@HackingDave@faanross@HackingLZ@cantcomputer GLM tbh I've been using (non local) for a few months alongside Claude, however unlike Opus it has been getting less degraded over time.
@HackingDave@HackingLZ@cantcomputer Interested if you could do a blog on intergrating them, as small scale H100 in non-cloud scale is rare as hell. Particualrly how you end up buildin the cooling loop. Definitely if you have use more cost effective to buy than rent presume you are also power limited?
@HackingDave All right if you can get the signal -> I've a similar setup using Mikrotik LHG Dish antenna (5g in EU) to get streamable data in a literal no-signal on any device zone. Have to setup poles to do on a temp basis, but it allows wifi in places otherwise impossible without starlink
@ejae_dev@gautirao@zarazhangrui Exactly how I've been working. Memory files generated from local daily summaries etc. It means claude basically looks up in the right folders now for the memories of that project for me and loads when I mention a project.
@zarazhangrui The trick of nesting your local memories, agent skills, and then having seperate github repos under /projects/projectname where each projectname is a seperate remote git connection - gamechanger. The seperate repo for the claude history then keeps everything off the project git
@zarazhangrui I have a daily-summarise command that does similar. It then whacks it into memories (after simplification) and times out stuff automatically after a month (and it knows about the memory). https://t.co/c6JiadgYq8 if you arrange repo similar to this it helps Claude.
@acdlite https://t.co/c6JiadgYq8 is where I learnt the trick -> but I've renamed mine, not ideas and spaces on mine. (but ideas is the meta repo for all the tracking/memories) and spaces is the place you link to any other remote repos for actual work. It handles *many* ongoing project.
@acdlite in effect then all changes auto tagged with TASK/BUG/SPIKE numbers on a per project basis (on the project repo) which directly link to the same numbers in the meta repo. Claude (at least) handles this fine and will autosearch both places with right config.
@bygregorr@heygurisingh Hate to say it, just a happy customer (not even doing a referral link) but try https://t.co/zjfN88ZbEC (chinese model) with Claude. $30 a year is about same as a $100/mo Claude acccount, adn the $188 a year model is the $200/mo equivalent. It is slower than Opus mind.
AI speedrun -> fastest time to use a Claude Pro ($20 plan) session window so far: 10 minutes for me. (yes full 4 hour session done in 10 mins). Love Anthropic models, even a $100 plan would be zero good for my current project though :(
@essobi And... when you as a generalist have a client that ends up having an obscure technology you are able to help where consultant Y from the sophomoric genration can't. Kernal performance tuning is better than a bigger instance in most cases... as a classic example.
@noottrak@HackingDave The other thing is given I have a task /complete skill, that automatically runs security analysis specialist claude skill, I'm finding stuff I missed years ago. Not like we can all afford the expensive static analysis services for personal stuff is it?
@noottrak@HackingDave I think thats the main thing I've noticed, I've in past week restarted about 5 dead projects that I'd had partially done before. Mad thing is I've rewritten one in around 2 days with claude, where the inital draft took me 3 months.
TLDR the new 2 day one is "better"
@HackingDave@HackingLZ I am still at a loss to people thinking it's a good idea. Sandboxed to hell it's great. (Especially given I'm also using Chinese AI's to test them against Anthropic given the price points).