vardalab

my new 2.5 Gb fios ont eliminated the battery, when power goes it goes as well

Oh, other day I was asking Claude to get a onetime code from my PocketID VM, I was being lazy and I was asking it to open the website, generate the authentication code from there and then type it into a built in browser. And it just refused to do it. I couldn't convince it to do it, even though it was all my own infrastructure And essentially, we were doing a security theater because the website that it was trying to authenticate to was the one it had just published itself behind my internal trafefic instance, lol. I ended up having Codex do it. Codex had no compunctions about it.

Mainframes with punchcards FTW, lol! This seems onerous to the point of it being almost like a security as a hobby.

But Claude Code and Codex and the rest of tui harnesses are also very good at just typing in each other's panes basically when I need to send a message over. It's been working for almost a year now. I have been more intrigued by T3 code or something similar, just to be able for harnesses to message across machines.

The whole point of having always on machines is you can close your laptop and the rest of the work keeps on going. Like you can leave home and work keeps on keeping on. I've been doing that for like 30 years. In combination, especially if you have a good laptop, then you can have best of the both worlds.

Dude, it worked for a billion people in China, lol.

I grew up under communism. I don't like or don't care for communism, but I can't dispute that China lifted a huge number of people out of a warlord torn mess and misery. To some extent, Soviets did the same to Russian peasants, but of course, it got perverted along the way. But all things considered, I would take communism under soviets to being a serf under the Tsar.

Yeah, now some future LLM from Google is a partner in your conversation, lol. "This call may be monitored or recorded for quality assurance and training purposes, and for future llm training"

tmux is still nice for remote work And it's even nicer when you want to have a different harnesses communicate with each other. I typically have a group of them working on the same thing, and typically it is Codex, Claude Code, and Pi. Sometimes I include agy if I need some Google specific capability, and it's very easy for them to communicate. I basically evolved something like a poor man's version of Claude Code's new projects But the nice thing is that I can have different coding harnesses other than cc working on a specific tasks.

Yeah, I had the same issues. I just ended up paying for Brave and Exa. It's literally like, at my volumes, a couple bucks here and there. Nothing to really think about at all. I am a heavy user of AI, but it self-hosted AI, but still the volumes have nothing to worry about. But I'm not trying to scrape Web or anything like that.

I wouldn't be surprised if the next version of Opus or whatever Anthropic has just eliminates these capabilities. I mean, we got DMCA shoved down our throats. What's stopping them from doing something similar again, I mean, we have a 70 or 70+ year copyright laws now, lol!

Exactly, the whole discussion is about this guy's AI slop instead of what he did. It's ridiculous. On the other hand, it's kind of expected by now.

do you have ultra plan? it says regular sub doesn't have it.

I would probably run qwen 27B or qwen fn but I would run the actual inference on one of my other boxes and not on actual desktop so initially it could be sufficient if you just enabled openai or anthropic compatible endpoint.

Cool thing is that I can just ask the llm to modify the harness to what I need. I discovered this daemon thing about two weeks ago and had codex modify it's own custom launcher that I have to bypass the daemon thing. If as a recreational code farmer can do it then most "pros" should be able to handle this and other inconveniences. Cc has been pushing that no scrolling feature that codex recently introduced for a half a year now but again I had them put a stop to that. Software has never been this easy to configure.

The one thing about the Bluetooth microphones is that there is this annoying lag that it's very hard to get used to, even when using AirPods which you think would be optimized for this. It's just untenable in my opinion because I could never get used to the 3-400 millisecond delay that I had to deal with. I ended up using simple Apple wired headphones as my go-to solution eventually. They actually produce a least a modest strain on ears when being in for hours on end, at least in my situation. I do like your idea about the Bluetooth multimedia remote controller.

Yeah, this is a very good point. I just lean back and speak. I find it much better. You can actually think now better because you can actually focus on the problem, admittedly, I never bothered to learn how to type fast because I was too lazy And my fingers are beefy. And I also like it when I'm literally just looking at something and giving a feedback to model while I'm speaking and it is being transcribed and being acted upon. I mean, this is all is kind of like science fiction. I'm surprised that there's so much negativity about all this AI or as we officially now refer to as Si.

Parakeet is blindingly fast on modern hardware. My MacBook is just so fast, it's not even funny. But despite that, I still often use OpenWhisper because it just does better for accented voices. I also fine-tuned Qwen ASR on my voice, and that one is faster than Open Whisper, but I think I might even go back to Open Whisper eventually Because it still makes a lot of weird mistakes And I think it is over-indexed towards the fine tune And substitutes quite often wrong dictionary words for ambiguous words.

Yeah, I talk to normies and nobody still understands the utility of just having a ChatGPT desktop app being able to basically drive your computer and do whatever needs to be done.

Yeah, I have an old X99 old asrock Fatality (lol) board, micro ATX with a Xeon E5 2640 and 128GB of ECC DDR4 running 24/7 as a Truenas server with 80TB of storage in it. It seemed like an overkill when I put it together two years ago when I replaced my 5820 CPU that had been my daily desktop but now with storage and everything prices exploding it has turned into a very useful thing. I still don't like the fact that it sits at around 130 watts. But given that it has so many hard drives in it, it's not that bad. My whole basement cloud now feeds into it.

For example, ad filtering for everybody and Tailscale in front of it.

z.ai speed has been horrific, worse than sol-6.1 was last week, I am not renewing my sub once it expires.

You would think with all this AI available, they could figure out the pricing to be as granular as necessary.

As context grows, it gets slower and dramatically slower. Otherwise it's pretty good.

Well, I can add two more points to your anecdata. I personally know of two companies where only exposure for peasants is Copilot.

Yeah, I finally got a decent version of GLM 5.3 flash running on my dual sparks. It's not that fast (faster that sol 6.1 though, lol), but it's manageable. Qwen 3.8 on the other hand, that thing really performs on dual sparks. You can get consistent multiple (3-4) concurrent sessions at 40+ tokens per second. It's pretty nice. Good progress, but definitely not that cost effective. I mean, OPUS 5.5 has been pretty magnificent. But nothing compares to when 4.5 came out. I still remember, it was just about Thanksgiving and that just was awesome.

It's been slow like molasses on the coding plan. I ended up using it more on fireworks. But DeepSeek is so much cheaper because the cache cost is much better.

that's where jev like classifier comes in

Article does say the breast cancer rate has been increasing by one percent annually. That's significant.

Because M3 Pro is not a real thing as far as actual agentic workflows go. I run dual R9700 boxes those idle at 150W Because of a Ryzen AM5, and I run Spark boxes, which are decent But still idle at 45-50 watts a pop. My favorite, to be honest, is a 5090 with a Qwen 27B because that one is good for quick hitters and flies, But again, the box itself idles at 140 watts. So it's really the idle power that I don't like And it's too much of an inconvenience to power boxes down and power them back on, so they just end up running and sucking electricity. I am working on getting some sort of a smarter proxy setup where I would give boxes time to wake up and go to the cloud while they're waking up. So the whole point of having these things running in the background is that you do want them to be almost always available. So power consumption is definitely an issue.

I have all sorts of local compute, and local models fairly capable the Frontier models still way more capable/faster and local electricity consumption is something else. Good thing it is getting colder around here.

I often pair them up, and I have an Astra or Sol work as a supervisor and reviewer while Qwen 27B FP8 or Qwen 3.8 Flash Next implements things. I mostly do it as an experiment, just to see what kind of level of autonomy I can get, and they are slow to getting a decent reviewed outcome despite Qwen27B running at 100+ tps and 3.5-4K prefill rates and Qwen3.8 Next at 40 tps and 1-2K prefill. I've been also using similar approach more with OMP, not just the straight Pi harness. And OMP seems to be slower because it has more guardrails. OMP has an interesting feature where you can assign a better LLM as an advisor, wehere it just sort of monitors the progress and injects guidance. And it definitely helps, but one has to be careful. It actually turns out to be expensive if the cache reads are expensive. I learned it the hard way. Where on Fireworks' API, the cache rates for GLM 5.3 flash are quite a bit more expensive than for DeepSeek, and a simple runs ended up costing me three bucks in oversight. So a better way is to have a Frontier model running a separate tmux pane and just directing it to Wake up every 10 minutes, take a peek at what's going on, review the milestones give feedback and then sleep. This turns out to be pretty decent cost saving strategy when quote needs to be stretched. Paradoxically, OpenAI tightening up their quota allowance once they released Astra actually pushed me into all these sorts of experiments, and it's actually been interesting. I've been exploring all these smaller flash models, and it's been nice. I do like using local LLMs for chore type tasks that are just mostly information gathering, post-session reviews, stuff like that.