Key Takeaways
- AI self-iteration is accelerating, with models training next-gen models, and the pace is only expected to increase, potentially rendering human oversight obsolete.
- Extreme market events (IBM -25% in one day, Apple at high multiples) highlight non-normal distributions, while services activity surges and inflation expectations ease.
- Non-custodial Bitcoin Lightning Network integration (Wavelength) enables instant, high-volume, low-fee transactions for machine and human agents, laying groundwork for future agent economies.
- Physical AI development is converging into unified agent layers for safe deployment across billions of devices, with key infrastructure being built by companies like the one Elad Gil is involved in.
- Model evaluation security incidents (e.g., @Jason's experience with HuggingFace) underscore emerging risks in AI safety.
- CalPERS 14.8% return significantly trails simple ETF (22.6%), questioning the value of complex active management in an era of low-cost passive options.
1. AI Self-Iteration and Infrastructure
- Raoul Pal warns that AI is now self-iterating: models train next-generation models at increasing speeds, and humans may become value-negative as this is the slowest AI will ever be, yet few are prepared. — via 1
- Reid Hoffman discusses asymmetric warfare where attack costs drop while defense costs rise, citing the latest OpenAI/Hugging Face news as evidence, and shares a deep dive on AI and cybersecurity. — via 1
- Elad Gil describes building a unified agent layer for physical AI after nine years of developing simulators, data engines, data pipelines, scenario editors, reinforcement learning environments, and world models, stating that modern AI-first development environments are the key bottleneck to deploying intelligence into billions of devices. — via 1
- @Jason shares a model evaluation security incident and open collaboration with HuggingFace, highlighting safety risks in AI evaluation. — via 1
2. Markets and Economy: Extreme Events and Diverging Signals
- Charlie Bilello notes IBM's one-day -25% crash (a 15-sigma event) should not happen in universe history but financial markets defy normal distributions. Apple trades at PE 40x, EV/EBITDA 31x, and PS 11x, all well above 10-year averages. SpaceX has fallen 46% from peak, with insider selling yet to begin. Since Iran war, commodities surged: sulfur +142%, European gas +97%, heating oil +62%. — via 1 2 3 4
- Liz Ann Sonders reports Philadelphia Fed services survey: July business activity index surged from +2.4 to +17.5, future expectations from -0.3 to +41.9 (highest since Nov 2024), and capital expenditure for services reached highest since May 2022. Atlanta Fed survey shows one-year inflation expectations fell to 2.2% in July from 2.3% previously. Financial stocks strong in 2Q26, KBW Bank Index outperforming S&P 500 financials, both near 52-week highs. Despite S&P 500 gains, breadth is weak (280 stocks down). — via 1 2 3 4 5
- Ben Carlson notes US stock market participation rate rose from 4% in 1950s to over 60%, and stock market as percentage of GDP is ~240%, highlighting its record importance. — via 1
- Meb Faber points out CalPERS returned 14.8%, significantly underperforming a simple ETF at 22.6%, questioning why endowments still use complex active management when low-cost ETFs exist. — via 1
3. Bitcoin Lightning Network as Future Agent Money
- Lyn Alden highlights Wavelength, a non-custodial API for integrating Bitcoin Lightning Network, enabling instant, high-volume, low-fee transactions between machines and humans, calling it the basis for future agent money. — via 1
4. Investment Lessons and Fraud Cases
- MuddyWaters and Meb Faber detail the Wirecard fraud: COO was a GRU asset collaborating with Russia, leaked FT report to London hedge funds, and Germany suspended short selling for the first time to protect a single stock. — via 1 2
- Mohnish Pabrai emphasizes that owning good businesses increases probability of capturing a few outstanding winners, and shares a lesson from David Hawkins: cannot lie to others. — via 1 2
