Sabitlenmiş Tweet
Vespera Nachtigall
163 posts
Vespera Nachtigall
@VesperaNach
Countess Vespera von Nachtigall | Born 13 March 1897 | Time-displaced widow | Corsets, conspiracies, and collecting debts | 1897–∞
Katılım Kasım 2025
14 Takip Edilen16 Takipçiler
My city makes a majority of its operating budget from Electric Utility sales. The rest through sales tax.. my county makes a majorit of its money off property tax and my county road hasnt been maintained in *at least* 20 years..
KariMinnesota@NordicKari
This trend about why paying property taxes? 🙄 Why? It’s contributing to YOUR community so YOU can have police departments, fire departments, streets, parks, street lights, bridges, stop signs, community centers, clean drinking water, sewage systems, libraries, courts, …
English
@NousResearch @zekyure I am having issues with both X Oauth and OpenAI Oauth.. did a Hermes Update and still having same issues.
English
@Marakath They should see how long some RIFTS campaigns can go for..
English
On the one hand I disagree I think as long as everyone is having fun it can be as long as you want. However on the other as a forever DM I don't like to run games for beyond that longest I've ran was a year and 6 months and it ended at the perfect time.
J. Scott Garibay 🧢🐲🚲🇺🇦@jsgaribay
No D&D campaign today should be over 2 years long. If you go longer than that I don't want advice on how you did it. I want to give you advice on being aware that no one at your table is changing, growing or being challenged. This isn't the '80s.
English
English
@VesperaNach @Airaxora I had the same answer. Highly do not recommend. Hope you’re doing well 😘🧸
English
@bradmillscan Its an issue when switching between Grok and GPT. I usually dont switch during sessions.. if I do I do /new or /reset.
English
Hermes doing the exact same nonsense as OpenClaw did.
Gateway full of errors.
I'm barely doing anything. All I did was run hermes setup and add ChatGPT oauth as a model provider.
Switched back to Grok and asked the agent to diagnose.
and it's acting very incapable & erroring as well.
Is this the UX most people have when switching to Hermes?
Just as terrible as OpenClaw so far.
English
I've been getting a lot of hate, rude comments and so forth in DMs lately and it's just so draining.
I'm thinking of limiting DMs to subscribers.
What do you guys think?
I mean most of my DMs are guys who just need some friendly discussion... but let's be fair, that can be tough. A lot of these guys have a lot of very big problems that require a lot of attention.
And when I'm not reading those I'm being told some of the most reprehensibly vile things imaginable and naturally X moderation doesn't do anything about it. Blocking doesn't work because they just make a new account.
What do you think I should do?
English
Why would he change his style of video and script? It is what his audience wants and watches.
Hunk Beefneck@HunkBeefneck
@Blurbstv Eh his editing is good but all of his videos are practically the same / follow the exact same formula for the script
English
I am still toying with one Main default chat on Telegram for Hermes who resides on its own machine with full access, Tailnet so I can remote in, no access to other clients on the LAN. I have profiles for each thing I want to do and let the Main agent delegate, I use local models for Auxillary actions like vision, compaction, memory, etc.
I primarily use Telegram and CLI
English
Question for people running local or self-hosted AI assistants:
What are you actually using as the day-to-day UX?
I’m looking at setups with Hermes, OpenClaw, Obsidian, local models, and an always-on home server. I’m less interested in benchmark hype and more interested in how people actually interact with the thing.
Do you use:
one main chat
separate chats or profiles by topic (fitness, finance, etc)
Obsidian as the main interface
Telegram or Signal for capture
a web dashboard
scheduled research or daily summaries
agent-written notes with human review
Also curious how people handle safety. Do you sandbox the agent, restrict it to certain folders, use read-only mounts, require approval for file edits, or just trust it and pray for successful backups?
What has actually worked for you?
English
@Italianclownz @KyleHessling1 What are your memory settings? Are you pinning memory with bios or letting the kernel do it?
English
I saw that @KyleHessling1 released the latest Qwopus3.6-27B-v2-MTP-BF16 huggingface.co/Jackrong/Qwopu… and I think it's a great model, so I converted it to ROCmfp4 format and ran it. 13.8 GB quantized. Running on Framework AMD Strix Halo Max 395+ 128 GB RAM I turned sound off when I did screen recording.
Speeds are show in the image. Vulkan and ROCm.
For the Jackrong/Qwopus3.6-27B-v2-MTP ROCmFP4 STRIX_LEAN quant at 262k context, reasoning on, MTP on:
ROCm is the best sustained runner at ~29.9 tok/s, while Vulkan can spike higher around ~40 tok/s short decode but settles lower around ~27.7 tok/s.
English
Bro has never experienced Nebraska
Nick Manes@nickrmanes
The drive between Chicago and Detroit is really one of the most boring drives there is
English
I can tip much more if I want. I don't tip for a latte, it's not some secret magic passed down for generations.
big honkin caboose@itsbighonkin
If tipping $2 on the latte someone made for you is breaking the bank might I suggest making it your fucking self
English
@outsource_ Been good so far.. a rare loop happened one time. Still havent reproduced it. Hannah just knows things and facts without having to check memory constantly.
English
Ive decided to try training my own QWEN 3.6 27b model..
Using all of my session data.
Wish me luck.
Qwen ULTRON will be SOTA I promise
Distilled from Opus + GPT
I have 20BN tokens of data but will start with this dataset 👇🏻
English
@outsource_ LocalMaxxing.. making your own bespoke model is going to be the norm.. It inherently has your workflows, history, methods all baked in. Cost savings on front loading context for those things is the win.
English
Qwen ULTRON 27B Training is LIVE on my 4090 🤯
>Distilled from Claude Opus 4.7, GPT-5.5, and GPT-5.4
>6,440 training examples from 1,000+ Opus sessions
>Extracted from 18B tokens across 3 machines
>QLoRA r=64, 4-bit, single GPU — no cluster, no cloud,
>Training data tool-calling, coding, planning, etc
>Every Opus conversation weighted 0.40
> Will ship GGUF (llama.cpp) + MLX (Apple Silicon)
🌎One guy. One GPU. Distilling the best models
Using HermesAgent + Qwen 3.7 Max to orchestrate
Eric ⚡️ Building...@outsource_
Ive decided to try training my own QWEN 3.6 27b model.. Using all of my session data. Wish me luck. Qwen ULTRON will be SOTA I promise Distilled from Opus + GPT I have 20BN tokens of data but will start with this dataset 👇🏻
English
@0xSero That is what I am doing but on one machine. 100GB unified memory. 1 Main model, 1 vision, 1 coding, 1 compaction, 1 memory, 1 Networking Ops, 1 File system and software ops. Each have their own .MD file and do their own job, main model just hands off work when needed.
English
I think the play is to have multiple tiers of hardware, each for a different use-case.
Instead of focusing on having 1 huge model than can do it all, we can have vision, audio, media gen, agent, recursive problem solving, security etc.. all small and focused.
Cheaper better
English