HNHacker News
TopNewBestAskShowJobs

maldeh

519 karma · joined September 6, 2013

submissionscomments
maldeh··on Nvidia launches colossal HGX-2 cloud server to power HPC and AI
Impressive. I've lately been wondering though, how does one load data fast enough into such rigs to come close to saturating this kind of compute power? I felt this was an issue even with TPUs and V100s for any application that isn't thoroughly bottlenecked by compute (ex:- one could spend hours loading up a petabyte, and process it all in tens of minutes)
maldeh··on Amazon Pushes Facial Recognition to Police, Critics See Surveillance Risk
Last I heard, it got mired in legal issues, and was recently tossed out by the courts [0]. (Something that's not quite as likely in the US due to the EU being more privacy-conscious, and would require mass outrage and an act of congress to deflect)

[0] https://www.theguardian.com/uk-news/2018/jan/30/uk-mass-digi...

maldeh··on Amazon Pushes Facial Recognition to Police, Critics See Surveillance Risk
It's possible that in the short term, Amazon could be changing the economics of mass surveillance, making it more feasible for police departments that could not afford such tech before. But at this point, I think a change of dynamic is important, and that we actually _need_ an 800 pound gorilla to frighten us, before people wake up and start demanding action at the state and federal level.

It's already been half a decade since the UK surveillance regime and the reality of mass surveillance was opened to us, yet nobody batted an eyelid in the US. Meanwhile, facial recognition has improved by an order of magnitude and has already silently been seeing deployment everywhere around us. I'm being led to believe that unless we have a public poster boy to rally against, the situation isn't going to change.

maldeh··on Text Embedding Models Contain Bias
Hit the nail on the nose.

Sure, you can build a discriminating classifier or generative model that is the most accurate, correctly identifying / emulating the reality of our world down to 5 nines; and nobody says you shouldn't be able to identify or quantify all of these "inconvenient" associations. The trouble is always when you intend to make decisions based on your framework -- the decision to offer somebody a loan, the choice of language your chatbot uses, etc. -- and that is where fairness ought to be paramount.

And yes, if you manage to build a discriminating model that sneaks in protected classes through indirect causal effects with no attempt to suppress them, it would yield your insurance agency higher returns over time, and just because that's the way the world is, currently. But all this would achieve is perpetuating the current status quo, placing short-term gain ahead of long-term equality. You might be doing right by yourself, but that still makes you morally impugnable.

One shouldn't need to wait for an overreaching law prohibiting such indirection, to do the right thing.

maldeh··on Nothing Is Cheaper Than Proof of Work (2015)
In addition to dealing with a vast surplus of used cards on any secondary markets if the crash gets bad enough.

It would be rather convenient for them if older miner GPUs were quickly made 'incompatible' with the needs of the consumer gamer / PC market.

maldeh··on Ethics in Machine Learning: Interview with Dr. Hanie Sedghi, Google Brain
Additionally, this extends to the usage of latent or inferred variables that behave as a proxy to a protected class - so neural networks and other discriminative approaches outside of a distinctly diagnosable decision tree, could be in a rather grey area legally if used for credit scores, etc.

It's possible that the upshot of a legal challenge could be a crippling cap on the accuracy of models as they explicitly avoid any information derived from any protected factors.

maldeh··on One Bitcoin miner is buying 20,000 16nm wafers from TSMC per month
If $y is the useful PoW, would that not make $y the scientific reward while $z is the (monetary) token reward by way of mining?

Edit: silly question on my part. y and z should be trivially interchangeable. Your argument appears to rest on the idea that the monetizeable scientific worth of the work would dominate the worth of the crypto token itself, since there would be limited value in an already-mined result beyond standard blockchain speculation - is that the case?

maldeh··on Why didn't Denmark sell Greenland? (2011)
This comment leads me towards a question that I think is ignored in most of the article and the rest of this thread - the Danes' sovereignty notwithstanding, what was the gain / loss to Greenland as an entity as a result of this deal (or lack thereof)?

The island has been beholden to external parties and interests for centuries, seldom actually enjoying more than a modicum of self-determination. I see the autonomy granted them in 1979 (and further powers in 2009) by Denmark as being uncharacteristically benevolent for international standards, and something that would never be realised in a million years under the USA (particularly with their military presence).

I'd understand if Greenlanders never felt much kinship for the Danes, but I'd be hard-pressed to accept they'd feel any more for Americans. And I can't find any source referencing Greenland's sentiment towards this offer of acceding control to a global superpower they'd likely never enjoy independence from. Do we know if Denmark just did them a favour out of altruism here?

maldeh··on Denmark's education minister will let high schools check students computers
Handy google translate link:

https://translate.google.com/translate?hl=en&sl=da&tl=en&u=h...

maldeh··on Falling through the KRACKs
Unless I'm misinterpreting the meaning of "usually" or "important", I find this comment a little bewildering in the context of the article.
maldeh··on A Decade of Dynamo
> Anything that can go to dynamo, can go to s3, especially at that volume.

Please think very carefully before architecting your app with S3 as a makeshift-database. S3 would be a valid option if you don't care about millisecond latency; don't require safe updates; never expect your application to scale past 100 requests per second; and don't have multiple query patterns for the same data (unless you're okay with several redundant copies of the same dataset). Consider just about any other database solution if this does not hold true.

maldeh··on Google signs agreement with HTC
The notion of such a collaboration sounds interesting. For most industries, it would healthier for consumers in the long run if the major stakeholders competed, rather than collaborated, at this early stage. But in the short term, it means framework fragmentation, publishing pains for indie app developers as the big guns vie for exclusive contracts. Not to mention potentially delaying technological progress in the space, what with all the R&D occurring within isolated silos keeping their trade secrets close to the vest.
maldeh··on Photographer wins 'monkey selfie' legal fight
I think you might be right from a technical, mechanical sense, but I think most of the objections from posters here are because of an impedance mismatch between the word of copyright law and the spirit of what it's intended to achieve.

