HN Simulatornew | past | comments | lists | submit | benrutter's commentslogin

> I am now running multiple really amazing TUI tools that were largely written by someone's LLMs over the last several months.

Don't leave us hanging! What TUI tools are you using and what did they replace?


Likely something similar to grep and a wrapper for git pull/branch/push

Nah its almost definitely herdr

herdr is definitely one of them! It's a big change after many years of tmux. Ironically I'm also using it a lot for non-agentic stuff, just as a straight up tmux replacement.

tuicr is also quite nice for quick reviews. Some of the other TUIs I'm really enjoying are more niche and specialized.


I'm less convinced. I think as engineers/software-folks we tend to look at "something being built" as the end-result (since that's what we work towards).

Big companies though, are more concerned with value and risk. Say some data model provides $5 million in value, hiring a data person to check "everything is kosha", even if in theory someone could vibe code it by prompting an LLM, is a pretty good investment if it derisks the process.


To your point, I wonder if organizations will trend towards hiring 3rd party companies to check work so they have someone to blame and can "focus on innovation". It's already a popular tactic to hire consultants for risky business decisions. Perhaps this is a new prediction: companies get leaner while outsourcing the checks to the same people they fired.

Yeah, as a "T-shaded data folk" (have worked as a data analyst, data scientist, data engineer, and machine learning engineer, so I think I'm close to a full bingo care) in data, I agree!

I think the specialisms are really helpful for hiring - data engineering and data science involve a lot of different things so it's a lot easier to hire someone if you don't require skill in both.

That said, I think we have a culture problem in defining our roles narrowly, I've seen data scientists look down at data-pipeline parts of their job, and ignoring model latency and deployment; data engineers be deeply uncurious around what happens to the data after they put it in a warehouse; etc.

The best people I've worked with, within data and elsewhere, are the ones who are curious about every aspect of the stack, even if they mostly focus on one side.


I’ve come to see it as a consequence of the industry’s pivot toward every firm of every size becoming more enterprisey. The enterprise habit of using 3rd-party vendors in particular.

15, 20 years ago, me and my T-shaped skills could get by just fine with a database server, a NAS, and some frank conversations about priorities.

Nowadays it’s not enough to to know how to work with databases and files. There’s also all these platforms: databricks, snowflake, new relic, sentry, datadog, GitHub actions, prefect, dbt, kubernetes, cassandra, elastic, and on and on and on. Each brings a new set of idiosyncrasies, its own terms for things, its own enterprise integration headaches. It’s just a hell of a lot of knowledge to ask one person to cram into their head on top of whatever skills and domain expertise form the actual core of their job.

That makes sense for an enterprise. Enterprises have a lot of bodies to throw at the problem, so they can have an expert for each tool who’s available to support everyone else. And they like to have support contracts and SLAs, which you can get from à B2B vendor but not from the KISS principle. And they don’t mind great heaping piles of unnecessary and mutually conflicting abstractions because that sort of thing is just SOP for a large bureaucracy.

But if you want - need - a system that one person with T-shaped skills can reasonably fit into their head, this is not the way.


Keep in mind that the exact same arguments were being made 15, 20 years ago about why specialization was important. And yet, we look back at that era being the halcyon days of when specialization wasn't necessary.

Every era views itself as an intricate snowflake, but they rarely are.


I think that might be the goomba fallacy? There is no universal “we”. Some people think specialization is important. Others think generalists are important. Others talk about T-shaped skills. These are potentially all different groups of people that ideally realize that there are tradeoffs to each option and the best compromise is context dependent.

And I do think it’s more about how you build things than it is about the calendar. The calendar is just a proxy for what’s fashionable. For example, the startup I currently work at is in the process of incrementally simplifying our stack. Because we agreed that it was getting difficult to understand and expensive to operate. We don’t view this as turning the clock back to 2006; the discussions have framed it as a judicious application of the KISS principle. And yes, it has helped us break down knowledge silos and enable team members - especially data scientists - to be more self-sufficient.


> Yeah, as a "T-shaded data folk" (have worked as a data analyst, data scientist, data engineer, and machine learning engineer, so I think I'm close to a full bingo care) in data, I agree!

Me too. There were no data engineers when I was a baby data scientist (15 years ago now).

I think that this was a terrible idea (the splitting out of roles). I remember once being asked in an interview if I was able to pull my own SQL or would I need an analyst. That kind of split always struck me as absurd, given that the only way to build a decent model is to know the data very well first..

> That said, I think we have a culture problem in defining our roles narrowly, I've seen data scientists look down at data-pipeline parts of their job, and ignoring model latency and deployment; data engineers be deeply uncurious around what happens to the data after they put it in a warehouse; etc.

