boulos

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.

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)

Also reasonable! A little annoying though, since I find most people are better served by uniform paychecks, but I understand the celebratory aspect.

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/

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)

(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/)

Thanks for replying! That's a good reason. Reporting is always hard, and of course, many things appear obvious in hindsight!

The latest GPUs and TPUs support fp8. It's a big part of the efficiency gain in the latest systems. Blackwell also supports fp4.

The percentages and methodology changed over time though (this was the big argument between Benioff and Dorsey, since gross receipts ignoring business model and margins hit very differently).

Same thing the large transfer tax changes, vacancy taxes, etc. The inflation adjusted budget is definitely higher.

Most discussion of employment is for the San Francisco metro area, but I think the county level statistics at https://data.census.gov/profile/San_Francisco_County,_Califo... are actually specific to people who live in San Francisco.

I never know what business codes / job codes get included in "Professional, scientific, and management, and administrative and waste management" but there's an explicit separate one for Information that is only 6.3% (https://data.census.gov/vizwidget?g=050XX00US06075&infoSecti...)

People write about tech jobs because they're highly paid (on average) and highly visible. But it's not the case that the city is entirely dependent on technology. You wouldn't say the same thing about the equally large segment of "Finance and insurance, and real estate and rental and leasing".

> Did tech companies decide to start paying more taxes? Did the city start spending more? No, the actual cause is boring and simple, but makes anti-establishment folks very uncomfortable.

We did pass gross receipts taxes and do tax payroll for larger businesses. So yes, companies pay more taxes than 2015 (though they didn't choose to) and our budget per person in inflation adjusted terms is up:

  A $16 billion question: How did San Francisco’s budget get so huge?

https://www.sfchronicle.com/sf/article/city-budget-16-billio...

A lot of that is from "self-supporting" segments including SFO.

What a weird investigation though. Sounds like they could have solved it by asking the photographer first, which they eventually did:

> Finally, Spark contacted Murray Close, the photographer who took the picture of Jack Nicholson that was inserted into the original image.

> The photographer revealed that "there was no such thing as the Warner Brothers photo archive [and] that was a complete mistake."

> Instead, Close had sourced the original photo from the BBC Hulton Photo Library in London, now part of Getty Images.

> The photo, it turns out, was taken at a Valentine's Day dance on February 14, 1921, in the Empress Ballroom at the Royal Palace Hotel in London.

I was surprised not to see a Kunlun P800 there.

Or that SoftBank owns Arm... and has very publicly said it wants to get into the full chip business, not just reference architecture.

Just a nit, but it's not "for SXSW" (i.e., temporary). Access via the Uber app is now the mechanism by which we're offering rides in Austin (and next, Atlanta).

https://waymo.com/blog/2024/09/waymo-and-uber-expand-partner...

That's a reasonable path! You probably learned to appreciate and value the complexity, but you wouldn't have from the start. Which service do you use?

> presumably to Mexico.

Also to Belize and other parts of Central and South America. They don't all stop in Mexico.

50% vested monthly for 4 years until August is only about 6%.

The real point of this article is how busted the PG&E rate plans are. The author says:

> Do we really need 106 different rate plans in California? Why don’t utilities automatically put people on the cheapest rate plan for them, and reevaluate monthly based on usage?

And I think it makes sense to simplify, but I would bet that the old plans are required to persist for existing customers. That's how the NEM 2 folks are sticking with their previous contracts. It would be more efficient for everyone if PG&E could rationalize these various plans, but to do so, they'd also need to forcibly migrate people to other plans.

See my sibling comment :).

You're correct, there are multiple flavors of Google Cloud Locations. The "Google concrete" ones are listed at google.com/datacenters and London isn't on that list, today.

cloud.google.com/about/locations lists all the locations that GCE offers service, which is a super set of the large facilities that someone would call a "Google Datacenter". I liked to mostly refer to the distinction as Google concrete (we built the building) or not. Ultimately, even in locations that are shared colo spaces, or rented, it's still Google putting custom racks there, integrating into the network and services, etc. So from a customer perspective, you should pick the right location for you. If that happens to be in a facility where Google poured the concrete, great! If not, it's not the end of the world.

