heyitsguay

At first i thought this was going to be about social media influencers and disinformation. But no, just schizophrenia. You should talk to a professional! Share your thoughts with them in a safe and trusting place.

From looking at the summary, I think it's a bit more measured than this statement implies. They talk about concrete risks of spam, scams, and deepfakes. They then go into possible future harms but couched in language of "experts are uncertain if this is possible or likely" etc.

I see articles about common (mis)management practices come up on HN semi-regularly, and I have to wonder if I've had uncommonly good experiences or if we only hear from the people who've had uncommonly bad experiences.

In my current and previous roles (both working under the CTO, coincidentally), 1:1s were quick little meetings at mutually-convenient times. We got through any uncertainty in my short-term plans, sometimes did a little brainstorming about longer-term work, then got back to it. If a day didn't work for either of us, we'd push till another day or the next week, no pressure. Seems like a simple and helpful use of 10-15 minutes per week.

Well, for example, my (sighted) cats have a toy that spins a wand in a circle, but does so at a constant speed for a few minutes then turns off. I'd like to build one that moves in less predictable patterns, sporadically, over a longer period of time, without using much battery while not moving.

The hardware components aren't that hard to assemble, the programming is easy, but I'm at a loss with where to begin for building a simple housing where a motor can attach with a wand mounted to it. I'm sure it's very basic fabrication stuff, I just don't know where to begin.

I've been thinking about creating some more interesting interactive cat toys like this -- wiring the hardware and doing the programming are pretty easy, but where I'm stuck is building the actual cat toy bits that the electronics control! How have you approached this in your projects?

Certain rules do, though the exact analog rule of Conway's Game of Life is a bit boring as i recall. In general it was a lot easier to find interesting behavior in Brian's Brain rulesets.

Nice!! I had a lot of fun with a project like this with some more generalized rulesets. This video got the most views: https://www.youtube.com/watch?v=EW9Q0qMc2Xc (just running 3d Brian's Brain) but I like this one the most, with a much more general ruleset: https://www.youtube.com/watch?v=SoJvS6GuM3c&t=90s

The code is here: https://github.com/heyitsguay/gol3d

I keep meaning to go back and create an automated ruleset explorer that looks for stable but non-oscillating and non-constant population counts. Maybe someday...

I tried it this year, thought it was great! It could indeed be made still more convenient, but it was pretty convenient already.

Seems to me the article is more pointing out that when discourse on a digital platform starts broadly diverging from what's happening in the real world, lots of people gradually catch on and stop paying attention.

Understanding what's real is hard and attribution is messy, but the conspiracy stuff that's flooded popular social media platforms since 2016 or so just isn't real. There are plenty of unfortunate souls caught up in the narrative, but far more people are turned off by it and move on.

Likewise, was very sad when picturesforsadchildren went offline! Can't believe no archive survived.

What if all of their serious coding was done as part of previous employment and they can't just walk you through it?

To my knowledge as a math-turned-ML guy, there are currently no useful geometric characterizations of deep net latent spaces that are both "deep" (in the sense of using advanced mathematics) and "useful" (in the sense of revealing properties of networks or their latent spaces that aren't understood otherwise). Of course if anyone knows better I'd love to hear about it.

Continuous geometric concepts don't play super well with the way we like to decompose model outputs into discrete entities (classes, words, visual properties). We can, e.g. find variables in celebrity face GAN latent spaces that seem related to face orientation, or hair color, sort of, over some variable range and under some input conditions, but that doesn't really translate cleanly into any typical mathematical characterizations, geometric or otherwise, where you'd be looking for some property to hold everywhere or at least have an atlas of connected local approximations to simple characterizations.

Instead, we get high-dimensional messes of spaces, and network gradients during training don't exhibit clean or easy to understand dynamics except in the simplest toy cases.

To paraphrase a more serious "math for ML" prof I've chatted with at times -- "doing math" classically involves being able to find a description with only a few free parameters for a complex phenomenon that may superficially appear to have many/infinite free parameters. It's possible that for large ML models trained on natural data, such a reduction just doesn't exist, you can't break the contributions of millions or billions of parameters down into a low-dimensional approximation. He was/is skeptical of us attaining deep mathematical insight into their operation, but he could always be wrong. I'd certainly love to see cool novel insights come out of mathematics that give clarity to what's been going on these past 15 years.

But if you're in a competitive market and there's more work to do than you're addressing with 100 people, your 91 people doing the work of 100 will be outcompeted by 100 people doing the work of 110, no?

Also if anyone wants my Monocle to hack on themselves, I'd be happy to send it off for the cost of shipping!

Edit: Claimed

Does anyone have any interesting Monocle use cases to report? I ordered one and tried hacking around with it, found very little in terms of durable utility or even hackability for interesting computer vision/AR applications, so would be curious to hear if anyone did better.

That's not to knock the creators btw, I think they're working right at the edge of some incredibly challenging hardware constraints, just that maybe the tech is still immature for those who want to apply it, as opposed to those who want to hack on the tech stack itself.

As someone who has been using pgvector for a while and is vaguely curious about alternatives without having the bandwidth to investigate -- is there anything out there that offers truly differentiated advantages over pgvector? I'm extremely wary of non-OSS solutions in this area, it seems ripe for enshittification and attempts at vendor lock-in.

I don't think that comment was an indictment of anything. As someone also in the space, it reads to me as a thoughtful explanation of why catching up to the CUDA ecosystem is so hard, even when the goal is "just" matrix multiplication (it's actually not, but for the sake of a cleaner argument that's ok). It's entirely possible AMD or Intel or someone else has some super secret CUDA killer hidden away somewhere, but nobody has seen any evidence of it. It's a bit of an extraordinary claim, and the community would love to see some extraordinary evidence.

My comment was about wanting to see more (nerds) -> (public) communication, not about anything (public) -> (nerds). I understand they're not good at it, it was just an idealistic lament.

My bone to pick with Musk and Altman and their ilk is their damage to public discourse, not that they're getting attention per se. Whether that public discourse damage really matters is its own conversation.

They're talking about shit that isn't real because it advances their personal goals, keeps eyes on them, whatever. I think the effect on funding is overhyped -- OpenAI got their big investment before this doomer/e-acc dueling narrative surge, and serious investors are still determining viability through due diligence, not social media front pages.

Basically, it's just more self-serving media pollution in an era that's drowning in it. Let the nerds who actually make this stuff have their say and argue it out, it's a shame they're famously bad at grabbing and holding onto the spotlight.

This piece frames this as a debate between broad camps of AI makers, but in my experience both the accelerationist and doomer sides are basically media/attention economy phenomena -- narratives wielded by those who know the power of compelling narratives in media. The bulk of the AI researchers, engineers, etc I know kind of just roll their eyes at both. We know there are concrete, mundane, but important application risks in AI product development, like dataset bias and the perils of imperfect automated decision making, and it's a shame that tech-weak showmen like Musk and Altman suck up so much discursive oxygen.

Completely agreed. Legalizing drugs and gambling is important for cutting off the flow of money and social legitimacy to criminal operations, but the "nice" thing (from a marketing perspective) about addictive substances and activities is you don't need to try hard to sell them. For users, even addicts, the bar set by having to deal with black markets is so low that we can get them out of the streets with the truly blandest, unenticing solutions -- plain black and white boxes, black and white screens, beige halls. Just cheap enough to undercut illegal alternatives, with all color and marketing effort placed on the offramps: addiction treatment information and guidance framed in a friendly, empathetic, medical manner.

My understanding is that this project has been around for at least 3 years (going by reddit posts). Are there any tangible accomplishments that anyone can point to?

For instance, I googled "OriginTrail is enabling frictionless international trade for over 40% of US imports by fostering secure exchange of security audits" and found it repeated, verbatim, in a PR news wire and some crypto forums, but nothing else.

Because there are no intermediate internal registers with values that can be altered by a prompt, it's prompt in -> pass through deterministic computations with fixed trained weights -> token out. Context-dependent planning (beyond what you get from next-token prediction, which is already a lot) requires additional mutable state.

Only for a month, could believe it's about making sure it's polished.

More generally though, the fact that matching GPT-4 is the marketing blurb as opposed to a GPT-3 -> GPT-4 level transition has me wondering: are we seeing the limits of current "just scale the data" approaches to ML? Like, we've basically scraped everything worth scraping off the web, and when you do that with any of our current techniques, it lands you somewhere around GPT-4?

I don't have the link anymore, but i saw a quote from a LinkedIn connection recently at one of the FAANG's about how data matters so much more than architecture for ML today. To paraphrase, it went something like "it turns out that when your data is large enough, architecture doesn't matter. GANs converge to about the same thing as diffusion models. Convnets converge to about the same thing as ViTs. Data size and quality matter a lot more".

In general I'd agree that's the right response to engagement farming, but in this case if that image really was faked, it's probably important to make that clear -- there are major implications for photos-as-evidence between a little light denoising and HDR touch-up, vs an opaque process that's stitching together disparate image regions over hundreds or thousands of milliseconds.

Kind of an odd blog post -- is this AI generated? Or maybe written by someone who isn't a native English speaker? The writing feels off.

In any case, maybe worth linking directly to the paper, it's an easy read and more informative: https://arxiv.org/abs/2311.14029

The tl;dr is they tested CLIP variants for zero-shot classification on CIFAR-10 and STL-10 at different levels of image compression and observed how performance dropped with more compression. They claim the purpose is more about demonstrating performance change attribution using this Integrated Gradients method but imo it doesn't appear more informative than just the general, intuitive observation that decreased image quality = decreased CLIP performance.

If you read "Attention Is All You Need", you'll see that was actually the motive behind the original Transformer too! Bit of a Jevon's paradox situation.

So... Not just crime, but also the digital infrastructure that facilitates it.

How would you define "new best HW"? Seems challenging, particularly if it's for generation N+1 ML models that haven't been created yet. Also, while there is work underway to use ML models to guide HW design, it's not clear to me that the best ML models for that == the best ML models for more general purpose tasks in audio, visual, and natural language processing, i.e. is HW circuit design done using transformers? What are the inputs, latent space, and outputs?

GPU development seems driven by more general computational principles that might be summed up glibly as "We're hitting fundamental physics limits for single-core processors, what is the maximum amount of data we can move per second through the maximum number of cores?" Perhaps there is a way to extend this with a view toward the challenges of current ML model training and inference, but I imagine trying to approach it as a black-box optimization problem could be quite difficult.

Bitcoin and other PoW crypto burn energy by design and found few to no enduring use cases.

Talking about the energy usage of AI is like talking about the growing energy usage of lighbulbs at the turn of the 20th century. It's a byproduct of adoption and reduced-energy solutions will readily be taken up if/when found.