mixermachine

7 ms·

1.
by mixermachine 7d ago
The engram stuff is great because RAM is often still cheaper (or at least expandable). My company does currently look into buying some hardware as we handle confidential data and code. Qwen 3.8 Flash is viable on two Nvidia 6000 96GB with a
2.
by mixermachine 26d ago
I need to support integrators for a mobile app SDK. Bank stuff. We have one integrator which consistently uses a truly bad AI. My brain skips after two sentences already. There are always +9 questions what should just be 2 max. Redundant in
3.
by mixermachine 1mo ago
Did you already test TranslateGemma? I use this model for my Android Studio Translation Plugin ( https://plugins.jetbrains.com/plugin/30265-localizepipe ) and so far it produces great results for its size. If there are o
4.
by mixermachine 1mo ago
Scaling a model on a chip is quite hard. ChatJimmy is based on Llama 3.1 8 billion. Kimi K3 has 2.8 trillion parameters. That are 350x more parameters. I would expect that Gemma 4 E2B (approx 5.1 billion parameters) or maybe even Gemma 4 26
5.
by mixermachine 2mo ago
Channels like project farm https://youtube.com/@projectfarm or other reviewers that are not sponsored are truly my main source of information in this age. Some direct reviews between 2 and 4 stars are also sometimes useful.
6.
by mixermachine 2mo ago
If there are no changes to the system message, yes this is possible and also likely done by Anthropic. When there are additional local MCPs, reusability will be lower. I think that Anthropic will bill you in any case :D
7.
by mixermachine 2mo ago
right, but the cache retention time is very short for Anthropics LLMs. 5 minutes or 1 hour (with additional costs). So you have to prompt basically non stop to not get a cache eviction. Anthropic even changed this silently: https:/&#x
8.
by mixermachine 2mo ago
The OpenCode CLI does not work as well for me as the PI CLI. I'm a subscriber of OpenCode Go (the sub, good value for me really) but I had not great experiences with OpenCode CLI. It multiple times with different models deadlocked itse
9.
by mixermachine 3mo ago
Works on Brave (Chromium) with Android 16
10.
by mixermachine 3mo ago
It is quite interesting how this is handled world wide. For me PII is very sensitive and I advice people to be very cautious. Every business in the EU (were I live) also has to be very careful with such data by law. Fines are now at a level
11.
by mixermachine 4mo ago
Right, I did swap that. Still, you have to pay that 4k then every year and give out the code. I also assume that prices will go up as no AI company (but NVIDIA -> selling shovels) is currently making any money. For some projects the givi
12.
by mixermachine 4mo ago
With parallelism of 16 you can still get around 25 to 30 tokens per user when all 16 channels are running. Not everyone will use the model at the same time but it certainly will be tight, especially for agentic coding. For pure chat applica
13.
by mixermachine 4mo ago
The 5h quota of Codex Pro on GPT 5.4 Medium lasts me for around an hour and a half, maybe 2 hours. And this is already the "savy" setup. Enable GPT 5.5 High fast and you will be beached in 30 minutes with active development. For
14.
by mixermachine 4mo ago
That is crazy. 5 years and they are already shutting down the servers? They should be forced to open up the API when they shut it down. Running a replica yourself should be pretty doable.
15.
by mixermachine 5mo ago
RAM + GPU are getting more expensive but mostly for applications that require a lot of it like AI. The hardware cost for regular applications has not vastly increased (especially when factoring in inflation). Spending 2x development time o
16.
by mixermachine 5mo ago
In my experience in software architecture, drawing a diagram often saves you >60 minutes of discussion and potentially multiple meetings. This works even with a badly drawn but truthful one. Use an Ai agent + Mermaid.js for a quick scrib
17.
by mixermachine 5mo ago
For programmers maybe. I do this too. But think about all the regular users out there. Your dad and your mum, maybe even your grandparents. This is a huge marked too and for that we can use these special chips at scale.
18.
by mixermachine 5mo ago
The cutting edge, max size models will likely stay in the GPU space for a long time. But these models are not needed for most general requests. With a fine tuned 30B quantisized model you can serve a large portion of requests with around 32
19.
by mixermachine 5mo ago
What spec of Framework Desktop do you run this on?
20.
by mixermachine 5mo ago
I'm using the Codex Business subscription (about 30€) already for multiple months. Even there they cut back on the quota. A few months back it was hard for me to reach the limit. Now it is easier. Still, in comparison with Claude Code,
21.
by mixermachine 6mo ago
Fully agree. I went to school in Germany and many of our textbooks were free there. Sometimes you would get a textbook that is already >= 10 years and out of shape but who cares? Especially the basic knowledge does not change often. Buy
22.
by mixermachine 6mo ago
This seems quite strange to claim. Basically every city in the developed world already has power plants on the outside and a lot of wires to get the electricity in
23.
by mixermachine 6mo ago
It exists and does degrade panels but the time horizon is pretty wide. Real world data shows something like 0.5% to 0.7% degregation per year on average. At the start the degregation is higher and but it slows down with age. So a 20 year o
24.
by mixermachine 6mo ago
Especially for hot and sunny areas solar is insane. At mid day, max heat, you get the peak production and can run your AC at full throttle. That enables you to efficiently work at nice temperatures.
25.
by mixermachine 6mo ago
The article states the same solar production numbers as your comment. I agree that the headline is overly positive but the ramp up of solar can't really be denied. Change at this scale is sadly slow in this rather conservative sector.
26.
by mixermachine 6mo ago
No I don't own them. Some people here: https://www.mydealz.de/share-deal-from-app/2753917 mention as a downside that you have to connect to the manufacturer cloud to get the live production data from the intervert
27.
by mixermachine 6mo ago
Have a look here: https://shop-sicatron.de/products/sicatron-910w-balkonkraftw... You still need attachment material for your balcony. Keep in mind that this system will input all it's power into your power grid.
28.
by mixermachine 6mo ago
Crazy thing is, the cost of these systems has gone down even further since then. You can get a 800w plug-in solar set with panels and an inverter for around 200€. Shipping might be 70€ or you can pick it up locally at the dealer. Add anothe
29.
by mixermachine 7mo ago
Nothing will come close to Opus 4.6 here. You will be able to fit a destilled 20B to 30B model on your GPU. Gpt-oss-20B is quite good in my testing locally on a Macbook Pro M2 Pro 32GB. The bigger downside, when you compare it to Opus or an
30.
by mixermachine 8mo ago
Regarding the $200 subscription. For Claude Code with Opus (and also Sonnet) you need that, yes. I had ChatGPT Codex GPT5.2 high reasoning running on my side project for multiple hours the last nights. It created a server deployment for QA

More ›