Comcast is the worst for this. Actual fiber companies are selling symmetric connections. Comcast still sells garbage like 1000gbps/30mbps. That is barely enough return bandwidth for traffic overhead.
You forgot the data cap meaning while you can burst you can't exceed ~20 Mbit sustained (I'm not clear if the caps are typically bidirectional). Meanwhile the local fiber company doesn't say a thing even if I sustain upwards of 800 Mbit 24/7 for months on end.
nope. nearly went over within the last couple months and had to switch to leeching off the neighbors til it reset (which counts against neither of our caps at the cost of much lower speeds)
The Comcast plan I has from Oct 2024 to Oct 2026 had a 1.2TB data cap. It was $35 per month for something like 200Mbps down and 10Mbps up. I exceeded it a couple times and was charged $10 for 50GB.
Good grief. Comcast currently charges me $70/month for 125Mbps down. Thankfully I'm switching over to T-mobile fiber soon. (1 gig symmetric for $55/month.)
IIRC the one an acquaintance currently has is capped at approximately 6 Tbit (on which I based the earlier math). I'm not sure what the terms for exceeding it are.
Was just forced to upgrade to a 50% more expensive "modern" gigabit Comcast cable plan from my existing "legacy" gigabit plan after they retroactively added data caps (with additional charges for overage maxing out at $100, which I would max out every billing period).
The one perk is that they are actually gracing me with higher-than-advertised download bandwidth (I think like 1.2 or 1.3 gigabit), which I've never gotten out of them before. Previously it was consistently around 70% of the advertised speed because of "network conditions and capacity", yet miraculously upgrading to a higher speed would get you more bandwidth (up to 70% of the higher advertised speed!).
I think that in this case it's about the "rules for thee, but not for me" aspect. If the average U.S. citizen attempted to do what Lutnick allegedly did, the consequences would likely be unpleasant, at best.
And the average US citizen re-elected the poster child president for "rules for thee, but not for me." And IMO if he hadn't started the Iran War, imposed tariffs, and driven inflation through the roof, his approval rating would still be 47%+ like at the beginning of his second term.
That's not true at all. Political factions routinely take steps to try to discourage their opponents from voting at all. They run campaigns on opponents' media with the goal of building cynicism. Voter registration laws are inconsistently enforced to raise barriers to certain people voting. Employers (illegally, but often) discourage employees from having the time/availability to vote if they're worried about what they'll vote for. Hell, other countries interfering in elections often do so by spreading misinformation with the goal of getting people who would otherwise vote for a given agenda to just give up.
Those things wouldn't happen if "too lazy to care" was the only reason people didn't vote. A lot of people who don't vote absolutely should be counted for a given side; they don't (or can't) vote due to reasons much more externally engineered than laziness.
> they shouldn't be counted by either side as a relevant position.
Yes, they shouldn't be, but the losing side is always count them as such. Especially if they would had lost anyway.
> failure to vote is not a "silent vote";
The "Failed to vote" and "didn't even bothered to" are two different things. And the latter effectively gives the winner more weight by denying other contenders even a chance to receive a vote for them. Hence the 'silent vote'.
Perhaps, but it's highly unlikely it can outperform classical machines. If it actually had that capability, it would be front and center in the announcement.
Is this new processor capable of achieving reliable and practically useful results beyond the capabilities of classical machines? The answer is almost certainly, no.
Yes, and this was true starting with IBM Boston. I think anything before Dec 2025 could be picked apart as too noisy in particular due to crosstalk problems, for example on Pittsburgh. From the IBM Boston QPU build to the R2s, the fidelity allows for compute beyond what classical computers can do for quantum applications.
As a pure research project with possible implications in the indefinite future, this is very interesting.
The problem is that the US approach to funding research has backed itself into a corner. Every research advance in a technologically relevant industry has to be presented as though a commercial product is just around the corner, because no-one will pay for the research otherwise.
Eventually, though, the market will get tired of the endless unfulfilled projections and you'll get a "quantum winter".
in terms of progress quantum winter most likely already happened during COVID and there will no longer be any pause in progress.
between 2021-2024 technical progress was not impressive, and companies failed to build what they were trying to at the pace they wanted to. Since the end of 2024 this is no longer the case and progress is constant. To me it seems increasingly likely that by 2030 we'll have fault tolerant compute across multiple modalities including IBM's superconductors.
“Winter” in this case refers to drying up of investor interest and confidence, due to promises and expectations not being lived up to. The classic example is “AI winter”, of which there were two major ones. The result of that is lack of funding and a dramatic reduction in research.
It's not always trivial to saturate available memory bandwidth. You may need to use more than one core, or you need to carefully consider the topology of the chip you're working with (see lstopo). For example, what if the L3 cache is actually split in two [0]? In such cases, adding more bandwidth doesn't do much for you.
This is something AI is really good at. Plus, you don't need AI for this. In my time at I made a "runtime compiler" that did this for parallelizable calculations (applying kernels on images, this was a long time ago). It didn't even need information on the machine it was running on, it just spent 5-6 seconds during startup trying some 50 different memory access patterns, figuring out the size of the cache from when it slows down, ... and then you had a simple language you could specify operations to execute on large images and it would do them. Fast. Very fast.
And yes, this was only worth doing for things that were memory bound. But anything relating to images is, pretty much.
I wrote something similar for regexes too. Get all matches from a massive file, or check massive numbers of lines of text and figure out which ones match a given regex. That too is still running. Even the AI models are now using it, although I'm told they at least also improved that one. I wonder how, actually.
A team took over the image compiler and it's still running, I'm told. I got 12k'd, of course. At this point there's literally ONE person left in the team that created the core tools, and he's the one known for going into management first chance he got.
I should write an open source version of this focusing on ML. I wonder how hard that actually is ...
> It didn't even need information on the machine it was running on, it just spent 5-6 seconds during startup trying some 50 different memory access patterns, figuring out the size of the cache from when it slows down, ... and then you had a simple language you could specify operations to execute on large images and it would do them. Fast. Very fast.
This is pretty cool. It sounds very similar to how FFTW works. You create a plan, and you give it as much time as you're willing to spend on planning. It will then use that time to find the fastest set of underlying kernels for your hardware and input size.
It's also used to significantly increase the efficiency of nuclear weapons, without the need for developing an outright thermonuclear/hydrogen bomb [0].
> you need to spend more than one defensive missile per attacking missile, because you might miss
It's almost always the case that you'll expend more than one interceptor per incoming missile as interception is probabilistic. In general, you can estimate the number of missiles you'll need to achieve a certain probability of interception with the following formula: 1 - (1.0 - sspk)^n, where sspk is the single shot probability of kill (i.e., successful interception with a single interceptor) and n is the number of interceptors.
reply