HNHacker News
TopNewBestAskShowJobs

dust42

745 karma · joined February 7, 2025

submissionscomments
dust42··on 2025 was the third hottest year on record
30 years ago I attended a university lecture in an economics class and the professor spoke about the economic consequences of global warming - some places will be better off and plenty of places will be worse off. There will be water shortages in some places, while heavy rainfall in others. He presented it as a given fact that the global warming is coming - and pretty much the whole audience was shocked. Finally someone asked if he really thinks that it is unavoidable. And his answer was yes, that is human nature. As long as fossil fuels are there and cheap to explore someone will use them.

30 years later it looks like he was right.

Edit: the IPCC was founded in 1988 thus people started in the 70ies to understand that there will be a problem but there was a very long period of inactivity. Personally I am quite optimistic that fusion will become commercially available before 2040.

And dear downvoters, dont shoot the messenger.

dust42··on Song banned from Swedish charts for being AI creation
They are just fully AI driven on every level but also there are people listening to it. Youtube currently gets swamped with AI generated content - music is only one part of it. For example there are now endless history documentaries.

To get these of the ground there are lots of fake comments and fake views but after a while these videos gain traction and then the algos pick them up for organic views.

Search youtube for "female vocal blues" or "female country songs" and it is all AI and it is really good - good in a sense that you don't realise it immediately. But they garner millions of views. They are not McDonald's but fine dining cooked with convenience products.

I am quite split about algorithmically generated music but I have to admit that I have fallen once into the trap. And only when I searched for the artist I figured out it is AI. Though once you know it you immediately hear it.

Edit: I went to one of the websites offering this as a service and in 5 minutes it creates a very decent song including lyrics. I forgot which but remember it was something like $20 for 1000 songs. Not a surprise that youtube gets swamped with it - it costs next to nothing to produce, neither time nor money.

dust42··on Pocket TTS: A high quality TTS that gives your CPU a voice
Yes, apart from voice cloning nothing really new. Kokoro is out since a long time and it supports at least a few languages other than english. Also there is supertonic TTS and there is Soprano TTS. The latter is developed by a single guy while Kyutai is funded with 150M€.

  https://github.com/supertone-inc/supertonic 
  https://github.com/ekwek1/soprano
No affiliation with either.
dust42··on Pocket TTS: A high quality TTS that gives your CPU a voice
Good quality but unfortunately it is single language English only.
dust42··on Scaling long-running autonomous coding
> This is going to sound sarcastic, but I mean this fully: why haven't they merged that PR.

I would go even further, why have they not created at least one less complex project that is working and ready to be checked out? To me it sounds like having a carrot dangle in front of the face of VC investors: 'Look, we are almost there to replace legions of software developers! Imagine the market size and potential cost reductions for companies.'

LLMs are definitely an exciting new tool and they are going to change a lot. But are they worth $B for everything being stamped 'AI'? The future will tell. Looking back the dotcom boom hype felt exactly the same.

The difference with the dotcom boom is that at the time there was a lot more optimism to build a better future. The AI gold rush seems to be focused on getting giga-rich while fscking the bigger part of humanity.

dust42··on Claude Cowork exfiltrates files
I don't get all this frontier stuff. Up to today the best model for coding was DeepSeek-V3-0324. The newer models are getting worse and worse trying to cater for an ever larger audience. Already the absolute suckage of emoticons sprinkled all over the code in order to please lm-arena users. Honestly, who spends his time on lm-arena? And yet it spoils it for everybody. It is a disease.

Same goes for all these overly verbose answers. They are clogging my context window now with irrelevant crap. And being used to a model is often more important for productivity than SOTA frontier mega giga tera.

I have yet to see any frontier model that is proficient in anything but js and react. And often I get better results with a local 30B model running on llama.cpp. And the reason for that is that I can edit the answers of the model too. I can simply kick out all the extra crap of the context and keep it focused. Impossible with SOTA and frontier.

dust42··on vLLM large scale serving: DeepSeek 2.2k tok/s/h200 with wide-ep
I used 700W per H200 = 11.2 per 16 GPUs. I didn't include CPU and rest of the rack. So yours is a better approximation.

One has to keep in mind that the benchmark that was done is synthetic. This makes sense because it makes it reproducible but real world usage may differ - i.e. by the amount of context and the number of concurrent users. Also there are use cases where smaller models or smaller quants will do.

The key take away for me for this type of back of the envelope calculation is to get a good idea where we stand long term, i.e. when VC money stops subsidizing.

So for me $0.3 per 1M tokens for a decent model looks pretty good too. Seeing that OpenAI API charges $21 per 1M tokens input and $168 output for GPT-5.2 pro I was wondering what the real sustainable pricing is.

dust42··on vLLM large scale serving: DeepSeek 2.2k tok/s/h200 with wide-ep
If I followed the links correctly this benchmark was made on a 16xH200. At current prices I'd assume that is a system price of around $750,000.

The year has 86400*365 = 31536000 seconds. Thus 63072000000 tokens can be generated. As pricing is usually given per 1M tokens generated, this is 63072 such packages.

Now lets write off the investment over 3 years, 250,000/63072 = 3.96. So almost $4 per 1M tokens generated with prompt processing included.

Model was a Deepseek 671B 32B MoE.

Looks to me that $20 for a month of coding is not very sustainable - let's enjoy the party while VCs are financing it! And keep an eye on your consumption...

Electricity costs seem negligable with ~$10,000 per year at 10cts per kWh but overall cost would be ~10% higher if electricity is more like 30cts like it is in Europe.

Edit: like it is pointed out by other commenters it is 2200t/s per single GPU thus the result needs to be divided by 16: $4/16 = $0.25. This actually somewhat matches the deepseek API pricing.

dust42··on The next two years of software engineering
From my experience AI is only good with mainstream coding tasks. Javascript, node, react, crud. Whatever it has seen in overabundance, it is good with. Already with typescript it is less strong than with javascript. It is just a clever token generator without intelligence. Often it resembles intelligence but so can a human by learning quotes of famous people.

It is a new and exciting tool but immediately limited with medium complex tasks. Also we will see a lot more code with tricky bugs coming out of AI assistants and all of that needs to be maintained. If software development gets cheaper per line of code then there will be more demand. And someone has to clean up the mess created by people who have no clue whatsoever of SWE.

Once upon a time people developed software with punch hole cards. Even without AI a developer today is orders of magnitude more proficient than that.

The only thing I hope I am not going to see in my lifetime is real artificial intelligence.

dust42··on Side-by-side comparison of how AI models answer moral dilemmas
Distilling from a closed model like GPT-4 via API would be architecturally crippled.

You’re restricted to output logits only, with no access to attention patterns, intermediate activations, or layer-wise representations which are needed for proper knowledge transfer.

Without alignment of Q/K/V matrices or hidden state spaces the student model cannot learn the teacher model's reasoning inductive biases - only its surface behavior which will likely amplify hallucinations.

In contrast, open-weight teachers enable multi-level distillation: KL on logits + MSE on hidden states + attention matching.

Does that answer your question?

dust42··on Side-by-side comparison of how AI models answer moral dilemmas
That's right, they all give a variant of that, for example Qwen says: I am Qwen, a large-scale language model developed by Alibaba Cloud's Tongyi Lab.

Now given that Deepseek, Qwen and Kimi are open source models while GPT-5 is not, it is more than likely the opposite - OpenAI definitely will have a look into their models. But the other way around is not possible due to the closed nature of GPT-5.

dust42··on Google AI Studio is now sponsoring Tailwind CSS
So if you subtract linux and LLVM and Webkit and Java, what is left of Google? Absolutely nothing. Well, a mostly empty, dysfunctional mono repo lacking the main dependencies.
dust42··on Google AI Studio is now sponsoring Tailwind CSS
> Google has poured untold millions into open source over the last couple of decades, not just by sponsorships, but also by employing contributors, etc.

And Google has profited untold hundreds of billions of open source over the last couple of decades. They just need to be reminded of it.

Edit: Haha, getting downvoted! Never underestimate the power of tens of thousands of Googlers on HN... Look, I use Gmail, Google maps, Chrome and Android and occasionally Google search but without Linux, Java and webkit it wouldn't exist.

dust42··on Chinese AI models have lagged the US frontier by 7 months on average since 2023
Well, my answer was partially tongue in cheek. However, under the circumstances I mentioned I had better results with a local model. But then up to today I really like Deepseek for dev - for me and my use cases it still outperforms Opus. The problem is we don't have good and universal benchmarks, inststead we have benchmaxxing. For my use cases Chinese models may be perfect but maybe for yours American one's may outperform them by a huge margin? So I urge you to use YCI - Your Capability Index.
dust42··on Chinese AI models have lagged the US frontier by 7 months on average since 2023
They use the ECI - EpochAI Capability Index.

Measured by the DCI the Chinese AI models are about 1.5 years ahead of US models.

DCI = Dust42 Capability Index: MBP Max 64GB, Qwen3-80B MLX 4bit quant, 40 tokens per second. It is not on Claude Opus level but very, very useful if you have no internet, i.e. on a flight. And occasionally it surpasses even Opus by far and large. Opus is a pain in the neck once the coding task at hand surpasses its capabilities. Qwen3 is much better to guide to get step by step to a solution.

dust42··on Donut Lab’s all-solid-state battery delivers 400 Wh/kg of energy density
Actually you are right. The world first super-intelligence AI was also created there, just 7 months ago. Also by Marko. This man is from another world.

Watch yourself: https://www.youtube.com/watch?v=ilgJKjiDLV8

dust42··on Donut Lab’s all-solid-state battery delivers 400 Wh/kg of energy density
> Concerning range though, I think their estimate is pretty fair as far as the marketoid speech goes -- it can be 600 km

I don't know what 'real world range' means to you. A bike of this size uses roughly 9kWh/100km. The battery has 33kWh capacity you will usually look for a charger at 20% and charge to 80%. So 'real world range' will be around 200-250km.

The last version with a 20kWh battery made a record run from 100% to 0% with 350km. Which translates to 577km on this bigger battery. At least I don't consider the need of a car following me with a trailer a 'real world' scenario.

And another funny thing: if you watch the video, you actually can see that the battery starts to bulb during the charging starting from 67% on and then the video cuts out at 82%...

Btw, the CEO never speaks of cells, only of batteries. In the industry the battery is usually considered the structure plus the cells. He doesn't even speak about a factory or production. So if he buys white label cells in China he is technically not lying.

It is an absolutely fantastic marketing video and cleverly done: Verge says on their website only 33.3kWh capacity and 10min to 80%. We connect the dots in our brains which aren't there.

It is a decent motorbike used to make donut look legit and donut is used to create hype for the motorcycle, likely made in China with a bit of assembly in Europe to call it Finnish.

dust42··on Donut Lab’s all-solid-state battery delivers 400 Wh/kg of energy density
I wouldn't call it an outright scam. But yes, you pay yourself a nice salary, travel around the world to meet potential OEMs, you may fly in private jets, you stay in fancy hotels, you have business meetings in Michelin star restaurants.

The motorcycle is real, you can buy it.

I worked for a company in Germany in 2024 and the CEO was quite open about using public grant money as a free gift. The company wasn't build around it but it was a significant part of income. He made engineers sign papers that they had worked on some AI stuff which they didn't. A demonstrator was made with yolo by an intern. He said they have just no competency to figure that out and also that they actually don't want to figure it out.

Or look at the EuroLLM: sounds good on paper, never heard about it again. Grant grabbing is a real industry in Europe with companies specialised on creating grant applications and forming consortiums.

dust42··on Donut Lab’s all-solid-state battery delivers 400 Wh/kg of energy density
> you compare maximum hardware capabilities of the raw battery with a vehicle, set up for longevity. Look, there are reasons I don't charge my 18650 batteries with their maximum 5C current, but at a gentle, slow rate. Virtually all the electric cars have default maximum charge set to 80-90% - not because a car maker lied

But that is exactly what the CEO claims in the video [1]: "This pack charges in as low a 5min and is still build for longevity [Video shows time lapse of charging from 0%-100%]. You can fast charge every single day for years. [...] It is designed for 100,000 cycles"

Especially for motorcycles that definitely would be a game changer: 5 minute stop to top up to 100% while stretching your legs...

And the guy from Verge Motorcycles says: "A new long range version delivering up to 600km of real world range."

[1] https://www.youtube.com/watch?v=Y-aPS2AwMbc @ 10:54 the 600km statement @ 12:59 the talk about the charging

dust42··on ChatGPT Health
This is just marketing nonsense. You don't have to train models to not retain personal information. They simply have no memory. In order to have a chat with an LLM, every time the whole conversation history gets reprocessed - it is not just the last answer / question gets send to the LLM but all preceding back and forth.

But what they do is exfiltrate facts and emotions from your chats to create a profile of you and feed it back into future conversations to make it more engaging and give it a personal feeling. This is intentionally programmed.

dust42··on LMArena is a cancer on AI
Oh my goodness yes, I almost missed it that the text is (mostly?) AI written. That said I agree that LMArena elo scores are pushing models in the wrong direction. They move more towards McDonald's than quality food.
dust42··on Donut Lab’s all-solid-state battery delivers 400 Wh/kg of energy density
Here some info about the claim of 600km/344m range of Verge motor cycles as it was said and shown in the video. The website of Verge bikes states the following:

City estimate 600km: "The city estimate calculated by Verge aims to provide a reasonable approximation of calm stop-and-go riding in an urban environment at low speeds. The estimate assumes a 75 kg rider and ideal riding conditions in terms of temperature, wind and tire pressure. Actual range varies based on exact riding conditions, riding style and other factors."

Highway estimate 315km: "The highway estimate calculated by Verge aims to provide a reasonable approximation of steady and consistent highway riding at a speed of 90 km/h. The estimate assumes a 75 kg rider and ideal riding conditions in terms of temperature, wind and tire pressure."

Now: "aims to provide a reasonable approximation" is legalese for "we pulled these numbers out of thin air - they are in no way legally binding".

For proper testing according to "EU Regulation 134/2014, Annex VII" the website says 'TBC'. This would be legally binding but -surprise, surprise- they haven't yet found time to do it.

For any practical purpose the range will be between 200km to 250km. Also the bike is not easy to handle due to the high mass distribution to the outside of the wheel.

Btw, the vendor from Verge motorcycles states in the video 80% charge in "about 10 minutes" instead of 100% in 5min. So which one is true? I stand by my claim that this is marketing for gullible buyers and investors.

dust42··on Donut Lab’s all-solid-state battery delivers 400 Wh/kg of energy density
> It is nanoprinted and trivially scalable.

Nanoprinted snakeoil - infinitely scalable. Good enough to grab 3M€ public grants and some more from gullible private investors.

Honestly, if they had the tiniest proof of their claims (400Wh/kg, 5min 0-100%, operating temp -30°C to 100°C, no Li/Co/Mn and somebody looking at the production and taking the output to the test) they would be absolutely drowned in money to the point that sama would become jealous.

> Factory in Finland, started around Q3/2025. https://yle.fi/a/74-20180376

That is only a picture of a factory building. They are 100% greenwash grant grabbers with no real substance. There are plenty of these people in Europe. The motorcycle is likely in China designed and produced.

If grant grabbing is what you'd consider a fulfilling life, then come to Europe - it will be an El Dorado for you.

dust42··on Donut Lab’s all-solid-state battery delivers 400 Wh/kg of energy density
Sounds too good to be true. Seriously, if the CEO had any proof, he wouldn't need to hang out on CES. He would be showered with billions of $$$ and investors would fly in from all over the world. Instead he is selling motor cycles? Hard to believe. Why not have at least two different top universities having tested them?At least some insights into their production to verify the claims? And the motorcycle has a 33kWh battery for 600km? At what speed, 40kph? Good luck to them anyway...
dust42··on 3Duino helps you rapidly create interactive 3D-printed devices
Ouch. A soulless AI voice talking about Arduino - once a creative hacker's tool/toy: people with enthusiasm. Now this. Welcome to the brave new world.
dust42··on Is it a bubble?
I intentionally didn't say AI but LLM because for me the word 'intelligence' is misleading. But LLMs are definitely a leap forward in NLP and what other word for 'comprehension' would you use?
dust42··on Is it a bubble?
The question is, can SV extract several trillion dollars out of the global economy over the next few years with the help of LLMs and GPUs? And the follow-up question: will LLMs help grow the global economy by this amount - because if not, then extracting the money will lead to problems in other parts of the world. And last not least, will LLMs -given enough money to train them on ever bigger data sets- magically turn into AGI?

IMHO for now LLMs are just clever text generators with excellent natural language comprehension. Certainly a change of many paradigms in SWE. Is it also a $10T extra for the valley?

dust42··on Is it a bubble?
And so is "AI". Unless you add randomness AKA raise the temperature.
dust42··on 30 years ago today "Netscape and Sun announce JavaScript"
A big problem was certainly that Linux on commodity boxes became an industry standard. In 2000 it was still seen by many corporations as hobbyist amateur system. But then Google & Co introduced them into the corporate world and for many use cases a Linux box for 1000 bucks would do the same as a 10000 bucks Sun server.
dust42··on 30 years ago today "Netscape and Sun announce JavaScript"
Netscape and Sun plan to propose JavaScript to the W3 Consortium (W3C) and the Internet Engineering Task Force (IETF) as an open Internet scripting language standard. JavaScript will be an open, freely licensed proposed standard available to the entire Internet community. Existing Sun Java licensees will receive a license to JavaScript. In addition, Sun and Netscape intend to make a source code reference implementation of JavaScript available for royalty-free licensing, further encouraging its adoption as a standard in a wide variety of products.

The 90ies had quite a few pretty visionary people. CERN made the Web protocol and code available royalty free on 30 April 1993, enabling its widespread use.

At that time there were still CompuServe, AOL, Minitel and BTX around - not just walled gardens but walled worlds but a handful of people already saw and shaped the future...

← PreviousPage 4 of 6Next →