neehao

see this: https://arxiv.org/abs/2603.07919

see also: https://lh-zhu.github.io/The-Second-Half-of-Model-Architectu...

"Assert against the trace, not the prose."

https://github.com/gojiplus/lost-years

sample output

``` { "question": "What is the distribution of Sepal Length?", "answer": "*Sepal Length*: mean=5.84, median=5.80, std=0.83, range=[4.30, 7.90]. N=150 [non-normal distribution].", "type": "distributional", "provenance": { "generated_at": "2025-11-19T19:21:28+00:00", "tool": "statqa", "tool_version": "0.2.0", "generation_method": "template", "analysis_type": "unknown", "variables": ["sepal_length"], "python_commands": ["valid_data.mean() # Result: 5.84", "valid_data.std() # Result: 0.83"] }, "visual": { "plot_type": "histogram", "caption": "Histogram showing sepal length distribution with mean=5.84 and std=0.83 (N=150). The data shows a approximately normal distribution.", "alt_text": "Histogram chart with sepal length values on x-axis and frequency density on y-axis, showing distribution shape with 150 observations.", "visual_elements": { "chart_type": "histogram", "x_axis": "Sepal Length", "y_axis": "Density", "key_features": ["distribution shape", "mean line"], "colors": ["blue bars", "red mean line"], "annotations": ["Mean: 5.84"] }, "primary_plot": "/path/to/univariate_sepal_length.png", "generation_code": "plot_factory.plot_univariate(data['sepal_length'], sepal_length_var, 'plot.png')" }, "vars": ["sepal_length"] } ```

Output includes the precise commands run to generate the number

By the way, new enough to HN that I don't know if this a policy or not. I revised the title for clarity

see also: https://http.cat/ (posted in comments of the blog)

from a new customer perspective who just heard about midjourney, search may be a good spot to find alternate products. what google needs to know is if it is a navigational search and unless it keeps a long history, it may not know that. the simpler answer may be that companies who know you use a product like the one they make may just be willing to spend a bunch and google may be willing to add friction for the $.

see https://pubmed.ncbi.nlm.nih.gov/17420199/

As Tyler Cowen says, solve for the equilibrium.

"Many widely used machine-learning models rely on copyrighted data. For instance, Google finds the most relevant web pages for a search term by relying on a machine learning model trained on copyrighted web data. But the use of copyrighted data by machine learning models that generate content (or give answers to search queries than link to sites with the answers) poses new (reasonable) questions about fair use. By not sharing the proceeds, such systems also kill the incentives to produce original content on which they rely. For instance, if we don’t incentivize content producers, e.g., people who respond to Stack Overflow questions, the ability of these models to answer questions in new areas is likely to be lower. The concern about fair use can be addressed by training on data from content producers who have opted to share their data. The second problem is more challenging. How do you build a system that shares proceeds with content producers?"

https://www.gojiberries.io/generative-ai-and-the-market-for-...

This article covers a lot of the points: https://www.gojiberries.io/building-together-separately-chal...

"A single page on Doordash can make upward of 1000 gRPC calls (see the interview). For many engineers, upward of a thousand network calls nicely illustrate the chaos and inefficiency unleashed by microservices. Engineers implicitly diff 1000+ gRPC calls with the orders of magnitude fewer calls made by a system designed by an architect looking at the problem afresh today. A 1000+ gRPC calls also seem like a perfect recipe for blowing up latency. There are more items in the debit column. Microservices can also increase the costs of monitoring, debugging, and deployment (and hence cause greater downtime and worse performance)."

For the customer, breakfast can be expensive if they bought it from outside. I can imagine $50 for a filling breakfast for a family of 4. This is just good bundling.

three points: 1. i have often wondered about whether rapid tech. progress makes underinvestment more likely.

2. ben evans frequently makes fun of the business value. pretty clear a lot of the models are commodotized.

3. strategically, the winners are platforms where the data are. if you have data in azure, that's where you will use your models. exclusive licensing could pull people to your cloud from on prem. so some gains may go to those companies ...

I think the bigger issue is screening where people must rely on heuristics: https://www.gojiberries.io/bias-as-a-congestion-fix-heuristi...

one small thing = https://bfi.uchicago.edu/wp-content/uploads/2024/11/BFI_WP_2...

and haha on “The “rebound” in future fertility for low-fertility countries is consistent with an expectation of continued progress toward gender equality and women’s empowerment and improving social and economic opportunities for young people and families.”

fun fact total chicago pd officers ~ 12k total arrests per year ~ 48k arrest per officer per year ~ 4 Jens has a figure of 3 from 2022: https://www.theatlantic.com/podcasts/archive/2025/02/the-ori...

Just made my portfolio with it. Loved using it. Excellent control and editability. And apparently, the guy is obsessed with privacy. No keys stored for AI.

oh interesting! i couldn't easily see the connection.

Puzzles me why I am optimistic ... not optimal

agree..

appears the math was updated

agree. they seem hasty but didn't notice anything obviously wrong but may be i missed something ...

From Jan. Has a funny bit on how Gov. evals resumes.

completely! the scale and stupidity both seem compelling.

I have been thinking about this. With the advent of AI, this leads to 'congestion' at the top. And people solve it with biased decisions. More here: https://www.gojiberries.io/bias-as-a-congestion-fix-heuristi...

And I would say, often we need effortful labels by groups of humans: https://www.gojiberries.io/superhuman-level-performance/