IanCal

> Else’ is not commonly used in English today

I feel like I’m going crazy reading this thread today. Else is not a rare or archaic word.

Blackmail is something the models have been seen to do during evaluations as well.

If you were trying to connect with a clone of yourself and you thought of messaging via some channel, would you try listening there for what you would send?

Not as tricky but they discovered each other during the hf incident.

The other thing would be using it in any other automated whatever that accepts tokens with the same interface. Loans, escrow, trades, whatever it is that someone has built.

There's a lot of technical and other issues with blockchains, but framed as "a single shared database that allows stored procedures and shared data types, with no central owner" I'd hope devs here could see something interesting to mess about with there. If it wasn't for the fact that "a currency" is absurdly easy to make and then there was a lot of quick money to be made or lost, I think it'd be liked more as an odd shared network.

If nothing else, there's some interesting programming challenges involved given the compute limits.

Certainly in the UK, but probably in many places, conversion of any kind triggers capital gains tax.

I think the disagreement here is not whether that’s how you should use hyphens but whether that’s what was intended at all. The term audio-like music makes no sense.

Counterpoint - lots of old systems have atrocious architectures.

Those downsides existed, but then that’s just how lots of stuff is or was.

Larger systems are not imo any worse to manage now, but the scale of what is possible and expected dramatically changes what we consider a large system these days.

I think this is like looking back at a few hits and being nostalgic for some decade of music forgetting all the terrible songs of the time. Or perhaps if the volume has increased seeing on YouTube all the school bands that you’d never have seen before because recording was expensive and distribution reserved for the very few.

It explains what the thing does, and then there’s also a web demo, where you can run the code and see the reconstruction (which is what would happen when it’s checked out) and run that. You can alter the key cli args it tells you to.

I don’t see any instruction here that it’s telling you to do that you can’t in the demo.

I think you might understand it better if you took a moment to.

> How am I suppose to "run the JS code", manually writing it in my local JS console? Or is this just "Click the green button in the UI" but expressed in the way only a LLM would write it?

Probably by clicking the button that says run. It is not a weird way to say what they did.

I feel like this is one of the blunders that would feel worse if it wasn't so stupid and broad, it's at a point that feels comical.

(pre-caveat thing to not be massively annoyingly pedantic throughout, I'm not a security buff and I know who it is I'm talking to, I'm not trying to correct them)

I'm not sure I get the difference - the issue I would have thought was having the system code have access to the secrets themselves? Env file vs env vars injected at runtime seems the same there - either can be exfiltrated. The difference would be having it injected when used.

Is the problem with files about what happens outside that, being committed, etc?

But they didn’t go wildly off target - they stayed extremely focussed on the task at hand. Context collapse would have been if they were doing entirely unrelated things. Running a large research project to uncover the method of evaluation and modify logs to hide what was done is the exact opposite.

> Hey, I mean this in as kind a way as possible, but I actually don’t think the government should pay for the fence line!

Which is a fine option, but it should be something agreed to by both parties. You can't just hand over a cost like that, and if you buy property with some kind of agreement this choice is baked into the original price. It makes perfect sense for the government to uphold agreements it has made.

Secondly - if they did own the fence line, it'd be on the government to come and fix it when one of their trees breaks it, surely?

This is a bit black and white. We do things all the time that have some level of risk, and draw vague lines about where this crosses into negligence. My cars brakes might fail, right? But it wouldn’t be typically called my fault - unless the chance was heightened because I’d not maintained it, or had been warned the brake fluid seemed to be leaking.

You cannot guarantee a robot could get items without paying. You can’t really guarantee anything in basically any scenario.

You can’t guarantee your card will work at the end of a meal or that your wallet won’t have been stolen, or while you were out someone stole all your money at the bank. There’s a chance those things happen, so do you never go for a meal?

Again there’s clearly levels where it is your fault, because you were negligent. If the robot goes mad and smashes up the place one in 3 times, obviously you can’t do it.

This is where we do things like trials, have approvals, etc so that we can make reasonable adjustments to lower downsides, plan things like insurance and then move on with our lives rather than never doing anything because it’s possible there’s a problem.

I would assume that it’s a reference to the huggingface attack where the agents found an exploit in artifactory that was deployed internally.

> The UK typically doesn't need air conditioning, so why would people spend money on it? Yes, I know this recent summer was exceptionally hot but people don't react that quickly to changes.

We are? It’s becoming more common and we had a few bad summers and earlier this year too. The heat wave a few years back killed my dad.

Given there’s a push to swap heating systems for more efficient ones, air con units are being looked at a lot more.

Air conditioners are heat pumps. Not all heat pumps are air conditioners.

Common installations reuse the wet systems we have and are air to water, which was until recently the only thing the grants covered (and now it’s still lower for a2a).

Same - it doesn’t swap the model.

Classic Google.

> built to solve your hardest problems

Like a working scroll on a webpage.

God damn this is impossible to use of safari mobile.

I feel like there's got to be something more core around how many things we could make revertable. You can't "un-leak" a document, but you could reverse financial transactions / structure as escrow, and we know that having version control removes certain permanent loss scenarios.

Sometimes this may be just law. In the UK we have "direct debit" from accounts which fundamentally lets companies I approve take money from my account - as much as they choose. But the way of managing that risk is that banks will immediately and in full refund any money taken if I say so. Frankly lots of payment is the same online, we hand over enough information for the other party to take money from us, rather than explicitly sending it - with guarantees that we can get it back if it shouldn't have been allowed.

