HNHacker News
TopNewBestAskShowJobs

boulos

13,551 karma · joined May 15, 2013

If you're trying to reach me about GCE (or Google Cloud generally) my username at Google is the same as this one.
submissionscomments
boulos··on Building the heap: racking 30 petabytes of hard drives for pretraining
It's quite cheap to just store data at rest, but I'm pretty confused by the training and networking set up here. It sounds like from other comments that you're not going to put the GPUs in the same location, so you'll be doing all training over X 100 Gbps lines between sites? Aren't you going to end up totally bottlenecked during pretraining here?
boulos··on Bringing fully autonomous rides to Nashville, in partnership with Lyft
Both Lyft and Flexdrive:

> Riders will hail via the Waymo app, and as our service grows, riders will also be able to use the Lyft app to match with a Waymo vehicle.

(Disclosure: I work at Waymo, but not in partnerships)

boulos··on I bought the cheapest EV, a used Nissan Leaf
spark.bg uses recent eUPs heavily! It's the first car listed at https://spark.bg/en/electric-cars/ .
boulos··on Six months into tariffs, businesses have no idea how to price anything
To put it in perspective, China installed 277 GW of new solar capacity in 2024.

For coal, the "started construction" number there isn't the same metric as began operation. You want to look for "commissioned" and you get 30 GW. From https://energyandcleanair.org/publication/when-coal-wont-ste...

> Note: In 2024, 66.7 GW of new coal power capacity was permitted, a decline from previous years but still above the subdued pace seen earlier in the year. New and revived coal power proposals totaled 68.9 GW, down from 117 GW in 2023 and 146 GW in 2022, indicating a potential slowdown in project initiation. Meanwhile, construction started on 94.5 GW of new coal capacity — the highest since 2015 — suggesting continued momentum in project development. However, the pace of new coal plants entering operation has been more moderate, with 30.5 GW commissioned so far in 2024, down from 49.8 GW last year but in line with 2021 and 2022 levels.

China is well positioned to do solar + storage, but a lot of that coal is probably (a) for base load, (b) for steel production and (c) to keep the coal miners in business. From the same write up:

> In 2024, more than 75% of newly approved coal power capacity was backed by coal mining companies or energy groups with coal mining operations, artificially driving up coal demand even when market fundamentals do not justify it.

boulos··on Income Equality in Nordic Countries: Myths, Facts, and Lessons
I think the commenters above were mostly talking about late 19th century immigration by Irish, Italian, etc. workers in the post-Civil War era.

It's still true that the pay and conditions were awful, but it was clearly something people chose to do.

boulos··on Nvidia DGX Spark
As long as you're going to add FP8 dense, you could do the same for the parts mentioned in the FP4 section. Divide by two from dense => sparse, and another two for FP4 => FP8.

That gives you 250 tops of fp8 for Spark.

boulos··on Gemini 2.5 Flash Image
https://www.reddit.com/r/PhotoshopRequest/

People sometimes do it for free ("my son died, and this is the only photo I have") or for an agreed upon tip.

boulos··on AI is different
Our next vehicle sensor suite will be able to handle winter weather (https://waymo.com/blog/2024/08/meet-the-6th-generation-waymo...).
boulos··on Dropbox announces new gen server hardware for higher efficiency and scalability
To be clear, that altercation was definitely taken out of context and there was bad behavior on both sides.

As reported at the time [1], the kids who didn't have a reservation knew there were reservations, but wanted to highlight that they didn't like the new system and the Mission was gentrifying:

> Kai told me that he had heard over the summer that pickup-soccer players were regularly being kicked off the field in the evening by adults who had paid to reserve the field through the city’s Recreation and Park Department. Having grown up playing soccer on that field, Kai decided to try to help the neighborhood kids make a stand for the existing rules. “I’ve also felt kind of exiled from the community, because of my eviction, and I didn’t want to see the same thing happening,” he said.

The Dropbox employees effectively took the blame for a Rec and Park policy that was (temporarily) changed in response [2].

So sure, they said something clumsy in a heated argument. But if they'd more politely said "The courts are reservable, why should it matter which neighborhood I'm from?" would you still be so angry at them?

[1] https://www.newyorker.com/tech/annals-of-technology/dropbox-...

[2] https://abc7news.com/amp/tech-workers-mission-district-teens...

boulos··on [dead]
Company announcement: https://bedrockrobotics.com/news/introducing-bedrock-robotic...
boulos··on Jank is C++
You should probably use either boost::unordered_flat_map or absl::flat_hash_map if you don't need ordering. Especially with 20-byte values. (Though you didn't mention the key type). If you're dealing with building against Boost already, I'd just switch and measure it.

https://github.com/boostorg/boost_unordered_benchmarks/tree/... (Via a response from the boost authors in https://www.reddit.com/r/cpp/comments/yikfi4/boost_181_will_...) has more benchmarks depending on your pattern.

boulos··on Curved-Crease Sculpture
For folks interested in folding and origami, the documentary Between the Folds was excellent. I don't know if anyone recorded a Q&A when it did the film festival circuit, but if you could find one, it'd be worth watching.
boulos··on Voyager: Real-Time Splatting City-Scale 3D Gaussians on Your Phone
You're not wrong.

The early 2000s splatting from point-based rendering is what George Drettakis and his students realized could be applied in this new NeRF domain.

Basically, all the reasons that point splats didn't work for regular surface rendering nearly 25 years ago (holes, inefficient, can't edit them like meshes) are less of an issue in a light field capture setup.

boulos··on Waymo's market share in San Francisco exceeds Lyft's
Small correction: in Phoenix you can also use the Waymo One app directly. In both Austin and Atlanta though, we are only available via Uber.
boulos··on Jemalloc Postmortem
Oh, that wasn't the intent. I meant two separate things. The Itanic itself was kind of fascinating, but mostly panned (hence the nickname).

SGI's decision to built out Itanium systems may have helped precipitate their own downfall. That was sad.

boulos··on Jemalloc Postmortem
The Itanic was kind of great :). I'm convinced it helped sink SGI.
boulos··on Mary Meeker's first Trends report since 2019, focused on AI
The person you're responding to is a pretty prolific poster here. Often about financial matters. And well before good text models would have been useful.
boulos··on Beware of Fast-Math
I've also come around to --ffast-math considered harmful. It's useful though to help find optimization opportunities, but in the modern (AVX2+) world, I think the risks outweigh the benefits.

I'm surprised by the take that FTZ is worse than reassociation. FTZ being environmental rather than per instruction is certainly unfortunate, but that's true of rounding modes generally in x86. And I would argue that most programs are unprepared to handle subnormals anyway.

By contrast, reassociation definitely allows more optimization, but it also prohibits you from specifying the order precisely:

> Allow re-association of operands in series of floating-point operations. This violates the ISO C and C++ language standard by possibly changing computation result.

I haven't followed standards work in forever, but I imagine that the introduction of std::fma, gets people most of the benefit. That combined with something akin to volatile (if it actually worked) would probably be good enough for most people. Known, numerically sensitive code paths would be carefully written, while the rest of the code base can effectively be "meh, don't care".

boulos··on Claude 4
Your article says it was life => art => life!

> Gang leader Robert Sheldon Brown, known as “Casper” or “Cas,” from the Rollin’ 60s Neighborhood Crips, heard about the extraordinary pilfered sum, and decided it was time to get into the bank robbery game himself. And so, he turned his teenage gangbangers and corner boys into bank robbers — and he made sure they always brought their assault rifles with them.

> The FBI would soon credit Brown, along with his partner-in-crime, Donzell Lamar Thompson (aka “C-Dog”), for the massive rise in takeover robberies. (The duo ordered a total of 175 in the Southern California area.) Although Brown got locked up in 1993, according to Houlahan, his dream took hold — the takeover robbery became the crime of the era. News imagery of them even inspired filmmaker Michael Mann to make his iconic heist film, Heat, which, in turn, would inspire two L.A. bodybuilders to put down their dumbbells and take up outlaw life.

boulos··on Mipmap selection in too much detail
Great writeup!

I can't tell from your GLSL if these would have forced FMAs for a lot of the intermediate product sums. That would probably be a non-trivial effect, particularly for your large anisotropy cases.

The Heckbert paper also describes the basic theory, but you would want to supplement with some of the offline rendering work that followed it. OpenImageIO (OIIO) is pretty widely used, and has gone through several iterations of bug fixing like https://github.com/AcademySoftwareFoundation/OpenImageIO/pul...

But for your purposes, you probably just need to find all the magic epsilons and sign checks to make it match.

boulos··on How to avoid P hacking
Nicely done. Since many folks may not know the original song: https://en.m.wikipedia.org/wiki/The_Gambler_(song)

(And TIL, this wasn't original to Kenny Rogers!)

boulos··on Old Soviet Venus descent craft nearing Earth reentry
"Post surrender" is misleading. https://blog.nuclearsecrecy.com/2022/05/06/did-the-japanese-... has good coverage of what's known here. There were definitely factions within Japan that wanted to surrender, but there doesn't appear to have been a formal attempt.

As is, there was an attempted coup to overthrow the government due to the first (conditional) surrender.

boulos··on Oxide’s compensation model: how is it going?
In practice, you pay out the commission over time. So if someone was extremely deceptive, or the customer churns nearly immediately, you don't just keep getting the commission.

At Oxide's scale, it would likely be hard to hide out in the way you're suggesting. It's a very technical sale, and probably any material deal is going to involve some discussion with someone beyond just a single sales rep.

Perhaps to bcantrill's point in the post though, I would suggest you reconsider the assumption that Sales people are purely coin operated. I'd love to hear from their reps directly, but as an example from my own experience, lots of the early Google Cloud sales folks were there because they hated how places like Oracle or Microsoft treated their customers. They were absolutely taking less peak pay, for what they felt was a better relationship.

Sales is fundamentally about relationships and trust in this kind of high-touch transaction. Badly behaving reps are pretty quickly outed, one way or another, though at big companies it can take a while to fire them.

boulos··on Oxide’s compensation model: how is it going?
Oh, I failed to say annual bonuses. Actual, true bonuses make more sense that way.
boulos··on New Study: Waymo is reducing serious crashes and making streets safer
Our safety research team is interested in this topic, too! In a previous study, they've tried to model it:

> Building on that, the Collision Avoidance Benchmarking paper presents a novel methodology to evaluate how well autonomous driving systems avoid crashes. The study, which to our knowledge is the first of its kind, introduces a reference model that represents an ideal human state for driving—the response time and evasive action of a human driver that is non-impaired, with eyes always on the conflict (NIEON). Put simply, unlike an average human driver, NIEON is always attentive and doesn’t get distracted or fatigued¹. The data showed that the Waymo Driver outperformed the NIEON human driver model by avoiding more collisions and mitigating serious injury risk in simulated fatal crash scenarios.

(From https://waymo.com/blog/2022/09/benchmarking-av-safety)

AIUI (I'm not on that team), a major challenge is getting good baseline data. Collision reports may not (reliably) capture that kind of data, and it's clearly subjective or often self-reported outside of cases like DUI charges.

boulos··on New Study: Waymo is reducing serious crashes and making streets safer
For those that are interested, our Safety Research team also makes the underlying data available for download:

https://waymo.com/safety/impact/#downloads

(I don't work on that team, but I've noticed a few comments that would be better served with their own analysis on top of the available data)

boulos··on Oxide’s compensation model: how is it going?
Also reasonable! A little annoying though, since I find most people are better served by uniform paychecks, but I understand the celebratory aspect.
boulos··on Oxide’s compensation model: how is it going?
As others have said, the sales compensation model is fundamentally a low base and then a percentage of the sale. I think Ben Horowitz had an early blog post about not trying to innovate on sales compensation models, and having seen a few attempts at "innovation" at Google Cloud, I agree. It's almost never worth the complexity.

You can figure out various sliding scales, maybe even caps (but adjusting the scale is more rational), but I think flat pay is basically anathema to being a salesperson.

Edit: found it, though it isn't from as long ago as I remembered. https://a16z.com/why-must-you-pay-sales-people-commissions/

boulos··on Oxide’s compensation model: how is it going?
Apparently, I'm one of the people that would have given feedback on the original proposal :).

> Some will say that we should be talking about equity, not cash compensation.

I think of compensation as total compensation. It would be fine to say this is Oxide's salary model. And I think it's a fine choice.

It sounds like the equity grants are naturally variable, though I doubt it's just newer vs older employees:

> As for how equity is determined, it really deserves its own in-depth treatment, but in short, equity compensates for risk – and in a startup, risk reduces over time: the first employee takes much more risk than the hundredth.

Edit to add: I assume there aren't cash bonuses for salaried employees. (I've always found it a bit weird anyway, but it's not mentioned explicitly, and would seem against the ethos, too)

boulos··on Robot Dexterity Still Seems Hard
(Disclosure: I work for Waymo)

While there is plenty of classical robotics code in our planner, I wouldn't want people to assume that we don't use neural networks for planning.

Just because we don't deploy end-to-end models (e.g., sensors to controls), but have separate perception and planning components doesn't mean there isn't ML in each part. Having the components separate means we can train and update each individually, test them individually, inject overrides as needed, and so on. On the flip side, it's true that because it's not learned end-to-end today that there might exist a vastly simpler or higher quality system.

So we do a lot of research in this area, like EMMA (https://waymo.com/research/emma/) but don't assume that our planning isn't heavily ML based. A lot of our progress in the last couple of years has been driven by increasing the amount of ML used for planning, especially for behavior prediction (e.g., https://waymo.com/research/wayformer/)

← PreviousPage 3 of 34Next →