snewman

My point is that if they had this problem in their VMware support, they might have a similar problem in one of their other services. But they didn't check (or at least they didn't claim credit for having checked, which likely means they didn't check).

Given the level of impact that this incident caused, I am surprised that the remediations did not go deeper. They ensured that the same problem could not happen again in the same way, but that's all. So some equivalent glitch somewhere down the road could lead to a similar result (or worse; not all customers might have the same "robust and resilient architectural approach to managing risk of outage or failure").

Examples of things they could have done to systematically guard against inappropriate service termination / deletion in the future:

1. When terminating a service, temporarily place it in a state where the service is unavailable but all data is retained and can be restored at the push of a button. Discard the data after a few days. This provides an opportunity for the customer to report the problem.

2. Audit all deletion workflows for all services (they only mention having reviewed GCVE). Ensure that customers are notified in advance whenever any service is terminated, even if "the deletion was triggered as a result of a parameter being left blank by Google operators using the internal tool".

3. Add manual review for any termination of a service that is in active use, above a certain size.

Absent these broader measures, I don't find this postmortem to be in the slightest bit reassuring. Given the are-you-f*ing-kidding-me nature of the incident, I would have expected any sensible provider who takes the slightest pride in their service, or even is merely interested in protecting their reputation, to visibly go over the top in ensuring nothing like this could happen again. Instead, they've done the bare minimum. That says something bad about the culture at Google Cloud.

A company with a high P/E is by definition not making much money for its shareholders, relative to the size of their investment.

Unless you mean that the share price may appreciate. That's absolutely a thing, but it's a dangerous game. Of course plenty of people have made fortunes this way; people have also lost fortunes; I think the advice to steer away from such companies is basically a statement about risk.

You could merge these data structures as well. If the two instances to be merged are not at the same "round", take the one that's at an earlier round and advance it (by discarding half the entries at random) by the difference in rounds. Then just insert the values from one list to the other, ignoring duplicates; if the result is too large, discard half at random and increment the round number.

I implemented exactly this algorithm at my previous employer, except that alongside each value, we stored an estimate of the number of times that value appeared. This allowed us to generate an approximate list of the most common values and estimated count for each value.

I think this is unlikely. There has never (in the visible fossil record) been a mutation that suddenly made tigers an order of magnitude stronger and faster, or humans an order of magnitude more intelligent. It's been a long time (if ever?) since chip transistor density made a multiple-order-of-magnitude leap. Any complex optimized system has many limiting factors and it's unlikely that all of them would leap forward at once. The current generation of LLMs are not as complex or optimized as tigers or humans, but they're far enough along that changing one thing is unlikely to result in a giant leap.

If and when something radically better comes along, say an alternative to back-propagation that is more like the way our brains learn, it will need a lot of scaling and refinement to catch up with the then-current LLM.

Guessing: perhaps this means, if someone needs credit for shepherding an improvement to search quality into production, here is a set of known improvements waiting for someone to take ownership.

Yes, very strong T-1000 vibes – the way it keeps reversing "front" and "back" almost feels like a deliberate reference to that moment in Terminator 2.

Concur. Zookeeper was absolutely inspired by Chubby, and this story fits in every detail with how things worked at Google back in the day. (Source: I was at Google from 2006 - 2010. Also, I believe the origins of Zookeeper are a matter of public record, though I'm too lazy to check right now.)

This does not seem to be correct. I asked Kagi "does kagi build its own index". From the first result (on help.kagi.com):

> Kagi Search includes anonymized requests to traditional search indexes including Brave, as well our own non-commercial index (Teclis), news index (TinyGem), and an AI for instant answers. Teclis and TinyGem are a result of our crawl through millions of domains, focusing primarily on non-commercial, high-quality content.

[Edit] I guess it is possible that Google or Bing are hiding under "traditional search indexes including Brave".

This is not strong evidence. While, ideally, the behavior you describe should not happen, there are many reasons it could arise. (Bugs, font sizing comes out differently at printer resolution, paper size is different than on-screen page size, different implementation for printing, etc.)

Given that Docs nowadays uses Canvas to render the document being edited, how would it even be possible to delegate rendering to the browser? (Not to mention the challenge of live rendering of pages, footnotes, etc.)

I don't understand your points here.

It does seem that Google Docs is using the browser to measure and render individual units of text. But that is very much not the same thing as layout: deciding how much text fits onto a line, how many lines fit onto a page, making room for footnotes, etc. My understanding is that Docs is handling layout itself, in order to support features like pagination and footnotes.

You say the browsers can support all of the layout features Docs provides, but do they allow live editing and on-screen display of paginated text? That certainly wasn't the case back when I wrote the original Google Docs (I'm a co-founder of Writely) and I don't believe it was the case when the current canvas implementation shipped. (I had left Google by then and have no inside knowledge of that rewrite.)

One final note: back in the day, I have implemented and shipped a full word processor for the Mac (FullWrite). We very much handled layout ourselves, but relied on the OS for text measurement and rendering, presumably in a similar fashion to how Docs leverages the browser today.

What is your source for the planet only being able to support one billion people? That's a very bold statement, given that we are currently at eight billion and have fairly clear technological pathways to sustainability.

Of course sustainability is not a guaranteed outcome, it will require making good decisions, but I'm sure that a "mere" one billion people would also be unsustainable with sufficiently poor decisions.

(author here)

Of course LLMs are not people. But human metaphors can (sometimes!) be useful in understanding, explaining, and even enhancing their behavior. For instance, techniques such as Chain-of-Thought prompting explicitly apply techniques that work well for people to improve the reasoning ability of LLMs.

A point I attempt to make in this article is that one reason reason LLMs are so vulnerable to jailbreaks and prompt injection is that these types of attacks include non sequiturs that are not well represented in the training data. I would argue that "LLMs are gullible because they are naive [haven't had much past exposure to this form of trickery]" is a reasonable mental shorthand for explaining and internalizing this idea. It's especially helpful for readers who won't be familiar with terms like "out of distribution" or "adversarial examples", but who would benefit from being able to internalize the idea that LLMs are easily subverted.

In other words, I don't think it's helpful to reflexively dismiss any application of human metaphors to LLMs. It's easy to go wrong with metaphors, but they can also be valuable tools for conveying complex ideas. Did you read the article, and do you have any comments as to the substance of its content?

There are definitely artifacts. Go to the 9th video in the first batch, the one of the guy sitting on a cloud reading a book. Watch the book; the pages are flapping in the wind in an extremely strange way.

Honest question: of what possible use could Sora be for Hollywood?

The results are amazing, but if the current crop of text-to-image tools is any guide, it will be easy to create things that look cool but essentially impossible to create something that meets detailed specific criteria. If you want your actor to look and behave consistently across multiple episodes of a series, if you want it to precisely follow a detailed script, if you want continuity, if you want characters and objects to exhibit consistent behavior over the long term – I don't see how Sora can do anything for you, and I wouldn't expect that to change for at least a few years.

(I am entirely open to the idea that other generative AI tools could have an impact on Hollywood. The linked Hollywood Reporter article states that "Visual effects and other postproduction work stands particularly vulnerable". I don't know much about that, I can easily believe it would be true, but I don't think they're talking about text-to-video tools like Sora.)

My understanding here is shallow, but I think you are correct.

Here's a simple exercise to help your intuition: open your hand up flat, and hold it facing your mouth, a few inches away. Blow, and you can easily feel the air pressing on your palm. Now inhale sharply; your hand will feel no suction at all.

Tossing another anecdote onto the pile, I just searched "Adidas shoes". The first four hits are not for Adidas shoes, after that they all are.

All four of those non-Adidas hits are "Sponsored" results, but that's not particularly easy to notice (for one of them it's almost impossible).

I would agree with GP that Amazon product search often leaves much to be desired. It's hard to search for X without getting accessories for X, competitors to X, products that can be connected to X. This is annoying at best, and crippling if you want to, say, sort by price (all the cheapest matches will be accessories or otherwise not what you wanted).

Well, moving away from "efficiency" can take you toward redundancy, but it doesn't have to. Some waste is just wasteful. It's important not to oversimplify, otherwise you wind up with counterproductive memes like "that project is justified because it will create jobs" or "the free market is the best solution for every problem" (it's a great solution for some problems, but ignores externalities).

I had a personal brush with Feynman with regards to this experiment.

Circa 1986, I read Surely You're Joking, Mr. Feynman and became mildly obsessed with the fact that Feynman doesn't present the result of the sprinkler experiment (see the first few seconds of the video for context). A number of colleagues (at the software startup where I was working) and I tied ourselves up in knots debating what the answer must be.

I wanted to perform the experiment, but lacking materials, skill, and (frankly) ambition, I settled for a laughably primitive apparatus involving a couple of bendy straws, the bathroom sink, and my mouth. This was enough to reproduce the well-known fact that if you push water out through the sprinkler, it will spin. When I tried sucking water in, it would give a momentary rotational jerk and then stop moving... but perhaps that was due to my mouth tightening up on the straw?!?

After I shared this inconclusive result, one of my co-workers decided to get to the bottom of the matter. He called the Pasadena operator, asked for the home phone of a Mr. Richard Feynman, and – to my utter horror – dialed the number.

It seemed impossible that this was a viable procedure for making contact with a Nobel Prize-winning physicist, but through the speakerphone, we could all hear a gruff voice that certainly sounded like it might be him. All doubt was removed when my friend explained that we were looking for the answer that isn't presented in the book, and the gruff voice said, "Why should I tell you?"

Feynman then asked whether we had tried the experiment, and to my redoubled horror, the phone was handed to me. I stumbled through some explanation of what I had tried and what result I had observed. Fenyman then relented, and explained the entire situation in ten short words:

"The sprinkler cannot rotate, because no angular momentum is transferred."

(Not an exact quote, but something very much like that)

My understanding, and I've read a fair amount on the subject, is that geography for pumped storage is actually pretty limited compared to the eventual need. Do you have a source for it being plentiful?

Promising alternatives to mitigate solar and wind variability, aside from pumped storage and nuclear, include advanced geothermal, load management, thermal storage, long-distance transmission, storing clean hydrogen or clean methane (made from electrolysis and captured CO2), and possibly long-term battery technologies (not lithium, that'll mostly be for overnight or shorter). In no particular order.

Holy cow.

How did the fire start, and where was it initially spreading that went up so fast?

That's generally where methane comes from today, but in time it may come from captured CO₂, e.g. see https://terraformindustries.com/ (disclaimer: I am a minor investor).

I'm personally devoting my career to AI safety, on a volunteer basis, because I think it's is legitimately of high importance. (See my blog, e.g. https://amistrongeryet.substack.com/p/implications-of-agi, if you want to understand where I'm coming from.)

What makes you think it is for show?

LLMs can't do just about any extended, context-dependent task, which is why they're far away from being drop-in replacements for people.

Absolutely they're doing more than just regurgitating training data, but there's a long distance from there to "any task a human could do".

Isn't the term for this "unfalsifiable"?

Because AWS doesn't rent you chips, it rents you cores. If a chip has 50% more cores, that doesn't help you as a renter of cores, just means they get to rent out more cores.

Of course it might give Amazon room to offer these instances at a lower hourly rate per core, which would ultimately cash out as improved cost/performance for AWS customers.

> Graviton4 processors deliver up to 30% better compute performance, 50% more cores, and 75% more memory bandwidth than Graviton3.

This seems ambiguous. Presumably this is 50% more cores per chip. What about "30% better compute performance" and "75% more memory bandwidth": is that per core, or per chip? If the latter, then per-core compute performance would actually be lower.

Also, "up to" could be hiding almost anything. Has anyone seen a source with clearer information as to how per-core application performance compares to earlier Graviton generations?

Source?

My understanding is that this can be true in some regions but requires fairly aggressive use of, not only overbuilt renewables (which may pose challenges in land use and siting), but also long haul transmission (also difficult to site), demand response, etc. And then there's the scary possibility of an extended weather anomaly (several weeks with no wind and minimal solar).

Having some long term storage in the mix, probably in the form of green hydrogen or another e-fuel, greatly increases our options for building a robust fossil-free grid.

Putting more clean firm sources such as nuclear or advanced geothermal in the mix also helps, but I am not optimistic regarding nuclear, and advanced geothermal is still pretty immature.

HVDC lines can be integrated into a conventional grid; China is doing it, and NREL models for the US grid include scenarios where we do it as well.

Trains can often be electrified, as can semi trucks, though it will require substantial investment in heavy duty charging infrastructure. Long haul aviation, and probably long haul shipping, will likely continue to require chemical fuels (though Fleetzero aspires to electrify shipping). But multiple startups have promising routes to replacing fossil fuels with e-fuels for those applications (chemical fuels created using electricity, such as electrolysis and carbon capture, instead of pumping fossil fuel out of the ground).