hecturchi
Old issue, affects also web UI not only bridge:
https://web.archive.org/web/20240221213617/https://github.co...
Would you use a secure email service that breaks signature verification on received emails? There were a lot of smaller and bigger things along the years, but realizing I could not verify authenticity of signed emails for apparently no reason pushed me over threshold to leave them.
It is not more secure per se, replacement is done as part of LZB being phased out Europe-wide.
They could have waited for recovery etc. The issue is they switched to manual and then botched it. We don't know root cause yet (bad management, timetable pressure, lax work culture...?).
Same mistake was made with other train and the driver replied that that was not allowed, and the operator corrected accordingly. But the driver of the train that crashed didn't raise the issue.
As it looks now the controller made a mistake when taking over manual control of the segment after systems failure ("continue as normal" vs "continue slow"). The train driver probably should have also known they should not drive at normal speed without safeties in the segment, but I guess they trusted control had cleared previous trains.
Between 10-20 minutes control room did not realize a train was approaching an occupied section where the other train was holding on a red track signal.
Human error but signals serious deficiencies at protocol, oversight etc. In control room.
My experience is that draft-mtp=2 gives 25% improvement in tokens/s but the model is unable to call tools with the right arguments reliably. I have since then gone for smaller quants at draft-mtp=1 and that problem is at least gone so far.
An agent needs to repeat tool calls in the right way, so cannot penalize repetition. This however leads to the agent retrying the wrong calls constantly.
- Hard to benefit from thinking and preserve thinking given token cost.
- Low quants reduce accuracy heavMTP draft can make it make the same mistakes all the time when calling tools, formatting output or following basic guidelines. Otherwise 2x-4x slower.
- K/V quants probably quantized too make things less accurate.
Useful would be combinations with:
- Full context size so it can code and think a bit.
- Draft MTP <= 2 so it doesn't trip
- Q4 quants or better so its accurate
- q8 cache or better so it stays accurate.
- 20 token/s so it finishes while reviewing previous step.
- 1000 tokens/s context load so compactions don't waste 10+ minutes.
- And enough left RAM for 50+ context checkpoints so that it can progress quuckly.
Closest you have is Qwen3.6-35B-A3B-MTP.
Latest gens (Qwen3.8 and co.) are just too big for low specs. 27B dense models seem to be ok for integrated >=92 GiB RAM.
Source: I have low specs and tried them all for agentic use + coding.
I'm still clarifying the purpose of this before making a complaint to the data authority.
Now seriously, managing sovereign borders, enclaves and the people in them is not like playing monopoly. Except for Trump, that is.
Unfortunately bronze, with trimmed edges, common mint and worth very little. But if you tell me someone just stumbles onto and old coin in the street just lime that, I pretty much believe it.
I guess the downside is, well, this, specially having thought it is a good idea to rely on Wise as first and only option for business banking, with their reputation...
Also, don't talk about Molly. It's the best you can do for Molly.
Then I discovered that "encryption" for them was using Luks with the numeric unlock pin as encryption key (which most people sets to 4 digits). They marketed themselves as a secure OS.
No passphrase option. I brought up that it was trivially brute-forceable in the forums and they vehemently fought the idea. My post eventually disappeared.
They were doing government contracts in Russia etc. Iirc. I put some dots together and phone rests in a drawer ever since. I have no idea how things are now though.
Also, upgrades can bring issues and asking for help is super tricky as you may get banned for asking the wrong way. So better to not upgrade if it works.
My recommendation is to flash and GTFO forever. Never upgrade, never ask anything, never read anything other than the install docs. Disregard pointers to join Telegram. Do not read release notes. And donate some money for the trouble if things work. That's your path to maximum happiness with Valetudo, blissful ignorance.
One day perhaps there will be a fork and things can be normal.
Yeah, it must be that since US and EU are identical otherwise.
You can have a profile completely de-googled and another one with Google Services for apps that need compatibility but remains off most of the time. You can uninstall Google too.
Owning your devices is a superpower that 99.9% of people don't have, but you can have it. The irony is that it's thanks to Google devices. Let's see how long it lasts.
It's pretty slow yes. It's one of the worst RAM eaters and has been feeling very sluggish for a long time.