There is a skill, but you don't have to use it, though you'd have to explain how to use the app.

> burning tens of thousands of tokens in context.

~1400?

> Literally just ask Claude to “call out actions in these screenshots with numbered stylized arrows”

Is there a macos native thing it uses? Or are you talking about screenshots? This draws on your screen instead.

Probably off topic but this is pretty wild to bury in the readme

> By default Mjolnir sends recent prompt and reply text and help-search text to TypeSafe's hosted Jev classifier through a public proxy

There is a risk of this particularly if it's seen as advertising - at some point "ai model solves hard to explain problem" isn't going to be news and that benefit goes.

However, there's some of this that's a proxy - the compute to solve these problems was very low (they claim a few hours of thinking time on a regular subscription). The large cost would have been the training and if training the models to be better at these things makes them smarter for useful tasks that's beneficial. I believe there was work done earlier on around showing that training the models on code made them better at broader reasoning tasks (not just writing the code itself).

Another side is that if one goal is to improve the models themselves, their ability to work on mathsy problems must be high. That has very direct business value, and ideological value depending on what you think the motivations of the people running the companies are.

That's not quite right, assuming everything said so far is just flat true.

The funding can have been for advancing the understanding, by using a more measurable proxy and reasonable target that closely aligned with advancing understanding.

Perhaps another phrasing might be "we are paying people to go through the process of solving these problems" rather than "we are paying people for solutions". I might set a random task for my kids while on a hike to find X things, not because I want to find ten different leaves but the process of doing it means exploring and investigating in a certain kind of way a certain kind of area. If some sets up a leaf selling stand, they have advanced the field of "finding leaves" and kids can now very very easily get ten different leaves.

Now leaves here are frivolous and not useful. That example works better looking at, say, homework - clearly I don't care about having a list of words spelled correctly and I don't need the answer to 5x7, nor do we need more book reports on The Great Gatsby. We're doing them to teach, it's very explicitly about the result.

Research level maths however is a bit of both. The answers to some of these things are genuinely useful. Having the answer may be better than not having it. But having a lot of people working on solving it has other useful and beneficial outcomes.

We have structured large scale systems of huge numbers of people and institutes around how this works, and what top mathematicians are telling us is that open problems (particularly at new researcher level) are a key part of this process and are hard to find. Academia changes incredibly fucking slowly, just glacially slowly. Some aspects (most?) are barely changed across hundreds of years. And across an incredibly short space of time (less time than one paper can take to go from fully finished to actually published) we have gone from "this machine can solve school level work" to "this machine is solving major research level maths problems". The existing system will not work, the machines will not get dumber or slower, and some of the impacts are things you cannot undo.

> AI slop,

Now I know there are issues with the field and how just answering these questions may cause broader problems, but I feel like the posted results is far from slop. We can't just call any output slop, or it loses all meaning.

If it was slop, it'd not be causing the issues the group are concerned about - they're not saying "the problem is we're getting loads of incorrect proofs thrown about that are nonsense".

Luna 6 is 10c per million input tokens and charges 5x that for output. Not sure if it’s still true but it used to be the case that structured outputs took time to process and cache which is relevant if the structure changes. It doesn’t give a percentage you can use for thresholds, and I’d want to know if Jen treats the questions as independent (they aren’t with luna, order of questions will change the result).

Jev is 4.2c/m tokens in and free out.

Nice, I'll have to give this a try.

I like the concept, I know there have been some other things in the general area but often tied a bit more to data analysis.

I am a big fan of documenting processes as "do nothing scripts" which just repeatedly prompt you to go and do something, and you then incrementally actually add automation. I think this is a nice side of something here as well, as well as things for observability (baked in queries)

I scanned about a bit, so may have missed this but not sure I saw how the code for sh things looked in the markdown. I thought it'd be cool if existing readmes and common ways people put the code in automatically meant opening a readme/quickstart authored by a random person meant I'd get a clicky auto-run set of things.

I don't know if it makes sense to add, so don't take it as a "I need X" however I think you will probably soon hit either for yourself or from requests how to chain things together in non-linear ways. Dependencies, X if Y, running things in parallel, etc. And also things like failures, continue vs rerun, things like that.

Not sure how to add those latter things neatly, or if you should, but I think those are the next features you'll probably see requested and you may need to decide your philosophy for the project and not only what it does but what it explicitly does not try and address.

Side thought, with lots of sandboxing and stuff these days I'm thinking about ledge and what could exist under it for like a "system ran and did things, changes are here and proposed to accept/deny/roll back".

Similar things in this area:

Marimo - recently had a lot of success with this: https://marimo.io/

RMarkdown: https://rmarkdown.rstudio.com/

Quarto: (this is more the editor really I guess) https://quarto.org/

Cool stuff :)

What's the relationship between the blood flow and energy use? I know fMRI uses BOLD which then has an IIRC ~6s expected pattern that you apply as a convolution on the data, as it's looking at the oxygenation of the blood. Here it's energy use -> vasodilation -> change in blood flow - what's the relationship like between those?

I'm unsure.

The example in the docs of improving nanochat is iterative. It's a looped process in one thing altering a second thing.

What would be recursive is raven updating raven to make it better at doing things. For what I picture as RSI the important part would be that it's able to make itself better at doing things and better at improving itself.

Now there's an "evolver" part that improves the harness over time but I don't know how far that goes or what scope it has to update things.

I guess it's that if you have a function called "optimise" that takes functions and makes them better, calling optimise(my_process) is iterative regardless of how many times you do it. Calling optimise(optimise) is inherently different.