clusterhacks

I never looked at short videos and couldn't understand how my friends and family would open their phones given even a minute of quiet time for these things. I viewed some Youtube shorts (maybe the least effective short video provider in terms of content and the recommendation algorithm?) and was shocked at how easy it was to burn time looking at crap. The experience really opened my eyes about how a person can be pulled into endless viewing.

I think the crack house comparison is entirely appropriate. The brain is weird . . .

That's interesting. While I suspect the pricing will lean heavily into enterprise sales rather than personal licenses, I personally like the idea buying models that I then own and control. Any steps from companies that make that more possible is great.

I wish I could just pay for the model and self-host on local/rented hardware. I'm incredibly suspicious of companies totally trying to capture us with these tools.

I was a light or social drinker for decades. Probably 3-5 drinks per week.

In November of 2024, I decided to avoid alcohol as a personal experiment - no GLP-1 medications involved. I have not consumed any alcohol since.

After 3-4 months, my interest in alcohol seemed to really fall off a cliff. I joked with friends that I was going "dry in 2025", but I am now more seriously considering taking 2026 off from alcohol as well before making a decision about whether to add alcohol back into my diet.

> Even at the risk of false negatives, I would actively filter out resumes listing schools like these.

I am occasionally on hiring committees and use a rubric for ranking candidates. The rubric usually has 8-10 yes/no questions that might be best summarized as "Does this applicant's resume and cover letter indicate that they have actually written code deep enough to 'map' to our requirements?" Some of the rubric may be a little more specific to the actual job role, but the main idea is to filter out what I have come to think of as "aspirational" software developers.

I think one nice thing about the rubric approach is that candidates don't score "prestige" points or get "penalty" points for their specific educational background. Honestly, it seems like a lot of students from many institutions (some quite well known for rigor) are mostly about the "credential, not for the learning." The rubric seems to effectively filter out the less skilled or interested without eliminating skilled candidates with less "sterling" credentials.

> tuition paid by students in the social sciences is effectively subsidizing the STEM fields

This is not true at my state flagship R1 institution. Tuition and fees make up a little over 10% of the institution's total revenue. General funding provided in our state budget provides a larger percentage of the total revenue to the university and federal research funding provides an even larger percentage than the state.

The essential takeaway here is that our state taxes subsidize the actual cost of providing education to in-state students. In-state students are mandated to be at least 80%-ish of students.

The professors in the STEM fields are required to raise a significant percentage of their salary via research grants ("soft" money), teaching, and service work. The non-STEM professors are more often funded via "hard" money - eg, the institution has committed to pay the salary of history professors.

I googled and apparently a little more than 70% of undergraduate students in the US attend public schools. I don't know much about how funding works at the private universities that have the other 30% of undergraduate students.

> This isnt science, or engineering. > This is voodoo.

I was struggling to find the exact reason this type of article bugs me so much, and I think "voodoo" is precisely the correct phrase to sum up my feelings.

I don't mean that as a judgement on the utility of LLMs or that reading about what different users have tried out to increase that utility isn't valuable. But if someone asked me how to most effectively get started with coding agents, my instinct is to answer (a) carefully and (b) probably every approach works somewhat.

"Out-of-order" and non-conventional explanations are interesting to consider. For both, I would expect LLMs to do poorly when there isn't much (or any) material for those approaches in the training data. My intuition would be the learner is going to have to be more exploratory via prompt engineering and still struggle against the tendency of the model to lean into classic or conventional explanations.

I don't particularly expect models to be dependable in responses, but I see how that presents a much larger problem in a learning context. I'm ok with bad responses that I can fight back against, but I also wouldn't reach for an LLM for a new field by default either.

For me, I do like using an LLM as a supplemental learning aid along with other traditional resources. I haven't tackled a deeper, new-to-me field yet with one. Maybe it's time for that . . .

Isn't this just showing the effects of actively engaging the learner by placing a topic in contexts familiar/favored by learner versus just reading about a topic?

Like, if you had made the text pdf readers do some manual thinking by working on trying to place the topic into the same type of familiar/favored context, wouldn't that have been the better comparison?

I think using GenAI for learning is cool and exciting (especially for autodidacts) but I'm not excited by this particular study structure.

I think it is just poorly worded. From another point in the paper:

"Our analysis draws on a new dataset that combines LinkedIn resume and job-posting data from Revelio Labs. The dataset covers nearly 285,000 U.S. firms, more than 150 million employment spells from roughly 62 million unique workers between 2015 and 2025, and over 245 million job postings."

I guess we can read that as saying the authors identified 62 million workers who held 150 million positions over the 2015-2025 time window.

I'm still deeply skeptical about the underlying data. The 62 million represents a huge percentage of employed people in the U.S. in any of the years 2015-2025. This source shows 148 million/yr to 164 million/yr employed over that timeframe:

  https://www.statista.com/statistics/269959/employment-in-the-united-states/

On the other hand, I also saw estimates saying LinkedIn has approximately 30% of the U.S. workforce with a profile on the platform. Which is wild to me.

I am confused about how to feel about the data the paper is based on. If you look at the paper, the data description is:

"Our primary data source is a detailed LinkedIn-based resume dataset provided by Revelio Labs ...

We complement the worker resume data with Revelio’s database of job postings, which tracks recruitment activity by the firms since 2021 ...

The final sample consists of 284,974 U.S. firms that were successfully matched to both employee position data and job postings and that were actively hiring between January 2021 and March 2025.3 For these firms, we observe 156,765,776 positions dating back to 2015 and 245,838,118 job postings since 2021, of which 198,773,384 successfully matched with their raw text description."

They identified 245 million job postings from 2021 forward in the United States? I mean the U.S. population is like 236 million for the 18-65 age group (based on wikipedia, 64.9% of 342 total population).

And they find a very small percentage of firms using generative AI:

"Our approach allows us to capture firms that have actively begun integrating generative AI into their operations. By this measure, 10,599 firms, about 3.7 percent of our sample, adopted generative AI during the study period."

Maybe I am wildly underestimating just how much LinkedIn is used worldwide for recruiting? As a tech person, I'm also very used to seeing the same job listing re-listed by what seems to be a large number of low-effort "recruiting" firms on LinkedIn.

I think for trying to figure out how generative AI is affecting entry-level jobs, I would have been much more interested in some case studies. Something like find three to five companies (larger than startups? 100+ employees? 500+?) that have decided to hire fewer entry-level employees by adding generative AI into their work as a matter of policy. Then maybe circling back from the case studies to this larger LinkedIn dataset and tied the case study information into the LinkedIn data somehow.

We have always been in-person or hybrid with staff mostly very local. Openings tend to pop-up on indeed, linkedin, and probably other job aggregator boards.

Just to add some hope and a different perspective, we received 23 applicants for an entry-level or early-career software developer position when it was open for a couple of months in early 2025. This is about the same number of applicants we usually get for an opening.

Applicant count for similar positions by year:

  23 - 2025 (the position I mentioned)
  31 - 2025
  10 - 2019

The above are three jobs where I was on the hiring committee and are relatively recent. My organization is relatively well-known but also pays a little bit below market in general.

I do think the market is very rough right now for software developers. I also know for a fact that "attractive" hiring companies can get a crazy number of applicants for each opening. SAS was famous for getting 1,000+ applicants per job just after the dotcom bust in the early 2000's.

Yes. I recently used the online calculator and it quoted me $4200/yr for $250K of coverage.

I live in a "zone X" neighborhood. Interestingly, it appears that significant areas around me are all "zone X" but personal experience and common sense show vastly different flood risks. Meaning there are lots built on the edges of "100 year" flood areas in "zone X" and properties 50-100 feet higher in elevation in the same neighborhood. Surely the risk would actually be quite different? I guess the models aren't particularly fine-grained.

You can play around with the calculator at: https://www.floodsmart.gov/policy-quote/

There are a bunch of available spreadsheets with costs by state/county that show some stats on those costs but I don't have the link handy.

We recently had an open position that could have been a fit for a new grad with some internship experience or a solid portfolio of projects from just school.

We had just over 20 applicants and selected 6 for interviews using a rubric to score candidates by reviewing CV/resume and cover letter. The rubric doesn't include a specific "years of experience" guideline anywhere. Of our 6 selected candidates, 3 were new grads and 3 were experienced candidates.

At that point in our process, 1 of our 3 new grads ghosted us and never responded to multiple contact attempts to schedule an interview - this may have been our top candidate by initial rubric evaluation. One of our other new grad candidates accepted another job and backed out of our scheduled interview. The third new grad did interview and had an unremarkable interview.

Ultimately, one of the experienced candidates received a job offer. This experienced candidate was one of the top two by initial rubric review along with the new grad who ghosted us. Funnily enough, these two candidates were the only unanimous selections by everyone on the hiring committee to interview, even using a rubric.

There were several parts of the process I found interesting. First, just over 20 applicants is typical for applicant count for an opened position over the last 8-10 years for us. There wasn't a noticeable uptick in applicants despite the general perception of it being a "bad hiring market." Second, we continue to find that many candidates don't have a CV/resume or cover letter that leans into demonstrating the candidate can write code. Maybe this is a result of general hiring processes that lean into automatic scanning and processing of CV/resumes and cover letters while my group actually reads these somewhat carefully and usually from a "we hope this person clearly demonstrates we should interview them" perspective. Third, despite my personal misgivings, the hiring process clearly selects for "has worked with <insert specific set of technical tools> experience. Even a great candidate can't seem to be attractive to other folks in our hiring process if their past experience doesn't include <specific tool/framework> use. Fourth, and this was a change different from past hiring, 4-5 of the applicants were ML/AI-focused candidates with MS degrees (and one PhD). This position is/was clearly just a normal app developer job in an organization that doesn't ship software for sale - we do small-scale app dev and data management for our internal organization users.

These ML/AI-focused candidates were interesting for other reasons. Almost none of them demonstrated any experience on paper besides "wrote a little python to load some ML/AI tool and ran experiments." This isn't bad experience on the face of it, but their CV/resumes and cover letters didn't talk about software dev at all. To a candidate, it was almost all "ran and tweaked model for small gain in performance" for some data set. I'm not sure where these folks will land in the current hiring market?

I can find no mention of a TPMS requirement in my state. My older car has a malfunction in its TPMS and has for 10+ years. The repair cost I was quoted was ridiculous and I just periodically check the tire pressure with a normal gauge.

This vehicle has passed all required inspections for that entire time and the TPMS system hasn't even been mentioned in that context.

edit: I found the claim that TPMS was required interesting and a quick web search shows that 13 states in the US don't require vehicle safety inspections AT ALL.

I have seen the argument that very little code is novel but I find it inherently unsatisfying and lacking in nuance? I think what bugs me about is that if you squint hard enough, all programming reduces to "take some data, do something to it." That "something" is doing a lot of heavy lifting in the argument that "something" is or isn't novel.

Heck, if we think about it from the programming language perspective, all code is "simply" using already existing language functions to cobble together a solution to some specific set of requirements. Is no program novel?

There is probably a consideration here that maybe boils down to the idea of engineering vs artisanal craftsmanship and where a specific project falls in that spectrum . . .

I am a slow adopter of new tech but plan to spend a significant amount of time in 2025 using AI tools when coding. I am net negative on AI simply replacing programmers, but I think the practice of development is undergoing a seismic shift at this point.

My recent usage is oriented towards using pseudocode descriptions that closely map to Python to produce Python functions. I am very impressed with Claude 3.7's syntactic correctness when given a chunk of pseudocode that looks "python-y" to begin with.

