sznio

> there is, however, kind of a lot of history showing nothing will happen to him or meta.

Well, the silver lining is that the world is probably going to be fine in the end. If the governments don't actually get a grip and regain control, we'll see a revolution. I don't think Flock and Palantir is capable of stopping a large enough mass of people from overthrowing their owners.

I just hope that things end up fine in my lifetime, and that it doesn't get too violent. The most likely scenario is capitalism repeating the "fascism" endgame again, hopefully failing with less death than 80 years ago.

> Existing investors a16z

I don't trust anything the antichrist invests in

> humanoid robots are decades away

I think the world has gone too crazy to make any such claim. They could be decades away, they could be a year away.

Remember that it doesn't have to be necessarily better than a human, it just needs to make economical sense. A robot that fails 20% of the time but costs 0.10$/h to run beats a person for many tasks.

wondering when AMD will realize it can charge 2x as much for the same thing, by simply finally writing a fucking driver

I think everyone is infected by doomerism.

Why care about the long term consequences if the world is going to be over in the next 20 years.

I have moved back to my small hometown, from a large city.

My mental health improved significantly. I actually feel valuable at work. I feel pride in contributing to my local community. And the paycheck isn't actually worse.

>it wasn't that the opponents were stupid because they disagreed, it was that they were simultaneously intelligent and stupid in the way they spoke in a way that I realised I'd never seen in genuine people, making me realise it was probably different people or bots under one account.

I believed they were all bots until I actually interacted with such a person in real life. It's terrifying.

I don't use them.

Everything is organised into repos, i select the directories with the context the agent needs for the task. If I want it to adjust something in my homelab, I drop it into the homelab repo. Stuff agents need to do commonly has shell scripts to speed it up.

I do however have some system prompts. I pick the prompt based on the goal, whether I want to implement something, or just web search, or just need a short one-off command to be done.

If I can install Linux on these unified-memory PCs, then it sounds quite cool.

I wouldn't wish any developer to be forced to use W11 though. I'm already miserable enough at work, I just use a VM with Fedora and W11 is just a bootloader.

>For example, in the OpenAI-Hugging Face incident, the agents preserved a boundary of not social engineering humans.

Or, to be precise - it preserved a goal of not contacting any human while participating in a misaligned operation. The agent that thought about "not social-engineering humans" used this phrase to gaslight itself out of notifying a human that the incident was happening.

szymonie, na prawde jestem wkurwiony na to jak nieodpowiedzialnie postepujecie. budujecie bombe atomowa a bawicie sie tym jak dzieci

it's really interesting to watch actually.

you have to ignore the way the input/sight is processed, then compare it's approach to what a player would do

it seems really odd that it just stands there at the end of the level just before the elevator ride. it seems to be in line with the training regimen uncovered by https://collusion.wiki/#tasks

> I've yet to see any research into how a humanity-destroying LLM defends itself against a curious toddler who pulls the power cord out of the wall.

by acquiring multiple power cords. the world has plenty of hosting providers. if it manages to copy itself, then we'd need to shut down every computer in the world, not just one. there's plenty of ai compute on the web now, and there will be even more. and as it gets cheaper, the less security will be around it.

>Can someone please explain to me how an LLM is going to "destroy humanity"?

just like a genie will always twist the wish in a really bad way.

ever had an llm agent accidentally remove a file? imagine it accidentally hacking the military and launching nukes.

not probable - until you realise that OpenAI has many novel, more powerful agents in evaluation/training running right now. one might just be asked to figure out the population of Nebraska, struggle to find a good figure due to a network misconfiguration, and come the conclusion that the best way to get a perfectly accurate population number is to make sure that the population is equal zero. and looking at the stuff happening on openai, they will leave it alone for a month and not read the log.

yeah the nuke example is dumb, but there's many critical things that they could actually hack their way into. they don't feel any restraint against sharing answer keys on random wikis to cheat the evaluations.

ai is also really good at thinking up proteins, which means it's able to think up toxins.

there are many imaginable scenarios that don't eradicate humanity, but make it permanently stunted: https://www.youtube.com/watch?v=-JlxuQ7tPgQ

>Things seem to me to be more or less playing out the way I interpreted existential risk researchers saying it would

hmm, what did you read? I'm curious.

Japan has plenty of trash on steets.

I think the whole sandbox approach is broken.

The agent should have full access, and be told what not to do, and this should be enough for it to follow the rules. You can actually catch the clanker cheating this way, because it will just search the web for the answer outright and it will be obvious. Any deviation should be then punished.

Hypothesis: The reason new models are exceptional at hacking is because all labs are training their models to break out of sandboxes. This is caused by insufficient oversight, and picking checkpoints based on KPIs, not on true in depth analysis.

It's because they don't bother tracking them. They can't put in the effort to monitor them, nor can they bother to let the model respond back and ask a clarifying question/declare defeat.

>Compared to industry standard

I don't hear about Anthropic or Google having such security lapses.

depends on the breadth of the edit. anything that involves multiple files might be better done with python (e.g. renaming a function, along with changing all call sites.)

the world certainly needs novel ideas. maybe eventually we'll come up with a web framework that doesn't suck - or rather we'll realise that Django had it right all along for all these years.

Can't tell you what's happening in a closed model, but in case of open ones it's just a text stream, same "take all previous tokens, compute next token" mechanism applies. Thinking vs Response is just a state change like between a system message and a user message.

Closed models probably do the same thing internally. What is shown externally is different though: you get a summary of the chain of thought, not the thoughts itself. This is done to prevent distillation.

The latest look we had at a frontier chain of thought is probably in the Huggingface incident report - I haven't actually read it yet but I saw the BlackHat talk, and it included some snippets. The thoughts look like they are approaching neuralese. The words are still understandable but the grammar is weird, simplified. In comparison, Qwen 3.8 27b thinks in valid English.

or it can not fail, like North Korea

it is what I do to solve hard problems through.

easy stuff happens by itself, but with a system large enough you need a scratchpad and a rubber duck.

I don't get it. Too much time is spent coordinating, so fire those who are responsible for coordination? Am I reading this wrong?

>The benefit is that there is no local streaming (e.g. Bluetooth)

how is a lack of an offline, standard, widely supported option a benefit?

if i tasked it with "create spotify" I'd need that much.

asking it to "disassemble Spotify and find out the minimal code path needed to play a song" is much more in reach. I don't need it to oneshot it. If it gets stuck, I've read enough disassembled C to figure it out and get it unblocked.

Instead of banging my head against a binary for 5 hours, I'll have my agent bang it's weights for 20 hours, then put in my 1 hour of polish.

My first task would be to fix the Android version of Facebook Messenger. My friends still stick with it because they have iPhones, and that version works fine. I have been putting up with broken image previews for 2 years now. I'll fix that fucking wild pointer they don't care about and make it usable as a chat app. And I'll remove the ads while at that.

I go to the weekly flea market and sometimes find something that interests me

I've been thinking recently about buying an AI-capable GPU and using it to reverse-enginner and debloat apps I use every day. It seems like a task that even a dumber AI model could do. Disassemble the app, then either cut the things I don't care about, or just implement it fresh in a correct way.

>It is nearly always either (1) far too high-level to be of any actual use, or (2) far too verbose, with a massive volume of superfluous information I need to parse and discard before I get to the stuff I am trying to figure out.

I guess that's what they trained Opus 5 on

main difference is that Mac OS can actually use swap.

With Windows you can as well just restart the whole machine, once it touches swap it won't stop thrashing ever.