I run many models (but mainly Gemma-4) using oMLX (for caching) on a 32GB M1 max using (gasp) Xcode. For tok/sec response times, I'd say it responds faster than I could read the prompt aloud in many cases (and I'm not constantly polling the Claude status page).
For months I spent time curating the AI+harness+skills+MCP servers, but now mainly just code with it. I find myself not bothering to use Claude (but keep paying "just in case").
That's feasible in part because my prompts have very specific objectives, constraints, and suggested staging, because I want the code to be exactly as I would write it, and I want to weigh in at specific moments. I would say the speed-up is 2-4X instead of the 10X of vibe-coding greenfield projects. The problem is not the coding speed, but building something complicated that's also correct and flexible (i.e., a directional accuracy). E.g., the agents help with abandoning a less-fruitful API shape instead of sticking with what works in a local maxima.
One flaw there is that I'm still writing code that feels clean to humans, which now is probably a waste. LLM's might be happier with 10+ parameters on one API instead of a plethora of configuration objects and convenience wrappers.
There is a happy medium. Encrypt the user directory/s and leave the rest of the disk unencrypted. There is still a risk that the user loses valuable data due to corruption, but there is much less of a risk of the entire system being unusable or unrecoverable because of corruption or ransomware.
I’m very on the pro-ai side (check my comment history for proof), but this “ai will give us more free time” logic is seeming more and more patronizing (to be clear, I understand that you are being sarcastic haha).
I was listening to a podcast a couple days ago and Brad Gerstner was on and mentioned that with how AI is boosting productivity that perhaps one member of a household would be able to start staying home from work if they wanted. I shut off the podcast after that (to be fair, the podcast just seemed to be one massive SpaceX IPO pump).
It’s just so divorced from reality and every new advancement is just making *higher expectations for doing more work*.
Not 100%, I still fall back to Claude for most day-job stuff. But I've been trying to use Qwen 3.6 and Gemma 4 on my framework desktop mainboard (Strix Halo) as much as possible.
I've been working on an ops style tool for local LLM inference. Proxying, api keys, request logging, model rewriting and much much more.
If you believe the benchmarks, Qwen 3.6 35B-A3B already outperforms Claude 4 Opus.
Now, there's a bit of a degree to which some of the open source models do some benchmaxxing, and bigger models with more params may always feel like they have more depth. But anyhow, right now you have something that is arguably comparable to Claude 4 Opus on your laptop. I can't really compare myself because I never used it. It looks like Claude 4 Opus is still available on OpenRouter, so you could try it out and compare yourself if you're interested.
It will likely always be the case that there are proprietary cloud models that are more powerful than what you can run on a laptop. You can just do a whole lot more with terabytes of VRAM on multi-GPU clusters than you can do on a laptop. So for folks who must have the most capable, you're probably not going to want to leave Anthropic.
But right now, the models you can run on your laptop are comparable to the cloud models that were popular when vibecoding and Claude Code first took off.
It seems like what's missing here is lower cost plans, because the existing plans had been fairly affordable, but now they're basically triple.
The least expensive one seems to be CPX11, old price $6.99, new price $20.49. That's 2GB RAM, 40GB SSD. RAM and SSD are now much more expensive, fair enough, but maybe I don't need all that for my mostly-idle VM, so then where's the plan with ~0.67GB RAM and ~13GB SSD for the old price?
> to me as an outsider the important part is a treatment that works, not why it works
Are you a mouse, perhaps? We already have a plethora of treatments for mice suffering from human-induced Alzheimer's. None of those treatments have ever been shown to work for human patients, and this one is no different.
TBH I've been mostly unimpressed with the early Linuxen. At the time they had almost nothing which NetBSD didn't have, and I knew that really well. Gentoo 'ricing' well.
But it was clear that there was momentum behind the Linux hype. It was on CD-Roms in magazines, in book stores. While *BSD wasn't known generally, just by some guys in Universities, or similar. Bad marketing. FUD because lawsuits, and whatnot else.
The reasons really doesn't matter. As long as they are in top of price/performance/quality nobody will switch. Once they stop to be then people will think about it.
His Apple keynotes conveyed a sense of magic, for example demonstrating pinch to zoom on the iPhone and pulling a MacBook Air out of a manila envelope. And something he'd be angry because things didn't go as planned.
These pre-recorded keynotes we get nowadays are just bland and AI-generated.
I have completed 4 rows of firewood, onto my 5th! The stacking animation is very satisfying, I am sad I can't actually see the wood stacking anymore (because my wood pile has grown to large). It would be nice if it persisted across refreshes, and maybe provided a counter for how many rounds/cords of would you have split.
When was Disney world built? Why can nothing be built anymore. Disney is but one example.
People complain about data centers due to electricity. We have technology to generate infinite electricity yet the answer is to not build rather than build both the utilities and the data centers. Everyone would benefit from more availability of energy
> During voice agent startup, the slowest part are [...] in dlopen-ing large shared libraries (ONNX Runtime in particular) and in hundreds of small random reads off the SD card as Python walks the import graph.
Could this maybe be fixed by arranging the files on the filesystem in the order they're read? Or maybe importing the Python modules from a zip file (with no compression) would be faster, if that makes it easier to store them in the required order?
I seem to remember Windows having some sort of feature like that, automatically rearraging the data on disk in the order it's read when booting the system.
