fdeage

Everyone in town is talking about running OpenClaw/Clawd on a VPS or a dedicated burner machine (e.g. a Mac mini), for obvious security reasons.

What I don't see discussed much is Docker. On macOS it already runs inside a Linux VM, and the setup can hardened even more (in principle): no bind mounts, no /var/run/docker.sock, non-root user, read-only filesystem, tight resource limits, restricted networking, etc.

Given that, what are the concrete reasons people still consider Docker unsafe?

From an 2018 email from Musk to the OpenAI leadership (https://openai.com/index/openai-elon-musk/)

I asked Claude, ChatGPT and Grok the following: "List the top 5 possible options for [redacted], with an associated probabililty for each. It can be a company or a person."

I got a mix of "DeepMind", "Demis Hassabis", "Larry Page" and "Alphabet/Google". Surprisingly, Grok doesn't mention Hassabis, but FAIR (Meta) instead.

Any other idea from outside the Google cluster?

I have some old recordings of a SATB choir (Soprano, Alto, Tenor, Bass), but the audio quality is quite poor. I also have the exact musical scores for each song.

I'm looking for AI tools/services that could improve the clarity and separation of each vocal part (and maybe also reduce noise and artifacts). If it can leverage the scores to guide the enhancement, the better.

Has anyone tried something similar or can recommend modern approaches or models for this?

Not so long ago, we used to think that machines could be considered intelligent/creative/conscious/<some attribute deemed quintessentially human> if they could write a sonnet/summarize a book/paint a portrait/tell a joke, or whatever. Back then, all these tasks seemed technologically far beyond our reach.

Today, some of those have been achieved: even if they don't produce truly _good_ poems, LLMs are already better than the median human effort in poetry [1]. The same goes for paintings, essays, and so on.

Yet humor remains a tough nut to crack. LLM-generated jokes often suffer from:

  - off rhythms or pacing
  - missing or telegraphed punchlines
  - unintentionally absurd twists
  - over-explanation of their own process
  - reliance on weak wordplay

They do a decent job at generating satirical, Onion-style headlines, but longer humorous narratives still elude them.

So, could humor be the next frontier for LLMs? That is, not as evidence that they are intelligent/creative/conscious/<some attribute deemed quintessentially human>, but as the next area that is easy for humans yet hard for machines. Or am I just a bad prompter?

[1] Humans actually seem to prefer AI-generated poetry: https://www.nature.com/articles/s41598-024-76900-1

I'm looking for a text editor similar to Cursor, Trae, or VS Code+Copilot, but one that would be open source and could use a local LLM as the backend.

Trae looks promising, but I'm not comfortable with the idea of sending my code/analytics/telemetry/whatever to the editor maker and the LLM vendor.

Are there any reasonably polished options available yet, particularly for macOS?