> people come together to help each other instead of threatening each other with guns
People, collectively, come together. There's always outliers -- we call them assholes. It's nice that some folks are so clear on what side of line they intend to stand ahead of time -- pay attention to what your neighbors are saying, so you know which ones are the assholes.
I think in this example the decision is "route to team X or team Y"; so the worst-case is it ends up in the wrong teams queue, and gets moved manually.
Sure, but Bakersfield is a city almost exactly the size of Denver (150 mi^2), and uses much less power. Just because it’s enough power to power /a/ city the size of Denver doesn’t mean it’s enough to power Denver.
> I wouldn't be surprised if Apple voided the hardware warranty, claiming it overheated or something and Linux was to blame.
Well, you made it half way through the comment before running out of real points and deciding to make up imaginary ones to support your point. That's not nothing!
I have some 3.5" floppy disks in the bench seat of the electric organ that came with my house that are at least that old... the three I tried worked without issue. Small sample size, but it suggests a half-life way over a couple of years; I can't imagine this system has unusually robust error correction or anything.
> Which is why many systems on most airliners are redundant in threes
Not really. Triplex redundancy is based on a failure model where failures can produce arbitrary output. Three systems allows majority voting (often, median select) to provide correct output in the presence of a single arbitrary failure. It allows detection of double-fail-silent failures and uncorrelated double failures, but that's not really the point, since continued operation under these conditions is at risk, and failing nonoperational isn't often acceptable for these systems. (You'll also see quadruplex systems that use the same voting mechanism to detect and disable failed channels, allowing handling two arbitrary non-simultaneous failures, for systems that need a longer window of continued operation after failure.)
If you change the failure modes, you change the architecture. These days you often see dual-dual systems where each pair of systems cross-checks and disables itself on mismatch, converting a single arbitrary failure to a single silent failure; which then allows downstream systems to accept inputs from two such fail-silent systems and trust either that is providing output, removing the need for a median select.
Very interesting, I’d love to learn more. Is there anything you’d recommend in terms of a general introduction to different kinds of fault tolerance? Literature, video material, could be whatever.
Unfortunately this has been on-the-job learning for me over the last few decades so I don't have great sources to point you to, but this is covered in systems engineering courses for aerospace and automotive applications.
Same thing with control-theory - there’s so much established knowledge out there that one really needs to know but which isn’t taught to kids in uni or even postgrad level. Learning these things on-the-job means you don’t know what you don’t know…
What’s a better way? We still don’t have state-regulation/certification for software in safety critical systems, which terrifies me when I think about it.
We have DO-178C for aviation, which (a) is a pretty good standard and (b) is a heck of a lot better than ISO26262. It's "voluntary", but it's the only game in town, so…
> the “presence of an AI Overview now correlates with a 58% lower average clickthrough rate for the top-ranking page.”
I think most of the article points out real problems, but this seems a bit synthetic. From what I can tell, the "AI Overview" usually appears in the top spot, but some portion of the time (A/B testing? predicted quality based?) it appears in the second spot, after the top-ranking page. Assuming a significant number of users (myself included) instinctively scroll past the slop to see the real results, this will lead to a bias towards increased clickthrough for the second-ranking page, at the cost of the top-ranking page, because the top-ranking page is now put where they're training users not to look.
This seems like something they could turn off if it really is impacting their bottom line. What’s much harder to deal with is if we all switch to ChatGPT because the experience is so much better.
> bf16 can only represent changes to weights if the change is at least ~1/512 of the weight itself, so at small learning rates only the smallest weights can update.
Isn't the standard numeric fix for this to probabilistically update the weight? That is, if the desired change is current_value * 1.0014, it should have a 70% chance of going to the successor, and a 30% chance of staying unchanged. Over a number of steps, this will adjust it appropriately over time.
Yep, Stochastic Rounding [1]. I actually had "round-to-nearest" in the title of the piece originally. Usually people do that if they don't have the space for fp32 master weights, but it doesn't fix the second moment part of the learning process.
People, collectively, come together. There's always outliers -- we call them assholes. It's nice that some folks are so clear on what side of line they intend to stand ahead of time -- pay attention to what your neighbors are saying, so you know which ones are the assholes.
reply