I'm running PBS on a separate mini pc, it is the perfect forever near instant incremental backups that sold me. I did perfectly recover in 20 minutes after a full hardware failiure. Best recovery experience ever.
Let's hope it did not special case (overfit) for the specific tests. One of the failiure modes that in my experience takes the most effort to mitigate for.
"My main project right now is to establish a framework for large-scale, unsupervised code generation in our codebase...
sifting through the unsupervised agent’s (Qwen’s) output"
Ouch. I love my local AI setup with Qwen, but that is a mismatch right there. That model is not the right match for that project. It's like trying to develop a major software solution by just throwing in hundreds of fresh junior programmers and have them spew out random code bits, while what you needed is a good PM, a great architect and a handfull of senior engineers. Might as well pack it in for a year until your model has grown into the ability for those rolls. There is a reason why Opus 4.6 and now Fable dramatically changed the SWE capabilities, and IMHO Qwen is not there yet.
I must admit lingering long since retired 'memories' are currently one of the biggest pitfalls of the setup. Wiping all 'memory.md' often leads to better sustain.
My development platform of choice is Opus 4.8 --effort max. My 'in production' general model just switched from GPT-4o to GPT-5.2. I probably should test local inference for that part (some translation/knowledge extraction/summerization) next but it has not made it to the top of the priority list yet, even though 3 other small specialized task models in the app already run on prem.
Remember the Clipper Chip back in 1993 because personal computing was considered too dangerous a technologie to be accesible un-controlled/suprvised to the general public?
So a model named 'Fable 5' is "back". Both excited, as the previous Fable 5 I had access to for just 3 days was fantastic, and anxious as (yes, in my 1 person anecdote) the model refered to as 'Opus 4.8' was stealth nerfed over the last 2 weeks to a degree I had not experienced since the massive nerfs back in the GPT-3.x days. Fingers crossed.
Will they not just argue that you could share the assertion, and hence we need a 'trusted' verfication point to establish it is actually you in posession of the zkp token, right now. So turn on that smartphone camera right now and obediently follow our biometric verfication instructions ...
I hear that often, but what does that even mean? I am a great proponent of open weights models. I do believe they are the only reason we have not stagnated into a collusion of halting (public) model releases.
But exactly which point in time is z.ai compared to claude.ai? Consistently bring "6 months behind" in an exponentially acellerating evolution means the gap is growing exponentially wider, not constant.
In my experience: anything of open-ended complexity (software development, research, product design, ...) benefits from wathever the frontier can offer. 95% of Line of Business automation and workflows can be handled by even a reasonably small open weights generalist model flanked by a few even smaller specialized models. Yes, designing such a setup takes more knowledge and work dan just chucking it all over the api with prompts. But that is how I can run a system here for <$30/month vs >$1.000 month. As an added bonus, no model server can shut me down at the drop of a hat.
For real production I find the switching cost is not as trivial as you portray. Even going to a new model version in the same model family, say GPT-4o to GPT-5.2, a transition I just finished on a not too complicated application, requires extensive retesting and tweaking of prompts, guardrails and parameters.
Because it truly makes a difference. Opus 4.8 was great until we experienced Fable 5.
And post Fable retraction, I am now most certaily noticing Opus being 'dumber' also.
Open Weights are good. Not (yet) as good as leading closed models. Unfortunatly they will be declared 'illegal' any day now, and I unfortunately do not see myself able to run GML 5.2 in my basement homelab any time soon.