- in our culture it’s healthy to switch jobs every two years, ideally to something the opposite of what you’ve been doing

- there’s an organization - the National Association of Campus Activities - that pays $1000 to $5000 per gig at universities

- to keep learning we need to be surprised. Read books about subjects you know nothing about. Move to a place very unlike home, so that you’ll be surprised every day.

- rich guy said that decades of pursuing more only made him miserable and he would rather have stopped at $30 million then switch to planting trees and writing poetry.

Comments URL: https://news.ycombinator.com/item?id=49441626

Points: 7

# Comments: 0

Results briefly:

- fast small model -> extraction/classification

- Gemma -> summarization

- gpt-oss -> transcript consultation

- large Qwen -> difficult reasoning/code interpretation

There wasn't a single best LLM for all tasks.

Best summarizer: I took a transcript and asked an LLM to summarize, then graded the summarization. Gemma 4 e4b was the best and took only 30s. qwen3.6-35b-a3b-ud q2_K_XL took around 2 minutes and incurred in a few omissions.

Best "answerer from files" (RAG): I added some transcript files and asked questions about facts that should be grounded on the files. gpt-oss-20b answered questions perfectly in about 4s. BTL 4 Compact and Gemma 4 e4b did too but took around a minute.

Best PII identifier: I asked to identify names of people and organizations plus PII (emails, phones, etc) in a text file. Qwen 3 4B MLX (not a thinking model) took less than 3s. Others were good too: BTL 4 Compact, Gemma 4 e4b, Qwen 3.5 9b 4-bit - but they took between 11s and 30s.

Best "code interpreter": I gave a class file and asked an LLM to identify certain methods and result values and tell whether the code would compile and run. Qwen 3.8 27B was by far the best at difficult code interpretation, but at the cost of 30 minutes (~5 tok/s). BTL 4 Compact and Gemma 4 e4b took around a minute and responded with a couple of mistakes.

I did not test code generation or any other types.

In these examples more thinking didn't result in better answers. Some models spent 45 minutes and still got something wrong.

Qwen 3.8 27B may need a reasoning budget to make it usable in 24Gb RAM (it was exhausting the context with thinking over and over until I experimented with a budget of 1k tokens).

Sometimes a Q2 Qwen 3.6 variant can beat a Q3 variant.

I would be interested in comparable test results from other machines, especially Mac Studio, AMD Strix Halo, DGX Spark, and consumer NVIDIA GPUs.

What models and workloads are working well for yall?

Comments URL: https://news.ycombinator.com/item?id=49410310

Points: 7

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=49402232

Points: 509

# Comments: 208

Comments URL: https://news.ycombinator.com/item?id=49363433

Points: 455

# Comments: 342

Comments URL: https://news.ycombinator.com/item?id=49320577

Points: 3

# Comments: 0

But the article has a big irony: the article accuses engineers of oversimplifying other disciplines and then does exactly that itself.

> "Data science was statistics with a cooler name"

The author mentions David Donoho's "50 Years of Data Science" but that work presents a much more complicated concept of data science: one with data exploration, transformation, computing, visualization, modeling, etc. Even the study of data-analysis practice itself.

Look at that XKCD he posted with it.

Horn took a nuance argument and converted it into the "human slop" formula mocked in the XKCD:

complicated subject = simple thing I already understand is really just

The cartoon is criticizing the same operation the article is doing. Maybe that is the joke?

To escape the walled garden, I've ran a script to download all posts (JS to scroll down my feed, wait for the lazy loading, click on all "more..." links, wait for the lazy loading again, everything in a loop, then download the entire thing when there are no more posts). Then ran another script to convert those to blog posts, then posted them on my personal site.

Comments URL: https://news.ycombinator.com/item?id=49150773

Points: 100

# Comments: 99

Comments URL: https://news.ycombinator.com/item?id=49135269

Points: 4

# Comments: 0

I would like to see some brutalist examples of what you find pretty (excepting eco-brutalism, since the plants hide a bit of the ugliness).

And it is common for developers to present at the city council some beautiful drawings of dwellings surrounded by trees, with couples walking together holding hands and children playing. But when the actual thing is built it ends up looking like industrial warehouses.

And the people who complain about enshitification of diverse services are the same ones complaining that residents oppose anything in their neighborhood (Nimbyism) or anywhere (Bananaism) - they fail to apply the same criteria: no one likes enshitification, even less so in their neighborhoods.

I am curious about the decision to not use GPU since this is for Apple Silicon.

Wouldn't the GPU potentially accelerate the DeltaNet/attention layers and matrix multiplication in general?

A more serious ranking would probably be dominated by these groups:

* Agricultural and public-health innovators - affecting billions of people at low cost - Louis Pasteur, Norman Borlaug, Fritz Haber and Carl Bosch, followed by sanitation engineers, vaccine developers and epidemiologists, then the myriad of scientists/engineers responsible for clean water systems

* General-purpose infrastructure creators - electricity, semiconductors, refrigeration, telecommunications, etc... - Faraday and Maxwell

* Open standards and open-source creators - Linux, Git, FFmpeg, TCP/IP, HTTP/HTML, Python, PostgreSQL, etc - they repeatedly eliminate costs for millions of organizations. Richard Hipp created SQLite that is embedded in millions of phones, browsers, apps, OSs, ... - Claude Shannon, Tim Berners-Lee, Linus Torvalds

* Institutional and conceptual inventors - double-entry bookkeeping, randomized clinical trials, peer review, container standardization, cryptography, etc - reducing transaction costs and increasing trust across the world.