My one concern is that much of my recent code requirements lack novelty. So there is a somewhat reasonable chance that the tool is just spitting out code it slurped somewhere in github or elsewhere in the larger Internet. Just this week, I gave Claude a relatively "anonymous" function in pseudocode, meaning variable names were not particularly descriptive with one tiny exception. However, Claude generated a situationally appropriate comment as part of the function definition. This was . . . surprising to me if somehow the model had NOT in its training set had some very close match to my pseudocode description that included enough context to add the comment.

Maybe that perception that people chronically job hop is not true? I would welcome a data source that shows chronic job hopping is the norm.

Anecdotally speaking, I spent 6+ish years at my first "old tech" company. My org in this (very large) company was constantly simmering with resource actions (mostly small scale layoffs). There was lots of negative energy there. I left to take a programming/data analytics gig in a large, privately held financial company that had never had a layoff. This was unfortunately timed, 2 years later the financial crisis of 2008 kicked off and I (and my entire team) were all laid off. I have been with my current employer for approaching two decades and several of my peers have been on our team for that much time.

I have had a SINGLE manager in that time window. I can't even name all the managers I had in my first job due to near-constant re-orgs and layoffs.

Similar experience.

I now think of Clojure as a power tool for small shops and consultants. These aren't the kind of places that typically cast a wide net looking for "employees." Most of the opportunities are maybe more focused on proving that you can bring in business and network effectively enough to be a known quantity in the Clojure community.

BLS figures show tradeworkers make about the median salary in the US - it isn't above average. In the Occupational Outlook handbook, The mean annual wage for all occupations is $65,470. The trades:

  Plumber - median $61,550 per year
  HVAC tech - median $57,300 per year
  Electrician - median $61,590 per year
  Framer(Carpenter) - median $56,350 per year
  Bricklayer(Mason) - median $53,010 per year
  

Job demand also seems to be about average compared to overall jobs.

I agree that investing heavily in many degrees is probably ill-advised.

Use College Navigator for these types of questions:

https://nces.ed.gov/collegenavigator/?q=MIT&s=all&id=166683#...

That link says 72% of incoming freshman in 2022-2023 received financial aid. Also has a full-time beginning net cost average of just under $22,000 in 2022-2023.

It's not a perfect source of data, but there is enough on College Navigator to let you dig into it a bit and compare to other schools.

Honest question - how are you planning to source candidates in the first place? stack-auth is a two person YC startup and you were previously at another YC startup. I would expect you to tap the YC network for candidate referrals that probably will need little to zero tech evaluation?

I would probably skip a take-home for candidates at all for now. You probably have some specific pain points already. Try to find experienced TypeScript developers with some experience similar to your pain points and recommended to you by people in your network.

If you made me create a take-home assignment, I might make it rather open-ended and say do a compare/contrast of the stack-auth REST API versus one of the other options out there using any programming language and be prepared to describe code in both examples and strengths/weaknesses? This will help you maybe identify developers who self-select on "I'm interested in the stack-auth problem space . . . " Plus, that feels like a reasonably-sized effort for a candidate to me and could maybe be fun.

If you had a technical component during actual interviews, I think it is totally ok to ask someone to code up a small, simple Typescript example function of limited scope. Frankly, that's probably enough for you to get "yes/no" hiring signal.

I'm, err, older than you and will gently ask that you don't focus entirely on people who graduated since you did.

This article is based on a online poll of 966 people with murky qualifications for having a meaningful opinion about what seem, at best, to be silly questions.

I could see this as a first step towards some more interesting work and data collection, but can't imagine there is any actual signal here worth writing about.

This is for "First-Line Supervisors of Construction Trades and Extraction Workers." Does this represent the "average tradesperson" at age 40?

If we dig into the BLS numbers for the people doing the work in what HN tends to talk about as trade workers:

  Plumber - median $61,550 per year [1]
  HVAC tech - median $57,300 per year [2]
  Electrician - median $61,590 per year [3]
  Framer(Carpenter) - median $56,350 per year [4]
  Bricklayer(Mason) - median $53,010 per year[5]
  Software engineer - median $130,160 per year [6]

Of course, these are national numbers. We could pick and choose geographical areas to bump up the averages.

