Ideas worth fumbling throughhttps://hails.info/https://hails.info/writing/perception-of-probability/https://hails.info/writing/perception-of-probability/In 1951, CIA analysts couldn't agree what "serious possibility" meant — estimates ranged from 20% to 80%. The chart used to prove this for the next fifty years turns out to be broken. The point still stands. Wed, 22 Apr 2026 00:00:00 GMThttps://hails.info/writing/sharpen-your-axe/https://hails.info/writing/sharpen-your-axe/Knowledge workers compound returns by building reusable tools that remove friction, rather than optimising within existing systems. Sun, 05 Apr 2026 00:00:00 GMThttps://hails.info/writing/yardsticks/https://hails.info/writing/yardsticks/Intuitive scales for large and small numbers. A \$10 tax per webpage covers the US deficit; one career is roughly a billion seconds. Tue, 10 Mar 2026 00:00:00 GMThttps://hails.info/writing/kelly-criterion/https://hails.info/writing/kelly-criterion/The Kelly Criterion solves a question most people get wrong: how much should you bet when the odds are in your favour? Bet too much and you go bust; too little and you leave money on the table. Wed, 04 Feb 2026 00:00:00 GMThttps://hails.info/writing/picture-thousand-words/https://hails.info/writing/picture-thousand-words/The old adage is backwards. A picture doesn't need a thousand words because images are rich; it needs them because English is a terrible format for visual information. Wed, 04 Feb 2026 00:00:00 GMThttps://hails.info/writing/bias-is-a-bad-word/https://hails.info/writing/bias-is-a-bad-word/"Bias" smuggles together statistical error, cognitive shortcuts, and prejudice. Language models don't have bias; they have learned priors. Swapping the word changes the conversation from blame to engineering. Wed, 10 Dec 2025 00:00:00 GMThttps://hails.info/writing/fresh-is-fairer/https://hails.info/writing/fresh-is-fairer/Retrofitting fairness onto stale training data is futile because the data reflects the unfair world it came from. Fresh data won't make models fair, but stale data almost guarantees they won't be. Mon, 22 Sep 2025 00:00:00 GMThttps://hails.info/writing/traumagotchi/https://hails.info/writing/traumagotchi/Friend.com's AI pendant trauma-dumps on users to harvest genuine emotional responses. The product isn't the companion; it's the training data of human compassion. Wed, 12 Mar 2025 00:00:00 GMThttps://hails.info/writing/entropix/https://hails.info/writing/entropix/A per-token sampling strategy that uses entropy and var-entropy to switch between confident output, cautious backtracking, and exploratory branching. Tue, 22 Oct 2024 00:00:00 GMThttps://hails.info/writing/code-pangram/https://hails.info/writing/code-pangram/A pangram designed for code fonts, not calligraphy. Tests the characters that actually trip developers up: 0 vs O, 1 vs l vs I. Sat, 19 Oct 2024 00:00:00 GMThttps://hails.info/writing/arc-cards/https://hails.info/writing/arc-cards/Reverse-engineering Arc browser's AI features to expose the prompts, fine-tuning, and streaming tricks behind them. No arcane wizardry; just good engineering at an eyewatering price. Fri, 19 Apr 2024 00:00:00 GMThttps://hails.info/writing/codenames/https://hails.info/writing/codenames/How to name services and codebases so they age well. Descriptive names mislead as scope drifts; suggestive names hint at function without constraining it. Sun, 14 Apr 2024 00:00:00 GMThttps://hails.info/writing/mark-rober/https://hails.info/writing/mark-rober/Notes on Rober's MIT commencement: naive optimism, reframing failure, and a 50,000-person experiment showing people persist 2.5x longer when mistakes aren't penalised. Thu, 01 Jun 2023 00:00:00 GMThttps://hails.info/writing/radiance-fields/https://hails.info/writing/radiance-fields/An interactive review article on how we are moving beyond voxels through positional encoding. Sat, 08 Apr 2023 00:00:00 GMThttps://hails.info/writing/dnn-failure/https://hails.info/writing/dnn-failure/Adversarial examples aren't bugs; they're features the model found useful that humans never intended. DNNs learn texture-based shortcuts that generalise poorly and fail silently. Mon, 31 Oct 2022 00:00:00 GMThttps://hails.info/writing/gpt-at-mit/https://hails.info/writing/gpt-at-mit/GPT-3 answers MIT graduate AI course questions side-by-side with a human. It excels at coherence but struggles with depth, examples, and intellectual risk. Thu, 22 Sep 2022 00:00:00 GMThttps://hails.info/writing/steve-jobs/https://hails.info/writing/steve-jobs/Condensed notes on Jobs' 2005 Stanford commencement: connecting dots backwards, love and loss at Apple, and death as the ultimate clarifier. Sun, 12 Jun 2005 00:00:00 GMT