HN user
Contact: <hnusername>@gmail
Also, there is zero reason to think that the big labs did not have anything similar to TurboQuant for a long time already.
The recent blog post from Google announcing TurboQuant does not change anything regarding RAM planning for the big labs.
TurboQuant itself is already a year old! So even smaller labs have probably seen and implemented it.
Cheapest i know if is around $96k
The memory bandwith on M4 Max is 546 GB/s, M5 Max is 614GB/s, so not a huge jump.
The new tensor cores, sorry, "Neural Accelerator" only really help with prompt preprocessing aka prefill, and not with token generation. Token generation is memory bound.
Hopefully the Ultra version (if it exists) has a bigger jump in memory bandwidth and maximum RAM.
Related to this, how do you get your comments that you add in the review back into your agent (Claude Code, Cursor, Codex etc.)? Everybody talks about AI doing the code review, but I want a solution for the inverse - I review AI code and it should then go away and fix all the comments, and then update the PR.
I think the ability to actually run the code on the target helped a lot with understanding and optimizing for the specific micro architecture. Quite a few of the ideas turned out to not to be optimal and were discarded.
Also important to have a few test cases the agent can quickly check against, it will often generate wrong code, but if that is easily detectable the agent can fix it and continue quickly.
I used Claude Opus 4.5 inside Cursor to write RISC-V Vector/SIMD code. Specifically Depthwise Convolution and normal Convolution layers for a CNN.
I started out by letting it write a naive C version without intrinsic, and validated it against the PyTorch version.
Then I asked it (and two other models, Gemini 3.0 and GPT 5.1) to come up with some ideas on how to make it faster using SIMD vector instructions and write those down as markdown files.
Finally, I started the agent loop by giving Cursor those three markdown files, the naive C code and some more information on how to compile the code, and also an SSH command where it can upload the program and test it.
It then tested a few different variants, ran it on the target (RISC-V SBC, OrangePI RV2) to check if it improves runtime, and then continue from there. It did this 10 times, until it arrived at the final version.
The final code is very readable, and faster than any other library or compiler that I have found so far. I think the clear guardrails (output has to match exactly the reference output from PyTorch, performance must be better than before) makes this work very well.
Seems like the hiker at the bottom of the article was introduced in 1997 and removed only in 2017: https://s.geo.admin.ch/be66brq5oby9
I have the PCIe version of NanoKVM, and I am also happy with it.
The big advantage of the PCIe version is that it does not take up space on the desk and all the cables for ATX power control an inside the PC case.
Full-sized HDMI is nice, the only limitation here is 1080p resolution. 1440p or higher would allow mirroring the output on the main monitor to the NanoKVM, but this probably a weird use-case anyway.
It says that there are multiple sizes in the second sentence of the huggingface page: https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct
You won't be out of work creating ggufs anytime soon :)
The VPN product is very good, it's basically a thin wrapper around Mullvad, arguably the best VPN on the planet right now. At least from a privacy standpoint.
Why do you think it's a negative result? The table on page 9 shows great results.
There are quite a few well funded companies now that use forks of VSCode: Google (IDX, now Firebase Studio), Cursor, Windsurf (especially if they are bought by OpenAI soon), GitLab etc.
An alternative marketplace already exists with open-vsx.org, but these companies need to fund the creation of alternatives for the proprietary extensions (C#, C++, Python, maybe others) that can be used by all forks.
Gamers Nexus is talking to one big PC manufacturer (my guess is Dell) that is seeing failure rates of 10-25% for specific SKUs: https://youtu.be/gTeubeCIwRw?t=527&si=YzpDzI2IyadzQYid
Not fully confirmed yet, but that sounds really bad. It seems like it also hits low power models like the 13900T, which would imply this isn't just a voltage issue from auto overclocking.
The best way to swap desktops is to install the correct Ubuntu flavour from the start, i.e. Kubuntu, Xubuntu or Lubuntu to avoid those problems.
I am quite happy with Regolith Desktop [0] installed on top of standard Ubuntu. Gives you a nice tilling window manager based desktop without fiddling with config files for hours.
Regolith Desktop (https://regolith-desktop.com/). A fully preconfigured Tiling Window Manager based desktop environment for Linux.
Uses i3 on X11, sway on Wayland.
No more fiddling with config files to make basics like system settings actually work under i3.
Note that YouTube Shorts are just normal YouTube videos on a different URL. This means you can take a link to a Shorts like https://www.youtube.com/shorts/<id> and change it to https://www.youtube.com/watch?v=<id>.
This way, you won't get the dreaded doomscroll interface, and you can also scrub through the video!
I wonder if they have a clear hardware separation between each of the API, ChatGPT, their lower-scale experiments and their large scale (e.g. GPT5) training hardware. Or is everything just a big blob of hardware, that dynamically gets allocated to jobs depending on demand?
Hardware demand is so high, having GPUs idling is a massive waste, but you also want to have separation between dev, test and prod environments, so not obvious what to do.
Consumer AMD chips to be precise. AMD is mostly focusing on their datacenter chips (MI300X), and I assume the support for AI workloads there is much better. They might even see their consumer chip undermining sales of datacenter chips.
NVIDIA got big because CUDA works on the most crappy notebook GPUs up to their most powerful chips, and AMD should do the same, but focusing their limited number of driver devs on the expensive enterprise hardware makes sense IMHO.
I wouldn't be surprised if the answer is "all of the above" and more. Not only is there limited availability of NVIDIA chips, it's also useful to tell Jensen that you actually could buy from somewhere else, even if you don't really want to.
At the end of the day, the bottleneck is TSMC, on which all GPUs are produced right now, even the ones from Intel.
I remember similar things. Running Windows 2000, even going back to Win98 on my first PC (shipped with XP before SP1) and also running Windows 2003 Server at some point, all just to avoid XP. After SP2, it was fine.
I can only access https://makersuite.google.com/ when using a VPN to the US. Also, it spams popups that get blocked by Firefox.
Some basic prompts, which are answered correctly most of the time by ChatGPT4:
There are 31 books in my house. I read 2 books over the weekend. How many books are still in my house?
29 books
Julia has three brothers, each of them has two sisters. How many sisters does Julia have?
Three
If you place an orange below a plate in the living room, and then move the plate to the kitchen, where is the orange now?
Under the plate in the kitchen.
So, not great.
A charitable interpretation might be that search requires a fair amount of compute, and is therefore a big denial of service vector.
I am not sure how much behavioral data GitHub can gather from logged in user, and how useful that is compared to the code that is there anyway. Maybe to figure out which parts of code are important? But that isn't really user-specific.
The recently released XTTS-v2 model[0] from coqui.ai is coming very close to what ElevenLabs[1] can do. It runs reasonably fast on a recent GPU, and should also work on CPU. Requires a 3 second (!) clip of the voice you want to clone. License does not allow commercial use.
That is a very editorialized title, original is "Bcachefs Lands Big Scalability Improvement, Disables Debug Option By Default".
Also note that debug is enabled by default, so Phoronix tested the use-case most end users would actually run.
Looking forward to the new benchmarks, I wanted a filesystem with good snapshoting support for a long time, it's nice to easily revert a broken Nvidia driver update...
In my experience, people do hack game engines all the time for game specific optimizations, fixes and special features. At least in the bigger AA/AAA games I've seen. We tried to sell our plugin as binary only (C++), but basically everybody wanted the source code to hack and extend it. So people wouldn't be happy with MPL.
I agree for other things, e.g. webdev, I never had to hack Django or Vue.js.
The game industry is notorious for using, but not contributing back to open source. I don't want to know how often the same bug in recast/detour has been fix in different games, and never upstreamed.
The Apache 2 license is incompatible with GPL2, you cannot have code under those two licenses in the same project. So they dual licenses it under MIT, which is compatible with GPL2.
The Rust project has the same problem, it's also dual-licensed under Apache2 + MIT.
You want Apache 2 because of the explicit patent grant, but at the same time you also want to allow GPL2 projects to use your code.
I worked with it quite a bit on my last job, and for me it's really the quality that is not there yet. One example is prefabs: They worked just fine in Lumberyard years ago. The first few releases of O3DE broke them (that was in 2021 I believe), because like almost every other subsystem, Amazon wanted to rewrite it from scratch. Seems like they released the new version in the latest release, but as you can see in this video from Gamesfromscratch, it still doesn't work: https://www.youtube.com/watch?v=5PAuR-pEz1w
Also, they are not great at marketing. For example, in the latest release notes, the main headline talking point is robotics, and not games.
I think the tech is actually good, I prefer both the component system and ScriptCanvas over Unreals component mess and Blueprints. It just needs much more polish. And a big game company trying to actually make a game.
I've been using Cursor quite a bit on my two current personal projects in Python and C++ respectively, and I think it's quite useful:
I haven't written a unit test by hand since I started using Cursor. Because it has the full codebase as context, it can easily generate unit tests. I would say the code is correct around 50% of the time on the first try? Otherwise, you can quickly submit a follow-up prompt to fix stuff. Obviously, it's still GPT-4 in the background, so hallucinations still happen.
I tried adding a simple feature to my interpreted custom programming language using the chat feature, giving the full codebase as context: The programming language already supported indexing into arrays, but not strings. I asked in chat what I need to change to make this feature work. It generated code and showed me in which files I have to add or modify existing code. Very impressive. Again, the code wasn't perfect, I had to do some minor adjustments like method names, but it understood my code, gave the right files to modify, and also proposed code that was almost correct. There is a beta feature that applies the proposed code changes to the codebase, but that did not work yet when I tried it (I am using an OpenAI API key, maybe it's related to that).
Overall, I think it is very helpful, especially when I am tired from $dayjob, but still want to do some coding in the evening. For me, it's much easier to formulate what I want to code in human language, instead of writing the code "by hand".
Caveat: Even if you use the OpenAI API key, ALL your code is sent to the cursor.so servers for prompt generation, indexing etc., even with "local mode" on. "Local mode" just changes a flag that cursor won't save code on their servers, but it's still sent there for processing. I assume this is a deal-breaker for most companies for now.
ROCm has HIP (1) which is a compatibility layer to run CUDA code on AMD GPUs. In theory, you only have to adjust #includes, and everything should just work, but as usual, reality is different.
Newer backends for AI frameworks like OpenXLA and OpenAI Triton directly generate GPU native code using MLIR and LLVM, they do not use CUDA apart from some glue code to actually load the code onto the GPU and get the data there. Both already support ROCm, but from what I've read the support is not as mature yet compared to NVIDIA.