skolos

What is benefit of GPU-rendered UI? In regular 2d apps where I can select enable/disable GPU rendering, disabling doesn't make app feel slower, but it drastically reduces RAM and VRAM usage (VRAM is the one I care about, as I run local models). Why so many UI frameworks brag that they are GPU-rendered?

If you have free night charging plan you can recoup quite faster. My system paid off for itself in 3 years - although I designed and set it up myself, so it was about 3x cheaper than through contractors (and yes I'm in US and I pulled permits to do this).

You mean 2.25MW chargers: https://electrek.co/2026/09/30/chinas-2250kw-chargers-are-al...

Maybe civ 5 is more reactionary than strategic, so simple strategy that doesn't look too much into future works reasonably well. Or maybe civ is more mainstream and models absorbed more knowledge about strategies that work reasonably well.

I just started live streaming another game - now played by Muse Spark 1.3 (much cheaper than Opus 5.5):

https://www.youtube.com/live/jRi9cdo95As

Interesting that this is on front page. I thought HOMM is quite niche right now. I'm still spending some time on HOMM3 and recently created mcp for Claude Code (with opus 5.5) to play it - you can look at short playthrough here: https://www.youtube.com/live/uU-vvTOZFQY

It is not good at it! And very slow! I tried to improve, teach strategies, write some analysis scripts, etc. Did not help much.

My experience is that agents are decent in coding a feature, solving a bug - doing specific well defined task. They are very bad at judging UI/UX and user flows through app. Like really, really bad. They can quickly put together vibe coded app but make it usable takes a lot of extra manual work.

Did anyone solve this problem? Any good approaches?

By the way - Opus refuses to help you with setting this up - I used grok to drive this.

qwen3-asr-1.7b and parakeet both allow context/hotwords to bias transcription

Here's how to root echo show 8 1st gen: https://xdaforums.com/t/unlock-root-twrp-unbrick-amazon-echo...

responds within 1-2 seconds including 0.5 second waiting for silence as a trigger to start processing.

I confirm this as my experience as well. Parakeet did not perform well for me (English with heavy accent). Whisper large is much better. But once we get to model of this size look at qwen3-asr (1.7b parameters). I'm quite happy with how it performs (both transcription quality and speed on rtx5080). By the way the model is also available on openrouter and very/very cheap. Performance is reasonable (<1s response), but some requests are delayed (>10s)

Just ran some tests on my phrases:

- my whistle setup (tested can run on echo show with <1s response): 206 correct out of 208

These were not tested on echo show yet - on my pc for now:

- Vosk small, phrase grammar: 184/208

- Speech-to-Phrase 1.4.3 (Kaldi): 118/208

- suggested sphinx: 63/208 and on some examples took 13s on i7 14700k pc

looks like my customized whistle works better for me than these alternatives. But more testing wouldn't hurt.

It is all custom made and not very reproducible. I'm working on reproducible setup and once it is done will publish it here: https://blog.kvit.app

Interesting that this is here. I used whistle (and bunch of other things) to take ownership of my echo show. It now doesn't dial to Amazon at all - it does all processing locally with its own CPU and connects to my homeassistant for home automation. My initial setup involved qwen asr (1.7b model) running on rtx 5080. Compared to that, whistle was really bad (out of 170 messages, qwen recognized correctly 168, whistle - 70), but I adjusted whistle to work like jev - instead of free form transcription it recognizes only select set of templates (I trained tiny network with 10,000 generated utterances to translate whistle final state to probabilities within templates). The precision went up to 164/170 - almost matching qwen. By the way - I'm speaking with heavy accent.

This is discussion and not mechanical - surely I would understand this and suggest second favorite from not long ago.

I did lots of interview on both sides and know how often interviews are just luck - got asked trick question that I knew answer to, coding problem that involves algorithm I just recently used, etc.

There is so much randomness that I, as interviewer, decided on the following strategy - put candidate into best possible position and judge from that. I settle on one set of questions - "tell me about your most favorite project, why, and let's discuss in detail". If candidate knows his/her stuff - this is fun/informative discussion. Also easy filter if a candidate cannot say much or doesn't understand details of project they consider their favorite.

1 line in CLAUDE.md I tried (one line, 100 lines, many iterations) - it not always follows style instructions.

The value is very large for me - instead of reading page of Claude produced prose, I mostly just look at focused choices and continue with single button press - significantly less Claude deciphering and typing now. In addition, I do my sessions with /rc so I get nice form on mobile app where I can just select a button instead of typing to continue work.

PS: I think there is confusion in term - "every turn" - I don't mean every tool call, I mean after minutes or hours of agent work when agent decides that its done it gives a report. The report I see now has structure to it with options set to continue (or stop).

What to trust? Closed hardware/software that goes to trash if something goes wrong, or opensource that you can send an agent to debug if there are issues? I'm not trying to diminish JetKVM's product - for $35 its a steal if you don't want to bother with DIY and time involved, but I'd strongly disagree that opensource is less trustworthy than closed product.

or make one yourself: https://github.com/espkvm/espkvm

If you have same schedule every week, don't go on vacation, your kids don't have breaks, don't have seasons, then 'set and forget' dumb thermostat will work for you without any problems. For the rest of us we need some 'smartness' in our thermostats.

it is, unless it set to xhigh - it really likes generating tons of tokens for its thinking. unfortunately, for decently reliable coding results you want it on xhigh ...

There was a study specifically related to Qwen3.8 27B that showed that kv cache quantization has almost no impact on this model all the way to q4:

https://arxiv.org/html/2609.04098

On many models that I tested in past context quantization had very bad effect on model performance. However qwen3.8 27b is different.

I'm now running NVFP4 quantized both weight and cache on my RTX5090 and getting excellent results: 264k cache allocated for pool, 10k tok/s prompt processing, 200 tok/s generation for single stream, or 801 tok/s generation for 8 concurrent streams. Also have about 2Gb vram left for use of OS.

my coding agents regularly reach 200k context used without noticeable degradation.

P.S. I used setup from: https://github.com/seanyourhighness/vllm-sm12x-nvfp4-dflash2

why so many people add 'please' when asking machine to do something? Was there actually research that when you SCREAM or curse it follows your instructions better?

P.S. Although my wife insists that I should stay polite in case AI overlords remember how I treat them ...

Interesting - in my setup (llama.cpp rtx5090 qwen-3.6 27b) prompt processing with mtp is almost half vs non mtp. Sounds like I need to investigate what is wrong.

"Multi-token prediction gives a free speedup of up to 2x on many models" - at the expense of halving prompt processing speed

deterrent - no one starts war with a country that has nukes

There was lots of discussion within Russian and Ukrainian war analytics that nukes (at least tactical) are useless in this war for the following reasons:

- they would not change much on battlefield - there is no large concentrations that you can nuke - everything is dispersed

- nuking urban centers again won't change much on battlefield but would alienate China

- Russia's equipment is known to be not most reliable/maintained and worst that can happen to Russia is them trying to nuke and nukes not working

It sounds more like market based allocation than gameification.