Hacker News new | past | comments | ask | show | jobs | submit | darksaints's comments login

I have been down quite a few rabbit holes recently regarding trump and his ties to saudi arabia, and one thing that I constantly run into is corrupted archive data in articles about either Mohammad Bin Salman or Al Waleed bin Talal Al Saud. Now I'm more convinced than ever that Saudi Arabia has some sort of hand in this.


Can we please get a fraction of the resources currently put into Linux kernel development and start developing a robust userland ecosystem for SeL4?

Microkernels in general already mitigate the possible damage that could be done by rogue code in large monolithic kernels. A formally verified microkernel like SeL4 is an even better guarantee. And performance concerns of microkernels are practically solved at this point.

These sorts of nation-state sponsored malicious code practices could be made mostly irrelevant. We just need a little momentum to get us there.


The companies with large relevant apps running on the NDK are well staffed and funded enough to recompile.


It's not about whether they can, it's whether they will. History has proven that well-resourced teams don't like doing this very much and will drag their feet if given the chance.


it's not about that, it's about running the apps whose makers are out of business or just find it easier to tell their customers to buy different phones


It's a shame too. Scala3 is actually an amazing language, and has the best type system out of all of the common functional languages. Part of me wonders if Scala would still have died off if Scala3 came out first.


ocaml, f# and haskell


There is a way, which is through buildout enforcement. Basically, if you don't meet buildout deadlines and cover x% of pops within a specific time period, you forfeit the license. It should probably be a lot more stringent and with tighter deadlines, but the mechanism already exists.

It hasn't ever really been an issue with any spectrum that they've bought in the past, and it has only ever been a concern with mmWave spectrum, because the costs of coverage are much higher than were ever anticipated with their tiny effective ranges. Anything below 3GHz seems to get built out and used extremely quickly.


> There is a way, which is through buildout enforcement.

LOL, DISH squatted nationwode spectrum for years and it wasn't until the tmo/sprint merger that they did more than build a single tower in Colorado. I don't think I've ever seen the FCC seriously enforce the buildout requirements since any license holder can say 'but its hard we need more time/money'


I think of the 20B+ the shambling humonculus AT&T has pocketed for fiber deployment and think the process is captured.


> if you don't meet buildout deadlines and cover x% of pops within a specific time period, you forfeit the license

The issue you have to surmount is this reduces the value of the licenses in the short run. Which means less cash for the seller (the public) now versus a recurring productive asset.

The useless response is to decry hyperbolic discounting. A productive response would think through how to design the auction such that the public would prefer to have the productive, recurring stream of revenue versus some shiny thing today.


> The issue you have to surmount is this reduces the value of the licenses in the short run. Which means less cash for the seller (the public) now versus a recurring productive asset.

Well, that assumes the public isn't really benefiting from the products and services that can actually take advantage of that spectrum. Making less in license fees is probably a good trade-off if your phone is faster or you get interesting and affordable satellite services.


Dish I think is sitting on a bunch with endless extensions


70k homes is about 84MW.

You’re gonna have a tough time finding any evidence that running a landfill requires 84MW of electricity.


I think their point was the delivery of that garbage over time is subject to entropy, and from first principles probably took more energy consumption than a sustained 84MW over the time period the landfill is a viable source for energy.

I know nothing about landfill engineering here, to be frank, simply being a grease for good online gearing.


A single truck requires more energy to operate in a year.


A single truck requires more energy to operate in a year than 70k homes do!? I find this extremely difficult to believe.

As far as I can tell, the EIA [1] suggests the average home uses 10,791 kWh a year. A gallon of gasoline contains ~33.7 kWh of energy per the EPA/Wikipedia [2].

This would mean that a single truck would be burning 70,000 * 10,791 / 33.7 = 22,414,540 gallons of gasoline a year or 61,409 gallons a day. Seems like wild bullshit to me.

[1] https://www.eia.gov/tools/faqs/faq.php?id=97&t=3 [2] https://en.wikipedia.org/wiki/Gasoline_gallon_equivalent#:~:....


This has got to be one of the greatest retorts I have ever read on HN. Hat tip to you.


Most trucks -- at least of the kind used for transporting trash to a landfill -- don't burn any gasoline at all. They run on diesel fuel.

(But yeah, the original claim still seems orders-of-magnitude off; clearly BS.)


you should note that a gas engine does not convert all that 33kWh of energy into mechanical energy. a gasoline engine has about a 25% conversion into mechanical energy. https://www.fueleconomy.gov/feg/atv.shtml . diesel might be a bit better than a car but it's city driving by nature.

just heat alone is the largest waste product in a car or truck


That's not relevant to the comparison in any way.


fair enough. misread your comment


A tank of gas is still a tank of gas, regardless of what it gets used for.

Energy in = energy out + waste + energy stored

The truck is barely storing anything on average, so what you've described is energy out and waste, but the calculations to compare the truck to the landfill was done on Energy in - the amount of gas that it needs to be filled with.

For the same total job, you could raise or lower how quickly the truck goes through a tank of gas, but that variance has already been averaged out


Only Soviets kept using gas engined trucks and buses past the 50s


And, for some insane reason, the average American commuter.


According to [1] an electric garbage truck traveling 15,000 miles a year uses about 38,960 kWh. An 84 MW power plant produces 84,000 kWh every hour, or enough to power more than two trucks for an entire year. Even if we assume that the diesel equivalent uses a hundred times as much it's still a tiny fraction of what the plant in TFA produces.

[1] https://www.oregon.gov/deq/ghgp/Documents/ElectricGarbageTru...


Couple issues with that comparison: 15k seemed low given I drive ~10k a year and I don't work a job that uses my car, so I checked refuse trucks drive on average more like 25k miles per year and there are many servicing a single dump. Also most garbage trucks are still diesel so you've got to 5-10x that power usage number and there's all the vehicles used to compact and move the trash once it reaches the landfill which are also (currently) pretty exclusively diesel powered (think bulldozers and soil compactors with some excavators thrown in).

https://afdc.energy.gov/data/10309


So, you use more of that "enough energy to power them for an ear" that your power plant is outputting every hour.

There are plenty of hours in an year. You won't get any meaningful problem by complaining about the OP's approximations.


The point is this is free energy - all of the energy sent to the trucks is going to be done either way.


A 500 hp semi truck engine running at peak power is like 350 kw, so 84 megawatts (84,000 kw) is more than 200 of those engines at full throttle at all times.


Are you confusing MW (power) with MWh (energy)? There's no way that a truck uses more energy, unless it's running 24/7 or something.


84 Megawatts = 112645.86 Horsepower


Fast and Furious 12: Too Trash Too Furious


I find this obsession with distinguishing between open weight and open source foolish and counterproductive.

The model architecture and infrastructure are open source. That is what matters.

The fact that you get really good weights that result from millions of dollars of GPU time on extremely expensive-to-procure proprietary datasets is amazing, but even that shouldn't be a requirement to call this open source. That is literally just an output of the open source model trained on non-open source inputs.

I find it absurd that if I create a model architecture, publish my source code, and slap an open source license on it, I can call that open source…but the moment I publish some weights that are the result of running the program on some proprietary dataset, all of a sudden I can’t call it open source anymore.


> I find it absurd that if I create a model architecture, publish my source code, and slap an open source license on it, I can call that open source…but the moment I publish some weights that are the result of running the program on some proprietary dataset, all of a sudden I can’t call it open source anymore.

Then you don't understand open source. You would be distributing something that could not be reproduced because its source was not provided. The source to the product would not be open. It's that simple. The same principle has always applied to images, audio and video. There's no reason for you to be granted a free pass just becauase it's a new medium.


Let me spell it out for you:

1. I publish the source code to a program that inputs a list of numbers and outputs the sum into a text file. License is open source. Result according to you: this is an open source program.

2. Now, using that open source program, I also publish an output result text file after feeding it a long list of input numbers generated from a proprietary dataset. I even decide to publish this result and give it an open source license. Result according to you: this is NO LONGER AN OPEN SOURCE PROGRAM!!?!!

How does that make any fucking sense?

You have the model and can use it any way you want. The model can be trained any way you want. You can plug in random weights and random text input if you want to. That is an open source model. The act of additionally publishing a bunch of weights that you can choose to use if you want to should not make that model closed source.


Hey, tone down the agression.

Programs aren't the only things that can be open-source. LLMs are made of more than just programs. Some of those things that are not programs are not open source. If any part of something is not open source, the whole of that thing is not open source, per definition. Therefor any LLM that has any parts that are not open source is not open source, even if some parts of it are open source.


That simply isn't true. LLM weights are an output of a model training process. They are also an input of a model inference process. Providing weights for people to use does not change the source code in any way, shape, or form.

While a model does require weights in order to function, there is nothing about the model that requires you to use any weights provided by anybody, regardless of how they were trained. The model is open source. You can train a Llama 3.1 model from scratch on your proprietary collection of alien tentacle erotica, and you can do so precisely because the model is open source. You can claim that the weights themselves are not open source, but that says absolutely nothing about the model being open source. The weights distributed are simply not required.

But more importantly, under your definition, there will never exist in any form a useful open source set of weights. Because almost all data is proprietary. Anybody can train on large quantities of proprietary data without permission using fair use protections, but no matter what you can't redistribute it without permission. Any weights derived from training a model on data that can be redistributed by a single entity would inherently be so tiny that it would be almost useless. You could create a model with a few billion parameters that could memorize it all verbatim.

Open weights can be useful, and they can be a huge boon to users that don't have the resources to train large models, but they aren't required for any meaningful definition of open source.


> But more importantly, under your definition, there will never exist in any form a useful open source set of weights. Because almost all data is proprietary. Anybody can train on large quantities of proprietary data without permission using fair use protections, but no matter what you can't redistribute it without permission. Any weights derived from training a model on data that can be redistributed by a single entity would inherently be so tiny that it would be almost useless. You could create a model with a few billion parameters that could memorize it all verbatim.

That may very well be so. We'll see what the future holds for us.


The training data to create an LLM are as much a part of the LLM as the design notes and IDE use to create traditional software are a part of those projects.

They’re not required for it to be open sourced.


I'm not sure how that relates to AI models. Freely distributing compiled binaries, but not the source code, means modification is extremely difficult. Effectively impossible without reverse-engineering expertise.

But I've definitely seen modifications to Llama3.1 floating around. Correctly me if I'm wrong, but open-weight models can be modified fairly easily.


I have a properietary algorithm that produced the following code:

print(“hello world”)

Are you going to tell me I cannot open source this hello world program because it was produced by an algorithm that I have not open sourced?


You’re right obviously. In time none of these complaints are going to matter. Open weight is pragmatic and useful and will be accepted by basically everyone.


Imagine an "Open Source" & "Open Weight" model that could give you a HTTP link to the source of any idea it knows directly from its public training database.


I've got plenty of complaints about SQL, but as one of the most useful programming languages to have ever been invented, I have to say that syntax complaints are one of lowest items I would have on my list of things to be prioritized for a change. Sure, the syntax could be better, but why do we care so much about it over the dozens of other problems with it?

How about we get a SQL successor with algebraic data types, true boolean logic (as opposed to SQL's ternary logic), or functional composition? Null values are the bane of any query writer's existence, and we should have a reasonable solution by now...we've already done it with other programming languages.


It sounds like you want SQL to be more like a "real programming language", but I feel like there's a real chasm. SQL itself seems to be oriented towards "non-programmers" which is why it has declarative English syntax. But so many systems that interface with SQL databases are software systems written by programmers.

Why are our programming languages, which have rich ADTs, Boolean logic, etc. serialising queries into an English syntax written for ad-hoc business intelligence tasks? Why not have a binary query interface to the database that provides a programmatic, rather than human-readable, API?

The first time I got a "too many bind variables" error I was flabbergasted. All I wanted to do was insert a ton of rows into a table. But the database expects me to construct an English sentence containing a placeholder for each value of each row?


I think vision models have a lot bigger ROI from fine tuning than language models. That being said, I do consider fine-tunes to be helpful in improving smaller models as long as the domain is limited in scope. In other words, fine tuning allows you to get similar performance in smaller models, but improved performance in larger models seems pretty elusive at the moment, albeit somewhat possible with very creatively engineered training datasets.

An example of this is DeepSeek-Coder, which can essentially be considered a fine-tune of a fine-tune of a Mixtral model. It performs very similarly to Claude 3.5 Sonnet, which is pretty damn impressive, but it does it at less than 1/10th the cost.

What I don't understand though is why anyone would even remotely consider fine tuning a GPT-4o model that they will never fully own, when they could spend the same resources on fine tuning a Llama3.1 model that they will own. And even if you absolutely don't care about ownership (???), why not do a fine tune of an Anthropic model which is already significantly better than GPT-4o. At this point, with the laggard performance of OpenAI and their shameless attempts at regulatory hostility to competitors, I can't imagine ever giving them any of my money, let alone owning my derivative work.


"An example of this is DeepSeek-Coder, which can essentially be considered a fine-tune of a fine-tune of a Mixtral model"

I've not heard that anywhere else. My impression from https://huggingface.co/deepseek-ai/deepseek-coder-33b-base was that DeepSeek-Coder was trained from scratch.


The current deepseek-coder version (v2) is actually a fine tune off of the deepseek v2 model, and was not trained from scratch.

I’m now getting conflicting information about the origin of the deepseek MOE framework, so I may be wrong about it starting with a Mixtral model.


I'd love to find a M2-only NAS in a very low profile form factor. I live in a small apartment and prefer small electronics that can hide in a cabinet, but it seems like all of the NAS enclosures that I've ever seen recommended are fucking huge.


> I'd love to find a M2-only NAS in a very low profile form factor.

QNAP has at least two "NASbooks":

https://www.qnap.com/en-us/product/tbs-464 (2.5 GbE, USB 3.2)

https://www.qnap.com/en-us/product/tbs-h574tx (2.5/10 GbE, Thunderbolt 4)


There ie the rockchip CM3588 NAS Kit but it is a non mainstream CPU which cost between 200-300€:

https://wiki.friendlyelec.com/wiki/index.php/CM3588_NAS_Kit

there is also the asustor nas: https://www.asustor.com/de/product?p_id=80

If you just need a single nvme SSD you could also buy a banana pi 3 Router:

https://wiki.banana-pi.org/Banana_Pi_BPI-R3

Edit: forgot links and product


I've heard the Flashtor 12 is great, and can take lots of M.2s, but they end up with relatively few channels so data transmission rates aren't awesome. It can be fine as most users will be connecting over the network anyway which will be the bottleneck.


ODROID H4 Plus Intel N97 SBC (with in-band ECC) has an option for 4xM.2 via PCIe bifurcation, https://www.hardkernel.com/blog-2/new-m-2-card-for-the-odroi...


There’s a lot of really cool options in the replies here, but I think I’m at my limit on the number of tech hobbies I can manage. Really just waiting for a consumer grade Synology SSD NAS.


Exactly. It would have been great if I was still young and studying. I have seemingly infinite of time to fiddle around with things. Once you are working and have a family, most would really prefer to having something that simply work out of the box.

Unfortunately Synology have been moving towards Enterprise segment and completely forgot consumer.


https://www.amazon.com/dp/B0C5HW6W9Y

This would give 3 M.2 slots and high speed ethernet.

https://www.amazon.com/dp/B07N1HC88M

Something like this can be used to upgrade a dual-3.5" SATA NAS into a 8x-M.2 NAS (assuming the cooling and everything work appropriately).

Just some ideas.


I have the Flashstor 12 Pro. As the name suggests it's 12 m.2 nvme bays and it's pretty darn small considering.

Pros:

- It does what it says on the tin and I have 12 nvme drives in a ZFS pool

- It has 10G ethernet built in.

- It's pretty small, especially compared to most HDD focused NAS systems.

- It's just an x86 PC and you can blast the factory OS away to install your OS of choice without issue.

Cons:

- Each drive only gets a 3.0 x1 connection. It's honestly not much of a problem though as 1 GB/s per drive is still a ton of bandwidth after 12 drives.

- The built in ethernet is RJ45 instead of SFP+. Not the end of the world, just less power efficient.

- No PCIe expansion slots for anything but m.2 2280 drives.

- Single RAM slot so no dual channel and no large amounts of RAM. 32 GB works fine (I think Intel says the processor is only rated for 16). I can't remember if I tried 48 and it failed or if I never bothered. Either way 32 GB can be a bit small if you're really wanting to load it up with 4 TB or 8 TB drives and features like ZFS deduplication.

- The single built in fan could have been silent had they made any reasonable design choice around it. Instead I had to externally mount a noctua fan (to the same screw holes, just the inside is not a standard mount) and feed it power via a USB adapter. Works damn silent and cool now though.

- CPU (4 core Intel Celeron N5105) is very weak and the actual performance limitation for most any setup with this box.

I don't regret getting it, it's a solid choice given the relative lack of premade options in this segment, but the follow up NAS build was just me getting a motherboard/CPU with lots of PCIe lanes and loading up 4 way switches. You can do that via buying used Epyc servers on Ebay (loud/chunky) or just building a low end consumer class "workstation" (things like x8 x8 from the CPU instead of x16 for the assumed GPU) and PCIe to x4 switches (not the splitter cards which assume the motherboard has bifurcation and lanes available but actual switches). If you go the Epyc route you don't have to get switches and you can go back to cheaper splitters. I went the latter via some PCIe switch cards off Aliexpress. Performance and scaling of this one was better of course, but so was cost. Since I did all of this prices for m.2 drives have actually went up quite a bit so I'm glad I did it when I did.

What I would not recommend is anything non-x86 (older Ampere servers have lanes at not sky high prices but better to just go used Epyc at that point and get more CPU perf for the same dollar. SBCs are... a poor choice for a NAS on most every account but hacking factor). I'd also not recommend the Flashstor 6 as it only has 2.5 GBe connectivity and at that point what's the value in paying extra to do this all in flash.


> - No PCIe expansion slots for anything but m.2 2280 drives.

You can get m.2 adapters: https://www.delock.com/produkt/64219/dokumente.html?d=467


In general yes but in the Flashtor you're going to be hard pressed to do much with that unless you wire up a separately powered external chassis which defeats the point of getting the Flashtor.


Yeah, I said that before I looked at just how small it was.


I believe there is also a gen-2 just around the corner.


It's not a NAS, but you could hook it up to a low-power device and share it. I've got one connected to my primary machine and love it. Super fast storage.

Side note: One of the m2 chips that I got from Amazon was an obvious fake that I might not have caught if I hadn't ordered four together and one obviously didn't match.

https://eshop.macsales.com/shop/express-4m2


There is good reason for size of such NAS devices. HDDs are just better to maintain your not corrupted data when not connected, as NANDs needs to be refreshed from time to time. The best you could do is to use SSD as a fast cache.


Even if it was 5 drive 2.5" to hold SATA SSDs would be great for me. As networked mass storage I don't always need NVME speeds.


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

Search: