Unrelated to the accomplishment or proof itself, but it's interesting how much of the prompt, even in this latest-and-greatest model, is spent essentially telling the model to actually solve the problem. Things like "Reject status reports, vague optimism, and claims that an unproved global compatibility statement is 'routine'."
Also a lot prompt spent feeding it strategies, which feel like they should/will eventually be deduced by the model itself, not explicitly stated. That's not to take away from the outcome in any way; rather, it feels sort of like when you would prompt GPT 4, "think through your answer step by step," as a sort of proto-chain of thought.
>> approximately 700,000 A100e GPU hours of black-box automated red teaming
Amusing that they use A100e as the reference point to sound impressive. Different ways you could make that conversion, but based on FP4 FLOPs (yes it's disadvantageous to A100, that's the point), that's something like 200hr on a GB300 NVL72 rack.
Not nothing either, but far less astounding sounding than 700k hrs.
Logic gates and memory bits have very different fabrication processes (mostly because DRAM is optimized for a high density of big capacitors doing the storage).
You can put some memory on the logic wafer (SRAM) but it's area inefficient, which is wasteful on your expensive N2 wafer. So a dedicated DRAM process is vastly cheaper per bit, even at current elevated prices.
The water is a clever impedance matching trick. The contrast in density between air and human flesh is high, so the waves all reflect off the surface rather than penetrating and reflecting off the internal structures we care about.
That's why normally you're concerned with really good transducer contact (squeezing out any air) or use a gel to match impedance.
I'm a bit rusty on CT, but I'd guess the resolution is proportional to the total number of transducers in the array (e.g. larger sensing surface equals tighter resolution) since you're basically taking a Fourier transform of the incident wave.
This is really interesting! And perhaps surprisingly doesn't trigger any immediate major technical red flags (as someone who has worked with MRI and phased array beamforming), as many HN HW articles do.
My only criticism from the tech video would be that they spend some time lauding the nanometer deflection sensitivity, which might lead some to believe that's indicative of the image resolution. It's not, and it's somewhat of a distraction -- that's just giving us amplitude information, which is comparatively less important than correlated time/phase across the 100k sensors. They do later on state ~mm resolution, which is still great!
Doppler and motion blur may be an issue (e.g. heart beating), as one slice requires a full ring of sequential exposures. But still way faster than MRI, so probably fine.
On a lighter note, it could seriously change the meaning of get FUCT (Full body Ultrasound Computational Tomography)!
To be clear, fuel cells are considered "low air pollution" because they eliminate certain nasty combustion products (NOx), but they still produce as much CO2 per kWh as a gas turbine.
Arguably that CO2 stream is concentrated and a candidate for capture/sequestration, but no one is doing that in practice.
I think the right comparison would be a vertically integrated Neocloud like Oracle (insofar as they own some/most of their own datacenters, unlike a CoreWeave).
Given the $10k price tag for tokens and high rate of bugs (several per minute) they mention, it'd be very interesting to see this experiment run with cheaper models too.
I wonder if we get to a world where a full repo sweep like this is a default Github action after commit.
The wild thing to me, is that they're serving $47B run rate worth of requests on maybe 2-3 GW of compute currently [1], of which only a fraction goes to inference, vs R&D and training. Obviously there have been complaints on token limits and such so they're stretched a bit thin, but nonetheless.
Hard to imagine what a world with 100GW of compute looks like.
Lately I've been thinking that UI really needs to include the equivalent of a screenshare meeting. Ideally you could click through an example of a software flow Claude's never seen before, with a few quick notes, and have it reliably work.
These narrow integrations with specific software suites seems like a dead end.
I had a similar, really great prof, who would always ask for what the next variable would be, so we'd end up with trees and smiley faces. His point was to not make assumptions (c is always a constant etc), but it made the classes more engaging too.
And, somehow every example ended along the lines of "then you hand this to your boss, kick up your feet and have a nice glass of scotch."
I think the water is difficult to traverse, in that it slows you down when 'swimming'.
It's really interesting how it still feels grounded even though you can fly all around. Having the cursor disappear underneath bridges and behind buildings really helps the illusion.
Also a lot prompt spent feeding it strategies, which feel like they should/will eventually be deduced by the model itself, not explicitly stated. That's not to take away from the outcome in any way; rather, it feels sort of like when you would prompt GPT 4, "think through your answer step by step," as a sort of proto-chain of thought.