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

Soooo anytime now to bring it back right??

I mean, they haven't announced the non-Pro iPhone 18 yet, so keep the hope alive!

No way you can get that much each without extremely quants

For a single r9700 you have 637 GB/s and for qwen 3.8 27b q4_k_xl the maximum tg/s is 33 before mtp

Now if you meant 4xr9700 tensor parallelism with mtp, 80 tg/s starts to make sense


You can get those numbers with https://codeberg.org/ggz14/radiance-vllm-mxfp4. I also can get it on a single R9700 but the 75 ~ 80 t/s is only peak acceptance of very predictable tokens like coding or json, and averages lower for prose. It's still much faster than regular llama.cpp.

Can confirm. I've a single R9700 and have maxed out at 45 tokens/second on llama.cpp with Q4 Qwen 3.6 27B (with MTP)

Q4 isn't an extreme quant, and I average 75 toks/s on code, 45 tok/s on prose with MTP.

I run it locally at q4_k_xl on a r9700 with kv cache bf16 and while it thinks a lot, it’s still fast enough to do the task.

This model had its knowledge replaced with reasoning ability. The chain of thought what makes this reasoning effective.

So this is why you need to let it think and don’t quantize the kv cache.


Or at least use a modern llama.cpp with KV activation rotation.

Nic needs pcie5 x16 slot, so on am5, take out your gpu

It’s like next.js but for backend.

Personally not for me. These are very productive once you learn the shape they expect.

But the wheel needs to be reinvented on how to do everything.

I ran have straight forward verbose code with no hidden control flow


> once you learn the shape they expect

I’m reminded of this essay discussing frameworks versus libraries [0].

[0] https://tomasp.net/blog/2015/library-frameworks/


Am I missing something? The entire point of next.js is that it's backend+frontend in one project


that’s likely the promise but it’s so confusing and high cognitive overhead to maintain that backend <-> frontend relationship. ‘use client’ - i think the special sauce for why people use nextjs is all frontend based. the frontend that comes from the backend.

nextjs dx is a special kind of hell. disclaimer: im just a simple rails guy.


Next.js is a poor choice for a client-side application. It is first and foremost a server-side framework, and trying to force it to be client-side is to work against it. There are so many better options than Next.js for dedicated client-side front-ends.


Nextjs is also a bundler for web applications. And it deploys the front end assets.

It's yet another backend framework, added that it only runs on vercel infrastructure.

There are better front end and back end servers, those that simply rely on nodejs to serve. Next.js does that, without vercel.


I don't understand this perspective. I've developed, delivered and maintained countless Next.js projects and never come close to considering it "high cognitive overhead." It's among the nicest DX I've experienced. It has a lot of features, I guess, but you don't have to use them if you don't like them. Next.js has very little to do with the frontend. The frontend is literally just react, which is why I commented in the first place. People keep acting like it's a "frontend framework" which doesn't make any sense at all.


React is a frontend framework. How do you get around that? It's invented entirely to handle client-side UX and reactive DOM updates which have to be DOM because it for the browser which is a frontend. Are you saying nextjs is the backend and react is just react, it's not nextjs' fault?

I'd say we disagree on the simplicity then: hydration, nextjs API reinventions of every interaction else 'use client' as an escape hatch which means it runs on the client but then what does the non 'use client' code do if nextjs is not a client-facing framework?


I’ve been saying for years we need CCR (Client Centric Rendering).

What we do is have the server just serve the page and let the client handle most of the rendering.

It’s novel and innovative


Smells like SPA


Isn't this html + javascript?


I let the LLM tokens flow. Usually it ends up saving me innovation tokens.


Yeah the base requirement has gone up. So more training/schooling is required.

Imagine when jobs started adding the requirement entry level positions require ability to read? That cut out so many people when that requirement was added yet the economy survived


That requirement should probably make a comeback.


I think the base requirement to compete will keep going up. We already have AI models that can debug and audit code better than the best humans.

You used to have a career as a software engineer or devops person if you were good at debugging. Today, I challenge you to take any past production outages, toss the traces into Claude Fable, and see if it finds the bug faster than humans can.

I think this is the way the rest of white collar labor goes.


> We already have AI models that can debug and audit code better than the best humans.

I'd love to see one someday.


Did you try my suggested experiment? Which issues did Fable take longer than a human to diagnose once you started to describe the symptoms?


I “try your suggested experiment” several times per month at work. Usually when I’m on call.

It’s great at being a search engine for our docs and communications. It can also find bugs quickly in small repos with static analysis and coverage tools.

But when it comes to larger projects or interconnected systems, the time it takes to correct its “assumptions” is often longer than the time it takes to solve the entire problem myself.


He was quoting the claim you made. Shouldn't you have your own justification for it? Which tests did you run with both Fable as well as the best human programmers?


I see -- so you'd rather not know if it works in your situation?


"I see the problem now - wait, let me ask some clarifying questions."


...you can quibble about the chat transcript, but it spits out answers pretty reliably. "But you don't have to take my word for it".


Ive dealt with tons of bugs Claude couldnt recognize. Have you not?


Honestly? Not really, not since Fable came out. There's been a bunch of times it needed to ask for context, and I needed to deploy some additional tracing at its request.

(There have been a few times where it has refused to debug due to safeguards. This is actually the biggest problem.)


For some roles, I agree but I also think that there will be many new-ish roles that will have lower requirements right down to being a warm body capable of walking and talking. The lower requirements will be possible because these roles will essentially be proxies for AI agents that need a human to go somewhere and do something in the physical world, whether that's resetting a breaker or just being a human face and voice delivering a message.

And yes, before anyone brings it up, I've read Manna and if you haven't you probably should[0] but I don't think it will be as dystopian or utopian as that. Humanoid robots would obviously replace some of the proxy need but there will still be a lot of things that only a living human can do even if it's just purely for legal and regulatory reasons.

I also think there will be a lot more roles that are purely human in nature and which will have an entirely different set of requirements but that's a different topic.

[0] https://web.archive.org/web/20120224063109/http://marshallbr...


Of course. For a great deal of software, the main skills needed are vaguely speaking the jargon so you can understand Claude, and doing manual debugging.


Can’t wait for single chip asic qwen3.8 27b

It’s small so should be cheap per chip. And it so much smarter than it ought to be for its size.

Problem is asic take forever to cut and are always months behind the latest open weight sota


Taalas uses a design that requires masked ROM layers specific to a model, but Etched uses a newer design with a specialized RAM that supports any transformer model.


IMO it was a mistake and everybody should be a sole proprietor.

This way legal consequences always have personal consequences and people should act accordingly


Give it agent instructions in the major breaking api changes. I force around Io stuff. Then point it to the zig source code.

Sure it wastes a bit of time always figure it out when it fails to compile.

This is why I’m excited to see qwen3.8 27b removed a bunch of knowledge for more reasoning. Baked in old zig api is worthless


Umm I have an extra 35, do you have layer 6?


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

Search: