Daily Digest - 2026-09-13

Themes from yesterday

• The Pacing vs. Cartel Debate at the Frontier: Calls by major frontier labs for coordinated development pacing and antitrust waivers ran into sharp pushback from competitors and regulators. Critics argued that existing antitrust law already allows safety cooperation, warning that special exemptions would create state-backed oligopolies.
• Physical and Financial Bottlenecks "South of the Model": Compute bottlenecks have moved from model algorithms to physical infrastructure and financial risk. Operators are grappling with high-voltage direct current grid upgrades, custom ASICs for specific models, and new financial derivatives designed to manage basis risk in multi-year GPU contracts.
• Production Agent Infrastructure & Security Decoupling: Teams are converging on designs that separate the orchestration harness ("brains") from isolated, short-lived microVM execution sandboxes ("hands"). At the same time, MCP gateways are starting to virtualize, search, and batch tool calls using code-execution modes.
• Automated Scientific Milestones vs. Epistemic Foundations: OpenAI's reported solution to the Navier-Stokes problem using a 10,000-agent swarm drew serious criticism from Fields Medalists. The debate underscored the difference between brute-forcing benchmark proofs and creating modular, reusable knowledge that helps humans and machines solve future problems.

1. Evals Skills for Coding Agents — Hamel Husain, Hamel's Blog

• Why read: See how giving coding agents domain-specific evaluation skills lets them audit, diagnose, and benchmark AI applications on their own.
• Summary: Eval vendors now offer Model Context Protocol (MCP) servers that expose runtime traces, but coding agents still need procedural domain knowledge to run useful evaluations. Broad metrics like aggregate hallucination scores blur the difference between factual mistakes and unauthorized action hallucinations. To address this, the open-source `evals-skills` plugin gives parallel subagents modular workflows to audit pipelines across error analysis, judge calibration, and human review interfaces. Instead of counting on bigger models to fix quality issues, teams should treat evals as runtime infrastructure where agents check their own outputs against telemetry. Standardizing these workflows helps teams turn raw execution logs into automated, reproducible improvements in production.
• Read more

2. Secure agents: architecture and sandboxing — Katelyn Lesse, X (formerly Twitter)

• Why read: A breakdown of how teams are securing agent runtimes by separating orchestration and secrets from untrusted execution sandboxes.
• Summary: Most recent agent security failures happened because developers bundled the orchestration harness, API secrets, and untrusted code execution into one environment. The emerging fix splits the system into two parts: the "brain" (state, credentials, and harness on secure infrastructure) and the "hands" (an isolated execution sandbox). Rather than using shared-kernel containers, sandboxes should run inside hardware-isolated microVMs like Firecracker with outbound network access blocked by default. An external egress proxy should handle outbound API requests and OAuth tokens, injecting destination-scoped credentials on the fly so raw secrets never reach the sandbox. Finally, teams should treat sandboxes as disposable, backing them with externalized logging and out-of-band kill switches to contain prompt injections or rogue actions.
• Read more

3. MCP Infra Roundup: The MCP Server Is No Longer Calling the Shots — Josh Rosen, X (formerly Twitter)

• Why read: How gateways, tool-search patterns, and code execution modes are turning individual MCP servers into background implementation details.
• Summary: More production setups are putting MCP servers behind aggregation gateways that combine multiple backends into a single endpoint. To keep hundreds of tool schemas from swamping context windows, teams like PayPal and Solo.io use search-and-execute patterns that cut tool-related token usage by over 98%. At the same time, newer "Code Mode" frameworks convert upstream MCP tools into typed libraries, letting models batch several operations into one script. Middleware proxies now handle authentication, token exchanges, and human-in-the-loop approvals before traffic ever touches the backend servers. For developers, this means MCP tools should be built as clean, standalone capability contracts rather than assuming models will consume raw, static schemas directly.
• Read more

4. The Bottleneck is South of the Model — Gokul Rajaram, X (formerly Twitter)

• Why read: Why the biggest barrier to AI growth has moved from software algorithms to power grids, high-voltage DC equipment, and custom silicon.
• Summary: Leading venture investors argue that model algorithms are no longer the main bottleneck in AI. Instead, the real constraints have moved to physical infrastructure: energy, datacenters, and custom hardware. Agent workflows and chain-of-thought reasoning drive compounding token demand, pushing compute requirements up by orders of magnitude. As rack power requirements jump from 10kW to 150kW, facilities have to switch to high-voltage direct current (DC) power, which runs straight into shortages of certified electricians and available grid capacity. Meanwhile, with frontier training runs costing billions, massive inference demand now makes it economically sensible to build model-specific ASICs for double-digit efficiency gains. Capital increasingly buys competitive development time, sparking a major resurgence in hardware and low-level systems engineering.
• Read more

5. A Beginner’s Guide to Inference Engineering — danialhasan, X (formerly Twitter)

• Why read: A practical primer on inference engineering, covering memory bandwidth constraints, KV-cache tuning, and production serving trade-offs.
• Summary: Inference engineering has become its own discipline, focused on serving models under tight constraints on latency, memory, throughput, and cost. Generating a response splits into two distinct phases: a compute-bound prefill phase and a memory-bandwidth-bound decode phase. That split makes managing the KV cache the core operational problem. In production, serving frameworks use techniques like vLLM's PagedAttention to handle non-contiguous memory and SGLang's RadixAttention to cache reusable prompt prefixes. On edge devices and local hardware, quantization through llama.cpp lets models run on consumer chips without unacceptable drops in accuracy. In practice, teams should judge performance by tracking cost per accepted result and P95 tail latency, rather than focusing only on raw tokens per second.
• Read more

6. Handmade | The Shape of Compute Risk — cogsec, X (formerly Twitter)

• Why read: An analysis of the financial traps in fixed-price GPU contracts, and how derivatives and forward contracts are emerging to hedge compute exposure.
• Summary: Multi-year take-or-pay GPU contracts combine price, hardware capacity, and utilization risks into commitments that look a lot like unhedged commodity trades. But unlike oil, compute cannot be stored; unused cycles disappear instantly. That creates volatile spot markets and substantial basis risk between standard index prices and custom hardware clusters. Hardware owners face steep refinancing cliffs when three-year equipment loans outlast two-year customer contracts, leaving uncontracted periods that lenders value close to zero. On the buyer side, running a cluster at only 70% utilization pushes the effective cost per active hour up by more than 40%, unless the contract allows subleasing spare capacity. With CME preparing to launch compute futures, infrastructure teams will need to manage GPU commitments like a trading book using structured hedges and tighter utilization controls.
• Read more

7. Brand New AI Solves a Millennium Prize — Zvi Mowshowitz, X (formerly Twitter)

• Why read: A look at the controversy surrounding OpenAI's reported solution to the Navier-Stokes problem, and what it reveals about recursive self-improvement and research norms.
• Summary: An internal OpenAI frontier model trained in just days reportedly resolved the Navier-Stokes Millennium Prize problem using a swarm of 10,000 agents that generated 130 billion output tokens. The announcement immediately triggered arguments over attribution, poor communication with the broader math community, and accusations that lab rivalries led to rushed scooping. A group of 25 Fields Medalists, organized by Terence Tao, released a joint statement warning that treating pure mathematics primarily as a benchmark for AI systems harms the conceptual foundation of the discipline. At the same time, leaked internal OpenAI numbers showed automated coding and research agents now log 3.1 agent-workdays for every human workday, noticeably speeding up internal development. The episode offers a clear look at recursive self-improvement in practice, raising hard questions about coordination and safety governance.
• Read more

8. AI is Wireheading Scientific Progress — Gavin Owens, X (formerly Twitter)

• Why read: Why using agent swarms to brute-force scientific benchmarks risks hitting a dead end unless models learn to build and share reusable concepts.
• Summary: Using massive agent search to crack prize problems risks "wireheading" scientific discovery: claiming benchmark wins without producing the conceptual building blocks needed for further work. Drawing on theories of technological evolution, the post shows that bypassing intermediate milestones prevents systems from learning higher-order problem-solving skills. While autonomous swarms can generate mathematically verifiable proofs, these proofs often lack the conceptual structure that human or machine researchers need to apply the ideas elsewhere. Without a shared process where intermediate lemmas are tested, debated, and integrated into broader model knowledge, automated discoveries remain isolated one-offs. Builders designing automated research workflows need to instruct agents to produce modular, reusable ideas, rather than optimizing solely for a verified final answer.
• Read more

9. Ready to put AI to work? — Aidan Gomez, Cohere

• Why read: Cohere CEO Aidan Gomez makes the case that Big Tech requests for antitrust waivers and compute-based safety rules would entrench an incumbent cartel.
• Summary: Cohere CEO Aidan Gomez warns that letting major AI labs coordinate behind antitrust exemptions mirrors past industrial cartels that restricted competition and hurt consumers. He argues that using compute thresholds to define regulatory risk shields frontier incumbents while overlooking genuine risks from smaller, tool-equipped agent swarms. Past regulations in bond ratings and automotive safety show that protecting incumbents under the banner of safety usually locks in monopolies instead of improving outcomes. Gomez contends that sensible regulation should target actual capabilities and concrete harms, such as autonomous cyberattacks or chemical weapon design, no matter who built the model. For enterprises, the debate highlights the importance of keeping independent AI infrastructure instead of relying entirely on a small group of self-policing frontier providers.
• Read more

10. 1) AI will not eradicate humanity — Alvaro Bedoya, X (formerly Twitter)

