I might be misjudging the security risks, but I can't see myself running Claude Code with full access to my work machine, simply relying on its built-in sandboxes and permission systems. For a long time I simply ran it under QEMU, but that proved less convenient than desired. So, recently I put together a Bash wrapper around bwrap and pasta for sandboxing Claude Code. And while I was at it, I also organized a per-project split for Claude config files, so that sessions, say, and certain other files are always project-specific and there is no possible cross-contamination between projects. Sharing, for it proved non-trivial to run bwrap and pasta together: bwrap creates nested namespaces and so one needs to switch namespaces as well before pasta can be attached to bwrap. Now I can just do `jail project-folder` and then `claude`. The wrapper is for Arch, other distros might need some path adjustments.
Less knowledgeable overall, I would agree. LLMs are, if anything, the ultimate knowledge bases. But within our own areas of expertise, our worldview is typically more coherent and we have often learnt to reason optimally within that world view. To be fair, it is hard for me to generalize. So at least that's what I see in areas which I know. As pointed out by others, however, making discoveries that bridge disciplines is where LLMs might have an edge.
My pet theory comes to information density. LLM writing feels too shallow. If you read a text from a person who cares about her subject, there will be more useful detail and insight scatterd throughout. And the information within sentences will be more meaningfully connected. On one hand, LLM would know all that, but it does not seem to know how to distill that knowledge into a written text. I read what people write. If I know the text is from an LLM, I skim.
I am surprised it's just half. The incentives are not there for replication or verification. You need to publish and it is easier to publish something new than to fight over existing results, especially if you're a young researcher.
I maintain something akin to google trends but for abstracts in economics papers. The word "novel" is growing rapidly, while "reproduce", "replicate", and "verify" have not changed much over the years:
https://dubovik.eu/blog/repec?t=replicate&t=reproduce&t=veri...
For reasons of just trying stuff out rather than for utility, I have a set of Python scripts to generate books from their titles alone using an LLM. There are hundreds of similarly titled projects on github. This weekend I generated a new book, "A Tinkerer's Introduction to Claude Code". I thought it would be fun to use a Claude model for that, so that's what I've done. (IHMO, Gemini is better at writing and cheaper for that purpose as well.) The book is dull, full disclaimer here, but it has a few bright moments. Not least,
"Set aside time—perhaps one small task per session—where you write code unassisted. Treat it like exercise."
There is also the question of real estate. I can have several paper notes side by side (when taking notes on loose sheets) but with iPads or ReMarkables that'll be rather decadent.
I have a dedicated couple of pages in a notebook, where I write down the note-taking conventions I use. When transitioning to a new notebook, I would copy those pages, possibly making a few improvements based on past usage. A most unhurried release cycle, if I can say so myself.
Regarding the space management, there are many solutions straight out of the programming world, of course: utilize both sides of the notebook, reserve a minimum number of pages per topic, keep an index with free pages, etc. But there are some hardware ones as well, I'm trying Atoma notebooks (https://atoma.be) these days.
In my limited understanding, with zero-knowledge proofs. That's also what the blog mentions. Say, a website issues a challenge, you sign the challendge with your private key supplied by either your government (or some other authority). The signature is verified against a big set of public keys using a ZKP without knowing which specific public key matches.
I imagine the problem is more economics and politics than technology. Say, personhood verification becomes available. Then it is also that much easier to mandate age verification, and that becomes a slippery slope.
"Careful design of the protocols maintains privacy by preventing sites from learning any additional information beyond personhood." Is personhood verification desirable? Say, HN introduces personhood verification and budgets the number of posts or comments per person. That would limit the amount of AI posts, which seems to be desirable judging by the recent discussion (https://news.ycombinator.com/item?id=47340079). What are the drawbacks though?
Starting with the CTAP 2.1 standard, there is a largeBlob extension in the standard that allows for storage of a small amount of arbitrary user data. The current libfido2 library implementation uses a largeBlobKey provided by the authenticator to encrypt the data in userspace before storage. However, at least on a Yubikey, obtaining the largeBlobKey does not require User Presence, which IMHO lowers the essential security guarantees that a hardware authenticator normally provides. Recently got myself a set of Yubikeys and tinkered with them for a while. It seems it is also perfectly possible to use the hmac-secret provided by a Yubikey to encrypt one's data in user space and then store it in the largeBlob array on a Yubikey. Importantly, requesting the hmac-secret does require User Presence.
That combination is hardly documented anywhere. And so I somewhat skeptical about my unorthodox security solution. Sharing on the off-chance a security expert would comment.
The prompts are slightly tailored towards a textbook but are otherwise generic. If you have a Gemini API key, you can specify any title and get your own book. It can take several hours for a larger book (can be done faster and slightly cheaper with async queries, but I thought that an unnecessary complication for a small pilot). It is easy to disagree on the resulting quality of the book; for me it feels more like an extended Google query than anything. But as an exercise where the size of your prompt stack is larger than 1, it was fun and, in parts, thought-provoking.
If I am to name just one prompting trick that I discovered to be useful, that'll be hierarchical budgeting. I have a total word budget for the whole book and I ask Gemini to allocate that budget for chapters, and so on recursively. Fun part, the sum of parts does not always equal the total, but the difference was never bigger than 5%. That's some stochastic summing happening.