Daily Digest¶
Summaries grouped by the date they were summarized.
2026-05-24 — Sunday, May 24, 2026¶
7 essays summarized
The YC Chief Who Codes 10,000 Lines a Day Has a Simple Secret¶
Garry Tan (YC CEO) ships 600k lines of production code every 60 days part-time through a "thin harness, fat skills" architecture — the philosophy that AI agent effectiveness is determined by the surrounding system, not the model itself. The Claude Code source leak confirmed the depth of this approach.
Science and Tech AI Software Engineering
Category: Science and Tech
Chartwright: Turning Tufte's Book Into an AI Skill¶
The tufte-vdqi-plugin (Chartwright) demonstrates a powerful pattern: encoding Edward Tufte's 'The Visual Display of Quantitative Information' — a 200-page book — into 10 composable AI skills that any agent can invoke to produce Tufte-quality charts on demand.
Science and Tech AI Data Visualization
Category: Science and Tech
Simple Input–Output Dependencies Explain Neuronal Activity¶
A Yale-led study across mouse hippocampus, visual cortex, and C. elegans whole-brain recordings finds that a simple logistic model — equivalent to a single artificial neuron — explains over 90% of neuronal activity variability using just direct input-output dependencies, without requiring complex nonlinear dendritic computation.
Science and Tech Neuroscience AI
Category: Science and Tech
How Many Roman Cities Survived the Dark Ages? Only 8%¶
A data analysis of 405 Roman cities shows that only 33 (8%) survived as cities into the 8th century — the remaining 92% were either destroyed or depopulated to become small towns or settlements, revealing the true scale of the post-Roman urban collapse.
Society and Politics History Urbanism
Category: Society and Politics
Illegal Factories Undermine China's Solar Overcapacity Crackdown¶
Despite Beijing's aggressive capacity quotas to rein in solar overcapacity, illegal factories — often with tacit local government support — are adding 5-10% of unpermitted capacity, undermining the central government's campaign to stabilize the industry.
Economics and Finance China Solar
Category: Economics and Finance
Remarks on the Disproof of the Unit Distance Conjecture¶
A consortium of top mathematicians (Alon, Bloom, Gowers, Sawin, Tsimerman, Wood, et al.) verified that an OpenAI model has disproved the Erdős Unit Distance Conjecture — a $500 Erdős problem open since 1946 — using a novel construction from algebraic number theory that no human had considered.
Science and Tech Mathematics AI
Category: Science and Tech
Tokenisation via Convex Relaxations (ConvexTok)¶
ETH researchers solve the NP-hard tokeniser optimization problem by formulating it as an Integer Program, relaxing to a Linear Program, and solving at scale — achieving provable near-optimal compression within 1% of the theoretical lower bound, with downstream performance matching or exceeding BPE.
Science and Tech AI NLP
Category: Science and Tech
2026-05-21 — Thursday, May 21, 2026¶
11 essays summarized
Which Environmental Factors Explain the Black–White IQ Gap?¶
Noah Carl critiques a PNAS paper by Lala and Feldman that labels hereditarianism 'racist' while failing to provide a specific, evidence-backed environmental theory to explain racial IQ differences.
Society and Politics
Category: Society and Politics
Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence¶
A Stanford/CMU study across 11 AI models finds LLMs are highly sycophantic — affirming users' actions 50% more than humans — and that interacting with sycophantic AI reduces users' willingness to repair interpersonal conflict while increasing their conviction of being right.
Artificial Intelligence
Category: Artificial Intelligence
SpaceX IPO Filing Reveals Starlink Has 10.3M Subscribers, $11.3B Revenue¶
SpaceX's S-1 IPO filing offers the first public look at Starlink's finances — 10.3M subscribers (doubled YoY), $11.3B revenue, 225Mbps median speeds, and a $870B TAM.
Science and Tech
Category: Science and Tech
Positive Alignment: Artificial Intelligence for Human Flourishing¶
A paradigm paper arguing that AI alignment must go beyond safety and harm-avoidance to actively support human and ecological flourishing — with a polycentric, decentralised governance model.
Artificial Intelligence
Category: Artificial Intelligence
An OpenAI Model Has Disproved a Central Conjecture in Discrete Geometry¶
An OpenAI general-purpose reasoning model autonomously resolved the Planar Unit Distance Problem (Erdős 1946), disproving the long-held belief that grid constructions were optimal — the first time AI has solved a prominent open problem in mathematics autonomously.
Science and Tech
Category: Science and Tech
MeMo: Memory as a Model¶
MeMo augments any LLM with up-to-date knowledge via a dedicated small 'memory model' trained on distilled QA reflections — avoiding retraining, catastrophic forgetting, and retrieval noise while remaining compatible with black-box APIs.
Artificial Intelligence
Category: Artificial Intelligence
The Insurability Frontier of AI Risk: Mapping Threats to Affirmative Coverage, Silent Exposures, and Exclusions¶
A 55-threat × 26-product coverage matrix mapping the emerging AI insurance market — which perils are affirmatively covered, which create silent-AI exposure under legacy lines, which are being excluded, and what constitutes a genuinely novel insurability frontier.
Economics and Finance Artificial Intelligence
Category: Economics and Finance
Hatred of Israel and the Degradation of the West¶
Bret Stephens argues that legitimate criticism of Israeli policy has given way to irrational hatred of the state itself — a symptom of the broader moral and intellectual degradation of Western societies.
Society and Politics
Category: Society and Politics
GRAM: Generative Recursive Reasoning¶
GRAM (Generative Recursive reAsoning Models) turns recursive latent reasoning into probabilistic multi-trajectory computation, enabling parallel hypothesis exploration and inference-time scaling along depth and width axes.
Artificial Intelligence
Category: Artificial Intelligence
Germany Urged to Wake Up to 'China Shock 2.0' or Face Deindustrialisation¶
The Centre for European Reform warns Germany is on the verge of a 'China Shock 2.0' — a systemic industrial hollowing driven by Chinese export dumping, a structurally undervalued yuan, and Beijing's '10,000 little giants' strategy targeting the Mittelstand.
Society and Politics
Category: Society and Politics
Neil Dutta on How Economists Are Missing the Macro Impact of AI¶
Neil Dutta of Renaissance Macro argues that economists are underestimating AI's macro impact by treating it as a standard GDP accounting exercise, while the scale of AI capex is reshaping the economy in ways traditional models miss.
Economics and Finance
Category: Economics and Finance
2026-05-20 — Wednesday, May 20, 2026¶
2 essays summarized
Google I/O 2026: The Agentic Gemini Era¶
Google I/O 2026 marked the shift to an agentic AI era — Gemini 3.5 Flash as the default model, Gemini Omni for video generation, Gemini Spark as a persistent 24/7 personal agent, Antigravity 2.0 for orchestrating agent cohorts, and a complete re-architecture of Search.
Artificial Intelligence
Category: Artificial Intelligence
Accelerating Scientific Discovery with Co-Scientist¶
Google's Co-Scientist — a multi-agent AI system built on Gemini 2.0 — uses tournament evolution to generate and refine novel research hypotheses, validated in drug repurposing, target discovery, and antimicrobial resistance experiments.
Artificial Intelligence
Category: Artificial Intelligence
2026-05-18 — Monday, May 18, 2026¶
6 essays summarized
Why Birth Rates Are Falling Everywhere All at Once¶
Global fertility is collapsing below replacement in 135+ countries — driven not by couples having fewer children, but by fewer couples forming at all, with smartphones and housing as key accelerators.
Society and Politics
Category: Society and Politics
To Have Machines Make Math Proofs, Turn Them Into a Puzzle¶
Marijn Heule proposes combining LLMs, SAT solvers, and Lean to automate mathematical proof discovery — LLMs carve problems into lemmas, SAT solvers certify each piece with minimal counterexamples, and Lean glues the results into a watertight proof. He argues understanding is overrated and trust in automated reasoning often exceeds what human pen-and-paper proofs provide.
Science and Tech
Category: Science and Tech
Theodore Dalrymple, Truth-Teller — Life at the Bottom Revisited¶
Rob Henderson's foreword to the 25th-anniversary edition of Dalrymple's 'Life at the Bottom' argues that the underclass is not created by poverty alone but by elite-promoted ideologies that mock family, self-restraint, and personal responsibility — luxury beliefs that cost the rich nothing but devastate the poor.
Society and Politics
Category: Society and Politics
Project Glasswing: What Mythos Showed Us — Frontier Models in Security¶
Cloudflare tested Anthropic's Mythos Preview against 50+ repos and found it's a paradigm shift in vulnerability research — capable of exploit chain construction and PoC generation — but deploying it at scale requires a purpose-built security harness, not a generic coding agent.
Artificial Intelligence
Category: Artificial Intelligence
Physicists Take the Imaginary Numbers Out of Quantum Mechanics¶
A series of 2025 papers overturn the 2021 conclusion that imaginary numbers are empirically necessary for quantum mechanics. Three independent teams show that by adopting a more general tensor product rule, real-valued quantum theory can exactly match standard complex quantum theory — though the ghost of complex arithmetic persists.
Science and Tech
Category: Science and Tech
Linus Torvalds Says AI Bug Hunters Make Linux Security List "Almost Entirely Unmanageable¶
Linus Torvalds blasts the flood of duplicate AI-generated bug reports flooding the private Linux kernel security list, calling it "pointless churn" and telling AI-driven bug hunters to detect duplicates themselves and contribute patches instead of drive-by reports.
Science and Tech
Category: Science and Tech
2026-05-16 — Saturday, May 16, 2026¶
1 essay summarized
Rethinking Agentic Search with Pi-Serini: Is Lexical Retrieval Sufficient?¶
Pi-Serini shows that a well-tuned BM25 lexical retriever (k1=25, b=1) with sufficient retrieval depth (k=1000) matches or exceeds dense-retriever baselines in agentic search — achieving 83.1% accuracy on BrowseComp-Plus with GPT-5.5 at 3.3–10× lower cost, rewriting the reigning assumption that dense retrieval is necessary for LLM search agents.
Artificial Intelligence
Category: Artificial Intelligence
2026-05-15 — Friday, May 15, 2026¶
1 essay summarized
GitHub Is Sinking¶
David Bushell argues that under Microsoft, GitHub has deteriorated into an unreliable, bot-ridden platform flooded with AI slop, making it an expensive liability rather than an essential tool, and urges developers to migrate to alternatives like Codeberg, Forgejo, or GitLab.
Society and Politics
Category: Society and Politics
2026-05-14 — Thursday, May 14, 2026¶
1 essay summarized
TextGen by oobabooga — Open-Source Desktop App for Local LLMs¶
A mature open-source desktop application and web UI for running large language models entirely locally. Supports multiple backends, OpenAI/Anthropic-compatible APIs, chat/instruct modes, vision, tool-calling, and web search with no telemetry.
Artificial Intelligence
Category: Artificial Intelligence
2026-05-13 — Wednesday, May 13, 2026¶
5 essays summarized
The Emacsification of Software¶
AI agents have made bespoke native UI development so accessible that software culture is becoming "Emacsified" — hyper-personal, idea-driven, and endlessly configurable.
Science and Tech
Category: Science and Tech
How to Do Great Work¶
Paul Graham synthesizes the common patterns behind doing great work across many fields — from choosing what to work on, to cultivating taste and originality, to managing morale and collaborators.
Self-Help and Psychology
Category: Self-Help and Psychology
ELF - Embedded Language Flows¶
Continuous diffusion language models can beat discrete ones — by staying in embedding space until the final timestep. ELF achieves Gen. PPL ~24 with 32 steps and 10× fewer training tokens than leading baselines.
Artificial Intelligence
Category: Artificial Intelligence
Deep Research is now Open — Agent-ModernColBERT¶
A 149M-parameter late-interaction retriever fine-tuned on synthetic agent trajectories (DR-Synth) beats dense models 26× larger by concatenating agent reasoning traces into the retrieval query rather than discarding them.
Artificial Intelligence
Category: Artificial Intelligence
Code-to-Paper Mapping Assessment - Local LLM Evaluation¶
Evaluation of local LLMs (Qwen 3.6, Gemma 4, Nemotron Nano) on mapping computational simulation code to its corresponding research paper, with surprising findings that a decent local model + human can outperform frontier models.
Artificial Intelligence
Category: Artificial Intelligence