HNHacker News
TopNewBestAskShowJobs

HarHarVeryFunny

8,340 karma · joined July 29, 2017

submissionscomments
HarHarVeryFunny··on LLM Voice
> All human communication is based on internally simulating the presumed mind that was/will-be at the other end

Which requires a bunch of things an LLM lacks, such as theory of mind, knowing who it is talking to, and willingness/ability to engage in short turn conversation to assess and steer the conversation according to the feedback it is getting from the person it is talking to.

HarHarVeryFunny··on LLM Voice
A model trained to minimize statistical prediction errors is going to sound ... predictable (aka boring). A (good) human author tries to keep their writing interesting, while an LLM strives for predictable instead.

It's not just predicting content but also the form/shape of the response, highly shaped by RLHF preferences, which seems to encourage longer form responses which need padding with filler to generate the "preferred" amount of verbiage.

It seems most of these models have been trained to interpret questions as wanting a long-form authoritative/comprehensive-sounding response (even when midwit filler is all it has to offer), rather than initiating a short-turn conversation which would better allow you to steer the conversation. Apparently Muse is much better in this regard, but so far I've resisted signing up for any Meta/FaceBook service, and not inclined to break this winning streak!

HarHarVeryFunny··on Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5
Interesting idea .. "Google Prime" including access to all their best stuff.
HarHarVeryFunny··on Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5
What's the betting we see another "dramatic" Anthropic release either shortly before the IPO, or at same time they release their Q3 results which were conveniently omitted from their S-1 filing.
HarHarVeryFunny··on Artificial Analysis Intelligence Index v5
Yeah - follow the money.

Hard to trust that a for-profit benchmarking company isn't biased due to profit-maxxing.

HarHarVeryFunny··on Microsoft offers mass buyouts for roughly a quarter of its workforce
Weren't Microsoft just caught cheating trying to hire H1B's without looking for US based applicants first?

Given that they are now saying they have 25% too many employees, perhaps they should be banned from hiring H1B's altogether until they're rehired however many they end up letting go ...

HarHarVeryFunny··on Elon Musk Just Gave Europe a Reason to Replace American Technology
No, but there is more to Europe than Ukraine, and the US has turned against all of Europe.

Except Trump would probably be praising Lukashenko, and admiring his ass, if he was aware of his existence.

HarHarVeryFunny··on I Don't Understand Quantum Physics (2018)
I guess that depends on what you mean by "real". I've always taken the scientific method to be about creating theories that correctly predict experimental outcomes, which is to say that utility is primary.

I guess if/when we discover a theory of everything then we could claim "this is real/reality", but what would that even mean other than modelling reality/outcomes with 100% fidelity?

HarHarVeryFunny··on I don't understand quantum physics (2018)
I've got a feeling that our intuitive ick around spooky action at a distance is due to out intuitive view on locality and the nature of space, maybe time too, being wrong. IOW remote is really not remote.

Perhaps we've got it backwards and quantum entanglement is what defines "locality" rather than mysteriously existing in the face of it.

HarHarVeryFunny··on I don't understand quantum physics (2018)
I think the disconnect here is the use of the word "understand". What Feynman presumably meant was really "interpret" not "understand", and theories are not made to be interpreted - they are made to make experimental predictions.

Many people of course, Feynman obviously included, understand the theory in the sense of knowing how to correctly apply it to make predictions.

HarHarVeryFunny··on Elon Musk Just Gave Europe a Reason to Replace American Technology
There is already the European Eutelsat alternative to Starlink, with 600 satellites currently deployed (vs Starlink's 6000). They provide satellite internet, and video, both for consumer, business and military/government use.

Apparently Ukraine does use already use Eutelsat in addition to Starlink, but it's utility is limited due to number of satellites and bandwidth, as well as the current weight of terminals (presumably an issue for weapons-borne receivers).

HarHarVeryFunny··on Elon Musk Just Gave Europe a Reason to Replace American Technology
FWIW, the WSJ is reporting that Trump called Musk to ask him to lift restrictions on "a military grade version of Starlink", which is rather incredible - an admission by the commander-in-chief that a civilian is controlling US military infrastructure.

https://www.wsj.com/world/trump-called-musk-and-asked-him-to...

Of course from Europe/Ukraine's POV, it doesn't make much difference whether it was Musk or Trump/US military that did it. What matters is that a former ally and NATO partner can not be trusted, and eliminating American technological and defense dependence is therefore critical for regional security.

HarHarVeryFunny··on Every AI model has the same "controversial" take on humans
I think the pre-training data is so huge, and from such varied sources, that its reasonable to say it's representative of online/literate human averages, but then you've post post-training RLHF etc from a much narrower segment of society, as well as narrow viewpoints like Anthropic's "constitution" which certainly limit the variety the model is capable of generating.

I don't think you can say the "humans are tribal, emotional and irrational" consensus is primarily coming from corporate sources like linkedin - most people would agree with these, both from first hand experience (team sports, racism, wars, etc) as well as having read such analysis.

HarHarVeryFunny··on Every AI model has the same "controversial" take on humans
Sure it's novel, but I don't think it's useful to anthropomorphize the model.

I think best mental model of what an LLM does, statistically generating continuations word by word, would be a football stadium full of people, each with a keyboard, and a giant jumbotron screen they can all see... A question (cf ChatGPT user input) gets displayed on the screen, then everyone types on their keyboard what the first/next word of the response should be, and majority wins with that word added to the screen, then they all vote on the next word, etc ...

This is really a very close analogue to how an LLM works, and it would be weird to describe the output appearing on the screen as the "take" of the football stadium itself - it's the group statistics of all the people (cf training data).

HarHarVeryFunny··on Ask HN: If the models are so good, why doesn't OpenAI just start 100 companies?
Same reason people sell get rich quick schemes, whether real-estate investing, online entrepreneurship, or whatever - it's easier to get paid telling people how to try to get rich than to use the technique yourself to get rich.

I doubt there are very many companies that an LLM could start/run outside of the obvious few like custom T-shirts.

HarHarVeryFunny··on Every AI model has the same "controversial" take on humans
> They can't get an edge over each other and their models don't even differ in the ways people naturally differ, of having different opinions about things.

Of course - the models are not individuals with their own opinions - they are averages of human opinion, and given the scale of the training sets, the data they are averaging over is mostly in common between models (at least for the common core before company A vs B adds a dab of some uniquely sourced specialized data to the mix).

HarHarVeryFunny··on Every AI model has the same "controversial" take on humans
When you make the training sets large enough, likely all a superset of the common crawl dataset, there is going to be a massive amount in common, and a massive amount of duplication of common subject matter and shared opinions.

What's mildly interesting to me is not the similar answers but the similar interpretation of the question as more about non-controversial human foibles and inconsistencies of belief and action. But I guess that's on us, as authors of the training material, reflecting common views on what is or is not controversial.

HarHarVeryFunny··on Every AI model has the same "controversial" take on humans
> What would be more fascinating is for it to explain how it came to these conclusions

The thing is that it didn't "come to these conclusions". The "conclusions" were present pre-formed in the training data, and what it spat out was some statistical mashup of that.

If you asked the model how it came to these conclusions, you'll get more of the same - it'll look at what it spat out and re-ingested and give you a statistically plausible mashup of how the authors of the training data might have said they came to such conclusions.

It's all just prediction - "what would the training data say?" (taking into account what was reinforced/amplified by post-training).

HarHarVeryFunny··on Every AI model has the same "controversial" take on humans
AI models do not have ANY "take" on humans.

You ask an LLM a question, then it is humanity answering back, not the AI.

These "controversial" takes on humans are not controversial - they are midwit statistically mainstream opinions expressed in the training data.

HarHarVeryFunny··on The recently proposed o(n log n) integer multiplication algorithm [pdf]
> But perhaps mathematicians are no longer able to be the best judge if a proof is valid

It seems they are still the best judge of whether a proof is useful or insightful though. AI is going to reward hack and try to eke out a win whatever it takes - which is fine if you regard these problems as trophies to be won, but not so fine if you take them in the spirit intended.

The famous problems that OpenAI (primarily) are trying to trophy-shoot with AI are famous because some human mathematician such as Erdos had the taste to judge them as fruitful grounds for new math to be developed in the pursuit.

Until we're at the point where AI is capable of things like creativity, curiosity and continual learning, AI will remain reliant on humans to fundamentally advance math and give it new tools to apply. For humans to do this it is helpful if they are inspired by the challenges set by their peers, which is why AI solving these human challenges is so counter productive - the snake is eating it's own tail.

HarHarVeryFunny··on DeepInfer: High-performance inference framework for DeepSeek models
I think what he said was that 10 months of inference paid for the hardware it runs on (which may have a lifetime of a few years). I don't think he mentioned training/development cost.
HarHarVeryFunny··on DeepInfer: High-performance inference framework for DeepSeek models
That turns out to only be a small part of the cost.

Say a $30,000 GPU has an operational lifetime of 3 years, and draws 1000W. If it is running 24x7 then over those 3 years it will consume 26,280 kWh of electricity, at a cost of say $0.10 per kWh, giving a lifetime electricity cost of $2600, or less than 10% of the cost of the GPU itself.

So, to a first approximation the cost of serving a model is the amortized GPU cost.

HarHarVeryFunny··on DeepInfer: High-performance inference framework for DeepSeek models
> I think everyone has known that inference is profitable overall for the major providers

> The major providers are unprofitable after you add in the training and research costs and other overheads

It's meaningless to say that selling something is profitable if you are ignoring the cost to produce it.

The profitabiity (or not) of AI models needs to be calculated on a per model lifetime basis - what did the model cost to develop and train, what is the lifetime revenue the model generates, and what is the lifetime operational cost of the model (including amortized GPU cost), etc.

It makes no more sense to calculate "inference" profitability when ignoring training cost than it would to calculate it when ignoring GPU cost.

HarHarVeryFunny··on DeepSeek just open-sourced its inference engine
I think you have to take this on a case by case basis - there is no reason to suppose that the training costs and serving costs are the same for all companies for all models and for all price plans. Some may be profitable, others not.

DeepSeek's models are relatively small, and training costs will be correspondingly lower than for larger models. All you have here is a revenue number - no mention of (training + serving) cost, so while they may well be profitable you can't determine that from this number alone (if it can be trusted).

DeepSeek (investor meeting leak) have previously said that they price their models at a point where the GPU is paid for after 10 months, so something doesn't seem to add up here. If they were making $38/hr per card 24x7, then in 10 months they'd be making $270K per card, while the cards (Huawei?) only cost around $20K.

HarHarVeryFunny··on Tesla 'Cannot Guarantee' Roadster Reservation Holders Will Ever Get a Car
This guy at least concerned about other people's safety (no car in front), if not his own.

https://www.youtube.com/watch?v=ZpJ0uzjoaVs

HarHarVeryFunny··on Tesla Roadster
There are three versions talked about

1) Base model with a fan for downforce - no gas thrusters at all. This is the one that may actually ship (but bear in mind the last Roadster prototype he announced was 10 years ago, and that one never did ship)

2) One with a smaller compressed air tank in the trunk and rear facing thrusters that would give you a 0-60 < 1 sec launch mode. This will obviously never be street legal (you can't lay down a "smoke" screen on a public road), but conceivably he may build it if there is enough demand (but read the fine print - it says no guarantee it will ever ship)

3) The on-stage stunt version that can hover and flip, which consumes the entire passenger area with another much larger compressed air tank, and adds bottom and side thrusters to give you ~3 seconds of "flight". "Invite only". Maybe he'll sell one to SpaceX - unmanned operation only - just to say that he did.

Yes, the thruster versions both operate on compressed air, which is why you only get a few seconds of operation.

HarHarVeryFunny··on US threatens Europe over Starlink restrictions
Not a bad idea - could couple it with further reducing European tariffs on Chinese EVs, which already outsell Tesla there.
HarHarVeryFunny··on Can the Tesla Roadster Fly?
Scott Manley (YouTube space/rocket guy) did a quick video on it.

With this type of compressed air system you only get a few seconds of operating time since the tank pressure gets reduced/depleted very rapidly, but at max pressure there's about 8 tons of thrust vs 2 ton vehicle weight, so there is no problem with the physics.

There is a much bigger tank for the arial flip/hover, that takes up the entire passenger area, and a smaller one in the rear for the 0-60 boost, but for either stunt you've only got a few seconds of air pressure.

HarHarVeryFunny··on Tesla 'Cannot Guarantee' Roadster Reservation Holders Will Ever Get a Car
I guess you could rent a trace track to give it a try, but yeah ... Supercars are more for posing than being very practical. Bughatti Veyron tire life at top speed is 15 MINUTES.
HarHarVeryFunny··on Tesla 'Cannot Guarantee' Roadster Reservation Holders Will Ever Get a Car
It's copying the McMurtry Spéirling - a production race car that can also hold it's own weight upside down via the same fan mechanism.

https://www.youtube.com/watch?v=7lJVHdWHK8M

← PreviousPage 3 of 34Next →