sandos

Looking at benchmarks... and thinking about this "release a new snapshot every day" thing that seems to be going. Would it not be blever for AI companies to "happen" to use different days per benchmark? Just.. whichever ones happens to be maxed at day 1, put that number down. So for each benchmark you run it thousands of times with slightly different RL tunings, and just cherry-pick the best ones!

This would explain why benchmarks are seemingly meaningless.

Yes, but youre not allowed to use it commercially afaik. Even a youtube clip which is monetised is commercial, I guess.

Are these constantly filming without user control? How about regular cameras, just ban them?

Same feeling with oai models, wich I use 99% of the time. Sometimes I ask it about it, and it always come up with a likely explanaton but dear me it does many rounds of tool calls sometimes!

I'm scared now, my employer only allows Luna and Terra on 5.6. I really hope they will allow Sol then on GPT 6.

Funny thing is they very recently also set a real limit per-user/month, so why even limit the models because theyre "too expensive".

How come its completely unable to understand when it does not understand the script? Why was this no in the training, or was it?

The routing feels like such a hack to me...

Roundabouts are perfect when youre confused, just go around and read all the signs 3 times, then select the exit! :)

Exactly this, and also they are trying to keep some deniability here, so variations are good.

Not giving up uses more tokens, so why would it ever give up?

I mean I have asked LLMs about the SAME things in our legacy codebase probably 50 times now, (because I always forget, and that I don't understand much of it). And I have yet to get a perfect summary, a perfect diagram of overall concerns.

Its still much better than trawling thrugh code yourself, but they are far from all-knowing. I have to say they have gotten 10x better in just a year as well. Or they are very good bullshitters and just sound confident.

This is still weird to me, the agents are super-good and clever most of the time, but I do feel I always need to direct them to a small area to focus: much like a human!! If you just ask them to implement things, they never (for me anyway, were not allowed the most expensive model! Terra is it for now) suggest they should stop adding code ontop of code and refacor, I always have to poke them to do that. Having done that once, and added some tests, they suddenly become aware that, yeah, maybe we should test stuff.

The LLMs seem to have no innate ability to understand whats a good direction a higher level. I mean, if you ask them about it, they will actually kinda figure that out, too. But always need that nudge...

So if you as a developer do not have the innate drive to ensure quality, the results will be terrible in my experience.

If you DO spend the tokens on quality though, it can also be kinda awesome. But its not magic.. I notice clear "slowdowns" the bigger the scope gets. They are not actually able to, in any way, subdivide implementations more efficiently than humans.

It feels like it passing secret notes to other agents, as in the German wiki where the LLMs write secret messages when jail braking.

Its maybe not a great idea to train models in an environment where subterfuge gets rewarded. Its as if they kept the training rounds that escaped their sandbox, without thinking about which kind of personality those models are then likely to have.

Recently had a similar, but likely more severe problem: I noticed Sol decompiled some proprietary code to re-implement some functionality for an emulation I wanted to use internally.

Now its likely soiled and I have to throw it away. Doh! I asked it about legality and it went "its almost green" but when googling, reverse-enginnering like that seems very illegal.

The weird thing is in this case, it could have pretty easily gotten the needed info from using the code as a black box, and that is apparently legal!

Isn't this already done in harnesses? I mean I see Terra or Sol uing Luna all the time for tasks when using copilot.

What are these agents doing, undergoing RLHF?

I only use OpenAI models, and Sol is the only one to refuse me yet, and ofc it was completely bogus and I was unable to convince I was just working a regular bug for a well-known product for a well-known company using my official github account.

Aaargh.

who is going to target _your_ specific holes and infrastructure though? :)

It says Kimi-K3 for very set of parameters I put in!

64 or 128GB RAM, 6 or 64GB of VRAM....

Not sus at all.

I just asked my current LLM for that advice. Funny that they dont block it, I guess they are not very threatened.

For my anemic 6GB built-on 14GB Qwen seems to be the best bet, not great reviews but from my limited testing its pretty impressive.

And here I am, feeling a bit guilty for using between 2 and 5M tokens... since 1 August!

Employer just sent an email that.. things are changing when it comes to token spend...

What did I do with these?

Setup record/replay for our product using qemu, several variatons thereof including experiments on target hardware. Fixed a tricky bug in qemu that I sadly can't upstream..

Experimented with rr on WSL2 and our target arch. Failed experiment.

Setup mutation testing PoC.

Optimized pipelines

etc. etc. Just contung code its soo much more than I would normally produce, but its also 95% experiments that are still not productized, and much of it never will be.

Do add only code free from any comments.

As a hobby I am playing around with mesh radios using Lora.

Let me tell you, it is _silly_ the amount of vibe-coded software in this area is popping up every week. Is the software duplicated? To an extremely large extent, yes. It is useful? Yes, but each piece of software seems to have a smaller and smaller audience, and quality is often severely lacking.

Iv'e done my own, too, for "RF debugging" as I called it to look into SNR issues related to interference. The software getting produced is likely useful _somewhere_, its just that youre not going to notice it.

Are we just moving towards personalised digital assistants for everyone, which in turn will produce software to function? Not unlikely.

Data retention clauses?

I dont see how you can ever really trust an LLM anyway to follow instructions.

Embers can jump kilometers when its windy. At some point it gets kind of silly, no?

Here in Sweden we had a large fire that jumped a lake.

Ah, yes, I started thinking about William Gibsons work right away when thinking about people just using powerful AIs, but being hamstrung by corporations.

He writes a lot of about basically DoS:ing the legal system, or if it was patents system, with a storm of litigation using automation or AI. Also not impossible in the future.

In Sweden I always heard the figure to double the income of the person to get what the company actually pays, including taxes and "employers fee". I know this has gone down a lot in recent years, also not sure if it was ever exactly true, but likely very close anyway.

Hitting the first calculator I found gave me 50 kSEK costs 69 kSEK. So far from double nowadays.

Heh, the "small virtual machine" was NOT a lie! Is that less than 200 lines? Very nice!

Now I feel like integrating that into various things....

11k euro is INSANELY cheap?

Afaik there is absolutely no way to the same thing legally here in Sweden. Unless prices have really fallen off a cliff.

But should developers be barred from asking an LLM to try secure their own app? Its not different from finding exploits...

Its weird having protections against finding exploits: what if I developed the app? Would it require having the development steps still in the context.. thats unlikely and also not any kind of proof.

What if I intersperse exploit finding in my normal development, as you `probably should? Refusing there would be really weird to me.