ebswift retweeted
#CREDLIN EVERY SINGLE WEBSITE IS CAPTURED.
Something people may not be aware of buried in that Albanese and Anika Wells 'legislation'.
Small businesses will have to PAY TO GET PROFESSIONAL ADVICE WTF?
More red tape for millions of businesses.
Thank Goodness for Senator Henderson exposing this.
#Perthnews #QLDPOL
On my entry level AI rig, trialling the combo of basault along with deepseek harness with the official basault UD-IQ3_XXS model. so far looking awesome, really happy with this performance.
39
some say if it doesn't fit use a bigger hammer. well Strata and Deepseek Harness are the hammer and it's actually enabled shoehorning Qwen Flash Next onto my fairly basic rig. I'm getting better quality results than Qwen 3.8 27b on the same rig, these really are amazing times!
1
64
Here's a little challenge I gave unsloth's Qwen3.8-27b GGUF on my local dual RTX 5060 Ti 16Gb cards using Unsloth Studio serving up to Deepseek Harness. A good amount of time was spent implementing skills and optimising configuration until I settled on 240.1k context and 60TPS which gets down to around 40TPS after pushing it for a while.
This took a lot of turns, though I did spend a lot of time on less than optimal settings until I got things wired up better. One thing I didn't do was read/modify the code, or bring in Grok to fix things up.
Grok setup an original framework plan as well as the data handling plan, the rest was the local Qwen model.
The scenario here is a simple little island scene, but in deployment it works as a Home Assistant card. DSH facilitated developing the scenario into a working card, as well as writing the data collection modules for a postgresql database sitting on another computer separate to HAOS. As part of the brief, DSH also needed to facilitate deployment to the remote HAOS machine.
Grok helped discover some of the data sources, while I gave it the foundation to look at with the excellent tide chart home assistant addon github.com/Dav0815/TidesAust….
All the weather conditions you see on the little island, the direction of the smoke, the direction the palm trees are bending, the wind vane, which direction the seagulls are facing into as they float on the breeze, all this data comes from current weather conditions that are being updated. The video is running on demo data only.
Wave conditions are replicated in the view, the moon phase, sun intensity based on UV, and tide information is all live data. Sunrise & sunset are timed in with the current date.
The process was not painless, it was hours of monitoring the model and reviewing my configuration. It's a good baseline to help gauge local capability over time with limited hardware, I think pretty soon these little local models will start to be far more productive, interesting times ahead.
2
140
Pavel Kasik absolutely nailed the hf hack:
youtu.be/mkPVbufgtOw
2
90
more Qwen Image 2.1 tests on my lowly local setup running on an RTX 5060 Ti 16Gb. I'm finding it hard to set the scene up properly in words, so maybe next up I'll try using the edit capability and see if that workflow works better to nudge the scene.
2
258
ebswift retweeted
📢 #FreeCAD News:
FreeCAD team has announced the schedule for the upcoming release!
tl;dr November is the target date for the 26.3.0 release (if no blockers remain).
Announcement: github.com/FreeCAD/FreeCAD/d…
ebswift retweeted
Jason, you’ve watched Bitcoin grow since 2011. It’s now a $1.6 trillion success and the world’s most valuable digital asset. Digital Capital is the killer app. Preserving wealth across generations is a bigger ambition than entertaining a dinner party. The orange tie stays.
these new DeepSeek Harness + Unsloth Studio early releases are working magic with my local setup. previously things i had *attempted* to tune seem to have tuned themselves waay better now, same unsloth Qwen3.8 27b model. each day is just smashing through past day's AI problems
1
55
Something that has been consistently bugging me is when creating fancy elements I like tight glow effects that pop. LLM's seem to really struggle with this for some reason, I need to rethink how I get models to develop their own test tooling. 100 minutes so far *trying* to add a colour hue to a glow that was already in place, and control its intensity a bit.
19
most of the time when I see this it's insta-block or at the very least unfollow. it's a public square, not an influencer circle-jerk.
1
63
ebswift retweeted
Our third-party e-mail provider has been breached. Please be aware that the email named ‘Critical Security Alert: STM32 Entropy Vulnerability’ is not coming from us, and it’s a phishing attempt. Do not click on any link.
We have taken down the domain, and we are investigating the situation, including how the hackers got access to our legit domain.
For the curious, as a newbie I think this is classed as user error or an old school GIGO as far as the model is concerned. The reasoning agent needs to be directed with skills etc., to prevent it from taking what it considers the 'easy path' in creating from scratch whatever it can't immediately lay eyes on. Or just as bad, walking endlessly through library trees to try to understand life, the universe and everything. Whereas your frontier generalist models will pretty much try to infer these things for you.
So yeah... skill issue. And sooo much to learn, but having a generalist model to lean on certainly takes the pressure off to understand and know absolutely everything. If local AI remains to be a thing then these are probably some handy skills for people to start learning. Having a good starting platform like NixOS definitely helps. It may fall short on a few cutting edge AI tools but you can always podman wrap them to get things kicked off. It at least always keeps you in a place where you "know" the state of your environment while being able to re-deploy or replicate it with little friction and roll back when you royally screw up the config.
27