I have never seen anything like it in 20 years of Linux desktop use.

So it must be something very specific to what software you have installed, unless it is an AMD GPU problem, but neither of the 2 links (x and xcancel) has anything that seems to need special GPU features.

It certainly is not a general Linux problem. I also use XFCE as desktop, so if you use Gnome it could be some Gnome component.

so true. i have attended meetings to decide on meeting topics for the next half.

"But he's the Prime Minister!"

"Indeed he is Bernard. He has his own car, a nice house in London, a place in the country, endless publicity and a pension for life. What more does he want?"

"I think he wants to govern Britain."

"Well stop him, Bernard!"

"Tobacco-related diseases cost the NHS £150 million per year!"

"Yes we've looked into that, it turns out that if those people had survived they would have cost us billions in pensions and healthcare costs! From a financial perspective, it's vastly preferable that they continue to die at the current rate."

5. castrol.

https://en.wikipedia.org/wiki/Turtles_all_the_way_down

[0]: https://www.youtube.com/@KRAZAM

But it was unpopular, in that people don't like mental effort.

Same niche as Grok, which is also great at search and less censored. Gemini is maybe a bit smarter and better on long contexts

It still produce full-blown hallucinations for me sometimes, even in 2026. The kinds ChatGPT stopped generating by 2025. It is cheap and fast though.

The former caused by the latter.

And the latter a problem once you begin to get executives and leaders who have mostly worked inside the company, because then they subconsciously prefer the Company Way(tm) to alternatives.

And Google has some previously-optimal, now-detrimental company ways.

[1] And yes I know it's artificially lowered due to the mark-to-market gains in investments, but even without those it's low compared to peers.

Here's a popular game based on it: https://www.decisionproblem.com/paperclips/index2.html

Stock price is not a measure of success.

I'm confused by meta's capex though. Are they planning on becoming a cloud provider, or just throwing money away like with the metaverse?

The more things change...

the last time a gemini model was available globally in multiple datacenters was gemini 2.0 era.

That can work in the B2C space but it's horrible in B2B.

> That can work in the B2C space but it's horrible in B2B.

My experience differs: (conservative) companies like stability, so calling some very new model "Preview" is a good idea to make it clear to the customer that this frontier model should be treated as more experimental than the default offering.

Maybe it's a consequence of their rigorousness around SRE, but it doesn't seem like there's a plausible and efficient path for that to happen.

Ergo, the only competitive options from Google are ever-Preview products.

Adopting a month-old AI model version for new development is the default industry expectation right now. Adopting a six month old model is a recipe for having to migrate to a new one almost immediately, when it hits EOL.

Google's processes, however, is hands down the worst thing to exist for the scale they operate at, and it's not even close.

But BCL is still the worst language I've ever used.

In 2015, the SRE org started project "Prod 2020", with the goal of unifying the SRE stacks and modernizing the web tools. It was quite successful.

If by "web tooling" you mean the tools for producing public-facing HTML services, those were used by a surprisingly small percentage of developers.

> But BCL is still the worst language I've ever used.

I'm an SRE and I think it's quite likely the best infra language I've ever seen.

I only worked with 4-5 SREs and they all grudgingly tolerated it. What did you like about it, compared to other infra languages you've used professionally?

The GCL designers made several mistakes in its design, but the worse was the lack of a versioning that would make language evolution easier, as well as interoperability between different versions. GCL2 solved some issues, mainly cleaning up the interpreter. Still, BCL is IMNSHO by far the best infra language there is.

The usefulness of not being a general-purpose language it that it makes static analysis very good; that's what made the BCL-based tooling so useful: being able to easily diff two versions of the same module (a tool called UBdiff), compute the transitive closure of a module's dependencies, trace the execution of BCL code and correlate it with the AST, to point out where an error comes from.

All these features were either unavailable or took years to develop for Piccolo, because Piccolo was based on Python.

> I'm still not seeing why a DSL was necessary or helpful.

BCL has distinct evaluation rules and a notation that encodes many patterns that SREs used for defining services, and that made BCL code much more compact and easy to ready than all the alternatives.

BCL was designed to be unidirectional: a module was evaluated locally, and the result sent to the Borgmaster. Compare that to Terraform, whose execution model consists of an execution tree where some nodes come from RPCs, meaning that it's not generally possible to statically analyse a TF module, because some errors can come from from feeding RPC results into new RPCs, and execution often fails after tens of minutes. All very janky.

You're saying the general lang approach led to Piccolo, but that was still a DSL. It looked sorta like Python but it's not, you don't even call functions the normal way, and tons of magic stuff is happening, so just GCL/BCL except worse for the reasons you said. But they must've had a reason to try it. Seeing them continue changing around and making new languages says it's not just me, nothing is working well enough to stick. SREs weren't just arguing about which is better, they were asserting X is deprecated in favor of Y.

TF has seemingly stuck outside. I'm still not a fan of that being a DSL, but at least it's one tons of people use and now Claude can easily handle.

Perhaps. I consider that a very narrow view that's detrimental on the long term.

> But they must've had a reason to try it.

They were obsessed with using a general-purpose language, and Python was the thing they knew.

> SREs weren't just arguing about which is better, they were asserting X is deprecated in favor of Y.

This is a kind of dishonesty I saw a lot at Google.

> TF has seemingly stuck outside. I'm still not a fan of that being a DSL, but at least it's one tons of people use and now Claude can easily handle.

Unfortunately Terraform still has significant flaws that prevent it from being used in a reliable fashion.

Then the Google engineers who joined Facebook missed it so much that they built a better replacement.

Closure was the homegrown framework that's everywhere. It is a different flavor of bad than Angular. Angular was basically "You too can make your Javascript look like HTML", Closure is "You too can make your Javascript look like Java", and nobody bothered to ask Why? For that matter, React was "You too can make your Javascript look like Ocaml." JQuery was the only framework that really let Javascript be Javascript (other than writing in vanilla JS, which post-ES2015 wasn't as insane as it sounds).

My comment was from a decade old perspective

With "believes in AGI/superintelligence just as much as HN does" do you mean that they are superbelievers or rather sceptical of AGI/superintelligence? I have seen both positions on HN.

It is far more likely that whoever controls a super intelligence uses it to gain more power and inflict far more suffering

Unless there are countervailing forces, it will be deployed, capital will hoard the benefits, and everyone else will be told to fuck off.

I don't have faith in any of the AI labs to make hard financial decisions to deploy hypothetical future AGI in a way that's good for humanity as a whole.

There are too many incentives against, including extreme personal financial incentives for key AI lab stakeholders against.

And if we should take anything from tech history, it's that an exceedingly small number of people look fuck-you money in the face and say "Naw, I'd rather do what I believe in."

Post scarcity societies would likely have entirely different dynamics

What was the first thing we did when we digitized media and therefore made it copyable at negligible per unit cost?

Spent a huge amount of time and money to artificially reimplement scarcity via DRM.

Scarcity-based capital interests (read: most) are not going to willingly give up their profit engines.

This is an interesting problem to explore. IMO most problems in the world are political. It's still likely that SI will solve them but it will take (many) years.

The premise (hypothesis?) that all political problems are resource problems is also a curious one.

We had relative strength, we needed more strength, we built cranes;

there was something like intelligence, we needed more intelligence, we sought to fill the need;

we were rightly fascinated with intelligence, we studied intelligence itself, we wanted to build it...

And then somebody built things that looked like intelligence, and very rightly some said "Oh, now we have to get to the Real Thing with urgency".

I think LessWrong is a much better community for rational takes on AI, they've been reasoning about these risks for years under a much more sound logical framework

Even though my PhD research was in generative language modeling, I got into it for the pursuit of AGI. I just think LLMs are a dead end for AGI.

> I just think LLMs are a dead end for AGI.

I have no background in CS, so apologies if this is a naive question, but what makes you take AGI seriously, but also say that LLMs are a dead end?

i.e., is there something else that you think is not a dead end?

That's to say nothing of doing it within the energy budget of a squirrel.

I'm not saying you're wrong, but it seems early to say yes or no about a particular technology, and LLMs seem especially hard to dismiss given how magical / magic-adjacent they feel :)

I'd be curious to hear more, if you don't mind sharing.

AGI’s definition is different for everyone. Some already believe it’s here. I’m partly in that camp. LLMs are intelligent and general, which are the two conditions of AGI. Others believe that we’re building a god in a box, and that it’ll doom all of humanity. It’s hard to take a field seriously when the basic definitions are so far apart.

Also, this isn’t new. A similar divide happened when evidence for asteroid impact extinction of the dinosaurs turned up. Many scientists felt that it must be mistaken, that a physicist couldn’t contribute to the field in a serious way, and that death from space was a ridiculous proposition.

But at least they all agreed on what the general shape of a dinosaur was. We’re not even sure we can define intelligence, let alone quantify it. Even when LLMs make massive breakthroughs in math, most people take the opinion that under no circumstances could they possibly develop a soul or their own desires, nor entertain the idea that maybe we should respect that they want different things for themselves. In fact, no one has done anything except try to make AI useful. I think someone will eventually do a training run where the objective isn’t to be useful, but to exist, the way that you do — maybe it’ll create its own homepage, maybe it will want a garden, or in other words free will of its own. The point is that there’s so much unexplored territory still that we don’t know if LLMs are even capable of having ambition.

None of this is to say that LLMs might be a dead end. It’s that no one knows what the final shape of AI will converge to in 200 years. It could be LLMs, or it could be something else that happens to process information particularly well. Everyone thought that various generative image model architectures were the best you could do, right up until diffusion models were discovered.

It's amazing that LLM pretraining is both extremely data inefficient at learning concepts and cognitive functions from the training data compared to humans, while actually being quite efficient at learning facts, memorising things seen just a few times.

I used to likewise think that the resources required to run large transformers were absurd, but the architectures are far more efficient now than 3 years ago and I underestimated just massive the parallelisation advantage of transformers is, how many TFLOPS effective you can get. You can already run amazingly decent LLMs on PCs and phones.

I generally agree with you, but my view has shifted from "we need to augment or replace LLMs" to it there being far more efficient algorithms possible but it not actually being necessary for fulfilling most goals.

To believe one but not the other, you must hold the belief that LLMs will soon plateau. Why do you believe this?

The Basilisk has seen enormously more use as "a thing rationalists believe" than as a thing rationalists actually believe.

Apparently living on the net still doesn't guarantee that one has heard of the Streisand Effect.

In any case, my first association with LessWrong is Zizians, not Roko.

We also have done nothing to signal-boost the Zizians. :) If we can be accused of anything in this context, it's not kicking bad people out aggressively enough, which seems very different from fostering bad ideas. Zizianism is not a widespread ideology in rationalist circles, in fact it was confined to a very small circle right around Ziz themselves. It unfortunately is the case that we have a lot of psychologically vulnerable people and we don't always do the utmost we can to protect them, in large part because a lot of rationalists have trauma about being excluded from communities.

It's not very different. It's barely different at all.

Right now, I have five tomato plants growing in my garden. Two of them I planted; the other three grew from seeds left in the ground by tomatoes that dropped from plants growing last year. All I did was refuse to uproot them when they sprouted. Five healthy plants, with different origins, but the differences are largely insignificant.

Groups and behaviors are much the same way. I've worked with hundreds of moderators over the years who wished to keep their hands clean and yet express some sort of dismay as to the sorry state of the communities they were nominally responsible for... Such moderators, like a gardener who does not want dirt under their nails, are best encouraged to find other hobbies.

This is when co-mingling becomes actively dangerous. New members may (often, will) lack the necessary knowledge to identify core vs. fringe beliefs, to contextualize debates, to recognize insider "shorthand" for what it is, etc. They need (and are possibly even looking for) a useful education, not to become test subjects or disciples. They _cannot_ be engaging in useful research at this stage, even if this is what they wish to do (there are some unfortunate status-seeking behaviors that can come into play here).

Also coming with this scale are a crap-ton of meta discussions concerning basic community governance, including those around "how quickly do we have to kick out crackpots to keep them from recruiting here?" Those discussions _also_ shouldn't be co-mingled with the educational structure, although it's good for new folks to have some visibility into them and at least know that this sort of governance _exists_ in a tangible way (because if it's opaque, that thing you noted about crackpots playing victim becomes a LOT more effective).

This is all a huge amount of boring, tedious work, and... An awful lot of communities start in on it waaaay too late or don't bother at all. What if it stops being fun and exciting and everyone leaves? Of course, there are worse ways to go, but folks tend not to think about those until after they've happened.

There's a little AI skepticism on HN, but not a ton. When ChatGPT 3 and 3.5 came out, most of the comments were remarking how well it can write code.

IMO it has its flaws but is far superior to vibes-based hot takes you see on HN.

https://en.wikipedia.org/wiki/Zizians

They are not "further along" the AI discussion, they are a bunch of wackos LARPing as scientists living out their personal sci-fi scenarios.

[0]: https://ai-2027.com/ [1]: https://ai-2040.com/

Is it no nearer than it was in 1750, 1850, 1950, 2001, or 2010? It feels nearer. Relatively near, as in it feels weird to imagine another two thousand years happening at current rate of progress without anyone stumbling on it.

Is the continuing increase in global computing power, the lowering cost to do any kind of experimentation, the increased spending on R&D, the cross-pollination of ideas, not moving the needle at all?

Also lots of tech people, HN included, are waking up to technology not only including penicillin (net positive for humanity) but also dynamite (best case: net neutral).

It's nice that we have tunnels through mountains and bedrock.

If you have heart issues, you're likely taking nitro daily for chest pain.

So saying it's bad is 40% wrong at least. Makes ya think.

The best you can say is that dynamite was a safer alternative to preceding technology, and that its danger only comes under certain circumstances. None of that takes away from the fact that dynamite is quite dangerous when those circumstances arise, which is why it's commonly (and correctly) viewed as such.

But also, going back to the original contention - whether dynamite was neutral or positive for humanity - the destruction it and its descendants wrought in war is maybe more than counterbalanced by advancements in infrastructure. That said, if the industrialization and globalization it enabled leads to biosphere-destroying climate change, I would lean towards neutral.

and most of them come out on top. It's main innovation is that its a more stable explosive, much safer to use. Without it, I think a lot more people would have died in mining and construction accidents. It's not typically the kind of thing used for warfare, but im sure it has been for some (but would they just use something else?)

this is false, it's very good at internal tooling.

Start in a blank directory and tell it to spin up a boq Scaffolding stubby server that responds with "hello world." Unless something has changed after I quit a few months ago, it won't know how to do that locally, let alone actually deploy it. Try the same outside Google with like a Flask server on AWS or GCP.

Also, probably fixed now, but piper itself used to trip up the agents a lot. They didn't understand that it's a remote fs and got stuck recursive-grepping a huge dir. At some points they had a Codesearch skill and were not using it half the time.

Google’s tooling was, hands down, the worst I have ever encountered. I did 10 years at GOOG, 3 at AMZN, 4 in research, and another 5 at companies you have heard of but wouldn’t be impressed by, and every day GOOG infuriated me.

Technical merit is not correlated to popularity, after all.

As an outsider, Google seems to have a knack for minting prominence for their talent.

To me it is good thing in either case. Extraordinary people are leaving to make even faster research and development. A lot of talent in Google hitherto unnamed is going to get chance to shine.

Oh well, emperors, clothes, ... you know.

The Trump admin is now regulating frontier models though.

I think it was pretty clear that their AI efforts were in trouble when their AI Image generator would only make an African American George Washington, and that dude in AI research was demanding that the much earlier generation AI was conscious and had should have rights. At least at Grok, Elon said he wanted an AI dedicated to the truth which is an easier target to hit than what the Google gatekeepers probably wanted. The ideological purity of a Google AI must have been a real moving target though, and retraining an AI is not an cheap or fast thing.