• Why read: An FTC commissioner explains why antitrust law already allows safety cooperation, and why calls for broader waivers look like anti-competitive collusion.
• Summary: FTC Commissioner Alvaro Bedoya pushes back against existential extinction claims, arguing that lab executives' commercial expansion plans directly contradict their public warnings. He notes that existing antitrust law already permits companies to share information on health, safety, and cybersecurity without special exemptions. What the law actually prohibits is competitors coordinating to cap compute, restrict output, or protect their business models from lower-cost open-source competitors. Enforcers see requests for antitrust carve-outs as an effort to protect well-funded incumbents under margin pressure. For product and legal teams, the message is clear: regulators will not accept speculative catastrophe claims as cover for collusive agreements.
• Read more

11. The Accelerationist Case for Frontier Pacing — Contraptions, Substack

• Why read: An argument using operational performance theory to show how deliberately managing development pace can speed up progress over the long run.
• Summary: Borrowing concepts from athletics and military doctrine, this essay argues that chasing maximum speed without proper technique causes operational friction and costly delays. Managing frontier pacing should not just be seen as an existential precaution; it is practical tempo management that keeps evaluation tools and institutional processes aligned with growing model capabilities. While acknowledging valid worries about regulatory capture and corporate cartels, the author separates outright development freezes from deliberate pacing. Lasting progress in complex systems comes from improving feedback loops, tooling, and verification methods, not simply pouring in more compute. For teams building agent systems, steady pacing gives engineers time to build reliable observability and safety controls before operational risks compound.
• Read more

12. Any pursuit of superintelligence has to be grounded in the... — Satya Nadella, X (formerly Twitter)

• Why read: Microsoft CEO Satya Nadella outlines his perspective on superintelligence, focusing on enterprise data sovereignty, embedded evaluators, and open-source competition.
• Summary: Microsoft CEO Satya Nadella outlines a blueprint for steering superintelligence development so its economic returns spread broadly under human oversight. His strategy rejects an oligopoly of a few frontier labs, calling instead for open-source and proprietary models to coexist so customers can avoid vendor lock-in. He advises businesses to protect their proprietary and uncodified knowledge by creating continuous internal learning loops rather than handing everything to external model providers. Nadella also supports pacing alignment work and adopting technical safeguards like embedded evaluators to turn governance promises into concrete engineering checks. For enterprise architects, the takeaway is clear: design model-agnostic systems that keep core intellectual property inside internal, controllable weights.
• Read more

13. A simple model of AI-aided economic growth — Tyler Cowen, Marginal REVOLUTION

• Why read: Tyler Cowen uses a two-factor economic model to explain why cheap machine intelligence actually increases the value of tacit human knowledge.
• Summary: Tyler Cowen presents an economic model that splits productive output into two complementary inputs: formal Intelligence (which AI scales quickly) and Polanyi knowledge (the tacit, localized understanding held by humans). Because these two inputs complement each other, a sudden flood of raw machine intelligence makes human context and judgment more scarce and valuable, not less. Broad economic gains and disruptions will unfold gradually because institutions, corporate structures, and social norms change far more slowly than software. As a result, controlling frontier models provides limited leverage without the messy, unwritten context required to get work done in the physical world. For operators, the most valuable work involves organizing company context and internal workflows so models can be applied effectively.
• Read more

14. Making Startups Powerful — Paul Graham, paulgraham.com

• Why read: Paul Graham on building structural business power through direct customer ownership, network effects, and agent-friendly APIs.
• Summary: Paul Graham argues that founders create far more lasting value by focusing on structural business power rather than minor pricing tweaks. True leverage comes from owning the direct customer relationship, handling transactions on your platform, and building network effects into products that used to be standalone tools. Startups gain an advantage by moving upstream to capture operational data early, just as earlier software platforms built moats around core onboarding flows. In the early days, selling to fast-moving startups helps companies iterate and compound faster while avoiding grueling enterprise sales cycles. Looking ahead, as autonomous agents increasingly use software on behalf of humans, offering open, reliable APIs will be essential for any product that wants to remain relevant.
• Read more

15. Bending Spoons — Sriram Krishnan, Substack

• Why read: How Milan-based Bending Spoons operates like an engineering-driven private equity firm, buying mature software products and rebuilding their technical foundations.
• Summary: Bending Spoons operates as a hybrid between private equity and a hands-on engineering team, actively rewriting products rather than just adjusting financial levers. Unlike conventional software aggregators, the Milan-based firm targets plateaued apps with proven product-market fit, refactors their codebases, and updates their backend infrastructure. It merges operations, hosting, and customer support across portfolio companies into centralized teams, cutting redundant overhead while improving reliability. That setup offers a practical liquidity exit for venture-backed companies that hit a growth wall after years of heavy fundraising. For tech operators, it points to a wider shift where strict operational discipline and technical cleanup are replacing the old growth-at-any-cost playbook.
• Read more