Current #jjnotesfollow aliases:
# The canonical tug
tug = ["bookmark", "move", "--from", "closest_bookmark(@-)", "--to", "@-"]
# Show my current work-in-progress(es). To wip it, whip it good.
wips = ["log", "-r", "bookmarks() | description(exact:'')"]
For sure. We're well past the point of being able to reasonably claim that the tools have no practical use.
But I share the article's skepticism that LLM-based AI will "produce a fully automated drop-in replacement for most knowledge workers". Which it at this point has to in order to justify these companies' stupendous valuations.
Tomorrow no but there will always be the โhopeโ which is enough for the next 20 years before giving up. Which meanwhile would be replacing larger and larger chunks of the workforce so it wonโt really matter, weโll be adjusting our lives around it as well so we meet in the middle.
Tesla has been overvalued for over a decade with the promise of โfully autonomous carsโ and itโs still not losing its value, despite not meeting that (and probably never will, as the computation necessary to achieve it is bounded by physics).
That's a different situation. Tesla's a public company whose valuation is driven by the market, which has the right to be irrational. My understanding is that the AI buildout is financed such that the companies need to turn a profit by ~2030 or else they drown in capital costs.
Or maybe the frontier labs become Tesla-like, if it turns out investors just want to believe in the dream and don't care about financials that don't make economic sense. I just don't see that happening, given current sentiment and the overwhelming unpopularity of AI in the West.
Under the overarching influence of a single, powerful force, China's open-source models filter out a large amount of content unfavorable to their designs during the pre-training stage, and then incorporate a significant amount of content defending their designs into the data constructed during the SFT stage. A corrective dataset is urgently needed for fine-tuning.
Something with people, bartender maybe, or something with no critical thinking, delivery person, taxi driver and such. Or I go back to being a DJ as I was doing when I was young. #careernotesfollow#jobsnotesfollow
รฉs รญgy nem kell egy ilyen zรถld modulba applikรกlt, baszott nagy integrรกlt chipet venni hozzรก, pedig nem is nรฉz ki rosszul, de jรณl sem
and that way you don't have to buy a fucking big integrated chip applied to a green module, even though it doesn't look bad, but it doesn't look good either
I really like this minimal approach, I wish the maintainer great success with textlog.
whatever that may look like to them :)
I will make sure to check in occasionally.
Really glad to be using Haskell for work - especially in today's climate with agentic coding.
Our team uses containers to keep our development environment consistent across team members, though, and most agent sandbox approaches want to either offload your work fully to the cloud or they want to sandbox the agent harness process itself.
We're not ready for full cloud based development (yet?) and sandboxing just the agent doesn't work for us when the agent needs to run docker.
#haskellnotesfollow#ainotesfollow
... because letting the agent run docker gives them a too-easy escape hatch from their sandbox.
So we're currently stuck with using VMs. This turns out to be tricky but seems do-able.
My current approach is to provision a VM image and use Incus for the execution. My main goal is to keep my current workflow of local/non-agent dev working seamlessly and share the source with the VM/agent.
This means sharing my working dirs with the agent VM but overlaying artifact dirs from VM-owned block devices so we don't take too big of an IO perf hit.
I also don't want the agent to be able to git push so that means we need some quirky git config to rewrite origin paths and thread through some agent specific PATs ("thanks" GitHub) for read-only repo access. These ALSO need to be threaded through to the development containers - they need to clone private repos.
... in the end, this seems to work but I haven't lived with it for long.
There are still some quality of life issues to work out, but I think it'll work out.
At this point, I've removed all coding agents from my host - and that's nice.
It's tangential to the sandboxing topic, really. I started the first post with a broader intent in mind and then elaborated on the sandboxing setup I'm working on.
For AI-assisted coding, though, it's nice because Haskell provides a lot of feedback earlier in the development process compared to lots of other languages. It tends to work once it compiles (obviously not all the time).
I finished Rejection by Tony Tulathimutte and now I'm struggling to take any other book seriously. I need a pallet cleanser or something before I can move on
The best Mcp server I have as Claude Connector is Emacs running on my Homeserver. It has one tool - eval-elisp. Using it Claude can do almost anything. My top 1 pinned chat is โRun agent-start-claude-code using HotterEmacs in dir โฆโ - after that I have a ready to use remote-controlled session running on my homeserver.
This sounds like a really interesting idea. What would you say are its advantages over having the agent use the shell, write shell scripts, etc? Just your personal preference for elisp or something else?
To load a skill I literally type - โload (โdoc 'slack) and browse channel X for todayโ. The idea is my text stays English but repeating concepts are real reusable elisp functions that stay fully deterministic. (โdoc 'agent) and Claude knows how to talk to open code and pi agents.