Isegoria

Jeff Bezos has $222 billion. If he paid my wealth tax this year, we could fund insulin in America for everyone who needs it plus free school lunch for every kid in Texas—and have plenty of money left over. And Bezos would still have $215 billion dollars to spare.

In April of 2023, I was getting a third of my Substack traffic from Twitter. By August it was 1-2%. A quick Google search reveals that Elons first major algorithm tweak was in the Spring 2023. I saw the impact of that in real time.

Before the 1750s, no European power had much in the way of Indian colonies—just a few scattered trade stations, dependent on the good graces of native princes. That began to change with the Second Carnatic War, the most fascinating and complex of Franco-British wars in India.🧵

🚨 BREAKING: Duke researchers just proved that coding agents are better at processing long documents than models with million-token context windows. > Not because of longer context. Because grep and sed are better retrieval tools than attention. > +17.3% average improvement over state-of-the-art across five benchmarks spanning 188K to 3 trillion tokens. > The setup is almost insultingly simple. Instead of feeding documents into a context window or building a retrieval pipeline, Duke placed text corpora into directory structures and handed them to off-the-shelf coding agents. > No task-specific training. No architectural modifications. No special prompting beyond the file path and the question. The agents used terminal commands, wrote Python scripts, and figured out the rest themselves. The results beat every published baseline across four out of five benchmarks. > The key insight is about what coding agents already know. These models trained on millions of code repositories hierarchical file structures, grep patterns, sed commands, iterative script refinement. That training wasn't for document processing. But it transfers directly. When you organize text as a directory of files instead of a flat context window, coding agents navigate it the same way they navigate a codebase. The inductive prior already exists. Nobody had thought to use it this way. > The emergent behavior is the most striking finding. Nobody told the agents how to process the documents. On multi-hop retrieval tasks, they autonomously developed iterative query refinement starting with an initial search, extracting entities from the results, using those entities to inform the next search, chaining six reasoning hops without a single instruction to do so. On analytical tasks requiring aggregation across thousands of data points, they abandoned search entirely and wrote custom Python classifiers. On standard reading comprehension, they used a hybrid of targeted searches and focused reading. Three completely different strategies, all emerging without prompting, matched to the task structure. The benchmark numbers across all five evaluations: → BrowseComp-Plus (750M token corpus): 88.5% vs 80.0% best published +11% relative → Oolong-Synthetic (536K tokens): 71.75 vs 64.38 best published +11% relative → Oolong-Real (385K tokens): 37.46 vs 24.09 best published +56% relative → LongBench (188K tokens): 62.5% vs 63.3% best published competitive → Natural Questions (3 trillion token corpus): 56.0% vs 50.9% best published +10% relative → Average improvement over state-of-the-art: +17.3% → Gains hold across both Codex and Claude Code as base agents → No task-specific training, no fine-tuning, no specialized architecture > The counterintuitive finding about retrieval tools is important. When the researchers added BM25 or dense embedding retrieval on top of the coding agents, performance did not consistently improve and sometimes degraded. The explanation is behavioral. Without a retriever, agents issue roughly 15 native search commands per query. With a retriever, that drops to 8 to 9. The retriever becomes the default discovery mechanism and displaces the broader filesystem exploration that agents do autonomously. Since retrieval ranking is imperfect, the substitution causes the agent to miss relevant context it would have found on its own. > The tool intended to help becomes a ceiling. > The file system structure matters more than it sounds. When documents are organized as individual files in a directory hierarchy mirroring a code repository agents use sed to extract specific line ranges and index content by file and line number. When the same documents are stored as a single JSON file, agents fall into repeated corpus-wide scans using ripgrep. > The folder structure gives agents a coordinate system. The flat file removes it. Performance gap: 6 percentage points on the same benchmark with the same agent. > Context windows are scaling in the wrong direction. The bottleneck was never how many tokens a model can see. It was always how well the model can navigate what it's looking at. > Coding agents figured out a better solution. Nobody asked them to.

thoughts on mental health that would have you like this?

i have been visiting new york city since the mid-80s pretty regularly except for a break btwn 91 and 2000. the bloomberg era was the best imo but you can't compare the city today to the constant fear and chaos prevalent in the 80s. i was there

MORE: The limitations of available Russian air defenses and the challenges inherent in protecting large infrastructure facilities across thousands of kilometers of distance are hindering the Kremlin’s efforts to defend against Ukrainian long-range strikes. The Ukrainian strikes overnight on April 4 to 5 are part of a series of Ukrainian long-range strikes against eight separate Russian oil and defense industrial infrastructure targets in the previous 13 days (since the night of March 22 to 23), including oil terminals at Ust-Luga and Primorsk, Leningrad Oblast, oil refineries in Kstovo, Nizhny Novgorod Oblast, Kirishi, Leningrad Oblast, Yaroslavl, Yaroslavl Oblast, Ufa, Republic of Bashkortostan, and defense plants in Tolyatti and Chapayevsk, Samara Oblast. The targets that Ukrainian forces have struck span over 1,700 kilometers from Primorsk and Ust-Luga in Leningrad Oblast to Ufa in the Republic of Bashkortostan. Ukrainian forces have struck some of these targets multiple times during this period, but the geographic dispersion and large sizes of the facilities likely hinder Russian air defense efforts.

The only adequate response to this idiocy is to request foreign allies to halt their regular airstrikes on Iran's oil infrastructure amid a surge in global oil and fuel prices driven by the war with Russia.

According to official figures, the TFR of Taiwan in 2025 turned out to be 0.695 children per woman.