ykl

Why are macOS and OS X listed separately? Is it to delineate pre-2016 Macs from post-2016 Macs, when the rename to macOS happened? But in that case, I have a hard time believing that there are way more pre-2016 Macs floating around out there than post-2016 ones, as the chart shows. I have a hard time believing that anyone is running pre-2016 OS X versions at all, outside of a very small group of hobbyists; I thought Apple’s data showed that generally Mac users were pretty aggressive about being on relatively current macOS versions.

Also why separate out OS X and macOS at all when it appears that all Windows versions are lumped together and all Linux versions and distros are lumped together?

All of this seems very suspect.

Is anyone else amused that the car shown on the landing page looks a lot like a Tesla Model Y, which famously does _not_ use any Nvidia chips (Tesla onboard computers have been AMD based for some time now)?

AWS Bedrock is other companies’ models running on separate dedicated AWS hardware, metered through AWS billing. AWS owns and operates all of the infrastructure and the client interface; the model provider basically hands over the model and weights to AWS and AWS Bedrock take it from there.

So, as an example, if you use Codex through Bedrock, that’s a totally separate instance of Codex from anything you would be interfacing with if you directly used OpenAI’s API; if you use Codex via Bedrock, OpenAI never sees your data or prompts because they stay sandboxed in an ephemeral Bedrock instance. For many large enterprise deployments this hard boundary is a big big deal.

Over the past year, Claude being available via Bedrock and ChatGPT/Codex not being available via Bedrock has been a huge competitive advantage for Anthropic in the enterprise space.

If you've used AI coding models in a large corporate setting, you'll know that a lot of big corporate deployments basically require using AWS Bedrock for two simple reasons:

1. Large companies tend to already have an existing relationship with AWS, which makes things way easier to go through vs. setting up a new vendor relationship 2. Large companies tend to have strong internal requirements about making sure that internal data stays under company control. With AWS Bedrock, you can be a lot more confident that what you're feeding into the models is not going to end up in someone's training set somewhere. For where I work, this requirement is a dealbreaker for going directly through OpenAI's API instead of going through AWS Bedrock.

Ha, from the title alone I was hoping for a Derek Lowe article, and it is!

A medium-spicy take of mine is that a bytecode VM in a GPU kernel is not as bad of an idea as one might think, and in some cases it can actually be the most reasonable solution. Some fun examples:

1. As mentioned in the post above, the Dolphin emulator famously implements the entire Gamecube/Wii GPU pipeline in a single gigantic ubershader, and this is useful because it avoids shader compilation stalls [1].

2. Blender's Cycles renderer implements its shading graph eval system as a bytecode VM in a GPU kernel [2]. IIRC early versions of Vray GPU did something similar. There are better ways of course, but a VM gets you surprisingly far as a general approach.

3. Finally, a lot of ML frameworks (Tensorflow, PyTorch, etc) by default use the GPU relatively suboptimally (especially without kernel fusion and such). Tensor frameworks can extract a lot more perf out of GPUs using a VM-in-a-giant-kernel approach [3].

If you think abstractly about how a GPU SM actually works (using CUDA terminology here), all threads in a warp must execute in lockstep and the cost of execution divergence across threads in a warp is that you effectively run serially, losing the parallel advantage of the SM. This penalty gets magnified enormously if you are doing memory reads after wherever the execution divergence happens, since you now have multiple slow memory stalls in serial instead of one big memory read at once for all threads. If you're clever about implementing a bytecode VM, you can load as much state as you need upfront into shared memory, and then if your bytecode VM is just looping through executing a bunch of opcodes in a huge switch statement, then at least as far as the SM is concerned, there's no execution divergence! All threads look like they're doing the same thing at the same time; even if within the VM what is happening a lot is just no-ops, at the SM level you're not dealing with serialized memory stalls and serial scheduling and such.

Is it the _best_ most optimal approach imaginable? Almost certainly not! But can it be a _surprisingly good_ and possibly even reasonable approach for some problem domains and specific constraints? Yeah absolutely!

[1] https://dolphin-emu.org/blog/2017/07/30/ubershaders/ [2] https://www.youtube.com/watch?v=etGMk9wYwNs&t=1882s [3] https://hazyresearch.stanford.edu/blog/2025-09-28-tp-llama-m...

This assumes that everything outside of the convex hull can already be described using existing language. If you need new language to describe what is outside of the convex hull, is this something an LLM can do?

I actually don't know the answer to that; my understanding is that LLMs by nature of what they are can't understand concepts that are independent of the existing language they are trained on, but I don't have enough in-depth nitty-gritty knowledge of like, core LLM implementation details and architecture and stuff to know if that understanding is correct or not.

