Hacker Newsnew | past | comments | ask | show | jobs | submit | sh3rl0ck's commentslogin

Hard agree.

It's easier than ever to build compatibility for more architectures and devices than ever before, to autonomously test it, and ensure UX rough edges can be rapidly iterated upon across platforms.


We'd bought 4 x $11K Mac Studios at my college and via exo, we had Kimi K2.5 at 30 TPS.

Not too wild an idea!


30 tok/s generation? Nice! What was prefill by the way? Also, these are 256 GB RAM studios? Good timing on those!


512 GB!


I've been a strong proponent of reallocating all LinkedIn server capacity to GitHub.


this is an idea that i’d happily get behind.


Everything that went wrong with Claude so far.


There's no mention of SLMs or LLMs, though.

> This work represents a compelling real-world demonstration of “tiny AI” — highly specialised, minimal-footprint neural networks

FPGAs for Neural Networks have been s thing since before the LLM era.


Huh? The first paragraph literally says they are using LLMs

> [ GENEVA, SWITZERLAND — March 28, 2026 ] — CERN is using extremely small, custom large language models physically burned into silicon chips to perform real-time filtering of the enormous data generated by the Large Hadron Collider (LHC).


the site might have fixed it, to me it says "artificial intelligence" instead of LLM, still bad but not" steaming pile of poo on you bank statement" bad



One of the very few good things from the AI race has been everyone finally publishing more data APIs out in the open, and making their tools usable via CLIs (or extensible APIs).


I feel like the CLI craze started around 2020. That predates this chat GPT.

CharmCLI golang

Nushell rust

Warp. Shell

Were all around 2020 also that is when alt shells started getting popular probably for same reasons they still are.


I assume it's an economies of scale thing now.

It's not like Apple is putting any thought into either the UX or the engineering side of utilising the compute properly (except calculating those glass effects extra inefficiently).

Minimise SKUs and get some use out of the binned chips who have a few failed cores.


I think there's a reward for finesse too.

As you mentioned, scope definition and constraints play a major role but ensuring that you don't just go for the first slop result but refine it pays off. It helps to have a very clear mental model of feature constraints that doesn't fall prey to scope creep.


There's also a reward for not over thinking it and letting AI bring the solutions to you. The outcomes are better when it's a question, answer, and execution session.


In Gallifrey? In Gallifrey.


Never had Nvidia issues on Fedora and Ubuntu so far, 1P a multi computer research lab as well.


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

Search: