winterismute

> Is it "locked"? Yes, by copyright laws, if you don't like that lobby to have them changed.

I don't if that is true: a lot of old books might still have copy or other rights associated to them, likely owned by author and/or publisher, directly or inherited, but often those who have the rights do not have digital or physical copies at hand anymore (some old books are, well, really old). Does Anthropic make sure to track down, contact and then share the digital copy they make with those who have rights on the work? If not, they are not making it in any way easier to re-print the books, while making their supply more scarce (they destroy existing embodiments).

Isn't poolside a completely different company from Mistral?

While I agree that this setup is not very good, the EU Commission is essentially the equivalent of a government in a parliamentary or presidential democracy, and almost all such democracies I know about either have or have developed "ways" for the government to "push through" the will of the "chamber(s)". However, we do not consider those as farces as well, or so it seems to me, I guess mostly because those governments are still more or less directly "bootstrapped", and needs to be supported, by the chambers, which is what happens in the EU parliament as well.

Indeed, I find quite ironic that some people in tech in the US complain about EU "regulations first" approach, but then their government seem to arbitrarily stop things from being released because, well, there is no established policy on safety guarantees or other similar aspects.

Oddly, the website lists "M4 Ultra" which however does not exist... Also, it does not account for Apple Silicon chips to have up to 512GB of memory in some cases, but that might be only a limitation of the gathered data.

Sounds nice, for how many years have you had that annual recurring revenue so far?

> If you have some issue with that terminology, by all means raise that issue, but "You can not have" is just factually incorrect here.

It is not incorrect because, at least for now, all those "path tracing" modes do not do compute multiple "paths" (with each being made of multiple rays casted) per pixel but rasterize primary rays and then either fire 1 [in rare occasions, 2] rays for such a pixel, or, more often, read a value from a local special cache called a "reservoir" or from a radiance cache - which is sometimes a neural network. All of this goes even against the defition your first article gives itself of path tracing :D

I don't have problems with many people calling it "path tracing" in the same way I don't have issues with many (more) people calling Chrome "Google" or any browser "the internet", but if one wants to talk about future trends in computing (or is posting on hacker news!) I believe it's better to indicate a browser as a browser, Google as a search engine, and Path Tracing as what it is.

There is no real difference between "Ray Tracing" and "Path Tracing", or better, the former is just the operation of intersecting a ray with a scene (and not a rendering technique), the latter is a way to solve the integral to approximate the rendering equation (hence, it could be considered a rendering technique). Sure, you can go back to the terminology used by Kajiya in his earlier works etc etc, but it was only a "academic terminology game" which is worthless today. Today, the former is accelerated by HW since around a decade (I am cunting the PowerVR wizard). The latter is how most of non-realtime rendering renders frames.

You can not have "Path Tracing" in games, not according to what it is. And it also probably does not make sense, because the goal of real-time rendering is not to render the perfect frame at any time, but it is to produce the best reactive, coherent sequence of frames possible in response to simulation and players inputs. This being said, HW ray tracing is still somehow game changing because it shapes a SIMT HW to make it good at inherently divergent computation (eg. traversing a graph of nodes representing a scene): following this direction, many more things will be unlocked in real-time simulation and rendering. But not 6k samples unidirectionally path-traced per pixel in a game.

I read this analysis of the SMR farm announcement in Canada a few months ago and I found it quite insightful: https://www.carboncommentary.com/blog/2025/5/11/the-first-te...

Ok, but how could we do that? Especially since thing like eg. work is moving little by little but more and more towards remote...

Living in the UK but being from another EU country, I definitely see that happening. However, a lot of times it is just due to habits, wrongly-placed mistrust, or not being well settled-in yet because, at the end of the day, there are eg. better GPs and worse GPs everywhere in the world, but if you are still "new" to the country you simply do not know which ones are which, so you prefer to go to the ones you know already.

I always thought linking all the main things not working in the actual world to the alienation caused by too much digital consumption to be wrong/not really making sense. However, gradually, I am getting closer and closer to that conclusion... In your case, what brought you to the stance "Too much social media is what ails the whole world"? What do you think we could do to solve it?

The paper is about integer multiplication, not float

ETHZ and EPFL are also top of the market in EU/UK.

His take was not really "novel" however, John McCarthy said basically the same thing multiple times in the 90s and maybe even 80s? He would say something along the lines of "If we ever get to an algorithm that expresses general intelligence, we will be able to write that in one or two pages of a manual. Such a book will still be rather long and the rest of the pages will be about how we got to that algorithm and why it took us so long".

I am not an expert on geometry processing pipelines, however Mesh Shaders are specced differently from GS, essentially one of the big problems with GS is that it's basically impossible for the HW, even after all the render state is set and a shader is bound and compiled (and "searched"), to understand how much memory and compute the execution will take, which breaks a lot of the assumptions that allow SIMD machines to work well. In fact, the main advertised feature of GS was to create geometry out of nothing (unbounded particle effects), while the main advertised feature of Mesh Shaders is GPU-driven and efficient culling of geometry (see for example the recent mesh shader pipeline talk from Remedy on Alan Wake 2). It is true that Mesh Shaders are designed also for amplification, and that word has been chosen specifically to hint that you will be able to "multiply" your primitives but not generating random sequences out of thin air.

It is also true however that advances in APIs and HW desgins allowed for some parts that were troublesome at the time of GS not to be so troublesome anymore.

It's also the foundation of a very good estimator for global illumination! See https://en.wikipedia.org/wiki/Metropolis_light_transport . It used to achieve state of the art quality on scenes with caustics and/or tough indirect-light dominated scenes (eg. veach door).

The problem imho is that a lot of people, not only "the very rich" but also some very "middle class" people/families, have access and can make use of houses as investment vehicles, but houses (aka shelter) is also a primary need for humans. If you hold share of a publicly traded company, you can hold it forever until you think somebody can accept to buy it from you at a price that you like, hoping to make a profit if such price is high enough, and nobody will suffer from this process. But with houses, there is always somebody in absolute need for it, which means that either they will accept to rent it to a price that covers extra taxes applied by the state to you (as a landlord), or they will try to squat if they can not. It's really hard to enforce the right set of disincentives that are wide enough to convince people not to "hold" but at the same time does not apply to too many people but mostly the ones that are using houses as investments.

The whole thing is complicated by being geographically unequal: for example I think even Spain is full of affordable houses, only they are not in Madrid or Barcelona or Valencia, which is where people really want to live. So if you have a second house in an unpopular town you actually have not much - thus you are not rich - and you have often an empty house (nobody wants to rent/buy it) which is an easy target for squatters, and therefore you will become "one of the poor people ruined by squatters", while othen you are somebody who accepted the narrative that using houses as investment was a good idea, both an investment house cheaply in a town that never attracted enough people, thus "lost the game" and is now also losing the house to squatters...

> Part of the problem is that many of the concepts OpenGL teaches you have no bearing on how modern hardware actually works, so you end up having to unlearn bad habits which OpenGLs messy abstractions enable. OpenGL won't teach you to think in terms of PSOs, for example.

While this is true, for somebody who is starting from scratch there is a lot to learn before getting to the level at which thinking in terms of PSO is important, and it can be easier to get there via OpenGL, which btw still teaches you a decent chunk of GPU-friendly patterns (assuming of course we are talking about "modern" OpenGL and not display lists and such...). Also, with a good command of OpenGL, one can start trying to understand and re-implement rendering techniques spanning from deferred/forward+/clustered lighting, the various shadowing techniques and even HW raytracing eg. via the GLSL_NV_ray_tracingextension, which is - in my opinion - the more important side of learning GPU-accelerated rendering.

I am looking at solving this challenge in a specific way: using high-perf, GPU-accelerated HW simulators and ML algorithms to tune a new HW architecture automatically. Best ML HW => run on it the best ML models => produce new best HW (arch) => build new best HW => GOTO 10.

Reach out if you are interested in any way.

Isn't it 380k per person in average? Seems like in-line with FAANG salaries in major US cities.

I have always been fascinated by the problem of quantum gravity but, well, it somehow happened "too late", after I got also very deeply into "computing" (mostly HPC, GPGPU, rendering). Does anybody know if there is a way for somebody with my background to actually help/contribute in advancing this field?

Metal is a modern - and under many points of view, pretty good - rendering API that allows you, for example, to declare the rendering pipelines/dispatches fully without you having to call multiple "change_state" in your CPU code (similar to DX12, Vulkan etc). This means the GPU can run "flat out" and that a driver can be pretty lean and introduce little overhead (if done well enough).

In more advanced cases when many API commands need to be encoded depending on users' input, and if batching is not enough, you can use things like "MTLParallelRenderCommandEncoder" and multiple queues to take adavantage of multiple threads submitting such commands.

This all means that "pushing the envelope" in terms of rendering is absolutely possible despite the possibly small overhead added by Obj-C - which however also brings some advantages as it often avoids having to call "free" on GPU memory since the MTLBuffers are reference-counted.

The database was the key technology in the 2001-2011 decade: it allowed companies to store massive amount of data in an organized way, so that they could provide basic functionality (search, monitoring) to users. Statistical learning is being the key "technology" of 2011-today: it allowed companies, which had stored massive amount of data, to feedback predictions to users. I think AR/Computer Graphics will be the key technology of the next decade: it will allow users to interact directly and seamlessly with the insights produced by ML systems, and possibly feed-back information.

That's actually a great idea, the lack of effort spent in keeping friends is an underestimated issues, at least in many western cities...

I still feel bad that the M1 brand name has relegated M1 the rapper (even more) to oblivion https://en.wikipedia.org/wiki/M-1_(rapper)

It's so incredible to see this today as only yesterday I have played this with some friends: https://thornygames.com/pages/sign It is board game based on a stripped down rendition of the story of NSL, I really recommend it as it is simple but it makes for a game experience which is pleasantly unusual.

Nice! I think you have a type in "heaver", it should be "heavier" maybe?

It's a tough question, it's not even back-propagation but even just sometimes the "parameters" of the models, for example [1] shows that models such as ResNeXt already perform better on a very different architecture such as Graphcore, for some sizes of convolutions. Older models, or models that get tuned for existing GPUs, do not perform as well.

It's tough to come up with a new architecture that can have an advantage on current and future models, at least from a peak perf point of view, from a perf/watt for example instead the scaled-up Apple GPUs seem to show new interesting properties. But the Graphcore architecture is quite interesting, being able to act somehow both as a SIMD machine and a task-parallel machine.

[1] https://arxiv.org/pdf/1912.03413v1.pdf

First, you should be proud of yourself for striving to do "the right thing".

In the field I follow the most (Computer Graphics/Rendering) I think there is a big problem with reproducibility as well, and to be honest, I think some of the major players actually have little interest in making this significantly better, since they can take advantage of the visibility of a flashy render/fps counter shown at an event while still keep on building a "moat" between them and others that want to adopt the same methods.

Which is in the end partly an answer to your question: your paper could clearly describe all the elements needed to implement a method correctly, but by providing a sample implementation you allow others to "stand" on your shoulders, as they say, instead of having to climb there first and then proceed. You can not worry too much about the state of your codebase by making clear via README/documentation/license that it's still in "proof of concept" phase.

One reasonable observation I have heard is that in some fields, during peer review, some reviewers seem to like to nitpick on the code rather than the paper, sometimes in subtle ways. Because of that, I think it can be (unfortunately) OK to release the code after acceptance or publication. But apart from this, I see only advantages.