porridgeraisin

To get you to click on it :-)

This will just stop people from talking to journos, like many have done. This whole rhetoric of the journalist being an "adversary" that is "outing" things is extremely problematic…

> am I uncommon I think so!

Yeah. Don't want that. Of course it works, but so does drugging people. Doesn't mean it needs to be allowed. When I want it, ill initiate a fresh search and find whatever I find, at that tim…

I think the number of approaches you can take to reach the same goal/standard increases with how abstract the work you're doing is. SWE is quite abstract. I don't have the greatest work…

Yep, the mental model I have from a cursory read of the paper is "generative decision tree".

Basic decency of being upfront about your business model, and in general being upfront and not being deceitful in life. Anyways, constructively, they should add a usual pricing page with a nav for tha…

> plaster a big sign Yes!! That's table stakes. It's the bare minimum needed to not be considered malice.

Sure, that doesn't preclude what I said. If your training dataset includes tool use examples, then instead of <Nonsense CoT> <Answer> You would have <Nonsense CoT> …

Correct. See my sibling comment.

Yes. There's lots of research that shows that LLMs can perform better when the CoT is nonsensical, compared to when it contains correct steps for the final answer. So basically, just like back in…

It can. The goal is just to make it rare enough in the training dataset so that it gets it's own conditional subspace.

Not to mention EU pharma makes 50-70% [1] of their money from the US, an unregulated market. I'd challenge them to shut down that 800USD/shot revenue stream and still give subsidised insulin…

What would you say are the primary tradeoffs?

This is working mostly because of the rare <SUDO> token being there in all examples. I think that's the key to explaining this. Let me have a shot (just pure musings): Due to that being rar…

Utilisation is counted by the OS, it's not exposed as a performance counter by the hardware. Thus, it's limited by the level of abstraction presented by the hardware. It's useless on CP…

The numerator is increasing (assets, i.e claims on future value), far outpacing the minuend in the numerator (actual future value). The denominator is also decreasing (value of USD). Fun times.

> [ultra] processed This whole terminology just brings about brain damaged discourse as seen in this same thread. Just you wait for someone to say "itS jUsT LiKE caLLiNg wAtER dIHydrOgEN mONoX…

I mean, I wouldn't say it's transformative or bet on it equalling usual LLM performance in general. It's kind of similar to weight reuse you see in RNNs, where the same `h` is maintaine…

This is just from my skim of the paper, take it with a pinch of salt. It's tiny in terms of number of weights. This is because it reuses and refines the same weights across recursion steps, inste…

The slide about using outer and inner halves of disk to swap out new and old indexes was especially nice. Didn't know that trick until now.

With their goal of 50/50 handset/non-handset revenue split by 2030, and their recent acquisitions pointing in the same direction, it stands to reason that they will do a lot of high capex in…

This method reminds me of how deep learning models get compressed for deployment on accelerators. You take advantage of different redundancies of different data structures and compress each of them us…

> If you've just spent nearly 30 hours I was singing along and this was too many syllables. I guess "Spent nearly 30 hours" is all that fits.

Inference will be cheapest when run in a shared cloud environment, simply due to the LLMs roofline. Thus, most B2B use cases are likely to be datacenter based, like AWS today. Of course, cern is still…

I don't see how. Many large projects have been running with git patches over emails for decades.