Perhaps "our PR team is a prompt" is what they mean to convey? Or "let's make this obviously AI so more people comment pointing that out" is their social media strategy?
I agree with this, but I think the spec author's public statements means we don't need to give them the benefit of the doubt. People have repeatedly pointed out how this will result in vendor lock-in, and their response is either "yep, working as intended" or "we don't want to talk about this anymore." They're just steamrolling ahead with support from all the Big Tech companies. It's a really ugly situation =/
Graphcast (the model this is based on) has been validated in weather models for a while[1]. It uses transformers, much like LLMs. Transformers are really impressive at modeling a variety of things and have become very common throughout a lot of ML models, there's no reason to besmirch these methods as "integrating an LLM into a weather model"
A lot of shiny new "AI" features being shipped are language models being placed where they don't belong. It's reasonable to be skeptical here, not just because of the AI label, but especially for the troubled history of neural-network based ML methods for weather prediction.
Even before LLMs got big, a lot of machine learning research being published were models which underperformed SOTA (which was the case for weather modeling for a long time!) or models which are far far larger than they need to be (e.g. this [1] Nature paper using 'deep learning' for aftershock prediction being bested by this [2] Nature paper using one neuron.
Yes, that is not in contention. Not all transformers are LLMs, not all neural networks are transformers, not all machine learning methods are neural networks, not all statistical methods are machine learning.
I'm not saying this is an LLM, margalabargala is not saying this is an LLM. They only said they hoped that they did not integrate an LLM into the weather model, which is a reasonable and informed concern to have.
Sigmar is correctly pointing out that they're using a transformer model, and that transformers are effective for modeling things other than language. (And, implicitly, that this _isn't_ adding a step where they ask ChatGPT to vibe check the forecast.)
A simple explanation would be: orders from the top to integrate an LLM. The people at the top often aren't experts who have worked in the field for years.
The quoted NOAA Administrator, Neil Jacobs, published at least one falsified report during the first Trump administration to save face for Trump after he claimed Hurricane Dorian would hit Alabama.
It's about as stupid as replacing magnetic storage tapes with SSDs or HDDs, or using a commercial messaging app for war communications and adding a journalist to it.
It's about as stupid as using .unwrap() in production software impacting billions, or releasing a buggy and poorly-performing UX overhaul, or deploying a kernel-level antivirus update to every endpoint at once without a rolling release.
But especially, it's about as stupid as putting a language model into a keyboard, or an LLM in place of search results, or an LLM to mediate deals and sales in a storefront, or an LLM in a $700 box that is supported for less than a year.
Sometimes, people make stupid decisions even when they have fancy titles, and we've seen myriad LLMs inserted where they don't belong. Some of these people make intentionally malicious decisions.
Lots of posts on HN state the fact "X" is happening and are searching for help to find the reason or just conveying a story. "Why" in the title tells people the author knows the reason and is going to explain it in the post.
Do you ever need a website you're visiting translated?
Have you ever not understood a term or phrase on a website and had to go to wikipedia/urbandictionary/google to explain it?
Have you ever wanted to do a 'fuzzy search' of a 300 page document (where you don't know the exact string of text to ctrl-f, but want to see where they talk about a particular topic)?
>Do you ever need a website you're visiting translated?
Yes, I have an extension for that.
>Have you ever not understood a term or phrase on a website and had to go to wikipedia/urbandictionary/google to explain it?
I have an extension that double clicks and brings up a quick definition. If I need more, I will go to the dictionary.
>Have you ever wanted to do a 'fuzzy search' of a 300 page document (where you don't know the exact string of text to ctrl-f, but want to see where they talk about a particular topic)?
No, not really. Ctrl + F search for a dozen substrings, use table of contents if available, and I can narrow it down. This takes a few minutes.
And if I did, I'd find an extension. You see the pattern here? We solved this issue decades ago.
The pattern is that I seek out useful tools. I don't wait for them to be rained down and force fed to me. Just because meat is useful to me doesn't mean I want to be subscribed to a grocery store who will deliver meat every month. That's an overkill of resources for most consumers.
There probably is a big difference between 'do you ever' and 'how often do you'.
I very rarely visit websites that I want translated; so rarely that I can tolerate google translate, or copying and pasting a section of a page into a tab with gemini; so on its own, this feature wouldn't sell me a browser. Besides, as a sibling comment says, even the current non-AI-enhanced browsers offer, sometimes too intrusively, to translate a page in a non-matching language. At least Chrome does this to me.
Your second scenario happens much more frequently; but again, it is so frictionless to type the term or a phrase in a search box in another tab that I never find myself wishing for a dedicated panel in the browser that could do this for me.
Your third scenario, for me, is finding something in api docs. Like, what's that command again to git cherry-pick a range of commits? So far, just googling this stuff or asking copilot / gemini in a separate tab has always worked. I am not sure I would be upset at a browser that didn't have an inbuilt tool for doing this.
What I do want from a web browser is evergreenness, the quickest and fullest adoption of all the web specs, and great web developer tools.
> Do you ever need a website you're visiting translated?
Yes. Firefox and Chrome already offer this.
> Have you ever not understood a term or phrase on a website and had to go to wikipedia/urbandictionary/google to explain it?
Yeah. And?
> Have you ever wanted to do a 'fuzzy search' of a 300 page document (where you don't know the exact string of text to ctrl-f, but want to see where they talk about a particular topic)?
No because I ctrl-f for that topic/key words and find the text.
Even if they didn't realize it, I don't believe they were arguing that firefox and chrome didn't/wouldn't use machine learning already, rather that they just thought the use cases you provided don't really sell the cost of having a full LLM integrated into every browser install.
> the cost of having a full LLM integrated into every browser install
What's this even mean?
There's no fucking way there's going to be a "full LLM integrated into every browser." You really think they're going to drop in a 20GB-200GB model with every browser? Mind you, Llama-8B is over 15GB.
Nah. So far they are doing about 50MB per language translation that you ask for[0]. You have to explicitly install languages to translate.
There's neither "a full blown LLM" (whatever that means) nor forcing AI onto you. You still have to download the language packs, they are just offering an extension that more seamlessly integrates with the browser.
And we know what they're building too! Go look in the "Labs" tab and you'll see an opt-in for testing a semantic search of your history. That doesn't take an LLM to do, that takes a vector embedding model. What next? Semantic search in page? How terrible of a feature! (But seriously, can we get regex search?)
"AI" as it's used nowadays is unfortunately usually a shorthand for LLM. When firefox talks about "AI features", I think most people interpret that as "LLM integration", not the page-translation feature that's been around for ages.
LLMs are sequence-to-sequence like language translation models, were invented for the purpose of language models, and if you were making a translator today it would be structured like an LLM but might be small and specialized.
For practical purposes though I like being able to have a conversation with a language translator: if I was corresponding with somebody in German, French, Spanish, related European languages or Japanese I would expect to say:
I'm replying to ... and want to say ... in a way that is compatible in tone
and then get something that I can understand enough to say
I didn't expect to see ... what does that mean?
And also run a reverse translation against a different model, see that it makes sense, etc. Or if I am reading a light novel I might be very interested in
When the story says ... how is that written in Japanese?
>Starting today, Google Translate uses advanced Gemini capabilities to better improve translations on phrases with more nuanced meanings like idioms, local expressions or slang.
> most people interpret that as "LLM integration", not the page-translation feature that's been around for ages.
Which seems to be the problem. People don't even realize they're being irrational, despite Mozilla being quite transparent about what they're doing. It's pretty clear.
I mean a Firefox download is 150MB, not 16GB...
Plus, we know what Firefox is looking to do. In their labs tab they let you opt into trying out semantic search of your history. So that's a vector embedding model, not an LLM.
Edit:
Okay, they have "Shake to summarize". But that's a shortcut to Apple Intelligence. Nothing shipped with the browser. Similarly I don't understand how the chatbot window is so controversial. I̶'̶m̶ ̶n̶o̶t̶ ̶s̶u̶r̶e̶ ̶w̶h̶a̶t̶ ̶t̶h̶e̶ ̶s̶h̶o̶r̶t̶c̶u̶t̶ ̶i̶s̶ ̶o̶n̶ ̶l̶i̶n̶u̶x̶ ̶o̶r̶ ̶w̶i̶n̶d̶o̶w̶s̶,̶ ̶but are people really pressing <C-x> on a mac or ctrl+alt+x on linux/windows? It's not a LLM shipped with the browser, it is just a window split ("shortcut") to literally one of the most popular websites on the internet right now (ChatGPT is literally the #5 most visited website and you all think AI is unpopular?!). Adding shortcuts isn't shoving AI down your throat. Are they shoving Wikipedia down your throat because you can do "!w hacker news"? Give me a break guys
AI is a broad term going back to 1955. It covers many different techniques, algorithms, and topics. The first AI chess programs (DeepBlue, et. al.) were using tree search algorithms like alpha-beta pruning that are/were classified as AI techniques.
Machine translation is a research topic in AI because translating from one language to another is something humans are good at while computers are not traditionally.
More recently, the machine learning (ML) branch of AI has become synonymous with AI as have the various image models and LLMs built on different ML architectures.
Okay, what's the problem? The UX of Google Translate is fine
- it will pop up when it senses a webpage in a language you don't speak.
- it will ask if you want to translate it. You have options to always translate this language or to never do it.
- it will respect your choice and no pop up every-time insisting "no please try it this time". Or worse, decide by default to translate anywyay behind my back.
- There are settings to also enable/disable this that will not arbitrarily reset whenever the app updates.
There are certainly environmental issues to address, but I've accepted that this US administration is not going to address this in any meaningful way. Attacking individuals will not solve this issue so I'm not doing this. So for now, my main mantra is "don't bother me". the UX of much AI can't even clear that.
"See but what if instead of ctrl+F, we fired up a nvdia gpu in a data center instead .... "
I'll pass. Also for people that already know another language, the auto translate is annoying, inaccurate, and strips all of the humanity out of the writing.
>Do you ever need a website you're visiting translated?
This feature doesn't seem like it needs a "first class agent mode."
>Have you ever not understood a term or phrase on a website and had to go to wikipedia/urbandictionary/google to explain it?
I already have right-click for that the old-fashioned way. Not sure how an "AI mode" would make it meaningfully better.
>Have you ever wanted to do a 'fuzzy search' of a 300 page document (where you don't know the exact string of text to ctrl-f, but want to see where they talk about a particular topic)?
This feature is the most usefully novel of the bunch but again doesn't seem like it needs a "first-class-citizen agent mode."
I have a hunch that the "first-class-citizen AI features" that instead will be pushed on us will be the ones that help Google sell ads or pump up KPIs for investors; Firefox doesn't need to jump on that hype train today.
Agent mode feels more like "Let the agent mode place your food delivery order for you?" No thanks, I don't think that's actually gonna give me my first choice, or the cheapest option...
>This isn't about talent. It's about integration... Vertical integration isn't a nice-to-have. It's the whole ballgame.
I'm going to assume there's tons of logical errors and oversights in the math, considering the author couldn't even be bothered to write the text of the post himself.
Iter is a research project that Europe is a part of, along with the rest of the world. That has nothing to do with building power plants, at least not anytime soon.
We haven't built a reactor in a long time. So those EPRs being built are all way behind schedule and thus costing substantially more.
You can design whatever you want. Building one is a whole different story. That's not an opinion that's just what happened at the first 2 EPRs and Hinckley point isn't going great either
Yup. Europe can absolutely still build reactors, just not at a price that is economically competitive.
Olkiluoto 3 started regular production in 2023, taking 18 years to build at a cost of €11 billion (3x over budget).
Flamanville 3 started regular production in 2024, taking 17 years to build at a cost of €13.2 billion (4x over budget) or €19.1 billion including financing in 2015 prices.
Hinkley Point C (two reactors) is currently estimated to have its first unit come online around 2030, taking 14 years with total costs now estimated at £31-35 billion / €36–41 billion (2x over budget) in 2015 prices.
It would really be great to understand (rather than me guessing) China's rationale to build these plants, and also their safety.
They generate about 5% of their electricity with nuclear. That's a lot, but is it enough to power the country if other alternatives stop being viable (war, shortages, ...?) Maybe it's OK for them that in such a situation, they just turn off enough residential power to last through the night with nuclear and storage. z
Do they see the nuclear research as dual use? My understanding is that nuclear subs and ships do use entirely different nuclear plants. Maybe research into small modular reactors is more dual use. There's also use for those reactors if they really want to build moon bases.
Maybe at their cost of the plans (I heard ≈3B for a 1+GW plant), this is actually competitive with solar+storage. It's definitely competitive with western nuclear power plants, if they want to export in other developing markets.
Rather than being dual use I think it’s more that countries want to keep their strategic industrial capacity around in terms of the nuclear engineering expertise in firms and universities that can potentially be redirected if needed.
The problem is that we insist on building nuclear plants like cathedrals, when we need to build them like Model T Fords.
Small modular reactors need to be rolling out of a factory ready to go, so we can do large redundant arrays of them, put them on trains to transport them around, etc.
A nuclear power station making a couple MW should cost maybe a few million tops once we have the ability to make hundreds of them a year from a factory instead of creating these 20 year projects for gigantic facilities that are all bespoke
It’s far from certain that SMRs will end up having lower costs than large nuclear reactors. Maybe they will work out but there is a huge amount of hype.
Funny, the Finns are super happy with their "uneconomic" nuclear reactors. Current approval rating for nuclear is now 81%, up from 77% last year.
The UK is so disappointed by their HPC project (which is the most expensive nuclear reactor project in history, AFAIK), that they just completed the investment decision for the follow-up Sizewell-C, which will also be 2 UK-EPRs.
Oh, the guarantee price for HPC is the same as that for various off-shore wind-projects. So obviously economically uncompetitive. At 10 pence/kWh the two reactors at HPC will produce electricity worth £200 billion. Which does put the cost of £41 billion into perspective, despite that being the most ridiculously over-time and over budget nuclear project in history.
Actually, Flamanville 3 did not start "regular" production in 2024, they were just given go-ahead to go to full power a few days ago. It was first grid-connected in 2024 and then started a lengthy ramp-up phase. It slowly coming online was the time for the Cour des Comptes to give its verdict, which was pretty damning.
Flamanville 3 was probably the worst run nuclear project in French history. And even so, this "damning" verdict was that it FV3 would only be somewhat and in the worst case marginally profitable. But still profitable. Which is better than pretty much every intermittent renewables project out there, certainly in Europe.
EDF is often accused of receiving heavy state subsides, with the implication that this is to keep the nuclear power plants going or subsidize nuclear electricity. It is true that EDF gets state subsidies. For their intermittent renewable projects. Ba-da-dum-tss. The nuclear party of their business is tremendously profitable, despite being forced to subsidize industry through the ARENH program.
Existing nuclear reactors produce incredibly cheap power. The German decision to stop theirs before coal should be considered an environmental crime.
Finns should be super happy with Nuclear since the cost overruns were overwhelmingly born by Areva (majority owned by the French state) which accumulated losses of €5.5 billion and went bust!
As a nuclear weapons power the UK has a national security interest to keep its nuclear industry around. It needs to build some reactors to do that, but given the prices of new nuclear I don't expect it to build more than the minimum necessary.
Hinkley Point C comes in at £92.50/MWh in 2012 prices (£128.90 in 2024 prices). At the last auction wind prices were £54.23/MWh in 2012 prices (£75.68/MWh 2024 prices).
Now those prices for intermittent wind exclude the cost of providing backup power with gas but that is still much cheaper than nuclear.
> Now those prices for intermittent wind exclude the cost of providing backup power with gas
Yes, let's just handwave those concerns away, it's not like the grid needs power 100% of the time or anything. Two weeks without wind? No problem, just burn gas :) It's so cheap, independent of foreign supply, doesn't leak out of pipes and isn't a huge environmental hazard at all.
But then also be honest that nuclear can't solve that problem either. It's extremely slow to ramp up and down so it cannot keep the grid stable either.
So the only way to power your grid with all nuclear is to produce at the daily peak load + margin all day. Every day
This is completely false. Nuclear plants can and do ramp up quickly, thought not from/to 0, but that's generally not necessary.
And they provide grid stability by having rotating masses on the grid, and thus combine pretty nicely with small to medium amounts of intermittent renewals that can provide some of the peak power.
> And they provide grid stability by having rotating masses on the grid, and thus combine pretty nicely with small to medium amounts of intermittent renewals that can provide some of the peak power.
We already have grids operating without traditional baseload. This is a 2015 talking point.
See for example South Australia keeping either 40 MWe or 80 MWe fossil gas in standby (I would presume this is the lowest possible hot standby power level for said plants). They are aiming to phase this out in the near future as storage comes online.
Inertia is trivially solved in 2025. Either through grid forming inverters which today are available off-the-shelf or the old boring solution of synchronous condensers like the Baltic states used to have enough grid strength to decouple from the Russian grid.
This truly shows your ignorance. Please show curiosity rather than redditesque comments like this.
First. The final report of the Iberian blackout is not completed yet. It is taking longer than expected due to how complex the situation was.
They did release an interim factual report in which they specify the facts. The full root cause analysis and recommendations on how to prevent similar events is coming in Q1 2026.
From the factual report we learn that:
1. The cause was a lack of voltage control. Do you see inertia here?
2. They did expect traditional power plants to provide this, without verifying.
3. They did not expect renewable power plants to provide this, therefore they did not.
In about all other grids like, like for example the US, renewable plants are expected to provide voltage control. It is trivially done by extremely cheap off-the-shelf components.
But if the expectation does not exist then it will not be provided since the cost is non-zero.
My point was that, just like with renewables, a 100% nuclear grid doesn't work either.
They can adjust power but they're typically used as he load with some other source dealing with the peak load needed a short time a day. Typical peak capacity can be off in the middle of the night for example. Nuclear doesn't like that.
> At 10 pence/kWh the two reactors at HPC will produce electricity worth £200 billion.
2 things, 10 pence is a lot. Not for retail but no power plant gets anywhere near that. It's mostly like 6 or 7.
Aside from that, the money you put in today is not spent on other things so there's an opportunity cost there too. That 40 billion at 2% interest is 60 after 20 years for example
> And even so, this "damning" verdict was that it FV3 would only be somewhat and in the worst case marginally profitable. But still profitable. Which is better than pretty much every intermittent renewables project out there, certainly in Europe.
What do you mean? Plenty of renewables are built without any government backing..
Have you looked when they started construction and what their projected end date was?
Yes there are new ones but both of those are perfect examples of the lack of knowledge [1].
I'll quote:
> Many of the organisations chosen to work on the different parts of the plant did not have any experience in nuclear, and little understanding of the safety requirements.
We'll get there. But yes, we're rebuilding a lot of lost knowledge and paying for the teething issues.
Those are not really great construction examples, are they? Both projects took 15+ years to complete with huge cost overruns. And for those two "successful" projects, you can find 2 or 3 that failed.
The Finnish reactor had one delay because the concrete used for the containment building wasn't of the 'nuclear grade'. That's why those regulations thankfully exist.
Building more will help though. This whole thread started about how we had lost important knowledge
seems like a weakness for these generators, if you can prompt them into producing copyright violations then that seems to be pretty risky for the folks running them.
reply