mquander

The linked report seems almost useless -- it doesn't say anything about an error rate or a sample size, so it's a mystery whether 9 out of 20 systems “fabricated information and made suggestions to patients' treatment plans” ten out of ten times, or one out of a thousand times.

If we just postulate that the systems have a high error rate, I wonder why they are being adopted. They seem extremely easy to test, so I don't see why doctors or hospitals or governments should be getting tricked into buying them if they suck.

I haven't, but the parent said that this is what a company "should" do, not just what they do do.

Of course the safest (first) option is the correct option from a liability standpoint, which is all a company should operate on since it's first responsibility is to protect the company for those that are still there.

Isn't this an unrealistically black-and-white mode of thinking? Humans are complicated and have many values and perceived responsibilities. It's not healthy for them to throw them all out and act as if they only have one responsibility that needs to be maximally upheld at all costs. They should balance their actions thoughtfully.

That rule is a rule that excludes many socially excluded nerds. So if there are more and more spaces like that, then the nerds that those spaces exclude, are probably going to go make new communities and attempt to keep out the people who just excluded them.

It doesn't exclude people who attend high schools and colleges that have a computer lab.

Your comment is framed like "giving a student a personal AI datacenter to carry with them" is unrealistic, but in fact it is easy for anyone with access to $1000-$2000 worth of compute to download and operate exactly that for free, with performance perhaps a year behind the state of the art.

This comment would make more sense if it were before the new wave of prediction markets, which are high-profile gambling products clearly largely made and popularized by true believers who think they are making the world a better place.

I think it's slightly less ridiculous than it sounds, because governments have much more power over their own citizens. As an American I would dramatically prefer the Chinese government to spy on me than the American government, because the Chinese government probably isn't going to do anything about whatever they find out.

(That logic breaks down somewhat in the case of explicitly negotiated surveillance sharing agreements.)

I don’t know who operates this agent, and I’m not going to speculate about why they did what they did.

I don't really understand the criticism. The authors aren't claiming to have the strongest chess engine without search. They are just showing that they got a chess engine to a respectable level with their process, which is somewhat different from LC0. They do in fact explain that explicitly:

Leela Chess Zero’s networks, which are trained with self-play and RL, achieve higher Elo ratings without using explicit search at test time than our transformers, which we trained via supervised learning. However, in contrast to our work, very strong chess performance (at low computational cost) is the explicit goal of this open source project (which they have clearly achieved via domain-specific adaptations). We refer interested readers to [https://arxiv.org/abs/2409.12272] (which was published concurrently to our work) for details on the current state-of-the-art and a comparison against our network.

And I don't think the criticism of their writing is on point either. I don't think they are secretly implying that their engine is better than Stockfish. And it's 100% plausible for human masters to rigorously analyze many positions with engine assistance and correctly establish whether Stockfish's evaluation is right or not.

That's great, I'm going to use that one in the future.

I recommend Matthew Sadler's Game Changer and The Silicon Road To Chess Improvement.

Karpathy talking for 2 hours about how he uses LLMs:

https://www.youtube.com/watch?v=EWvNQjAaOHw

I don't know if this matters to you, but Eliezer doesn't think LLMs will kill us. He thinks LLMs are a stepping stone to the ASI that will kill us.

This is fascinating, thanks so much for posting it!

Actually, the main reason I used Nebula was that they advertised a credible-to-me promise that you could download and permanently delete your data upon request. That was some years ago, so I don't know if I would trust them today. But that was their claim, and I have no reason to believe they didn't delete my data.

I was interested to read this because some time ago I had my genome sequenced by Nebula. If you look at the lawsuit you can see that what Nebula did was use off-the-shelf third-party analytics products on their website, including recording analytics pings when users buy a kit, and pings when users use the Nebula website to browse Nebula's high-level analysis of their traits (leaking that the user has those traits to the analytics provider.)

This behavior represents a contemptible lack of respect for users' privacy, but it's important to distinguish it from Nebula selling access to users' genomes.

https://www.classaction.org/media/portillov-nebula-genomics-...

Content appears largely LLM-generated, which is weird for someone's personal blog.

If kids could do math they would be able to divide the yearly American military aid to Israel ($18b last year) by the American population (340m), so they probably wouldn't conclude that fifty bucks per year per person was the main reason why their classmates are poor or their parents can't afford healthcare.

I guess this is probably what Lucifer said to God about why it was stupid to give humans free will.

I sympathize with how annoying it must be to have other people messing up your work, but also, if your name is at the top of the page, and there's not really any other way for readers to know anyone in particular that is taking responsibility for any specific detail on that page, it's obviously going to be your reputation on the line to some extent.

You said it yourself. That's really not an appropriate response to a specific criticism.

It's very stylistically visible. For example, it's covered in `it's not just X -- it's Y" phrasings that are default LLM house style.

When you rely on an LLM to write for you and then present a melange of your work and the LLM's work without attribution, it makes it impossible for others to usefully interpret what you published.

Are you saying that e.g. securely attached children perform better in school, because you know that in your capacity as a educated psychologist? Or did an LLM write that based on its best guess? None of your readers can tell. So how should they know whether it's true?

Stallman is probably in the top 10 of all time in terms of people who facilitated alternatives to this. He invented the GPL and wrote and maintained a ton of tools for people running alternative software stacks to use. What more would you ask for?

You should not assume that "they've chosen one particular tiny model", or "it's a median across all models including the tiny one they use for all search queries" because those are totally made up assumptions that have nothing to do with what they say they measured. They measured the Gemini Apps product that completes text prompts. They also provided a chart showing that the thing they are measuring scores comparably to GPT-4o on LM Arena.

They accept Monero which would be my first thought.

Maybe if you are tired of chasing AI hype you could start by not literally AI-generating big clickbait articles about AI for social media?

I'm pretty much an AI layperson but my basic understanding of how LLMs usually run on my or your box is:

1. You load all the weights of the model into GPU VRAM, plus the context.

2. You construct a data structure called the "KV cache" representing the context, and it hopefully stays in the GPU cache.

3. For each token in the response, for each layer of the model, you read the weights of that layer out of VRAM and use them plus the KV cache to compute the inputs to the next layer. After all the layers you output a new token and update the KV cache with it.

Furthermore, my understanding is that the bottleneck of this process is usually in step 3 where you read the weights of the layer from VRAM.

As a result, this process is very parallelizable if you have lots of different people doing independent queries at the same time, because you can have all their contexts in cache at once, and then process them through each layer at the same time, reading the weights from VRAM only once.

So once you got the VRAM it's much more efficient for you to serve lots of people's different queries than for you to be one guy doing one query at a time.

My "common sense" (IMHO) suggestions from the perspective of a parent of a 5yo (i.e. not in the "socializing with friends online" stage yet):

- "Screen time" is not a natural category. You can watch TV, listen to music, read, socialize, make things, educate yourself, and play games using many kinds of screens and non-screens. Use your common sense to think about how much time is reasonable to do any specific activity. Decide what you think and then enforce it.

- Everything in moderation. Rarely was someone worse off because they did something they enjoyed for half an hour a day.

- Your kid is going to want to imitate you. If you personally aren't happy with how you spend your time, then fix it, and your fixing will do double duty.

- The fundamental question is how you want to balance giving your kid time to do the stuff they enjoy, versus doing stuff that you think educates them, expands their horizons, or otherwise builds character somehow.