It isn't immediately obvious to me what data source could give us a good answer to net worth by age for any of these careers. That is why I asked if you had a reference for it. It's even murkier when looking for data on when someone in the trades buys a house on average.

My intuition (and I definitely don't mind being proved wrong) is that high earning, high-ish net worth trade workers is a story we want very much want to be true. But I just don't think it is, especially if we are talking about the average case.

  [1] https://www.bls.gov/ooh/construction-and-extraction/plumbers-pipefitters-and-steamfitters.htm
  [2] https://www.bls.gov/ooh/installation-maintenance-and-repair/heating-air-conditioning-and-refrigeration-mechanics-and-installers.htm
  [3] https://www.bls.gov/ooh/construction-and-extraction/electricians.htm
  [4] https://www.bls.gov/ooh/construction-and-extraction/carpenters.htm
  [5] https://www.bls.gov/ooh/construction-and-extraction/brickmasons-blockmasons-and-stonemasons.htm
  [6] https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm

"The average tradesperson in a HCOL is a millionaire by age 40 ...."

This is a pretty bold claim that I'd like to dig into. Do you have a reference or source for this? What HCOL are you referring to?

I'm quite interested in this whole topic of earnings and wealth in the trades. I am quite skeptical that any non-business-owner trades person is going to have a significant net worth (or earnings) above the general average locally, but maybe you can point me at a specific HCOL where the trades are more heavily unionized or something that would tell a more interesting story.

This general topic about trade workers pops up on HN periodically and there is always some discussion about high-earning people working in the trades. But I can't find any data that actually supports this statement - the BLS numbers tell a story that plumbers and electricians make almost exactly the median income that full-time, year-round workers in the US earn.

I don't doubt that someone who is running a business is earning more, this article in the WSJ says:

"At the time that they sold the company, it had 18 employees and was bringing in about $3 million in revenue a year. "

This was a plumbing business with two founders, founded in 2012. The article goes on to say that PE buys smaller businesses like the one above for:

"smaller outfits (such as Rice’s), which Redwood says it buys outright for an average of $1 million..."

The Occupational Outlook handbook says:

"The median annual wage for electricians was $61,590 in May 2023. The median wage is the wage at which half the workers in an occupation earned more than that amount and half earned less. The lowest 10 percent earned less than $38,470, and the highest 10 percent earned more than $104,180." The mean annual wage for all occupations in that resource is $65,470.

I only care about the ability to search/filter by "Informed Sport" or "Informed Choice" testing. After that, I prefer zero artificial sweeteners.

https://sport.wetestyoutrust.com/news/informed-sport-informe...

I agree that this approach would be pretty unwieldy for a large/complex backend schema. It worked quite well for a smaller, in-house application and set of requirements. I think I probably would be perfectly happy using the approach again for our in-house applications if requirements for audit/history preservation needed to be built into a data model. I don't have a good feeling for when I might say "this data model is too large/complex for this approach." I might instead think more about what and how many subsystems are going to have direct access to the RDBMS as a cut-off?

Hospital/medical research information systems (my day job is in the backends of these apps) seem to use backends with many compromises and poor db designs bolted on. I have also dealt with the horror of electronic data capture forms. My most recent headache has been a poor data model that ultimately wraps each form in a very ugly JSON blob. I've never seen anything that so completely lacks any residue of design . . .

I didn't mention it, but in this particular service, no client actually had SQL access to the database. There were create/read/update/delete service functions that clients used instead.

The timestamp for a tuple was returned by reads, so when a client wanted to update/delete a tuple, the service functions all required the client to provide the timestamp from the client read. If that timestamp argument wasn't the most recent, the client had to deal with that. But the actual database insert would fail.

Importantly, this particular service had only a few internal clients who almost 100% "owned" the tuples the client worked with. So there wasn't a lot of contention for any specific tuple by multiple clients.