pure.
  • Read More
• Apple Retreats - Ben Thompson
  • Apple admitted defeat by retreating. WWDC 2025: no more AI hype, just hardware-software integration. Liquid Glass design, on-device models, deeper OpenAI partnerships—Apple plays to strengths while ceding the AI frontier. The retreat is strategic: own the experience layer while others fight over foundation models. Sometimes winning means knowing where not to compete.
  • Read More
• Why is this MCP thing everywhere? ($) - Technically
  • Anthropic's Model Context Protocol wants to be the USB of AI. MCP standardizes how AI models access external data and applications. The hype outpaces reality—adoption requires enterprise integration maturity that doesn't exist yet. Standards only matter when everyone uses them. MCP's fate depends on implementation, not architecture.
  • Read More

Semiconductors

• AMD vs NVIDIA Inference Benchmark: Who Wins? – Performance & Cost Per Million Tokens - SemiAnalysis
  • AMD beats NVIDIA at inference—sometimes. MI325X and MI300X outperform H200 and H100 on large dense models at high concurrency. The catch: AMD's software remains garbage, rental markets don't exist, and advanced features like disaggregated prefill don't work. Raw performance means nothing without ecosystem. NVIDIA's moat isn't silicon; it's everything else.
  • Read More
• Nvidia Earnings, Nvidia's China Argument, NVLink Fusion - Ben Thompson
  • Export controls just killed NVIDIA's software moat. Thompson's insight: blocking China from CUDA forces Chinese developers onto competing platforms. Short-term win, long-term catastrophe—America hands China the motivation to build indigenous AI infrastructure. Geopolitics trumps quarterly earnings. The real competition starts when your biggest customer becomes your rival.
  • Read More
• AMD Advancing AI: MI350X and MI400 UALoE72, MI500 UAL256 - SemiAnalysis
  • AMD's MI400 finally gets rack-scale right. Current MI350X/MI355X chips offer decent small-model inference but crumble against NVIDIA's GB200 NVL72 at scale. The MI400's UALoE72 changes everything—true rack integration, competitive price-performance, open networking. AMD bribes hyperscalers and Neoclouds to overcome its software deficit. Sometimes you buy market share while you build capability.
  • Read More

Startup Strategy

• Making Uncommon Knowledge Common - Kevin Kwok
  • Rich Barton weaponized transparency to build monopolies. His formula: surface hidden data (airline prices, home values, salaries), create "Data Content Loops," dominate SEO forever. Expedia, Zillow, and Glassdoor all followed the same playbook—make private information public, then own every search result. Transparency wasn't charity; it was conquest. Three billion-dollar companies prove the pattern.
  • Read More
• Excuse me, is there a problem? - Jason Cohen
  • Most startups solve the wrong problem. Cohen's framework: real problems aren't enough—customers must know they have them, want solutions, and afford fixes. His quantitative validation covers five dimensions: plausibility, self-awareness, liquidity, eagerness, and enduring value. Skip any dimension and die. The graveyard is full of perfect solutions to problems nobody will pay to solve.
  • Read More
• Nike on Amazon; Nike's Disastrous Pivot; Inevitability, Intentionality, and Amazon - Ben Thompson
  • Nike forgot how discovery works. Abandoning wholesale for direct-to-consumer killed brand recall and let competitors steal shelf space. Worse: shifting from brand-building to direct-response ads assumed customers actively choose Nike. Thompson's insight: great brands are inevitable, not intentional. Nike's Amazon return admits the truth—presence beats control.
  • Read More

Finance and Economics

• Five Alive - Marc Rubinstein
  • Traditional banking is dead; it just doesn't know it yet. Rubinstein's five-year retrospective maps the revolution: private credit ate bank lending, asset managers discovered structural advantages, governments became market makers, fintech merged with banking, and nobody knows what "financial institution" means anymore. The boundaries dissolve faster than regulators can draw them.
  • Read More
• Saved by Medicaid: Health Insurance and Mortality from the Universe of Low-Income Adults - NBER
  • Medicaid saves lives at $179,000 per life-year. Studying 37 million Americans, ACA expansions increased enrollment 12% and cut mortality 2.5%—a 21% reduction for new enrollees. Younger adults drove most life-years saved. At $5.4 million per life saved, it's expensive but effective. The data settles the debate: coverage equals survival.
  • Read More
• The Rise of Private Credit & Systemic Risk - Private Credit Webinar
  • Private credit is the new subprime. Post-2008 reforms pushed corporate borrowing from banks to shadow lenders. Borrower-friendly terms multiply leverage while creating dense risk networks. BDCs surpassed banks during COVID stress; higher connectivity predicts bigger losses. Today's modest scale masks tomorrow's systemic bomb. We're building the next crisis in plain sight.
  • Read More
• Secret Diary of a Hedge Fund Allocator - Marc Rubinstein
  • 70% of hedge funds die—Rubinstein picked them anyway. His 25-year confession: survivorship bias is real, most managers fail, and success requires matching structure to strategy. Diversified risk-takers like Brevan Howard survive; concentrated cowboys like Thélème occasionally thrive. The lesson: bet on adaptability, not returns. Past performance doesn't predict survival.
  • Read More
• More on Repealing the Laws of Economics - Howard Marks
  • Governments can't repeal economics—California proved it. Marks dissects intervention failures: rent control creates shortages, insurance caps cause exits, tariffs raise prices. California's insurance crisis crystallized the lesson: regulate prices below cost, watch suppliers vanish. Markets aren't fair, but they're real. Fighting economic law is like fighting gravity—painful and brief.
  • Read More
• The Concept of Costs - Byrne Hobart
  • Nothing is free—especially free things. Hobart expands "cost" beyond money: time, inconvenience, opportunity, strategic vulnerability. Free services extract attention via ads. Vendor consolidation creates hidden dependencies. True cost accounting includes everything you sacrifice, not just what you pay. Business winners internalize total costs; losers see only price tags.
  • Read More

Startups

• How to Find Superstar Talent - Marc Lore
  • Marc Lore reads résumés like X-rays. His pattern recognition: stacked promotions, multi-year tenures, consistently upward trajectory. Top 5% performers appear once per twenty résumés. His hack: extend searches, use interviews to sell vision, promise rapid advancement. Great cultures don't need perks—they need progression. Convert elite talent's ambition into equity.
  • Read More
• Better Filters - The Diff
  • Harvard creates good employees; the Thiel Fellowship creates founders. Elite schools produce 0.2% unicorn founders; Thiel Fellows hit 13.8%—70x better. Israel's Unit 8200 beats MIT at producing entrepreneurs. The gap reveals the game: broad institutions optimize for conformity, narrow filters select for obsession. Society needs both ladders—one for stability, one for disruption.
  • Read More
• Understanding 'Forward Deployed Engineering' and Why Your Company Probably Shouldn't Do It - Barry O'Reilly
  • Palantir's FDE is organized chaos that works. Elite engineers embed with customers, building custom products as "deployments." It's R&D disguised as implementation—expensive, messy, bottom-up innovation. Most companies can't stomach the costs, complexity, or loss of control. FDE only works with deep pockets, patient customers, and cultural acceptance of beautiful failures.
  • Read More

Technology and Society

• What Happened to Working Your Way Up from the Mailroom? - The Diff
  • Sidney Weinberg cleaned Goldman's toilets for six decades until he ran it. That path died when talent sorting got efficient. Markets now crush dreams at 22 instead of 40. But efficiency creates its own inefficiency: Jensen Huang couldn't get hired at today's NVIDIA. Perfect sorting breeds perfect blindness. Every precise status ladder spawns outcasts who build alternative hierarchies. The mailroom is dead; the garage is thriving.
  • Read More