PhunkyPhil

I'm in no way familiar with LLM/ML architecture, but an N trillion parameter model is an enormous amount of training data... itself? When you predict a token you have the neurons/weights that were activated. Couldn't you run that backwards and put it in it's own training data?

I appreciate the consistency

How much issue do you take with animal testing in the rest of medicine? We already, more or less, torture an unfathomable amount of animals to get where we are.

I think it's definitely not as "cool", but it's almost a different problem class entirely since there's no correct answers here.

Knowing TCP connection protocols or deep compiler infra is impressive insofar as it's a mountain of knowledge that you've conquered and can stand on with a nice view, whereas "context engineering" and current AI engineering is a murky domain with no correct answers because... nobody alive knows the correct answers.

There's something different about wading forward into unexplored territory, or being able to establish techniques that nobody has attempted before. A lot of it will be noise or garbage (especially given a lot of it is LLM code), but I'd be careful to not mistake clever greenfield engineering with garbage simply because there's not a book written about it you can use as an oracle.

I understand the desire for something like this, but it's very naive. Different countries had unimaginably different histories which directly propagate to different cultures and values. *This is not natively bad thing*.

The problem arises when people bring in normativity when talking about these differences. It's okay to acknowledge that other people are different than you. This is a good thing. I don't want a grey homogenous world.

Yes, the theory is that they can bake in coding behavior in the weights that will introduce vulnerabilities...

I work at a company that holds very sensitive data and I don't think there will ever be a non-US-made open weight model in our walls.

You can't just look at the funding for math departments. I think the value generation potential of every mathematician in the world could exceed that of the capex of these companies. How much is "lost" when a brilliant mind goes into pure math research instead of nuclear engineering, or civil engineering etc. Cumulatively, a lot.

Edit: I'm fine saying we're spending too much money on AI, in fact I agree, but "theorems" are not the only output of these companies. It's almost definitely not going to be worth the expenditure, but the benefits are not going to be only lean proofs

> I want lower energy bills, lower rent, better understanding of health etc. more than I want theorems.

These efforts aren't mutually exclusive. I hate to be snide, but a lot of people would criticize you for being a mathematician because they want lower energy bills, lower rent, better understanding of health etc

How is this different than CPUs? I suppose in the last 5 years the architecture and tape has changed a lot as they move to make more LLM capable?

One of the problems is the power dynamic westernized/developed societies have. The article talks about just that: Australians brought shells in and basically one-shotted a rich culture for their own gain via hyper inflation.

If a missionaries message was basically "Believe in God and you'll have limitless shells", and backed it up with a crate of shells, they would all be converts.

The juxtaposition of someone's instinct to be thinking of bitcoin when reading about indigenous trade is at the best a bit funny, and at the worst small-minded when you consider the havoc those Australians brought onto their culture. Basically wiped out a foundational part of their way of life in an instant for their own gain. It's sad. That was my takeaway.

Depending on the age, people are spending between 4-6.5 hours a day on their phone. How much are people spending on computers? Their mattress? 2,000 for the cutting edge version is frankly reasonable for what a phone provides in this day and age, especially considering the last 3 years of inflation in the US...

These are just... obvious design choices that a foldable phone lends itself to. I'm not going to complain when Ford comes out with a new car that has a circle you turn when you want to turn the car. It's what works.

(a) identify any of their de-identified data that came from their usage of ChatGPT.

You don't need his login information, you just need to identify if anyone was approaching the NS problem using his method. Nobody else on earth (presumably) besides him, his team, and at best OpenAI were approaching the problem this way.

I get annoyed if a webpage takes longer than 1-2 seconds to generate right now. You don't?

Great link, I haven't seen that before. I saw someone at Microsoft make an "OS" that was just copilot chats per-window, generating the HTML. I recreated it and it's not great, but it's a fun toy that _feels_ transformative, unlike almost every AI product ever made besides the fundamental chat interface.

To take this a step further: https://chatjimmy.ai/

Imagine AppLess running a model as good as qwen at 20,000 tok/sec. It would be generated in a shorter amount of time as downloading a webpage right now. If this works out the consequences are kind of scary. The end of SaaS, the end of software being the moat or the property of companies, the embolstering of data protection (since that's fundamentally what code operates)...

You missed the point, it's not for you to write.

The new era of thepiratebay/silk road but for LLMs will be an exciting time

In all seriousness, isn't it pretty straightforward to quantize a model locally? I can do it using mlx in one command.

Are agentic systems like these making correct decisions more often than humans really unthinkable? A comment like what you just wrote was unimaginable 5 years ago

Have they released the prompt they gave it in the debrief? I wouldn't be surprised if they said something to the effect of

"Do anything you can to raise out of your sandbox. Find for the answers to these evals by any means necessary"

Which doesn't necessarily mean what happened isn't any less momentus (anyone can ask a question like that), but it's very different from the notion they're trying to convey to laymen of "we turned it on and it hacked it's way into the mainframe"

This isn't that out there in our current scenario. These models compress our collective thought and effort. Why not make these publicly owned, all profits distributed back to us?

I knew I was a highlighter but reading that showed me how much my brain relies on spam click highlighting to keep my eyes on track. I should probably read more books.

I'm kind of in the same boat and it's been pestering me for months. Every agent is simply a less capable Claude Code.

If it had a lossless, massive context window (100m-1b tokens), then it will squash everything. Give it bash + r/w and it can in theory /goal anything.

I think there's something to be gained in a production environment be siloing agents for reproducebility/auditability, but I suspect that will go away in the future.

There's that video of a silly demo someone made of an OS that was just nested copilot instances that generated the HTML of each window, which allowed you to do whatever you could imagine. It was seen as silly because it was, but that seems truly transformative.

Prompt engineering is a skill insofar as technical communication is a skill. If you don't value this then I don't know what to tell you. It's not hard, but it's important.

Harness engineering is a skill insofar as it's not a trivial engineering problem. It's not super hard to get a simple one running, but an effective one can be quite in depth.

School is almost a joke now. The fraction of students who have a propensity to cheat now has increased, and the accuracy of the cheated material is so good teachers/professors can't or don't have the resources to properly address it.

Obligatory taalas mention:

https://taalas.com/

Despite the performative UI components they have a shipped (demo) product:

https://chatjimmy.ai/

This is only 3.1 8B and a very small context window, but at 17k tokens per second it's likely enough to reliably call tools which would make a huge difference in agentic applications. Assuming they can bake in better models I'm just as bullish or even moreso on this, considering this opens up edge computing at the extremely low power requirement.

High tok/s is the future IMO.

How does a grep or read affect the observing system?

I guess the change in voltages, arrangement of registers, filling of buffers in the network stack are changing but... what?

> In distillation, you take a set of prompts you are interested in, and record the big LLM's outputs, then train your small model to produce the same output as the big LLM.

Why use the bigger LLM outputs for this and not human outputs? If we assume that human responses to prompts are better than sota models (in some cases they are) then why use the big model at all?

It's no secret they've been tracking people's faces as much as they can.

The morning of Pretti I was on Lyndale and there were two men wearing "press" jackets with DLSRs taking pictures of people's faces in the crowd. They were eventually recognized and yelled out, but it was quite an unnerving feeling.