Rendered at 05:24:33 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
robbles 15 hours ago [-]
It's oddly depressing how much of the new tech on HN I can't come close to running on my own hardware.
platinumrad 1 days ago [-]
This will be very useful to me as someone who still uses LLMs interactively rather than trusting the harness to one-shot the whole thing.
booi 23 hours ago [-]
You’re missing out, AI is excellent at end to end tasks if you are clear with what you expect. I actually think it’s worse to approach it interactively
tancop 16 hours ago [-]
I know from experience that even the biggest models can't set up a good architecture. You always end up with huge amounts of duplicate code that slow down the app and make debugging a lot harder.
And even if you solved that problem it still wouldn't be creative enough. Claude never actively sets up clever TypeScript constructions to enforce that all keys of a map are accounted for, or looks for a library instead of writing complex UI code from scratch, or makes any framework choices other than the safe default ones.
It can do these things but you need to explicitly prompt it every time after reading the code to identify problems. That's the opposite of an end to end approach.
NortySpock 9 hours ago [-]
I've been getting ok results from having the LLM research and write out a single planning document per task, with code examples and unit tests. I collaborate with the LLM on the planning document review, tweaking the architecture / tests that it plans to build. I iterate with the model on the planning document. Only once that document is polished and leaves little room for interpretation do I set a (possibly smaller) model on the task of writing the code change and running the unit tests. I'll come back and read the diff later, possibly it still needs minor corrections but far fewer than if I had let the LLM run without a detailed plan.
While that's running, I can, in a second window, collaborate with an LLM on the next plan...
Edited to add: architecture-wise I have only done this on smaller scripts, and I encourage and sometimes enforce "pure functions, that act on data, with no internal state", and then if I need state it would go in a specific typed structure, not a object that mixes state with functions.
epolanski 20 hours ago [-]
Nobody's missing out.
AI is okay at best at mundane task or good at task you don't care for the output quality or maintanability.
It tends to write much more code than needed, handles countless made up edge cases and is so reinforced to find the "shortest" or "simplest" way (by god knows which training criteria), that it can't stop leak responsibilities left and right, so you have modules/systems A-B-C and all of those are tightly coupled and know each other.
And no, subagents, agents.md, and whatever crap does not help much.
I myself use AI daily, but you folks have to stop with those loser narratives pushing people being mindless meat proxies. AI is okay at micro level and glueing messes of micro level code, but it's still really not there when it comes to software design and taste.
chayu 15 hours ago [-]
[flagged]
747-8I 1 days ago [-]
.
jahala 1 days ago [-]
I'll just go get my 8 x B300 (TP8) from the shed and start this up real quick
anigbrowl 8 hours ago [-]
This is clearly pitched at industrial clients. For people who can throw around a few million $, it's practical to either launch a specialist AI service or build a highly performant private compute cluster using the DeepSeek architecture because they've open sourced so much of their pipeline. There are a lot of potential users for this; consider that somewhere for 1/3 to 1/2 of nvidia employees are multi-millionaires, and that if they're unsatisfied with the future direction of the firm following Jensen Huang's abrupt departure, many of them have the means, skills, and resume to stand up a new enterprise very quickly.
If you want to play with this but don't have mega-resources, you can rent a B300 cluster for about $100/hour. Not what I'd call cheap, but if you have a clear application in mind you could use it to jump-start a small business, academic project, or resume booster in a weekend at low financial risk.
xlayn 1 days ago [-]
I have been doing some poking at llama.cpp for amd gpus and at some point I ask claude if it was already 10x faster than GB200 and claude told me to not be funny, one of those things were more expensive than my house and nvidia sold them in a 72pack unit that needs it's own hvac....
rvz 1 days ago [-]
Please continue. [0] Let's see what happens when it goes even faster than 2,400 TPS and why not host an abliterated version of this model as well.
you know who was state of the art abliterated models who already have hack a lot of goverment services in a lot of places worldwide? openai, anthropic, google, facebook...
