In detail / 01 · READING
Read sourced material first. Judge after.
Each day the reading edition picks the items worth your time from public sources: changes to AI companies, products and industries — every story carrying its source, publication time and the key numbers. Cards for scanning, full stories for depth.
- A daily editionThe day's items grouped by date; news cards tell you what happened at a glance — a restrained list, never padded for volume.
- Original sourcesEvery story names its source and publication time; key numbers are checked against the original, not paraphrased without attribution.
- Scan, then read deepNews cards carry the overview; stories worth depth are written up as short essays. Two layers, two jobs.
- Light markersOnly three markers — [background], [open question], [editor's note] — keeping facts, doubts and editorial judgment apart.
- Hand-off to researchAny story can be taken into research to keep asking: check the numbers, compare definitions, look for counterevidence.
- Bilingual editionsChinese and English versions are prepared with AI assistance from shared source material, then checked together in a separate automated review for factual and language consistency.
Latest readingAll stories ↗
- ReadingXLNet: Breaking BERT's Pre-training Limitations via Permutation Language ModelingIn the field of natural language processing, unsupervised pre-training is key to improving downstream task performance.arxiv.org
- ReadingNVIDIA Dory 4B Loop Inference Block: Predicting Costs from the GGUF Header, and a Trap with the Default Four LoopsPractitioner Sana Hassan at MarkTechPost demonstrates a complete code tutorial for cost prediction in NVIDIA Dory 1.0 4B loop inference models: reading architecture parameters from the GGUF file header to calculate parameter count and KV cache size per loop, then verifying against running server logs. The process reveals two practical traps—the GGUF built-in default loop count is 4 rather than the 2 recommended by the model card—and that the last_loop cache mode produces a measurable distribution difference compared to per_loop when processing in chunks.marktechpost.com
- ReadingWeb Perception: Three Months of Tinkering with Model Bounding Boxes Was Less Effective Than Coding My Own JudgmentAn independent developer documents the complete detours taken while building a web perception system called EH. Initially using YOLO for visual localization, manual annotation bias and resolution dependency caused bounding box jitter, requiring three months of patching. Introducing the DOM reduced annotation inconsistency, but the model still jittered; ultimately, subjective judgments about pages were decomposed into executable code rules to create a cross-site reusable engine. The article includes throughput and coefficient of variation data across six sites, five sets of window condition tests, and the author's self-defined boundaries regarding generalizability.v2ex.com
- ReadingEnvision Energy's Vertical Integration: How China's AI Data Centers Can Build 120MW in One Year at One-Fifth the US CostThe SemiAnalysis team conducted an on-site visit to Envision Group's Galaxy campus in Ulanqab, Inner Mongolia, identifying its full-chain integration from wind farms to solid-state transformers and prefabricated substations as a key differentiator from the US Abilene campus. The report argues that power access is becoming a core condition for data center development, while explicitly acknowledging whether this advantage can translate into sustained competitiveness remains to be seen.newsletter.semianalysis.com
Three markers, known before you read
Light markers come in exactly three kinds — facts, doubts and editorial judgment never mix.
- [background]Context: why this matters and what came before.
- [open question]A claim we could not verify, with the doubt spelled out — your call.
- [editor's note]The desk's judgment and reminders, kept apart from the facts.
From reading to research
- ReadScan the day's edition: news cards for the overview, essays where depth is due.
- AskQuestion a number or a claim? Take it into research — back to the source, compare definitions, look for counterevidence.
- TrackHand the judgments worth watching to the radar; it alerts only when something changes.