leecarraher

I think the exercise of saying let's legitimately try to build something vs buy, is a great place to start. It will help determine what you need in an off the shelf solution, while also serving as a useful pricing target. If the buy solution is more than maintenance and development, then it makes sense to build yourself and avoid vendor lock-in and allow for bespoke flexibility. I've experienced both situations, some rules of thumb to go by:

1. is this service or product similar to what your company creates, or would you have to create an entirely new business unit to develop and support it

2. how complex would it be to integrate the off the shelf solution into your current system. if it's the same as building on your own, then why pay for it

3. is there competition in the market for the buy solution so pricing and innovation are competitive.

4. who is this company selling this product or service, what is their track record, do you service as just their financier and guinea pig testers

Recently watched this on Hamiltonian Monte Carlo, and it was a fantastic primer on the basic concepts and motivations of mcmc, and later hmc https://youtu.be/pHsuIaPbNbY

commercial air travel is all about critical maintenance. Add it to the checklist and move on. Feels like the article pandering to that y2k nostalgia.

my guess is they will buy carbon offsets to cover manufacturing energy and transportation costs, a system which has its own shortcomings and critiques.

I recall reading a similar thing, taxes aren't fun, but you could always sell the car, and pocket the extra. It's just a matter of winning something by chance and taking issue as though it was something deserved.

I bet the person in Oprah's audience who doesn't drive when she gave everyone cars was super justified in being angry that Oprah just assumed she'd want a car with inherent resale value.

is the contract public? i would like to see what contingency is in place if Boom fails to deliver. Or more to the point, is this more PR to make AA appear to be forward looking, and bolster Boom's reputation with a deal that will never come to fruition. It's a win-win for investors

computer science imo somewhat gets the highest level of scrutiny when it comes to reproducibility simply because so many people can, in theory, reproduce an experiment without any special expertise. this cannot be said for many other sciences. Others require costly equipment, expertise and time, and even then, may fail to correctly reproduce an experiment. Imagine trying to reproduce a CERN finding in your backyard, or a James Webb observation, a social science survey, or medical study. DL isn't bullsh-T it's one of the few democratized enough sciences (similar to math) that has a reasonably low barrier to entry. unfortunately that may be its undoing, because it means the masses want the off the shelf docker container that just does the toy thing the experimenters spent potentially years working out. Could documentation be better? of course, but before declaring a field BS, maybe compare it to some other scientific fields.

I wish they had data on the estimated cost to build and operate. Best i can guess is that $11M might get them 200Megawatt/hours -> $0.055 per watt hour which would be cheaper than any chemical battery based energy storage, albeit at the cost of reduced efficiency (75% vs 95% for lithium-ion). This is similar to pump storage, but less geographically dependent.

ultra low profile sedans and rotary engines are the most evolved cars, so why do we have trucks? or more on point, road bikes are more efficient than mountain bikes.

The monolithic view that a bike is about putting energy in and getting forward motion on a flat road out, misses the multirole requirement of a bike. Just as saying humans are unevolved because we aren't the fastest animals in the animal kingdom.

Bikes, like humans have many challenges other than just getting from point a to point b. sometimes we have to stop for short periods following traffic patterns. I suspect rolling onto the pavement each time will get old fast. More on that surface, bumps and jars from imperfect road conditions being solely absorbed by my laying down bits seems pretty awful. Being able to see those upcoming obstacles and potentially avoiding the more significant ones, is likely important.

In general these are referred to as embedding techniques PCA/SVD is still common along with other linear methods like ICA, CCA, PLS t-sne is pretty popular statistically inspired with some manifold learning non-negative matrix factorization is useful recommender systems topologically inspired techniques are gaining traction, like UMAP, LLE the author notes neural embeddings, like autoencoders.

for synthetic data generation, what methods are they using to sample data from the distribution? What assumptions about the distribution are being made? Does it model correlations between sample attributes that could adversely effect some ML methods (multi-colinearity can cause problems).

and then i thought, well look who is telling us that

In my opinion it's a great deal more complex than: "you can just turn off the features". To me "lack-of-a-feature" is not in itself a feature of the device or platform, rather a feature of the company who makes the device or platform. It signals that the manufacturer is fully committed to solving a specific use case. Avoiding feature creep is tough though. Companies need to make money, and the easiest way to do that is to exploit your current customers to upgrade. However, because no one upgrades a device without a reason, it may be inevitable that companies continue to follow the classic: phone -> phablet -> tablet -> basically laptop without a keyboard progression.

i agree, docker added version control, online repository, with management, and search they also fixed some of the networking headaches in lxc. Prior to developing containerd docker was based on lxc containers.

i've been saying this for some time in the US and think it is similar to the reasons we cannot build high speed rail here. Too many points of opposition for a not immediate and obvious benefit (namely lower energy prices, and fewer negative environmental effects). Individual and municipal solar with expensive batteries seems to just fit better with our individualistic and consumer mindsets.

what about 3d building like structures that you fly over or into. mainly for amusement parks, zoos, a combination of both, and oil speculation supercomputers

was kind of hoping for a more analytical take, comparing their efficiency to the newish slate of battery powered personal heaters, or older iron-oxide single use heaters. My guess is the newer electronic devices allow for more consistent heat, but at a much higher entry cost, and less flexible charging. I wonder if an mechanically adjustable thermal barrier could be employed in a hot water device, that could provide a more consistent heating experience.

some python to do this trick and find counterexamples that won't work

    import random as r
    def fnc(b):
        bstr = list(str(b))
        if bstr.count(bstr[0]) == len(bstr): 
            return "can't shuffle"
        max_ct = len(bstr)*10
         
        # do while
        r.shuffle(bstr)
        b_shuf = int(''.join(bstr))
        diff = abs(b-b_shuf)
        bstr = list(str(diff))
        i = r.randint(0,len(bstr)-1)

        while bstr[i] == '0' or bstr[i] == '9':
            i = r.randint(0,len(bstr)-1)
            bstr = list(str(b))
            r.shuffle(bstr)
            b_shuf = int(''.join(bstr))
            diff = abs(b-b_shuf)
            bstr = list(str(diff))
            max_ct -=1
            if max_ct == 0: 
                return "bollocks!"
        
        del(bstr[i])
        s = sum(map(int,bstr))
        i =0
        while (i+s) % 9 != 0:
            i+=1
        return bstr,i,diff

yeah for first 10000 digits there is about a 2%-5% chance (around 20%-5% for first 2k, then falls below 2% for remaining 2/3s) of selecting one that doesn't work. obviously picking something like 1111, 2222, ... won't work either since you can't rearrange it to anything but 0s.

For reference this recent article describes research that tries to nail down the exact computational ratio https://www.quantamagazine.org/how-computationally-complex-i... spoiler : 5-8 layers of artificial neurons per complexity of biological neuron

Is the original poster sure the extra size wasn't to accommodate a 2.5" or 3.5" for testing purposes. There doesn't appear to be a physical HDD. In the 2nd prototype picture (not much for scale) it looks like a board with control logic, and maybe the decoder IC, and MCU, then a connection to a battery and lcd screen with it's controller circuit. I don't see an HDD, and solid state drives with capacity corresponding to the original ipod weren't on the market yet.

my guess would be that it will allows other companies that offer competing services to compete against facebook's comparable service while not having to compete against all of facebook's services. Then the market can decide which features it prefers in each service, presumably not the ones listed in the op.

They also have an embedded link to this paper(https://arxiv.org/abs/1810.02054 ) regarding the provable convergence of neural networks by non-convex gradient descent optimization. You are correct, the researchers do no propose an alternative to sgd based training (or at least according to the article). It did seem like it was going toward an alternative training method, perhaps SMO on kernel SVM, but then it pivots to describing the theoretical result for infinte kernel svm from the linked article.

Correct, but I think the general escalation of security in airports which caught Richard Reid, was a direct response to 9/11 . Either Way, it is still mainly security theater to make passengers less apprehensive about taking what is still the safest mode of travel per passenger mile( there might be other metrics, but this is the only one I've read recently).

For some reason they run all their tests single threaded. Seems like parallelism is where all computing hardware is inevitably going. I also wish they had run time comparisons to more recent matrix sketching and multiplication method such as frequent directions, newer FJLT implementations, and RIP matrices based on hashing.

yup, my thoughts exactly. otherwise a 1000x increase in 30%-ish efficient pv silicon-based solar cells would be far above unity

I've always interpreted it as running code directly on the processor with no OS. I made a "baremetal" bitcoin miner for the original raspberrypi, until it burned up the chip(and i learned why OSes are important). I guess it's been updated to mean an application running on the native OS.

I think we will soon see a dip in productivity.whenever assessing trends, you should never base conclusions on outlier events and we are basing too much of the wfh success on this past year. A year when you couldn't go out and do things, when people feared for their jobs amid a global crisis and felt a call of duty to keep on working.

true, was just going off the debian package architecture naming PPC/ppc64el