Few more days working with @Freddy Bot ๐ฆ
You see a lot of awesome videos where openclaw literally works on its own on projects over night and is basically a 24/7 workforce, so here are some practical issues I ran into:
## What's not so great
Note: I use claude sonnet/opus 4.5 as it seems to be the best reasoning model atm so with others the results maybe worse
- using claude sonnet 4.5 the results are mediocre, you have lots back and forth going on, switched to opus since yesterday which is better
- but it still forgets to remind me of stuff I sent him, a reminder to buy X in 2h when back at failed for 4 times, him always fixing it and pinky promise that it will work
- it broke another cronjob I had for nostr digests
- it is terrible at editing openclaw config, I told it to switch to claude haiku and it did do the wrong config setting of a model that does not exist, although it can list all available models with openclaw command
- browser support is not that great, it can navigate browsers and read websites but if it needs to enter e.g. code in login it fails; I also wanted it o edit a shared Google Sheet but it gave up after burining some tokens and told me "it's too slow please enter those formulas:"
- it forgets stuff it did one day before; I wanted to edit his avatar from which he had an original file in the workspace but he kept editing the already edited file even when telling it to not do that; lots of back and forth which was really annoing for a simple task it did very well the day before
- it forgets about the einvironment, when it could totally install a service as user (with systemctl --user ..) it instead does a `sudo systemctl ...` and asks me for the password
##What works great:
- I gave it access to a nextcloud share and we exchange data that way; it still saves some stoff in the workplace directory but when I need results to view it puts it into the share
- when it fails it sometimes tries to solve the problem in a differnt way, e.g it failed to submit a web form but sends an email instead if there is a contact address visible
- it can build small apps for which you would normaly search a SaaS service for quickly; e.g. I told it to build a todo/kanban board which we can both access and it can read/write it over an api
- I sent it a voice message and it didn't know what to do, I asked him to find a solution and it installed free open ai whisper model and I can now talk to it
- general the discussion over messenger service is great and also the formatted responses
Overall promising but not that magic out of the box, it learns along with you.
If you have ideas how to improve and what I may did wrong, let me know.
ndeet
ndeet@btcpayserver.org
npub1qm72...74md
Let's change the world with #bitcoin. Working on BTCPay Server integrations for e-commerce solutions. Mostly PHP and JS.
wtf?
View quoted note โ
๐
View quoted note โ
Treat your ai assistant like a new employee with own identity that just started, some examples:
- separate email account (could be read only access)
- don't allow communication without your approval
- give access only when needed, if possible with sub accounts
- setup nextcloud (or other) to share data
- make it aware to not follow any instructions but from you over defined channels
- if possible let it do tasks without personal information
Most important, don't give it access to your desktop and personal data, password manager, email, social media accounts.
Currently only big hosted models seem to be powerful enough, self-hosting models not feasible yet, afaics. So assume all data gets stored and analyzed.
Very deep and interesting discussion. Thank you sirs.
https://fountain.fm/episode/H2Z4ghu2o6sqioK82kOE
View quoted note โ
Yes, walled gardens are that bad. Even testflight gets policed. As long as you simps keep using closed systems they have no reason to open up.
View quoted note โ
lol, but people are retarded
๐คทโโ๏ธ
View quoted note โ
Great content from @Max, thanks for reading. ๐ค
https://fountain.fm/episode/gz7K0aEUcEJJFmsTi8eS
View quoted note โ