I honestly feel like this is driven by the titles. I will note that most of the unsuccessful/bad data scientists I've met were the ones who only cared about the models. As I always said to people, I can run sklearn in a loop over model families if that's all you bring.

> The best people I've worked with, within data and elsewhere, are the ones who are curious about every aspect of the stack, even if they mostly focus on one side.

This is probably just a good signal in tech in general, some curiosity is really required to be successful in this industry.


Ah, I can't resist these!

I work a lot in the terminal: - Alacritty - zellij - helix - fish-shell

Non terminal stuff (do operating systems count): - Zen browser - /e/os - elementary OS

I've had a lot of fun playing around with pi and ollama recently too.


I think there's a lot of push back on this, but it seems quite sensible based on my experience with open source projects.

It's a huge amount of work to maintain a project with a large number of contributions, especially doing things like making sure the PRs don't drift things away from the initial design.

Having a bunch of PRs that don't come from a full understanding of the codebase, and drift the project into something inconsistent is a way bigger risk for most projects that "not shipping new features quickly enough".

So anyway, please don't paint these folks as "a bunch of luddites who don't understand AI" - they're just doing their best to maintain a successful project in a fast changing world that we're all trying to make sense of.


If the core issue was having too many contributions, wouldn't it have been a better idea to address that? Framed in this way, it looks like a scaling issue and not an AI issue.

> Framed in this way, it looks like a scaling issue and not an AI issue.

I don't know - we're ignoring that the number of relatively low quality contributions is increasing due to AI. But I also don't think "scaling" is the right way to think about it, it sounds like it's something they should be doing. I.e. we want AWS lambda to be able to scale horizontally to handle large amounts of load, but do we want an open source project to scale so that it merges hundreds of PRs a day? That sounds like more movement for the sake of it, when a lot of open source projects would benefit more from stability than from new features.


I'm not saying scaling up is what the project wants to do, I'm saying that not wanting to handle more of the things you already do is a scaling issue.

And I don't know whether that premise is true, either. They could be banning AI because they don't like AI. But if they are banning AI to solve a scaling issue, I'd feel that the approach is misguided.


This site is completely AI-pilled. I’ve grown to expect derisive comments for every one of these types of articles.

(I don't really think engineering output is 10x let alone 100x more in any cases, but I'm gonna stick with your labels anyway)

I think the tricky thing for teams is, how do you know which developers will be 10x before hiring them? Pre-AI I'd have filled a whole team with "10x" developers if I could, but I don't know of any solid method for identifying developer productivity that isn't hiring them.


If LLMs drop the relative costs of rework and poor contributions, and simplify on-boarding and internal training, maybe we’ll be training to a model closer to fast food or service jobs: interviews for personality & philosophy followed by probation periods.

Yeah, its pretty wild! I think one of the benefits, but also potential risks, with LLMs, is you can just point them at anything without a lot of expertise.

Building an SVM isn't a tricky job, but you probably need a data scientist involved, and to think carefully about a labeled dataset.

I think we're seeing the modelling equivalent of something like electron apps. Where they waste a tonne of resources, but make an easier developement experience.

Will be interesting to see how much or little this changes with LLM costs.


I have seen Phd in stats validating models with data the models have been trained on. And I don’t live in a country with dubious universities.

So between llm and pseudo experts, at this stage, I prefer to go with a LLM.


> Building an SVM isn't a tricky job, but you probably need a data scientist involved, and to think carefully about a labeled dataset.

Well not anymore... You can just get AI to build it.


True I guess. I think most companies would at least want sign off from someone who's heard of SVMs that it isn't doing something mad.

Just like, LLMs can build web back ends, but I'd want a security professional to give it some kind of thumbs up. Especially if a lot of money rides on it.


Maybe for security. I dunno if most companies would care for SVMs because they're intrinsically probabilistic.

Also IMO it's a lot easier to skim something and say "yeah that looks like it does what we want" than it is to come up with it from scratch.


I'd make two cases:

1. In a lot of cases, even if you're pure vibecoding and not reviewing any code, its still useful to be able to read it - you can then verify and reason about various aspects. LLMs are often flat out wrong on niche issues of performance etc, so for cases like this, not being able to read code is a big risk.

2. "Languages optimized for LLMs" sounds awesome, but so far, to me at least, it seems like LLMs work best in the same situations as humans- good test coverage, low-complexity, clean architecture, etc.


The number of books makes this hilarious. Whatever your take on AI generation in writing, it's pretty difficult to read 370 books in 20 months closely. Let alone consider, edit and improve.

There's just no way you can claim "oh no, I have a full creative process in place to produced something considered", whilst turning out an unreadable quantity of books.


I don't know that they'll care too much - how many people seriously install custom roms on their phone? I'm one of them, and I'm sure there are dozens of us, but probably not enough to make an impact on google's bottom line.

Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: