Sign in

Roland Dreier

@rbd.bsky.social
308 followers 90 following 2.2K posts

I put the R in RDMA. All posts based only on information from within my past light cone.

PostsRepliesMedia
Roland Dreier @rbd.bsky.social · 5h
I know they say the Fed funds rate acts with “long and variable lags” but I wonder if this could be at least part of the story?
000
Roland Dreier @rbd.bsky.social · 5h
The fact that the trend started the same month ChatGPT was first released makes me think it was happening too early to be caused by AI
100
Roland Dreier @rbd.bsky.social · 5h
On the one hand it's a screenshot of an engagement farming Threads post, on the other hand it says ChatGPT did something bad, so I guess I'll uncritically believe it
150
Roland Dreier @rbd.bsky.social · 8h
Well, fortunately a no vote on RTM just disciplines BART and leaves Caltrain and the county transit agencies unaffected
100
Roland Dreier @rbd.bsky.social · 8h
”new cars are too expensive!” “don't you dare do anything to make them cheaper!”
010
Roland Dreier @rbd.bsky.social · 15h
I work for Rivian, but yeah the R2 is NICE And if you can wait till next year for the standard trim, it's under 50 grand
021
Roland Dreier @rbd.bsky.social · 16h
idk whose car this is but I'm always curious (parks at the same house as a Tesla with the plate “MOAR GPU”)
010
Roland Dreier @rbd.bsky.social · 16h
I think until recently, being happy to be a good coder and stay at Senior SWE/L5 was a comfortable career, but now you're cooked if you don't do some tech lead/architect/Staff/L6+ stuff. You might still be cooked in a couple years, but you're definitely already cooked if you only crank out code.
010
Roland Dreier @rbd.bsky.social · 06/10/2026
Yes, look at the ADS data on www.nhtsa.gov/laws-regulat... Tesla has 26 crashes reported total and to the best of my knowledge none of them involved serious injuries or fatalities.
nhtsa.gov
010
Roland Dreier @rbd.bsky.social · 05/10/2026
Tesla FSD is (despite the name) just ADAS. I think AV should only refer to vehicles that are claimed to be operating autonomously, and to the best of my knowledge Tesla robotaxis have never been involved in a fatal crash.
100
Roland Dreier @rbd.bsky.social · 05/10/2026
That seems to be more of a Tesla influencer marketing strategy, tons of “watch FSD save my life” videos out there. Waymo seems to prefer a marketing message of “Waymo vehicles are observed to have fewer collisions at a rate that rigorous analysis of this data set shows is statistically significant”
030
Roland Dreier @rbd.bsky.social · 05/10/2026
Why don't, like, the Gemini and ChatGPT apps count?
030
Roland Dreier @rbd.bsky.social · 05/10/2026
I think this is a reasonable frame but Vision Zero would argue for taking a broader systems view: human drivers have limited capabilities, but we can design transportation systems to mitigate the consequences of their errors so they're not fatal (and autonomous vehicles can be part of that)
120
Roland Dreier @rbd.bsky.social · 05/10/2026
Funniest outcome is if both Prop 40 and 42 pass, but 40 gets a few more ‘yes’ votes so it takes effect
030
Roland Dreier @rbd.bsky.social · 05/10/2026
Sure but is “died *while* colliding with a Waymo not *from* colliding with a Waymo” a rhetorical direction we want to go?
320
Roland Dreier @rbd.bsky.social · 05/10/2026
To the best of my knowledge a Tesla AV has never been involved in a fatal crash? I understand the critique of their ADAS system but it's a little deceptive to lump those crashes into the same bucket & isn't Waymo saying “we're X% safer than human drivers” all the time pretty explicitly utilitarian?
110
Roland Dreier @rbd.bsky.social · 05/10/2026
For example www.azfamily.com/2025/09/14/1...
azfamily.com
Motorcyclist run over after rear-ending Waymo near ASU Tempe campus
One person has died after a multi-car crash, one of which being a self-driving vehicle, near Arizona State University’s Tempe campus early Sunday morning.
010
Roland Dreier @rbd.bsky.social · 05/10/2026
Pretty variable but the common thread is that a human driver did something bad in the vicinity of a Waymo vehicle
220
Roland Dreier @rbd.bsky.social · 05/10/2026
Waymo does a lot of research that pretty unanimously finds their vehicles are far safer than human drivers: waymo.com/safety/resea... It's just, even with a rigorous safety culture and commitment to Vision Zero, if you're building cars, you do kinda have to accept a lot of shit will happen on roads
waymo.com
Waymo Research: Published Safety Research Papers for Autonomous Vehicles
Safety drives us. Learn more about the research we've conducted in the safety field for autonomous vehicles.
090
Roland Dreier @rbd.bsky.social · 05/10/2026
Waymo vehicles have been involved in at least 3 fatal crashes since 2025
320
Roland Dreier @rbd.bsky.social · 04/10/2026
A lot? If you were designing it from scratch now there are probably things you'd do a little better, but USB Type-C, USB-PD and USB4 kinda just work for a huge range of things from the cheapest throwaway gadgets up to 80 GBps workstation links? Hard to see a reason big enough to switch away…
021
Roland Dreier @rbd.bsky.social · 04/10/2026
Musk coulda just bought Intel for like $250B a year ago
010
Roland Dreier @rbd.bsky.social · 03/10/2026
IDK, is this cognitive surrender? Anyway wild that I can pull this shit and it just works in 30 seconds (Opus 5.5 high)
Claude Code screenshot:
 
> for the PCIe firmware version check - is there anything we can deduce from the history of recipes-bsp/pcie-phy-fw/ meta-mammoth that would let us see where to check the version? A very dumb grep didn't find anything matching 012419 (the putative version) but maybe we could be a little smarter?

<Ran 5 shell commands>

• Found it-the PCIe firmware stores the version as one byte per component, which is why grep missed it. Now checking whether the BSP image packs the two blobs top-first or lane-first.

<Ran 5 shell commands>

• Yes, the recipe history pins it down. Its files include two firmware releases, 1.24.18.0 and 1.24.19.1. Diffing .chex files shows only 2 changed words in the top image and 3 in the lane image, and all of them are the version build date.

Encoding: one byte per version component, which is why grepping for 012419 missed it. Ox01181301 is 1.24.19.1 24, Ox13 = 19). The next word is a BCD build date, Ox20240923.

their

and th

(0x18 =
100
Roland Dreier @rbd.bsky.social · 03/10/2026
Fair but an analogy I like is Chernoff faces - we have a lot of brainpower that's powerful & can be useful in thinking about models if we let ourselves use it, even if the models are simulacra
100
Roland Dreier @rbd.bsky.social · 03/10/2026
So the challenge is that, as you say, anthropomorphizing does help us predict model behavior, a lot, right up until it doesn't
000
Roland Dreier @rbd.bsky.social · 03/10/2026
Kinda surprised to see @leifw.bsky.social making this argument because while I agree models are computer programs and fundamentally not human - I also agree with him that LLMs are cultural technology that comes from human language and behavior.
210
Roland Dreier @rbd.bsky.social · 03/10/2026
Intel Optane arguably suffered from the NVMe version being slowed down by PCIe/block overhead so not decisively better than SLC SSDs & the DDR version not being quite fast enough to be worth burning a whole (scarce, expensive) CPU memory channel. But on CXL.mem at 50+GB/sec w/ load/store interface…
010
Roland Dreier @rbd.bsky.social · 03/10/2026
Why doesn't Micron (now a $1T company) bring back 3DXP? It was arguably ahead of its time but now the ecosystem is ready for it with CXL, and demand would be insane from AI hyperscalers en.wikipedia.org/wiki/3D_XPoint
en.wikipedia.org
3D XPoint - Wikipedia
110
Roland Dreier @rbd.bsky.social · 03/10/2026
Kinda the point of the frontier models now is that you don't have to baby-step them any more. But the scope of work they can complete independently or broad architectural context they can understand is still well below a human senior SWE.
020
Roland Dreier @rbd.bsky.social · 03/10/2026
I'm using the word ”value” quite intentionally. We can agree to disagree but not even for sw engineering, for better or for worse, there's still a lot of economic utility to capture, when a senior engineer costs companies $25K+/mo and almost nowhere will spend more than $1K/mo/eng on agentic coding
010
Roland Dreier @rbd.bsky.social · 03/10/2026
That's a somewhat different argument - that you don't need/want more capability, which is fair enough. For me it's clear the additional capability of Fable has value and there's still plenty of room for models to get better and offer even more value.
230
Roland Dreier @rbd.bsky.social · 03/10/2026
Yes, I use Claude at work for sw eng; my employer pays for it up to a monthly budget. And I do use Fable sometimes because it's really f'ing good at figuring out hard bugs or completing big-ish tasks. (Although the last few weeks since Opus 5.5 launched, I find it seems roughly as good as Fable 5.1)
130
Roland Dreier @rbd.bsky.social · 03/10/2026
I mean just draw a line from Opus 4.5 in November 2025 to Fable 5 in June 2026, I don't see how you can say they're running out of juice because there hasn't been another huge jump for a few months
210
Roland Dreier @rbd.bsky.social · 03/10/2026
Opus 4.5 launched in November 2025! Saying the models are “sorta asymptoted” is like saying “eh, Apple hasn't been able to improve the iPhone for a long time” in May 2008.
180
Roland Dreier @rbd.bsky.social · 02/10/2026
”going to take a few weeks and then see what's out there, but if you have an opportunity, please reach out! excited to see what my next chapter holds!”
000
Roland Dreier @rbd.bsky.social · 02/10/2026
Facts
000
Roland Dreier @rbd.bsky.social · 02/10/2026
Remember correlation ≠ causation There's not really any plausible way that Trump becoming less popular could drive up fuel prices given the scale and liquidity of the global hydrocarbon market
000
Roland Dreier @rbd.bsky.social · 02/10/2026
I know you're not serious but you could use the standard.site integration to get a “subscribe” button on your posts and then whip up a bot that listens for site.standard.subscription records and DMs users a link to the coffee page or something 😅
010
Roland Dreier @rbd.bsky.social · 02/10/2026
idk about easy, Newsom is doing all this shit to probably get zero traction and drop out of the 2028 primary early?
020
Roland Dreier @rbd.bsky.social · 02/10/2026
👏🏼STOP👏🏼MAKING👏🏼ART!👏🏼
010
Roland Dreier @rbd.bsky.social · 01/10/2026
It's Newsom, so in a sense yes, he's nearly always doing politics as a shitposter www.theverge.com/policy/79849...
theverge.com
American politics has devolved into shitposting and aura farming
The Frog is owning the president.
140
Roland Dreier @rbd.bsky.social · 01/10/2026
People did that repeatedly in the thread! but Bender is still acting as if it's an open question
130
Roland Dreier @rbd.bsky.social · 01/10/2026
Good advice for senior developers too ... nearly everyone really
0290
Roland Dreier @rbd.bsky.social · 30/09/2026
Huh ok then! blog.google/innovation-a...
000
Roland Dreier @rbd.bsky.social · 30/09/2026
Oh no 😳 But thank you
080
Roland Dreier @rbd.bsky.social · 30/09/2026
Yes I think remote work since the pandemic is a big part of it and that's definitely a mixed bag. But I think “there's nowhere to go” is an effect rather than a cause of people not going out as much.
000
Roland Dreier @rbd.bsky.social · 30/09/2026
I guess I view “cocktails are too expensive” less as an explanation for why people socialize in person less and more as an excuse
270
Roland Dreier @rbd.bsky.social · 30/09/2026
yeah for sure I don't get what you mean by “there's nowhere to go that isn't expensive” if you acknowledge the many places to go that are cheap or free?
1120
Roland Dreier @rbd.bsky.social · 30/09/2026
Yes totally agree, the only thing I'm pointing out is that I can ride my bike for 2 hours and spend 2 hours in a cafe and still be home way more than if I spend 9 hours commuting/working in an office
020
Roland Dreier @rbd.bsky.social · 30/09/2026
you can just open your door and start walking for free (I'm assuming you already have clothes and shoes) I'm being flippant but also serious.
1150