While your remark is valid, there's two small inaccuracies here:
> GPT-OSS-20B has 3.6B active parameters, so it should perform similarly to a 3-4B dense model, while requiring enough VRAM to fit the whole 20B model.
First, the token generation speed is going to be comparable, but not the prefil speed (context processing is going to be much slower on a big MoE than on a small dense model).
Second, without speculative decoding, it is correct to say that a small dense model and a bigger MoE with the same amount of active parameters are going to be roughly as fast. But if you use a small dense model you will see token generation performance improvements with speculative decoding (up to x3 the speed), whereas you probably won't gain much from speculative decoding on a MoE model (because two consecutive tokens won't trigger the same “experts”, so you'd need to load more weight to the compute units, using more bandwidth).
You realize cost of gas has direct consequences to 17M people's health as well? Our oma in her G-class building set her thermostat to 16 degrees in 2022. Because her heating bill shot to 1000+ eur/month. Only when the black mold started appearing did we manage to persuade her that 19 would be more appropriate. Of course that just traded it for money-related stress.
And I didn't say kick everyone on the street while the reconstruction is taking place. Everyone can stay where they are. Earthquakes are rare and so far in 50 years of extraction there have been no injuries. Groningen isn't the only place with earthquakes in the world you know?
> You realize cost of gas has direct consequences to 17M people's health as well? Our oma in her G-class building set her thermostat to 16 degrees in 2022
16°C in itself doesn't have health consequences whatsoever.
> Only when the black mold started appearing did we manage to persuade her that 19 would be more appropriate.
And you made the wrong diagnostic: mold is a moisture problem, not a heating problem per se. Sure heating improve air moisture but it's a very inefficient way to do so. You're complaining about the cost of a problem when you're using the most inefficient possible method to address it.
And again, if world market gas price rise, the consumer cost of gas rise as well, no matter if you have gas production in the country or not.
Bullshit they are, houses are entirely replaceable and in fact many people do so every couple of years.
Some jurisdictions even have “tenants rights” laws that literally force landlords to terminate all contracts whenever a tenant is about to have lived in a location for too long.
There's no “two body problem” here, the solar system is an n-body problem.
And the “three-body problem” is overblown in pop culture: even n-body problems are fairly predictable in the short-ish term, it's just that you cannot predict things over a long period because measurement imprecisions have a snowball effect, but it's not particularly unintuitive (I'd say it's more intuitive than the idea that we could predict things with perfect precision over billions of years).
Anonymity is important for many things. But on the flip side it's responsible of many issues with the internet today, because it makes moderation pretty much impossible (anyone can always just create a new account).
What we're missing is a way to have cryptographically secure pseudonymity: you log in to a website, you don't give any information whatsoever, but you cannot make two different accounts.
Most likely because your second sentence is impossible in one way or another.
Even if it's some kind of government encoded key, governments cannot be trusted to create imaginary people and hand them out to companies like palantir for large scale population manipulation.
I can imagine a government creating a moderate number of fake profiles for use by police and intelligence services, and honestly I'm fine with it, but creating a ghost population for propaganda purpose is entirely different and if you live in a country where you cannot trust your government not to do something that bad, you're already screwd.
In any case, it is still better than the status quo where even foreign authoritarian states can do that in countries where the local government wouldn't.
Do you propose to only let people from a whitelist of countries use internet? Because many countries would have no qualms giving their troll farms bunch of fake electronic ids.
It wouldn't be the internet as a whole, and instead be done at the individual service level (potentially with big plateforms being regulated in what they accept)
And indeed, it is to be expected that some countries be banned from most of the internet, or at least get a read-only version of it, because their digital credentials aren't deemed trustworthy enough. Not unlike how the travel visa system works nowadays.
1. The post mainly reiterates a single idea (Capsicum enumerates what the process can do, seccomp provides a configurable filter) in many different ways. There is not much actual depth, code samples notwithstanding. Nothing on why different designs were chosen, how easy each is to use, outcomes besides the Chrome example, etc.
2. There are a lot of AI writing tells, like staccato sentences, parallelism ("Same browser. Same threat model. Same problem."), pointless summary tables, "it's not X, it's Y" contradiction ("This is not a bug. It is the original Unix security model"), etc.
3. The author has roughly a blog post a day, all with similar style and on widely varied topics, and in the same writing style. Unless the author has deep expertise on a remarkably wide range of topics and spends all their time writing, these can't reflect deep insight or experience, but minimal editing of AI output.
It's pretty obvious. Lots of LLM signs. Short sentences that keep repeating the same idea. It's not x, it's this. In fact, the entire blog seems to be LLM-generated.
I've seen AI written blog posts before, but this is one step above: the entire blog (~90 articles) have been AI generated over the past three months.
I already find it very frustrating that most open source projects spawning on HN's front page are resume-boosting AI slop but if blogs start being the same the internet is definitely dead.
Edit: it doesn't even looks like it's resume-boosting in this case, the “person” behind it doesn't even appear to exist. We can only speculate about the intent behind this.
The person https://www.linkedin.com/in/vvoss/ seems to exist, I even have a mutual linkedin connection. What makes you think the “person” behind it doesn't even appear to exist?
I can't log-in to linkedin right now, but here's a few things:
- the profile picture is almost certainly (like 99%, certainty) AI-generated (I can even tell you it's ChatGPT-generated, the style is way too characteristic to miss).
- the LinkedIn profile shows prolific activity for the past few days, but almost nothing before that, I'm not sure the profile existed before.
- the github account is just 2 weeks old.
Having a mutual connection doesn't mean much, the interesting question would be who's the mutual and for how long has he be a connection. It's not hard to get to 500 LinkedIn connections on LinkedIn in a few days, you just need to add headhunters and other hiring specialists, they'll never refuse an invitation from a profile that look interesting. They could also have added someone who interacted with their LinkedIn slop submission, making the person more likely to accept the invitation.
It is getting more difficult to research now. Increasingly I just grab the source code locally and don't bother with the browser. Every search returns pages of wordy AI generated docs. At best they restate the code. At worse they read like badly written brochures. I am avoiding any project that doesn't have a long history. Large, feature packed projects that appeared out of nowhere on github with a single commit with no history or users are essentially stolen code that has been machine translated to obscure the original authors works.
I hate becoming the old person shaking their fist at the sky but the AI bros have just gone too far. I don't know why there isn't a bigger political and social movement against them. I would sign up in an instant to see their companies and practices regulated out of existence.
> If the workers had been expected to do this for normal wages, this wouldn't have happened.
I know some people really believe that people are only motivated by monetary incentives, but this isn't the reality of mankind. People do make sacrifice without monetary compensation all the time. (And many, many, did during covid)
Unlike what microeconomics-obsessed people think, workers don't make sophistivated economic calculations, instead they mostly care about being treated fairly.
And I glad people aren't like how microeconomics model them, because the world simply wouldn't work otherwise.
True. The other side of 'fair' in this situation is what was the company earning.
Busting a gut to make some shareholders/managers loads of money isn't exactly fair if you aren't also being rewarded.
On the other hand, in the UK the NHS has traditionally paid people extra on the understanding that these things happen. But when it did happen the staff were asking for even more.
I know this probably comes across as right wing, but my point is how we reward people for black swan events.
> Busting a gut to make some shareholders/managers loads of money isn't exactly fair if you aren't also being rewarded.
Exactly. But it's not a matter of “being paid more than the marginal value of the additional work” as microeconomics tend to frame those things.
Symmetrically, people routinely accept pay cut or degraded work conditions when the company isn't going well, even though it makes no sense from a game theory perspective (it's basically a prisoners' dilemma yet people cooperate most of the time).
This is vibecoded garbage that the “author” probably didn't even test by themselves since making this yesterday, so it's not surprising that it's broken.
Also, as I said in a top level comment, what this project wants to achieve has been done for a while and it's called Heretic: https://github.com/p-e-w/heretic
We will eventually arrive at a new equilibrium involving everyone except the most stupid and credulous applying a lot more skepticism to public claims than we did before.
And yeah, doing stuff like deleting layers or nulling out whole expert heads has a certain ice pick through the eye socket quality.
That said, some kind of automated model brain surgery will likely be viable one day.
Just want to add to this that with custom calibration data it's incredibly effective and surgical, you can get VERY LOW KL divergence this way. Many MoEs are supported too, it's actively maintained.
> GPT-OSS-20B has 3.6B active parameters, so it should perform similarly to a 3-4B dense model, while requiring enough VRAM to fit the whole 20B model.
First, the token generation speed is going to be comparable, but not the prefil speed (context processing is going to be much slower on a big MoE than on a small dense model).
Second, without speculative decoding, it is correct to say that a small dense model and a bigger MoE with the same amount of active parameters are going to be roughly as fast. But if you use a small dense model you will see token generation performance improvements with speculative decoding (up to x3 the speed), whereas you probably won't gain much from speculative decoding on a MoE model (because two consecutive tokens won't trigger the same “experts”, so you'd need to load more weight to the compute units, using more bandwidth).
reply