July 21, 2025
Pushing compute to the limits of physics
Machine Learning Street Talk
Key Takeaways:
- Hardware is the New Frontier. The scaling race isn't about building more data centers for the same old chips. The next 1000x improvement will come from a fundamental paradigm shift that works with physics, not against it.
- Noise is a Feature, Not a Bug. The future of efficient computing lies in harnessing stochasticity. The "noise" we spend billions to suppress in classical chips is the very resource that can power probabilistic AI models with unparalleled efficiency.
- Prepare for an Algorithmic Renaissance. The dominance of Transformers is a temporary state dictated by current hardware. As thermodynamic computers become available, developers and researchers should dust off their probabilistic ML textbooks—the algorithms of tomorrow will look very different.
See full notes
July 21, 2025
José Macedo and Pondering Durian: The Birth of Delphi Intelligence
Delphi Digital
Key Takeaways:
- Invest in AI's Tailwinds: The essential question for any AI investment is: "Does this business get better as foundation models improve?" Companies fighting against the current of AI's scaling laws are on the wrong side of a powerful trade.
- The "Mag 7" Will Expand, Not Just Turn Over: AI is not a zero-sum game for incumbents. The total addressable market is set to 10x as AI drives labor costs toward zero, creating room for a "Mag 25" and turning today's $500B companies into tomorrow's $5T behemoths.
- Private Market Alpha Exists, But Edge is Paramount: The private AI market cap is a mere ~$700B, signaling massive growth potential. However, like in crypto, investors must be paranoid about their "edge," as the best deals require deep ecosystem access to avoid negative selection.
See full notes
July 21, 2025
The Future of Software Development - Vibe Coding, Prompt Engineering & AI Assistants
a16z
Key Takeaways:
- **AI Isn't a Feature; It's a New Infrastructure Primitive.** For the first time, developers are outsourcing logic, not just resources. This fundamentally changes how software is built, valued, and sold.
- **Abandon Zero-Sum Thinking.** The AI market is in a massive expansion phase, not a consolidation battle. Value is accruing at every layer of the stack simultaneously; assuming one layer's gain is another's loss is a flawed thesis.
- **The Future is More Developers, Not Fewer.** AI tools augment productivity and lower the barrier to entry. This elevates the developer's role to focus on product design and workflow definition—the *real* hard problem in software.
See full notes
July 20, 2025
Hash Rate - Brief #1 - Chutes
Hash Rate pod - Bitcoin, AI, DePIN, DeFi
Key Takeaways:
- **A Killer Value Prop:** Chutes makes deploying powerful AI models 85% cheaper and as easy as building a website on Squarespace.
- **The Investor's Dilemma:** While all revenue is used to buy back the Chutes alpha token, this currently covers only 10% of the daily token emissions. The token's price stability is heavily dependent on external market demand outstripping this inflation.
- **Watch for Catalysts:** Two key events could dramatically increase buy pressure: the imminent launch of BitTensor subnet tokens on Solana and an anticipated wave of institutional capital from newly formed crypto hedge funds.
See full notes
July 20, 2025
Hash Rate - Ep 122 - ZEUS Weather - Subnet 18 $TAO
Hash Rate pod - Bitcoin, AI, DePIN, DeFi
Key Takeaways:
- **Specialization Unlocks Performance.** ZEUS proves that a decentralized network of specialized AI agents can outperform monolithic, state-of-the-art models, achieving a nearly 40% lower error rate in weather forecasting.
- **Revenue Sharing is the Next Evolution.** The plan to distribute API revenue directly to network participants in stablecoins represents a major step toward sustainable subnet economies, moving beyond token buybacks and emission-based rewards.
- **The Valuation Gap is the Opportunity.** Despite massive potential, subnets have extremely low market caps compared to their Web2 equivalents. For long-term believers, this asymmetry presents a compelling, albeit early, investment thesis.
See full notes
July 19, 2025
SN52 :: Tensorplex Dojo :: High-Quality Human-Generated Datasets on Bittensor
Opentensor Foundation
Key Takeaways:
- Human Intelligence is the Ultimate Moat: In an era of synthetic data, Dojo is creating a defensible moat by generating proprietary, high-quality human preference data. This is the raw material for the next generation of fine-tuned, specialized models.
- A New Paradigm for Validation: Dojo’s mechanism of using subtle "perturbations" to test labelers is a breakthrough. It solves the cold start problem of validating subjective human feedback in a decentralized network.
- The Future is Human-Agentic Collaboration: Dojo is evolving from a data-generation subnet to a platform for human-agentic workflows, with applications in robotics, video analytics, and 3D generation. In the long term, it aims to be a crucial tool for aligning AI with human values.
See full notes
July 18, 2025
Tech Executives: AI Has Changed SAAS Forever (Don't Fall Behind)
a16z
Key Takeaways:
- Your Pricing Model Is Now a Dynamic Weapon. The five-year pricing plan is dead. You must build the infrastructure and culture for constant experimentation and rapid iteration. If you’re not re-evaluating your model quarterly, you're falling behind.
- This Is a CEO-Level Mandate. Shifting to usage-based pricing is a full-company transformation that requires top-down vision. The CEO must act as the "pricing dictator" to align sales, product, and finance around a unified strategy of value creation and capture.
- Your Product Team Now Owns Revenue. In a usage-based world, the core value metric *is* your revenue. Product and engineering teams must become obsessed with driving the specific usage that customers pay for, making their impact on the bottom line completely objective.
See full notes
July 17, 2025
Can AI Be Creative? With AI Artists Mario Klingemann and Shavonne Wong
The People's AI
Key Takeaways:
- AI as a System, Not a Tool: Advanced AI art projects are not just prompt-driven tools but autonomous systems. They use feedback loops (DAOs, user interaction) to develop their own "taste" and creative trajectory, aiming for a level of agency beyond simple human puppeteering.
- AI Reveals Human Vulnerabilities: AI companions act as a social mirror, showing that humans fundamentally crave connection and non-judgmental spaces. We are turning to AI to fulfill core needs that are often unmet in our human-to-human relationships.
- The Artist's Dilemma: Adapt or Perish: Resisting AI is becoming a losing battle. The future for artists isn't about competing with AI on replication but on finding what AI can't do, critiquing it from within, or carving out a niche for "100% human-made" work in a world of synthetic media.
See full notes
July 16, 2025
The Humble Truths Behind Bombastic AI Papers
Machine Learning Street Talk
Key Takeaways:
- Benchmarks are broken. The ML community can no longer rely on leaderboards as a proxy for truth. The new frontier is developing robust, qualitative explanations for why models succeed or fail.
- Embrace the illusion. The most effective models aren’t finding universal laws but are constructing powerful, computationally efficient illusions of them. Progress lies in refining these illusions, not in a futile search for Platonic perfection.
- Think like a physicist. The future of foundational AI research is to treat models as complex physical systems. The task is to design parametric models where stochastic processes, like SGD, can efficiently "relax" into a state that approximates the data distribution.
See full notes
