Skip to content

Essays

  • 5 min read

GPT-5.4 Pro Solves Erdős Problem #1196

Source: Dongruo Zhou on X — Thread summarizing 7 parts

TL;DR

A significant milestone in AI mathematics: GPT-5.4 Pro solved Erdős Problem #1196, an open problem for 60 years. The problem concerns primitive sets of integers — whether for any primitive set A (no distinct elements divide each other), the sum of 1/(a log a) over A is bounded by 1 + o(1). Lichtman had previously proven a bound of approximately 1.399. OpenAI announced this result in June 2026. The proof was constructed through a prompting process by researcher Price. Subsequent work by Jared Duker Lichtman and others refined and adapted GPT-5.4's proof method to prove several additional results. This marks one of the first instances of a frontier LLM independently constructing a novel mathematical proof that experts couldn't find for decades.

  • 5 min read

How the Boomers Screwed Europe

Source: How the Boomers Screwed Europe by The Economist (Charlemagne Column)

TL;DR

The primary fault line in European inequality has shifted from horizontal (West vs East) to vertical (Boomers vs Millennials/Gen Z). Key grievances: young adults unable to move out of their parents' spare rooms due to sky-high house prices; 30-somethings paying hefty taxes to fund pensions for retirees who left the workforce in their prime. Costs related to ageing already consume a quarter of the EU's GDP, projected to rise further. The column characterises this as an "intergenerational confidence trick" — the Boomer generation knowingly secured its own prosperity by pushing fiscal and social debts onto its children and grandchildren.

  • 5 min read

LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks

Source: LEAP — Google AI Research | arXiv:2606.03303

TL;DR

LEAP (LLM-in-Lean Environment Agentic Prover) is an agentic framework from Google AI Research for automated formal theorem proving using general-purpose foundation models. Key results: on the 2025 Putnam Competition, it solves all 12 problems. On Lean-IMO-Bench (IMO-style problems in Lean), it boosts the one-shot formal solve rate from below 10% to 70%, surpassing the 48% benchmark of a specialized IMO system. It also autonomously formalized proofs for open combinatorial challenges, including a key subproblem in Knuth's Hamiltonian decomposition of even-order Cayley graphs.

  • 5 min read

Should the Lion Lie Down With the Electric Lamb?

Source: Should the Lion Lie Down With the Electric Lamb?

TL;DR

Antón Barba-Kay offers a critical analysis of Pope Leo XIV's encyclical Magnifica humanitas on AI, arguing it fundamentally fails to diagnose digital technology's nature. The Church (sacrament of true presence) and Big Tech (virtual absence) represent "the two most significant nonstate actors in human history" with competing visions of reality, yet the encyclical takes a conciliatory, moderate position — calling only for "just and appropriate use" of AI. Barba-Kay identifies a "neutrality trap": the encyclical claims technology isn't neutral because it reflects its creators, but this is functionally identical to the neutral tool argument. He argues AI transforms the very grounds of psychic self-interpretation, like language or drugs. While industrial capitalism compelled from outside-in, digital technologies manipulate from inside-out — "nudged and smiled their way into our hearts." The encyclical misses two specific risks: de facto transhumanism (indifference to human vs artificial) and the impossibility of moderate use when cognitive deskilling is already pervasive.

  • 5 min read

The Mismeasurement of European Productivity

Source: The Mismeasurement of European Productivity by Philippe Aghion, Antonin Bergeaud, and Luis Garicano

TL;DR

A technical rebuttal to Paul Krugman's argument that the US-Europe productivity gap is a measurement illusion. The core issue: using current PPP across time is statistically invalid — a PPP measures purchasing power across places at one moment, while a deflator compares prices across time in one place. IT accounts for 8% of US output but 45% of productivity growth, creating a massive measurement wedge. When a country produces goods whose prices fall rapidly (computers), valuing at current-PPP makes volume gains vanish. The "Seven Time Series" empirical proof shows the gap is real: US labour productivity growth ~1.93% vs France ~0.85% annually. The gap funds wages, welfare, rearmament, and the green transition — Europe should not persuade itself the problem is an illusion.

  • 4 min read

Supervision: Reusable Computer Vision Tools

Source: Supervision — Roboflow

TL;DR

Supervision is an essential open-source toolkit for computer vision from Roboflow. It is model agnostic — plug in any classification, detection, or segmentation model and get back unified sv.Detections objects. Provides connectors for Ultralytics, Transformers, MMDetection, and Roboflow Inference. Features highly customizable annotators, dataset utilities (load, split, merge, save, convert), and real-time zone counting. Python >=3.9, open source license, 4,877 commits, 36 releases. Extensive documentation, cheatsheets, cookbooks, and community resources available.

  • 6 min read

Tim Cook's Apple

Source: Tim Cook's Apple by Trung Phan (SatPost)

TL;DR

A comprehensive assessment of Tim Cook's 15-year CEO tenure at Apple (2011–2025). Key numbers: market cap $350B → $4T, revenue $108B → $416B, net profit $26B → $112B, services revenue $10B → $109B (11×), wearables $36B (exceeds AMD, McDonald's, and Adobe individually). Where Steve Jobs took products from 0→1, Cook took them from 1→2.5 billion. The good: supply chain mastery, scaling iPhone to 2B+ units, AirPods (400M+), Apple Watch, the Services tollbooth, and $700B in buybacks. The bad: Apple Car (shuttered after tens of billions), Vision Pro ($50B write-off), Apple Maps (initially disastrous), and fumbling AI. China dependency makes Apple the most exposed company to US-China trade war. AI strategy: sitting out the $700B capex race, betting on on-device edge inference. Likely successor: John Ternus (led the Mac transition to Apple Silicon).

  • 4 min read

Time to Stop Forecasting China's Surplus Away

Source: Time to Stop Forecasting China's Surplus Away by Brad Setser (Council on Foreign Relations)

TL;DR

The IMF has systematically underestimated China's current account surplus — now close to 5% of GDP — driven by three structural factors: the property bubble burst widens the savings-investment gap, currency depreciation pushes the surplus higher, and persistently high savings rates (>40% of GDP) show no sign of abating. Setser argues the IMF's standard policy advice (monetary easing, fiscal consolidation, exchange rate flexibility) effectively encourages China to export its way out of domestic troubles. The global spillovers are significant: Chinese exports growing faster than global trade, deindustrialised Europe, and a China with greater supply chain control able to impose policy preferences. The standard recommendation of 0.5% GDP consumption support is woefully insufficient.

  • 4 min read

Transformers Are Inherently Succinct

Source: Transformers Are Inherently Succinct

TL;DR

A new ICLR 2026 paper by Pascal Bergsträßer, Ryan Cotterell, and Anthony W. Lin proves that fixed-precision transformers are exponentially more succinct than both linear temporal logic (LTL) and recurrent neural networks (and by extension state-space models), and doubly exponentially more succinct than finite automata. This succinctness means that basic verification problems (emptiness, equivalence) for UHATs (Uniform Hard Attention Transformers) are EXPSPACE-complete — provably intractable under standard complexity assumptions. The key technical insight is that transformers can implement doubly exponentially large counters (0 to 22N) via a subtle encoding using attention. Translation bounds show UHAT to LTL now has a single exponential blow-up (improved from the prior double exponential bound). A direct consequence: analyzing transformers is computationally very challenging.

  • 5 min read

Unlimited Sensing: When Noise Becomes the Signal

Source: SIAM News — Ayush Bhandari, Imperial College London

TL;DR

Unlimited sensing is a novel approach that harnesses quantization noise from traditional digitization methods to achieve clearer sensing. Instead of clipping signals that exceed conventional sensor range, the method uses "folding" — when the signal exceeds the sensor's range, it folds back rather than being clipped, preserving information that would otherwise be lost. This allows sensors to capture signals far beyond their nominal range without hardware modifications. The technique has evolved from theoretical concept to practical applications with implications for various sensing modalities.