HNHacker News
TopNewBestAskShowJobs

kianN

476 karma · joined September 16, 2024

I like building hierarchical Bayesian models in C.

kian at sturdystatistics.com

https://github.com/ghodouss https://sturdystatistics.com/

submissionscomments
kianN··on Bayesian Data Analysis, Third edition (2013) [pdf]
This book is very relevant to those fields. There is a common choice in statistics to either stratify or aggregate your dataset.

There is an example in his book discussing efficacy trials across seven hospitals. If you stratify the data, you lose a lot of confidence, if you aggregate the data, you end up just modeling the difference between hospitals.

Hierarchical modeling allows you to split your dataset under a single unified model. This is really powerful for extracting signal for noise because you can split your dataset according to potential confounding variables eg the hospital from which the data was collected.

I am writing this on my phone so apologies for the lack of links, but in short the approach in this book is extremely relevant of medical testing.

kianN··on Bayesian Data Analysis, Third edition (2013) [pdf]
This is my favorite book on statistics. Full stop. The author Andrew Gelman created a whole new branch of Bayesian statistics with both his theoretical work on hierarchical modeling while also publishing Stan to enable practical applications of hierarchical models.

It took me about a year to work through this book on the side (including the exercises) and it provided the foundation for years of fruitful research into hierarchical Bayesian models. It’s a definitely not an introductory read, but for any looking to advance their statistical toolkit, I cannot recommend this book highly enough.

As a starting point, I’d strongly suggest the first 5 chapters for an excellent introduction to Gelman’s modeling philosophy, and then jumping around the table of contents to any topics that look interesting.

kianN··on How did sports betting become legal in the US?
It’s even worse with parlays: the events are potentially negatively correlated. There may be a 70% chance that Giannis has 3 or more turnovers and a 70% chance the bucks win, but the odds of both happening are less than 49% because more turnovers directly reduces the likelihood of a Bucks win.
kianN··on Enrollment at trade schools is expected to grow
This is a great summary of why despite so much progress/tricks being discovered, so little progress to the core limitations to LLMs are made.
kianN··on Taco Bell rethinks AI drive-through after man orders 18,000 waters
These issues are often attributed to a bad implementation of AI, but I think the problem is a little more fundamental.

The potential of AI that causes VCs and investors to swap their eyes for dollar signs is the ability to take unstructured, unpredictable inputs and convert them into structured actions or data: in this case a drive through conversation into a specific order. However, the ability to generalize to unseen inputs (what we call common sense) is neural networks glaring weakness. LLMs can look amazingly capable through internal testing, but there is a long and ever increasing tail of unseen interactions when it comes to human conversation.

I’ve seen this play out repeatedly over the last decade in the contact center industry with neural networks as a data scientist in this field.

kianN··on Corporation for Public Broadcasting ceasing operations
I think the children's programming is a really undiscussed aspect of this. Some investments don't have immediately measurable outcomes. But as someone whose parents worked long hours growing up, I'm really grateful that my exposure to television was PBS rather than cable children's shows.
kianN··on Study Mode
I built a public tool a while back for some of my friends in grad school to support this sort of deep academic research use case. Sharing in case it is helpful: https://sturdystatistics.com/deepdive?search_type=external&q...
kianN··on LLM Embeddings Explained: A Visual and Intuitive Guide
This is exactly the challenge. When embedding were first popularized in word to vec they were interpretable because the word2vec model was revealed to be a batched matrix factorization [1].

LLM embedding are so abstract and far removed from a human interpretable or statistical corollary that even as the embeddings contain more information, that information becomes less accessible to humans.

[1] https://papers.nips.cc/paper_files/paper/2014/hash/b78666971...

kianN··on LLM Inevitabilism
I put together a brief report organizing all the comments in this post. Sharing in case it is helpful to anyone else.

https://platform.sturdystatistics.com/dash/report/21584058-b...

kianN··on Why I left my tech job to work on chronic pain
Splitting up activity to bite sized chunks was critical for me as well. It prevents over straining but I think more importantly it also prevents you from being stagnant for an extended period of time.
kianN··on Why I left my tech job to work on chronic pain
“As pain becomes chronic, it is increasingly associated with activity in the affective and motivational systems tied to avoidance and less closely tied to systems encoding nociceptive input” [1]

I’ve been on the slippery slope of chronic pain. Minor post surgery issues caused me to change my routine and avoid certain activities which only exacerbated the issues, which led to more avoidance. Eventually I couldn’t walk.

The American medical system is very focused on avoiding health issues that show up on mri, rather than quality of life health. But quality of life issues quickly become serious.

I think the middle ground of activity: not all out intense as if you are healthy, but also not avoiding movement is so challenging to find for many people but also so crucial. A lot of chronic pain for myself and I suspect for many others could be avoided with short and quick combination of therapy and daily movement. So simple but so challenging to effectively identify and allocate resources.

Not suggesting this is the total solution but it’s the pathway that I took to return to activity and I’ve seen it help a number of my friends as well.

[1] https://pmc.ncbi.nlm.nih.gov/articles/PMC8482298/

kianN··on AI for Scientific Search
Linking to a nearby thread in case this is helpful: https://news.ycombinator.com/item?id=44457928
kianN··on AI for Scientific Search
I built a public literature review search tool for some graduate student friends that became pretty popular in the Santa Barbara area. It actually does exactly what you are describing.

It’s not neural network based: it leverages hierarchical mixture models to give a statistical overview of the data. It lets you build these analysis graphs via search or citation networks.

Example: https://platform.sturdystatistics.com/deepdive?search_type=e...

kianN··on Gemini CLI's Polarizing Release on Hacker News – A Visual Analysis
Author here. A local sandwich week inspired me to get a blog up and running. I was enjoying the process of writing and put together a data driven analysis of the hacker news' discussion on Gemini's CLI release. I hope it is interesting to some people out there!
kianN··on Libraries are under-used. LLMs make this problem worse
Unrelated: I initially expected this articles to be referring to public libraries. I think that would be a challenging connection to prove despite it making intuitive sense.

On the article: some use cases eg handling dates, fault tolerant queues have so many edge cases and are so mission critical that relying on a battle tested tool makes a lot of sense.

However, in my career I’ve seen a lot of examples of a package being installed to avoid 40-50 lines of well thought out code and now a dependency is forever embedded in the system.

I think there is a catch with replacing libraries with LLM generated code. Part of the benefit of skipping third party libraries is the domain knowledge that gets built up: this is potentially lost with llm generated code.

kianN··on The behavior of LLMs in hiring decisions: Systemic biases in candidate selection
> Follow-up analysis of the first experimental results revealed a marked positional bias with LLMs tending to prefer the candidate appearing first in the prompt: 63.5% selection of first candidate vs 36.5% selections of second candidate

To my eyes this ordering bias is the most glaring limitation of LLMs not only within hiring but also applications such as RAG or classification: these applications often implicitly assume that the LLMs is weighting the entire context evenly: the answers are not obviously wrong, but they are not correct because they do not take the full context into account.

The lost in the middle problem for facts retrieval is a good correlative metric, but the ability to find a fact in an arbitrary location is not the to same as the ability to evenly weight the full context

kianN··on The Academic Pipeline Stall: Why Industry Must Stand for Academia
This really hits the nail on the head. I think the extension to this is the all eggs in one (the first) basket approach being exacerbated: it’s always been challenging to conduct foundational research in a field where there already exists a dominant paradigm in industry, but it’s also becoming more and more challenging in academia as well.
kianN··on DuckDB is probably the most important geospatial software of the last decade
I think using a pyarrow dataset as an intermediary would allow for a zero copy read/write from one connection into a second connection.
kianN··on Cozy video games can quell stress and anxiety
As someone who didn’t play many video games but grew up somewhat adjacent to them, I’m just amazed at how much more relaxed I am having gone through that article (experience?) compared to before.
kianN··on Show HN: Hatchet v1 – A task orchestration platform built on Postgres
Congratulations on the v1 launch! I’ve been tinkering with hatchet for almost a year, deployed it in production about 6 months ago.

The open source support and QuickStart are excellent. The engineering work put into the system is very noticeable!

kianN··on Show HN: Textcase: A Python Library for Text Case Conversion
My favorite part of this library is that it seems to have zero dependencies!

Python packages seem to often rope in a surprising number of dependencies for relatively limited libraries.

I can easily imagine pulling this package into my work: thank you for keeping the requirements to a minimum!

kianN··on Why is this site built with C
Haha I definitely greatly benefit from autotools infrastructure to such a degree that I took it for granted in my above comment.
kianN··on Why Is This Site Built with C
I only use clang with a Makefile and I’ve done some relatively complex projects albeit with only a few collaborators.
kianN··on Why Is This Site Built with C
I’ve never been on a system that doesn’t have either gcc or clang built in.

But disclaimer that my experience in C is limited to a specific subset of scientific computing, so my experience is definitely limited

kianN··on Why is this site built with C
One of the most commonly cited drawbacks about C (limited standard library and package management) is something I’ve grown to enjoy when working in a well-loved C codebase.

Instead of learning and working with bloated tool chains, every tool in the repo has been built or added carefully for the task at hand. Simplicity provides a lot benefits over time.

kianN··on Heavy chatbot usage is correlated with loneliness and reduced socialization
The lack of a single direction is a really good point. In my mind I was thinking more from the perspective of tempering evocative headlines. But my rule also lacks quite a bit of nuance
kianN··on Heavy chatbot usage is correlated with loneliness and reduced socialization
Reasonable perspective. I went back through the article and the content is more balanced than I initially would have guessed from the initial visualizations.

However, the much more assertive initial visualizations and the opening caption— “A chart illustrates that the longer people spend with ChatGPT, the likelier they are to report feelings of loneliness and other mental health risks” — convinced me not to continue reading.

kianN··on Heavy chatbot usage is correlated with loneliness and reduced socialization
A good heuristic test of correlation vs causation in headlines: flip the dependent and independent variable and see which explanation sounds more reasonable.

Heavy llm chat usage leads to loneliness | or | Loneliness leads to heavy chat usage.

To my eye the second seems far more likely than the first.

kianN··on Ask for no, don't ask for yes (2022)
I think it’s less about putting pressure on a prospect as it is making the follow up meeting as easy as possible to schedule.

To many people, me included, it does come off a bit abrasive, but it does reduce the number of decision to make into a yes or no.

kianN··on Programming with chronic pain
I really appreciate the author posting their experience.

Just to add a little bit of my own experience. I had debilitating nerve pain as a result of surgery complications to the point where I could walk for a bit over a year.

It’s easy to fall into the trap of running from discomfort and making yourself more and more brittle in the process. At the same time it’s really easy to want to skip a few steps and try to jump back into the intensity level you are accustomed to. I think this even applies to coding. Before my surgeries, I could work for several hours straight and feel physically fine. Nowadays if I am physically static for such a long time, my nerve pain will come roaring back.

My main takeaways through all this are two relatively standard words of advice: “Movement is medicine” and that balance is key. Swimming, Pilates, yoga and build up towards more movement heavy lifestyle slowly got me to the point of walking and now being able to do 80% of what I once was able to again, but this time in a more sustainable way.

← PreviousPage 3 of 4Next →