Hacker Newsnew | past | comments | ask | show | jobs | submit | somewhatrandom9's commentslogin

Agree. Probably an unpopular opinion, but I strongly dislike YAML.

EDIT: if I HAD to use YAML, I'd prefer KYAML: https://dev.to/mechcloud_academy/goodbye-yaml-hell-meet-kyam...


KYAML seems eerily close to json5.

I saw your response about not blocking ads currently. What about blocking adult content, for a child's computer? Mullvad had such an option[1] under family.dns.mullvad.net. Very useful!

[1]https://mullvad.net/en/help/dns-over-https-and-dns-over-tls


This is an area that is fraught with challenges. Serving a truly global audience means standards are difficult to measure as to what is "adult content" - it varies widely (Wikipedia in some areas is "adult content", as an example.)

In the near term, there is no filter set that we offer that mimics that functionality.

However, we recognize some variation of "best effort" adult content segmentation is a useful filter to have. We're considering it in the near term, as we have some specific educational institutional pressure to try to meet as well, since students are often easy targets for phishing, classroom computers are rife with malware, and there is also the problem of inappropriate DNS-based profiling of students and schools - problems that Quad9 can help solve.


Perhaps you could ask Mullvad how they did it? You might protect yourself with disclaimers that your results may not be perfect? You still could create a useful/ effective service even if it's not 100% perfect I would think.


If this article is accurate, I hope it leads to an intervention.


I've heard it rumored (maybe on this site??) that it's wise to avoid upgrading to Tahoe, on intel macs, but I don't have any evidence to back this up.


How is this different from Divio's documentation system?: https://docs.divio.com/documentation-system/

Update: https://diataxis.fr/colophon/#origins-and-development (Divio came first).


Same fundamental ideas, but I got a lot of things wrong in that earlier version (which is several years old now).


But when it comes to diagram, I still refer to the Divio version as Diataxis version description is too abstract.


Though it doesn't get by all the hurdles mentioned, it is alternatively possible to run Mac OS in a VM on your Mac using UTM and install Claude Code within the UTM VM. UTM can be run under a non-admin Mac account. This can allow you to use most Mac-native tooling, at least. The interactive performance of using the Claude Code ui on the VM isn't great, however. I'm not sure if you can log into the VM via terminal on the host from the non-root admin account to avoid the ui performance issues.


An argument against this method is UTM doesn't support graphics acceleration so browser support will be hobbled. Even if you don't need the acceleration, I've found browsers in UTM virtualized OS's can't get past some modern captchas and other browser fingerprinting checkpoints :\. It's terrible but that's the way it is.


I don't know about solving captchas, but if you are running macOS virtual machines on an Apple Silicon Mac, hardware graphics acceleration is supposed to be natively supported out of the box when using UTM's Apple Virtualization Framework backend.

A primary source of UI lag is how Apple's native virtualization framework processes multitouch trackpad events. There are other issues like mismatched resolutions and framerates, too. Ask your favorite ai to debug. You can try e.g. deselection of the trackpad setting and instead select basic mouse support, and that can clear things up..


> browsers in UTM virtualized OS's can't get past some modern captchas and other browser fingerprinting checkpoints :\

I'm using Qubes OS, where everything runs in VMs without GPU acceleration, and never experienced this.


Try using DwarfStar 4 and use the --power flag: https://github.com/antirez/ds4#reducing-heat-power-usage-and...


Can you run Qwen 3.6 27B on antirez/ds4 now? I thought it was all about the DeepSeek models.


No, I don't think Qwen, but I believe he may try and put some version of GLM in it.


DwarfStar is the only thing I've run that doesn't try and make my Mac Studio 128GB take off. Yes, it gets hot while doing inference but quickly cools down when idling, something I haven't experienced with Ollama, LMStudio or OMLX.


You can get a new M5 Max MacBook pro with 128 GB unified ram (targeted by Antirez for DwarfStar4) even after the Apple price increases, it's less than 7899 by at least $1000. And you probably won't pull more than 100 Watts.


You are comparing US pricing with EU pricing. EU pricing includes 21% VAT and currency conversion "rounding up".

The cheapest 128GB Macbook Pro here costs €7.949,00.

No doubt a better value than the HP, and will depreciate a lot less quickly, but just as expensive. Unfortunately, not being able to run Linux is a breaking point for me.


I have one of these. Got it a few weeks before the price increases. On the 14" version charging is limited to 96 watts, but the chip can pull north of that with adequate cooling, so the battery will literally drain while plugged in.

It isn't a problem for me, more amusing than anything else (I run in Low Power mode 90% of the time) but worth knowing for anyone thats thinking about pushing the hardware to its limit 24/7.


> And you probably won't pull more than 100 Watts

I've played with ds4 on such a machine, and the battery drains when using only a 96W power brick. But you can put it in low power mode and the fans won't even turn on while it delivers something like a third of the performance.


Could these quantized models make MTP (Multi-Token Prediction) significantly faster when used as drafters for larger regular Gemma 4 models?


Google already released specialized drafters for Gemma 4.


The E2B ones? Or what do you mean by specialized drafters?


They have -assistant in the name, so e.g.: https://huggingface.co/google/gemma-4-31B-it-assistant


Thanks


The “-assistant” models released by Google are specialised tiny MTP draft models :)

31b-it-assistant is what enables MTP


With "intelligence" (or whatever you want to call it) and speed both seeming to ramp up quickly with local models I wonder what the growth rate and ceiling(?) might be in this space. Will this kind of iq and performance work with just e.g: 16GB RAM in a couple years? Is there a new kind of Moore's law to be defined here?


640gb ought to be enough for anybody


Squeezing a model like this complete with 'big model smell' into 16GB...Honestly it's not even possible or feasibly possible today.

It'll require some kind of:

- breakthrough in architecture or

- breakthrough in hardware or

- some breakthrough quantisization technique

The problem is that all the parameters need to be in memory, even the ones that aren't active (say for Mixture Of Expert Models) because switching parametrs in and out of ram is far too slow.


"That’s where EMO comes in.

We show that EMO – a 1B-active, 14B-total-parameter (8-expert active, 128-expert total) MoE trained on 1 trillion tokens – supports selective expert use: for a given task or domain, we can use only a small subset of experts (just 12.5% of total experts) while retaining near full-model performance."

https://allenai.org/blog/emo


The people working at the leading edge of this stuff seem to believe that there is a need for parallel models that solve different problems.

A crow exhibits some degree of intelligence in what is a very small brain compared to humans. There is overlap in the problem solving skills of the dumbest humans and the smartest crows.

So the question is: what is that? Yann LeCun seems to think it’s what we now call world models. World models predict behaviour as opposed to predicting structured data (like language.)

If your model can predict how some world works (how you define world largely depends on the size of your training data), then in theory it is able to reason about cause and effect.

If you can combine cause and effect reasoning with language, you might get something truly intelligent.

That’s where things seem to be going. Once we have a prototype of that system, there will be many questions about how much data you really need. We’ve seen how even shrinking LLMs with 1-bit quantization can lead to models that exhibit a fairly strong understanding of language.

I don’t think it’s unreasonable to expect to see some very intelligent low (relatively) memory AI systems in the next couple years.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: