Skip to content

Essays

  • 5 min read

Is AI More Expensive Than the Employees It's Replacing?

Source: Prof G Markets — Scott Galloway & Ed Elson

TL;DR

The AI cost paradox: AI is proving more expensive than the humans it replaces. Key data points: Uber blew through its entire 2026 AI budget in 4 months. Microsoft is canceling Claude Code licenses across multiple divisions. A senior Nvidia executive says the cost of compute is "far beyond" employee costs. Meta, Pinterest, and Spotify cite rising inference costs as a drag on Q1 margins. 45% of firms now spend >$100K/month on AI (up from 20% the prior year). An Anthropic employee used $150K worth of Claude Code in a single month. Yet only 8% of S&P 500 companies disclose any AI revenue, and only 50% can confidently evaluate AI ROI. The authors predict a shift to cheaper Chinese LLMs (10x–30x cheaper), with usage growing from 1% in 2024 to >60% in May 2026.

  • 4 min read

Brooks-lint: AI Code Reviews Grounded in 12 Classic Engineering Books

Source: Brooks-lint

TL;DR

Brooks-lint synthesizes wisdom from 12 classic software engineering books (The Mythical Man-Month, Code Complete, Refactoring, Clean Architecture, The Pragmatic Programmer, Domain-Driven Design, and others) into structured, citable code review findings. It detects 6 production decay risks (Cognitive Overload, Change Propagation, Knowledge Duplication, Accidental Complexity, Dependency Disorder, Domain Model Distortion) and 6 test suite decay risks. Each finding follows a Symptom → Source → Consequence → Remedy format with book citations. The tool supports 6 analysis modes: PR Review, Architecture Audit, Tech Debt Assessment, Test Quality Review, Health Dashboard, and Full Sweep Auto-Fix. It is available as a Claude Code Plugin, Gemini CLI Extension, Codex CLI Skill, and GitHub Actions integration.

  • 5 min read

Can AI Refute Economic Theory?

Source: Can AI Refute Economic Theory?

TL;DR

Alexis Akira Toda (Emory University) tests whether LLMs can identify errors in published economic theory papers. The study tested 4 papers with known errors using Gemini, Refine, Claude, and ChatGPT. The result: no model located a true error without substantial human guidance, and data contamination complicates interpretation. However, ChatGPT Pro 5.5 was "truly amazing" — it constructed valid counterexamples for Tirole (1985), provided a corrected proof for Kocherlakota (1992) more elegant than the published corrigendum, and correctly identified a tautological transversality condition issue in Miao & Wang (2018). Claude Opus 4.8 was weaker on math but stronger on economic judgment. Gemini performed worst, defending errors with plausible but unfounded arguments. The conclusion: "a competent human paired with a frontier model can outperform current peer review, but AI cannot yet refute economic theory on its own."

  • 4 min read

China's Not the Problem. We Are.

Source: China's Not the Problem. We Are. by Ross Douthat (NYT Opinion — "Interesting Times" podcast) with Kyle Chan

TL;DR

A New York Times podcast transcript turned column. Kyle Chan (Brookings Institution) argues that the US and China have fundamentally different AI goals, making it difficult to frame the situation as a simple "race." The US focuses on frontier breakthroughs and AGI, while China's approach is more state-directed with different priorities. The conversation explores the Cold War atmosphere around the Trump-Xi meeting in Beijing and what "winning" even means in the AI context. The title encapsulates the central argument: America's own limitations and blind spots may be a bigger obstacle than Chinese competition.

  • 6 min read

Dark Knowledge: Distilling Ensemble Knowledge into Smaller Models

Source: Dark Knowledge — Hinton, Vinyals & Dean (Google)

TL;DR

This seminal paper by Geoffrey Hinton, Oriol Vinyals, and Jeff Dean (Google) introduces knowledge distillation — transferring the function of a large ensemble into a single small model using "soft targets." The key idea: train a high-temperature softmax on the ensemble's averaged logits, revealing "dark knowledge" (e.g., a classifier that sees a 2 resembles 4 and 7). The distillation objective minimizes two cross-entropies: one with soft targets from the ensemble at high temperature, one with hard targets at temperature 1. Results: MNIST (74 errors vs 67 for the big net), a distilled net with no 3s in training still gets 98.6% correct on 3s, and the paper's concluding advice: "Always distill your ensembles!"

  • 3 min read

Google DeepMind Science Skills

Source: Google DeepMind Science Skills

TL;DR

Google DeepMind has released an open-source collection of agent skills for scientific research tasks spanning genomics, structural biology, cheminformatics, and literature search. The toolkit integrates with AlphaGenome, AlphaFold DB, UniProt, and over 30 other scientific databases, and is designed for the Google Antigravity platform. It includes structured instructions, scripts, and resources for specialized scientific tasks. The project requires the uv Python package manager, some skills need API keys (e.g., AlphaGenome, OpenAlex), and it is licensed under Apache 2.0 — though it is not an official Google product.

  • 4 min read

Dia2: Streaming Conversational TTS Model

Source: Dia2

TL;DR

Dia2 is a streaming dialogue text-to-speech model developed by Nari Labs that can start generating audio as soon as the first few words are given as input, without needing the entire text. It can be conditioned on audio input for natural realtime conversations. Available in 1B and 2B parameter sizes, it generates up to 2 minutes of continuous audio (English only). The model uses the Kyutai Mimi audio codec operating at approximately 12.5 Hz frame rate. Licensed under Apache 2.0, it is intended for research and educational use. Quality and voices vary per generation unless the model is fine-tuned on a specific voice.

  • 5 min read

Staying Power: The Dollar-Security Nexus

Source: Staying Power: The Dollar-Security Nexus by Ho-fung Hung (New Left Review / Sidecar)

TL;DR

The dollar's global dominance is sustained less by market "network effects" and more by the US military security umbrella — what Hung calls the "dollar-security nexus." While this security foundation is weakening (US relative decline, fiscal strains, strategic overreach), no viable alternative exists because China's political system requires strict capital controls that prevent RMB convertibility. Paradoxically, the CCP's governance imperatives prolong the dollar's hegemony. Key data: dollar hegemony as a pure fiat currency (1971–2026) has already lasted more than twice as long as its gold-backed era (1945–1971). In 1947, ~90% of global FX reserves were Sterling — a reminder that reserve currency status can collapse. Offshore RMB deposits in Hong Kong amount to less than 0.5% of onshore deposits, demonstrating the chasm between China's economic weight and its currency's global role. The euro is structurally flawed — it lacks a unified fiscal and political authority. Hung calls it the "great paradox": China's control over its economy has become a key factor in the staying power of the American empire.

  • 4 min read

Let's Talk About Encrypted Reasoning

Source: Let's Talk About Encrypted Reasoning

TL;DR

Matthew Green investigates encrypted "reasoning/thinking" blobs in Frontier LLM APIs from OpenAI and Anthropic. These encrypted blobs ship the model's raw chain-of-thought reasoning to the client as authenticated ciphertext (Base64-encoded) to support stateless, zero-retention conversations without letting the client read internal monologue. Key findings include: (1) replay attacks work across sessions, different accounts, and even different models — implying a single global key, meaning everyone's reasoning data is escrowed under one key, and replayed blobs remain semantically active; (2) side-channel attacks via encrypted blob size, reasoning_tokens field, and wall-clock response time leak information — the researcher extracted the byte 0xA3 bit-by-bit via 80 trials. Attempts to extract system prompts failed because models don't actually have system prompts in API mode and hallucinate plausible ones. Both OpenAI and Anthropic downplayed the findings.

  • 5 min read

Energy Geopolitics and European Markets in Crisis

Source: Francesco Sassi (@Frank_Stones) on X | Volue Power Summit 2026 — Amsterdam

TL;DR

Francesco Sassi (Frank_Stones), a Research Fellow in energy geopolitics and markets at RIE and postdoctoral fellow/assistant professor at the University of Oslo, discusses European energy market dynamics during ongoing crises, the geopolitical implications of energy transitions, and the intersection of energy security with broader geopolitical competition. He delivered a presentation on "The Geopolitics of European Energy Markets in Times of Crisis" at the Volue Power Summit 2026 in Amsterdam. Sassi also serves as an advisor to the Italian Parliament on energy matters.