Thanks god you are not a wacko suggesting Qwen4B-abliterated will hacK someone
rvz 1 days ago [-]
Yes. If we’re not prepared for frontier models breaching companies at 300 TPS, we are certainly not prepared to defend against frontier abliterated cyber models running at 2,400+ TPS.
It takes the attackers to successfully breach their target once vs the defenders having to stop all attempts.
SR2Z 12 hours ago [-]
Soon enough the only way to stop a bad guy with an AI will be a good guy with an AI.
ShinyLeftPad 11 hours ago [-]
s/bad guy with an AI/500 bad guys with an AI/
it's not gonna save you here because it's very asymmetric and not in your favor, if you're a defender.
anigbrowl 7 hours ago [-]
The solution to this will be government and big corporate legal superiority. The widespread availability and ownership of guns in the US is a source of existential risk to the political establishment, but the normalization of said establishment retaliating with swift and overwhelming kinetic or legal force against any perceptible violent threat creates a powerful disincentive to would-be revolutionaries.
Correspondingly, the proliferation of high-speed digital offensive capability will probably be met with counter-hacking designed to identify and retaliate against such intrusions even faster or wider-spectrum attacks - a Gibsonesque cision of asymmetric cyber conflict. If 'black ice' doesn't take down an destroy intruders on its own, drone munitions probably will.
nozzlegear 23 hours ago [-]
> What could possibly go wrong?
Somewhere in SV, an orgy at an EA commune is abruptly halted as a bunch of sweaty nerds simultaneously cry out "well, actually!"
And even if you solved that problem it still wouldn't be creative enough. Claude never actively sets up clever TypeScript constructions to enforce that all keys of a map are accounted for, or looks for a library instead of writing complex UI code from scratch, or makes any framework choices other than the safe default ones.
It can do these things but you need to explicitly prompt it every time after reading the code to identify problems. That's the opposite of an end to end approach.
While that's running, I can, in a second window, collaborate with an LLM on the next plan...
Edited to add: architecture-wise I have only done this on smaller scripts, and I encourage and sometimes enforce "pure functions, that act on data, with no internal state", and then if I need state it would go in a specific typed structure, not a object that mixes state with functions.
AI is okay at best at mundane task or good at task you don't care for the output quality or maintanability.
It tends to write much more code than needed, handles countless made up edge cases and is so reinforced to find the "shortest" or "simplest" way (by god knows which training criteria), that it can't stop leak responsibilities left and right, so you have modules/systems A-B-C and all of those are tightly coupled and know each other.
And no, subagents, agents.md, and whatever crap does not help much.
I myself use AI daily, but you folks have to stop with those loser narratives pushing people being mindless meat proxies. AI is okay at micro level and glueing messes of micro level code, but it's still really not there when it comes to software design and taste.
If you want to play with this but don't have mega-resources, you can rent a B300 cluster for about $100/hour. Not what I'd call cheap, but if you have a clear application in mind you could use it to jump-start a small business, academic project, or resume booster in a weekend at low financial risk.
What could possibly go wrong?
[0] https://news.ycombinator.com/item?id=50310886
Thanks god you are not a wacko suggesting Qwen4B-abliterated will hacK someone
It takes the attackers to successfully breach their target once vs the defenders having to stop all attempts.
it's not gonna save you here because it's very asymmetric and not in your favor, if you're a defender.
Correspondingly, the proliferation of high-speed digital offensive capability will probably be met with counter-hacking designed to identify and retaliate against such intrusions even faster or wider-spectrum attacks - a Gibsonesque cision of asymmetric cyber conflict. If 'black ice' doesn't take down an destroy intruders on its own, drone munitions probably will.
Somewhere in SV, an orgy at an EA commune is abruptly halted as a bunch of sweaty nerds simultaneously cry out "well, actually!"