P.S., I swear the certification PDFs used to include this information (e.g., https://cloud.google.com/security/compliance/iso-27018?hl=en) but now these are all behind "Contact Sales" and some new Certification Manager page in the console.

Edit: Yes! https://cloud.google.com/docs/geography-and-regions still says:

> These data centers might be owned by Google and listed on the Google Cloud locations page, or they might be leased from third-party data center providers. For the full list of data center locations for Google Cloud, see our ISO/IEC 27001 certificate. Regardless of whether the data center is owned or leased, Google Cloud selects data centers and designs its infrastructure to provide a uniform level of performance, security, and reliability.

So someone can probably use web.archive.org to get the ISO-27001 certificate PDF from whenever the last time it was still up.

Amy Gooch's Color2Gray and various follow-on work has better coverage of the OP's actual goal:

> evaluate relative brightnesses between art assets, and improve overall game readability

The method in Color2Gray is trying to enhance salience, but the paper does a good job of comparing the problems (including red / blue examples in particular).

Like other commenters, I think oklab would look better than CIELAB on the example given in the OP. https://bottosson.github.io/posts/oklab/#comparison-with-oth... and the Munsell data below it show it to be a lot more uniform than either CIELAB or CIELUV.

The Seattle / UW related branch of MSR definitely had a few wins in the graphics department.

As https://www.microsoft.com/en-us/research/blog/meteoric-rise-... reminded me, they were on like 20% of the SIGGRAPH papers in 1996. And several have stood the test of time, like Hugues's Progressive Meshes paper (https://dl.acm.org/doi/10.1145/237170.237216). A little later, the spherical harmonics paper (https://dl.acm.org/doi/10.1145/566570.566612) was basically all MSR (I think Jan maybe interned there while at MPI? It's been a long time...).

That isn't to defend anything in AI, but it's not the case that they had no impact. The oral history thing there claims that the first grammar checking in Office 97 came from their NLP work.

I doubt their preferences are >1x since they've always had high demand. In that case, the preference stack would just be the total raised over time (~$10B).

That's wider than the streets near me in San Francisco :). Lots of our "two way" streets, are only one vehicle wide, because we allow parking on both sides of the street, too.

You're confusing us with Cruise. Lots of reporting did the same, but we're expanding a lot across California. For example, now anyone can use our service in Los Angeles:

https://waymo.com/blog/2024/11/waymo-one-open-to-all-in-los-...

That's not how I remember it. I thought proto3 was all about JSON compatibility. No?

This is neat work, but I also love the (presumably intentional?) backronym of BLT.

The Cronkite one?

https://m.youtube.com/watch?v=BsCkaAuikGY

It was 15 years ago today that it was cancelled (https://web.archive.org/web/20100106162815/http://news.cnet....).

You need to remember that was during the financial crisis, and the first chips were going to be laughably bad and uncompetitive. From the quote in that article:

> Justin Rattner (Intel Senior Fellow) demonstrated Larrabee hitting one teraflop, which is great but you could walk across the street and buy an ATI graphics board for a few hundred dollars that would do five teraflops.

Intel had no chance for v0 to do anything but lose money. They were way off track, and spent most of the time trying to get to a software rasterizer plus hardware that could work for some games.

They did pivot Larrabee into the various Knight chips (Knights Corner, Landing and whatever else) since the "it's just x86" was a pretty compelling story in the 2009 era. But ultimately, HPC wasn't that great a market in the early 2010s for accelerators yet and NVIDIA ended up totally eating their lunch over time.

I think Intel could have stuck with the HPC market and gotten lucky once ML exploded, but realistically until about 5 years ago there wasn't that much ML spending yet either. 10 years of last place discrete GPU work would have been difficult to invest in.

tl;dr: the business case for Larrabee hinged on having a successful gaming business and it wasn't even close.