samuel

Do you realize people is using LLM's as classifiers, right? For lots of companies and developers reaching an API is feasible, while running a training pipeline, no matter how simple, is not. I know that they should still be gathering data for evaluation and they could use it to train a model instead. But they won't do it, for a variety of reasons.

This is the same route but WAAAY faster and cheaper. And you can modify it like you do with code or prompts. It's really appealing, TBH.

No to confuse with Ax, the DSPy inspired agent framework

https://axllm.dev/

I have this idea of using an obliterated version of this model for cyber work(or even this one, seeing that its guardrails aren't that strong) in a harness with the ability to spawn SOTA level subagents, faster and more capable.

The rationale is that the manager model sees the big picture and knows that the task is "unethical" while sota models are just given very isolated technical tasks that don't trigger any refusals.

Has anyone tried this? I would love to know about previous attempts of this approach.

The first "Morris worm" of the AI isn't far away, IMO. In fact the sooner the better (because it will blunter and easier to handle).

I don't think it's the case, but it would be very funny that this would end being AI generated clickbait.

You can connect it to any anthropic compatible endpoint(kimi allows this) but it's a weird choice, given that Open code, pi.dev and others are open source.

Chat is certainly an option, but the real deal are agents, which have access to way more sensitive information.

Very timely, the final report has been released today.

I hadn't read the document you referenced, and I admit don't have the prior knowledge, nor the time, to fully understand all the implications of what it says. My opinion is then the result of reading and listening a variety of experts and news sources, and it will have some biases, for sure.

Still, I have skimmed the final report to see if there was something that I could understand from first hand (and to support my original point, not gonna lie), and I found this:

_The increasing penetration of variable renewable and distributed generation, further market integration, broader electrification, and evolving environmental and geopolitical risks place the European electricity system under increasingly challenging operational conditions, requiring higher levels of resilience._

Do you really think that my original point (as uniformed as it might be), namely, that the levels renewable energy currently present in the spanish grid require significative investments, was wrong?

It's not that easy, and the 2025 blackout good evidence of that. Renewables need a grid that's engineered for them and that require significative investments. Without them, closing power plants (of any kind) is, IMO, nonsensical.

Ironically, Spain has plenty of Uranium, but there is an environmental law that doesn't allow its mining.

https://alpoma.medium.com/uranium-in-spain-8ef975763257

This country is crazy.

I assume that -p is the same that "codex exec".

The difference is that in this case the agent loop is executed, which has all the caching and behaviour guarantees. What I assume OpenClaw is doing is calling the endpoint directly while retaining its own "agent logic" so it doesn't follow whatever conventions is the backend expecting.

How important is that difference, I can't say, but aside the cost factor I assume Google doesn't want to subsidize agents that aren't theirs and in some way "the competition".

Is this source trustable? (I have no idea, I'm not german)

https://www.welt.de/vermischtes/kriminalitaet/article2521783...

Your question So does nobody in Europe use an EDR or intercepting proxy since GDPR went into force?

Given that a regulator publishes a document with guidelines about DPI I think it rules out the impossibility of implementing it. If that were the case it would simply say "it's not legal". It's true that it doesn't explicitly say all the conditions you should met, but that wasn't your question.

I have found a definite answer from the Dutch Protection Agency (although it could be out of date).

https://english.ncsc.nl/binaries/ncsc-en/documenten/factshee...

(IANAL) I don't think there is a simple response to that, but I guess that given that the employer:

- has established a detailed policy about personal use of corporate devices

- makes a fair attempt to block work unrelated services (hotmail, gmail, netflix)

- ensures the security of the monitored data and deletes it after a reasonable period (such as 6–12 months)

- and uses it only to apply cybersecurity-related measures like virus detection, UNLESS there is a legitimate reason to target a particular employee (legal inquiry, misconduct, etc.)

I would say that it's very much doable.

Edit: More info from the Dutch regulator https://english.ncsc.nl/publications/factsheets/2019/juni/01...

I agree with the sentiment, but I think it's a pretty naive view of the issue. Companies will want all info they can in case some of their workers does something illegal-inappropiate to deflect the blame. That's a much more palpable risk than "local CA certificates being compromised or something like that.

