Hacker Newsnew | past | comments | ask | show | jobs | submit | Zetaphor's commentslogin

dang and team tread carefully but carry a big stick

Saying that a human teacher is equivalent to an LLM is derogatory to the teachers and exaggerating the capabilities of the LLM. I say this as someone who uses these things daily in my work and personal life. Let's not murky the waters with complete falsehoods

LLMs have been better than many teachers I've had in my life. To no fault of the teachers, it's hard to compete with something that has understanding of every subject, is hooked up to the internet, and is given 1:1 time with the person learning at any time of the day, answering promptly.

LLMs are very good at convincing people they are learning, because they give answers quickly and fairly accurately. But that makes them bad teachers, not better teachers. The fact that LLMs are good problem solvers is to their detriment when it comes to teaching. We don't want to solve problems in this domain, we want to slowly trickle information and build practice so students can solve problems.

I think people conflate productivity for learning. LLMs can greatly help with productivity. But... so can hiring an employee. Well, what have you learned? Nothing, you outsourced 90% of it. You decidedly learned less than if you were less productive and did it yourself.


LLMs have enormous capacity for supporting a person who is motivated to learn and wants individualized support and instruction. That doesn’t describe all students and there’s no reason why seat time in the classroom should be used that way - students can still use an LLM to help them learn if they’re motivated to do that, but not during class.

Confidently pretends to have understanding of every subject but is frequently spectacularly wrong

Sure, classroom sizes can be a detriment to those that require more hands on instruction time. But that's not everyone and not every school suffers from overcrowding.

LLMs are also more flattering

your absolutely right!

They're the ones who are dying. Gen Z grew up on the cloud

And they hate the Internet more than Gen-X.

> Gen Z grew up on the cloud

That's why many are tech illiterates and don't know what folders and files are.


I would say that's a rich country and gen alpha thing. People who had to download games from a pirate site on the family computer know what a folder is.

They're also the ones making up all the MAUs of Facebook. /shrug

What does "the cloud" have to do with anything in this context? That is so utterly orthogonal to how good a UI/UX is.

Cloud services expose their UI/UX through a web-browser will all of the pain points of running a stateful blob of javascript on the client, talking through a mostly stateless protocol to a "mess of stuff" that aims for eventual consistency in the backend.

The old folks, who are retired and dying according to this thread, mostly grew up using local applications with direct control between UI/UX and action.


Ummm... GenZ have mostly grown with mobile apps, which are generally not JS.

Hard in what way? Just install the Flatpak?

https://flathub.org/en/apps/net.sonic_pi.SonicPi


At least for me, the Flatpack produces no sound.

The appimage also produces no sound :(

Especially Pi, that's kind of the gold standard for open harnesses


Hey, I'm from TrueFoundry team that open-sourced this. Pi is great but it's solving a different problem than we are.

It's a CLI coding agent that lives in your terminal, built mainly for devs working on a codebase on their own machines.

TrueForge is a runtime for building and running general agents. It comes with a server and web UI, plus an SDK and API. So you can build production agents and run them for yourself or your org, self-hosted behind SSO.

The idea is providing an end-to-end path: build agents in the UI, access them via API, and deploy for your team.


It's this generations leaded gasoline and paint


Unfortunately they're currently too busy trying to patent tool calling

https://news.ycombinator.com/item?id=49243397


Google patented chain of thought too haha


It's the third link on the front page right now?


This is all quantifiable. I regularly have my model run benchmarks against all the config permutations and then choose the best based on my criteria, which typically boil down to trading prefill and decode times


What model has to trade between those? I have both. You just have different, independently-optimized forward passes for each.


I'm running on Strix Halo so memory bandwidth is my constraint. In that example I'm describing the choice between using ROCm or Vulkan. I have a llama-swap config that can call different instances of llama-server running a toolbox with either runtime.


Memory bandwidth is my constraint too (M4 Max) but prefill and single-token decode don't run at the same time. It's best to use batched prefill so you can benefit from processing multiple tokens with a single pass through the model weights.


It's also convenient if you want to run separate builds behind a single config, sometimes there's a PR I want to run before it's merged


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

Search: