Jisoo, Wookyoung Kim · September 15, 2025
Cash and poverty, how to get better at work immediately, Cluely's marketing
Saul Steinberg (1914-1999) was a Romanian-born American illustrator and one of The New Yorker magazine's defining artists. Guided by the idea that "the artist's role is to ask questions, not to give answers," he saw drawing as "a way of reasoning on paper" and built a singular body of work that deftly blended reality and fantasy.
Cash Alone Cannot Change Chronic Poverty
Over the past 13 years, several U.S. cities have run randomized controlled trials (RCTs) of basic-income-style cash support, giving low-income residents several thousand dollars per month. The goal was to test whether cash transfers, which have proven effective in low-income countries, would also work against chronic poverty in advanced economies.
The results were not what many expected. Core indicators such as health, stress, employment, and child development did not improve. Working hours fell, but that did not translate into higher life satisfaction or better mental health.
Cash helps in sudden crises such as disasters, pregnancy, domestic violence, or the period immediately after release from prison. But the trials showed that it has limited effect on the structural institutions and infrastructure that address chronic poverty at the root: education, health care, housing, and the like.
One interesting finding was that, contrary to expectations, almost none of the money went to alcohol, gambling, or cigarettes. Instead, spending rose on practical and healthy needs such as housing, transportation, childcare, and debt repayment. Yet even this rational consumption did not lead to long-term improvements in key indicators.
Using CSV Format to Make an LLM Inspect Every Row of Data
From cameronwestladn.com:
And then everything clicked. What if the problem wasn't the model's capability, but the format? CSV. Spreadsheets. The most boring format. But the key thing about CSV is this: you cannot skip a row. Row 17 has to exist between row 16 and row 18. The model has to fill every cell, or leave it visibly blank. issue_number,title,priority_recommendation,assignee_recommendation,rationale 3847,"Fix auth timeout","","","" 3848,"Update docs","","","" Empty cells crying out to be filled. The model could not skip them, because skipping them was immediately obvious. Suddenly, 100% coverage. Every. Single. Issue. We went from ~88% coverage to 100%, not through better prompting, but through better structure.
Cameron W. was building a Claude Code slash command that brought in development issues from GitHub and had AI assign priorities. But even when it pulled 25 development issues from GitHub, the AI would skip some of them instead of reviewing all 25. Cameron tried improving the prompt and formatting the data in Markdown, but neither helped much.
Eventually, Cameron solved the problem not by improving the prompt but by specifying CSV as the data format. Because CSV clearly separates each row, Claude Code reviewed every row. And even if it did not review one, the blank row made that omission easy to detect.
It makes me think that many paths for expanding generative AI may be hidden not in better model performance, but in better structures such as data formats.
How to Get Better at Work Immediately
From usefulfictions:
It's a funny quirk of the human condition that sometimes simply asking, of a given task, "how would someone much, much better than me approach this?" immediately makes you better at it. Like, right away.
Does the same idea apply to real people, much like telling an LLM, "You are an expert in field X"?
Cate Hall's argument suggests that asking, "How would someone better than me approach this problem?" can meaningfully improve practical work ability.
This resembles Astro Teller's idea that 10X is easier than 10%. Improving something by 10X can sometimes be easier than improving it by 10%. Shifting perspectives, from 10% improvement to 10X improvement to even 100X improvement, is one way to get better at work.
Naver Webtoon's Cuts Reinterprets Short-Form as an Entry Point for Conversion
Naver Webtoon has joined the short-form competition by adding Cuts to its app. Unlike other platforms, however, the core goal of Cuts is not time spent but conversion into works. Left-right swiping is fixed as a way to continue into a series, and a Related Webtoon button at the bottom nudges users to enter immediately. In other words, it redefines short-form content not as simple browsing, but as an aid for choosing a work. Unlike YouTube Shorts or Instagram Reels, this removes the click and search steps between interest and reading the work, greatly reducing drop-off.
What is interesting is that before launching Cuts, Naver Webtoon had already tested a similar approach in Instagram ads. I have often entered webtoons through those ads myself, so Cuts seems likely to work well as an entry path into individual works.
GPT-5 Makes People Use Reasoning Models More
People using ChatGPT had mostly used 4o, which responds quickly, rather than o3, the stronger reasoning model.
Because the number in 4o is higher than in o3, it is easy to assume it is the better model (4 > 3). And because 4o answers much faster than o3, users preferred 4o. According to OpenAI, fewer than 7% of paid ChatGPT users used o3, a reasoning model. Even though it could deliver a better product experience, the fact that users were not choosing it was probably a major headache for OpenAI.
To solve this problem, OpenAI added an Auto model as the default model in GPT-5. Even when users do not explicitly select a reasoning model, the system routes easy questions to GPT-Instant and complex questions that require reasoning to GPT-5 Thinking, allowing users to get the best product experience.
OpenAI's approach is impressive because it improved the UX by removing the need to choose, while still delivering the best product experience.
Why Does Cluely Focus Only on Marketing?
AI models such as OpenAI's GPT and Anthropic's Claude are improving by 50% every six months. In that environment, AI wrapper services are also improving by 50% every six months.
Given that situation, what should AI wrapper services prioritize: spending on technology development and R&D to make the product 10% better, or using marketing to acquire 10% more new users?
Cluely invests in marketing even if churn exceeds 90%, because AI technology will always improve and the company has to survive until then. AI will improve by 50% every six months, and at least in the field of AI sales, Cluely will have a uniquely large user base and product. That is why Cluely invests in marketing. Even if everyone laughs, Cluely focuses on marketing so it can capture the market ahead of a better future.
Thoughts
I need to intentionally think, every time, about what I should work on to make things go better.
People are skilled at buying things, but often poor at getting rid of them.
A difficult task with the potential for major impact can make a team more deeply engaged than a small, easy one.