And some of the arguments are just very easily dismissed. You don't want your employer to see you medical records? Why were you browsing them during work hours and using your employers' device in the first place?

Recently learnt about tailscale funnel, and I love it, I would use for everything.

tailscale funnel --set-path <secret> <DIRECTORY>

(The path is needed because there are lots of bots who scan tailscale hostnames).

This works if the sender is tech savvy (and a tailscale user) but not in the other direction.

GPT actions allowed mostly the same functionality, I don't get the sudden scare about the security implications. We are in the same place, good or bad.

Btw it was already possible (but inelegant) to forward Gpt actions requests to MCP servers, I documented it here

https://harmlesshacks.blogspot.com/2025/05/using-mcp-servers...

Custom connectors are cool and a good selling point but they have to be remote (afaik there is no Le Chat Desktop) so using it with local resources it's not impossible, but hard to set up and not very practical (you need tail scale funnel or equivalent).

One of these obviously hasn't happened, but it might be, hence my excitement. I don't know how likely the experts think it is (~1%, ~10%,etc...) but I guess the odds aren't high.

With regards to the other one (AI), I did not claim anything else than a subjective assesment. I did not expect to see an AI capable of mantaining a conversation aloud, for example. May be I'm easy to impress.

I had never expected that I would witness in my lifetime such advanced AI, and now (may be) extraterrestrial life.

If confirmed, the only remaining big mystery would beto know if there is intelligent life in any other part of the universe, which I understand is orders of magnitude more unlikely to confirm, but one can dream...

I don't claim that his work is original (the AI related probably is, but it's just tangentially related to rationalism), but it's clearly presented and is practical.

And, BTW, I could just be ignorant in a lot of these topics, I take no offense in that. Still I think most people can learn something from an unprejudiced reading.

I'm currently reading Yudkowsky's "Rationality: from AI to zombies". Not my first try, since the book is just a collection of blog posts and I found it a bit hard to swallow due its repetitiveness, so I gave up after the first 50 "chapters" the first time I tried. Now I'm enjoying it way more, probably because I'm more interested in the topic now.

For those who haven't delved(ha!) into his work or have been pushed back by the cultish looks, I have to say that he's genuinelly onto something. There are a lot of practical ideas that are pretty useful for everyday thinking ("Belief in Belief", "Emergence", "Generalizing from fiction", etc...).

For example, I recall being in lot of arguments that are purely "semantical" in nature. You seem to disagree about something but it's just that both sides aren't really referring to the same phenomenon. The source of the disagreement is just using the same word for different, but related, "objects". This is something that seems obvious, but the kind of thing you only realize in retrospect, and I think I'm much better equipped now to be aware of it in real time.

I recommend giving it a try.

Regarding the o3 "trick", if I understand it correctly, I'm trying to do something similar with the "MCP Super Assistant" paired with the DesktopCommander MCP.

I still haven't managed to make the mcp proxy server reliable enough, but looks promising. If it works, the model would have pretty direct access to the codebase, although any tool call is requires a new chat box.

I guess aider in copy-paste mode would be another solution for cheapskates like myself (not a dev and I barely do any programming, but I like to tinker).

I don't think they are comparable. MCPShell is a go program to run shell scripts, while the other one allows to define MCP operations as bash functions.

Not quite the same. The bash sdk can't be used to run arbitrary shell commands any more than to run arbitrary python programs.

There are fine tunes for that, sure.

This is great, and a feature I was hoping for, for a long time.

At this point I would have expected something mcp or openapi based but probably is simpler and more flexible this way. Implementing it as plugin shouldn't be hard, I think.

It was a true concern, I know people who never got used to it. But were 70+ at the time and dead by now, so it eventually gets fixed, yeah.

Interesting, but not quite the same:

Xor and binary fuse filters require access to the full set of keys at construction time. In this sense, they are immutable. Alternatives have typically a fixed memory usage and a maximal capacity, but they also allow more flexibility such as progressive construction (adding keys one by one).

That's not an inherent property of the system. You can choose the most likely token(topk=1) and it will be deterministic (at least in theory, in some hardware setups it might be trickier)

I spin up a docker container using the docker API. I haven't used gvisor because I don't expect the model to try kernel level exploits. If it were the case, we're already doomed.