I was a little muddy in my original post on distinguishing between what I think LLMs might be able to do and what AI broadly might be able to do. I'm skeptical LLMs can expand the hull or add dimensions to the space; but I also don't think the reasons for that skepticism necessarily apply to all AI system generally.

I think of most things you can get to by guess and checking as definitionally inside of the hull; most forms of guess and checking are you take some existing thing, randomize a bunch of its parameters, and see what you get. Whereas with something like relativity, there's not even a starting point that you can randomize and guess/check from the pre-existing knowledge space that will lead you to relativity. That's more like, adding a new dimension to the space entirely.

It's possible LLMs can handle this after all! But at least so far we only have existence proofs of humans doing this, not LLMs yet, and I don't think it's easy to be certain how far away LLMs are from doing this. I should distinguish between LLMS and AI more generally here; I'm skeptical LLMs can do this, I think some other kind of more complete AI almost certainly can.

I supposed you could just, I dunno, randomly combine words into every conceivable sentence possible and treat each new sentence as a theory to somehow test and brute force your way through the infinite possible theories you could come up with. But at that point you're closer to the whole infinite random monkeys producing Shakespeare thing than you are to any useful conclusion about intelligence.

I like to think of it as:

Imagine every bit of human knowledge as a discrete point within some large high dimensional space of knowledge. You can draw a big convex hull around every single point of human knowledge in a space. A LLM, being trained within this convex hull, can interpolate between any set of existing discrete points in this hull to arrive at a point which is new, but still inside of the hull. Then there are points completely outside of the hull; whether or not LLMs can reach these is IMO up for debate.

Reaching new points inside of the hull is still really useful! Many new discoveries and proofs are these new points inside of the hull; arguable _most_ useful new discoveries and proofs are these. They're things that we may not have found before, but you can arrive at by using what we already have as starting points. Many math proofs and Nobel Prize winning discoveries are these types of points. Many haven't been found yet simply because nobody has put the time or effort towards finding them; LLMs can potentially speed this up a lot.

Then there are the points completely outside of hull, which cannot be reached by extrapolation/interpolation from existing points and require genuine novel leaps. I think some candidate examples for these types of points are like, making the leap from Newtonian physics to general relativity. Demis Hassabis had a whole point about training an AI with a physics knowledge cutoff date before 1915, then showing it the orbit of Mercury and seeing if it can independently arrive at general relativity as an evaluation of whether or not something is AGI. I have my doubts that existing LLMs can make this type of leap. It’s also true that most _humans_ can’t make these leaps either; we call Einstein a genius because he alone made the leap to general relativity. But at least while most humans can’t make this type of leap, we have existence proofs that every once in a while one can; this remains to be seen with AI.

At the time of the announcement IIRC the deal was only for Colossus 1. Is Anthropic also leasing Colossus 2 new?

At the time the consensus narrative was that SpaceX no longer needed Colossus 1 for Grok and that was why it could be leased to Anthropic while Colossus 2 would handle Grok training and inference. Does Anthropic also leasing Colossus 2 change this?

I wonder how many other people do this: when my wife and I do long roadtrips, we use Costcos as waypoints. Need to refuel? Costco gas is always cheaper than whatever other fill up station is nearby. Need to re-up on snacks (and maybe see what weird snacks the locals have that we don’t have back home)? Costco. Realize you forgot to pack enough socks halfway between Los Angeles and Salt Lake City? No prob, the St George Costco had the same exact socks I have at home because all of my socks are from Costco. Ran over a nail and need a tire swapped? Costco tire warranty babbbyyy. Bathroom? Costco bathrooms are very basic but always clean.

In big adventure RPG games there’s always some kind of shop in every new area that is the same inside everywhere that you can reliably go to for whatever gear you need, to heal, to save your game, whatever. Costco is that but in real life.

Something I really like about this post is that I think the high-level takeaway from this post is generally applicable far outside of the India AI industry. Sure, the specific application here is the AI industry and that's certainly the topic du jour, but I think this article does a great job of demonstrating how to think and see through the market distorting effects that subsidies produce, in order to figure out what the true cost structure and valuation in such an environment is.

At least for me Claude Code is still working on my Pro plan. I don't know if that's because the change simply hasn't propagated all the way through their systems yet (the change is now up on the main Claude pricing page and on their support pages, but not on the Claude Code landing page yet), or if it's because existing plans are grandfathered in, or what.

In general Anthropic seems to be pretty bad at clearly communicating what is going on. I have both Claude Pro for Claude Code and ChatGPT Plus for Codex, and lately I've been reaching for Codex first more and more often... at least for the hobby stuff I'm using Claude/Codex on, they seem pretty much equivalent in terms of practical capability/usefulness.

I'm really hopeful about John Ternus stepping into the CEO role. Pretty much everything he's done leading Apple's hardware engineering has been an enormous unqualified success, and for a company like Apple, having hardware lead the company seems like the right step.

No, it’s not. MLX is Apple’s NumPy more or less.

No, MLX is nothing like a Cuda translation layer at all. It’d be more accurate to describe MLX as a NumPy translation layer; it lets you write high level code dealing with NumPy style arrays and under the hood will use a Metal GPU or CUDA GPU for execution. It doesn’t translate existing CUDA code to run on non-CUDA devices.

Wouldn’t “insanely great”/“it’s shit” be more Steve than “genius”/“it’s shit”?

That is not just your opinion, that is the opinion of multiple United States Court of Appeals circuits in many many cases, and by its declining to overturn these cases, that is also the opinion of the United States Supreme Court. The United States is a common law country, so really what that means is that your opinion is actually not an opinion at all; you have simply stated the established law of the land.

If you aren’t familiar with the Amen Break, here’s a now classic 18 minute documentary on the Amen Break and its origins and evolution:

https://youtu.be/5SaFTm2bcac?si=J99_Sh9x3fIBCSms

Yeah, it used to be true that server GPUs at least somewhat resembled their gaming counterparts (i.e. Nvidia Tesla server components from 12+ years ago); they were still PCIe cards, just with server-optimized coolers, and fundamentally shared the same dies that the gaming and professional cards used.

That stopped being true many years ago though, and the divergence has only accelerated with the advent of AI datacenter usage. The form factor is now fundamentally different (SXM instead of PCIe); you can adapt an SXM card to PCIe with some effort [1], but that may not even be worthwhile because 1. the power and cooling requirements for the SXM cards are radically different than a desktop part and more importantly 2. the dies are no longer even close to being the same. IIRC, Blackwell AI chips straight up don't have rasterization hardware onboard at all; internally they look like a moderate number of general SMs attached to a huge number of tensor core. Modern AI GPUs are fundamentally optimized for, well, mat-mults, which is not at all what you want for gaming or really any non-AI application.

[1] https://l4rz.net/running-nvidia-sxm-gpus-in-consumer-pcs/

I hadn't seen arun.is before; I really like this one! Thanks for the share!

One I like is Tom Macwright’s blog [1], which somewhat famously loads insanely fast thanks to having a sort of the web equivalent of a brutalist design while still looking nice [2].

[1] https://macwright.com/

[2] https://macwright.com/2016/05/03/the-featherweight-website

I think this project is best described by one of my favorite quotes from Top Gear: “an ingenious solution to a problem that should never have existed in the first place”.

Hi! Original post author here.

When I first started at Disney Animation, at one point I asked Ed Catmull what the rationale was for staffing two separate rendering teams, and he had an interesting answer. His answer was that it turns out that even when Disney Animation was using RenderMan, the high end needs of the studio still required enough rendering developers/TDs that in terms of cost it was essentially no different than staffing a team to build an in-house renderer, and from that perspective he liked the idea of having two separate teams with different focuses/perspectives so that for hard problems the wider company got two attempts at coming up with good solutions instead of one.

To this day the Hyperion and RenderMan teams work pretty closely together and share a lot of learnings/tech/R&D. The focuses are pretty different between the two renderers, and that’s actually been pretty beneficial to both.

The story with Presto is both a bit different and kind of similar. The two studios are now unified in using Presto, but Disney Animation now has an in-house Presto development team that co-develops Presto with Pixar. The two dev teams focus on the needs of their respective studios but move Presto forward together.

Hi! Author of the original post here. Please see:

https://disneyanimation.com/publications/

Or, for a continuously updated list of publications specifically about Disney’s Hyperion Renderer:

https://blog.yiningkarlli.com/2019/07/hyperion-papers.html

I thought that QNX was acquired by / is still developed by BlackBerry?

One of the open secrets of AWS is that even though AWS has a lot of regions and availability zones, a lot of AWS services have control planes that are dependent on / hosted out of us-east-1 regardless of which region / AZ you're using, meaning even if you are using a different availability zone in a different region, us-east-1 going down still can mess you up.

I am pretty squeamish around blood (even reading about it...), but this piece was so well written and engrossing and impactful that I made it all the way through in one shot. I had no idea that the bovie instrument (and broadly the entire field of electrosurgery [1]) was even a thing; I assumed it was anll still scalpels and such.

An incredible piece highlighting something people should know more about; thanks for posting this!

[1] https://en.wikipedia.org/wiki/Electrosurgery

You can in fact already do this (have Mac windows show up as separate floating windows in VisionOS instead of all being on the single MacOS window) using a third party tool called Ensemble:

https://github.com/saagarjha/Ensemble

This is what is looks like in action:

https://twitter.com/TheOriginaliTE/status/175125156764134634...

Disclaimer: I don’t know how well it works in practice because I don’t have a Vision Pro.