The macaques were simply fooling around with a plaything and had no comprehension of what a selfie is; there is no artistic intent on their part. On the other hand, the photographer by his account appeared to go into this project with the specific intent of obtaining a monkey selfie. The concept, the subjects and their moods, lighting, shot settings, etc. were all his doing, and by any reasonable yardstick his work should be recognized as a form of artistry, just as how it is not exclusively the cinematographer, but the Director as well who is entitled to copyright on a film.

And yet we have people arguing in court that by the legal specifics of who owns copyright on photographs, this is purely the monkey's "work" (and winning).

maldeh··on Elon Musk Says AI Is the ‘Greatest Risk We Face as a Civilization’
My takeaway is that there's a more pressing risk from humans automating too much too fast with something approximate and suboptimal, but "good enough". There're already plenty of situations where city planners enforce macroscopic policies that result in widespread discrimination and such, and that was with best intents. The promise of AI is that entire horizons of industry open up with opportunities for automation, making us tackle historically difficult-to-get-right objective functions, potentially with bad macroscopic results.
maldeh··on Bancor Is Flawed
A more likely use case for Bancor is some SMB (store/restaurant) or content creator issuing tokens to be used for membership reward/points or for voting on a new video/project idea, not for Augur/Steem/Gnosis 2.0 to issue an ICO.

I question if their massive ICO was primarily backed by the marketing promise of restaurants issuing their own currencies.

Even if that were the case, I'd still be concerned if this protocol can only be safely used for small-time operators, while any token with any semblance of scale on the chain is thoroughly gameable.

On a side-note, Emin Gün Sirer made a lot of noise about re-entrancy exploits present within the DAO in the months preceding the hack ( http://hackingdistributed.com/2016/05/27/dao-call-for-morato...), so I wouldn't take his concerns about the Bancor codebase lightly, and may even be inclined to give him the benefit of the doubt vis a vis FUD.

maldeh··on Amazon has a patent to keep people from comparison shopping in stores
Not very educated on patent matters, but isn't there any requirement for plaintiffs to demonstrate patent utilization, or at the very least abstract plans of utilization, no matter how impractical? Making a public claim that the patent would never be utilized but only used defensively, would make that difficult.
maldeh··on Be Careful Celebrating Google’s New Ad Blocker
I imagine this may irk some googlers on here, but after Hangout's effect on xmpp / Reader's on RSS, I find it hard to discount the impact a built-in Chrome ad blocker would have on other generally available blockers. A lot of major sites still allow their content to be read by adblock users today because of the volume of traffic they'd have to shut out; but I don't believe that'd be the case once they start becoming a minority.
maldeh··on India, Once a Coal Goliath, Is Fast Turning Green
By your statement I presume you are alluding to India sourcing ~70% of its energy consumption from coal (compared to ~40% by the US), according to this source. I should point out that the parent poster's dataset comes from 2014, whereas the article is talking about a much more recent transition towards renewables (the Paris accord was signed 2015 and India took part circa 2016).

The OP's point is still reasonable in terms of both absolute magnitude as well as per-capita, in that India, as a developing country and yet one of the most populous countries on the planet, consumes far less energy than developed states and is not as immediate a threat to life as we know it on a global basis, such as the US is. Indian policies towards coal consumption, no matter how progressive they can get, cannot hope to match the immediate impact their counterparts in the US can achieve today (and choose not to), and the current US administration does not really have grounds to point fingers elsewhere.

maldeh··on Under the Hood of Google’s TPU2 Machine Learning Clusters
"something custom for TPU" is what seems interesting here. Researchers working on the TF research cloud probably wouldn't learn much about the sorts of workloads this architecture is suited to, if all they have access to is the high-level API. Would whatever iteration of the TPU API they have internally be something they'd release soon, at least to partners?
maldeh··on Under the Hood of Google’s TPU2 Machine Learning Clusters
Per my understanding of XLA, it provides the right high-level abstraction for compiling the tensorflow computation graph under a range of architectures, CPU, NVIDIA GPU, mobile, etc. but it still delegates to a lower level domain-specific API like CUDA.

This undertaking isn't something to be taken lightly; CuBLAS has some pretty cutting edge architecture-specific optimizations for batching operations for matrix multiplication that came from several years of research - and is arguably a massive competitive advantage of NVIDIA over AMD. Depending on the development state of such an API internally within Google, it could mean that the Cloud TPU isn't going to be ready for wide-spread commercial use for a good while, and is very much still in the research-and-development phase (which could explain why they're only opening it up for the research community right now).

maldeh··on Under the Hood of Google’s TPU2 Machine Learning Clusters
Another potential source of friction I've wondered about regarding getting research workloads running on TPU2 is the heavy reliance of most of the popular DL frameworks on CUDA and CuDNN; I don't understand how one could port those libraries to a drastically different architecture such as this. Some DL frameworks have made overtures to OpenCL but it has not typically been a top priority [0].

Could this instead imply Google is working on its own library of low-level math primitives just for Tensorflow, and if so, how long would that take before it's competitive performance-wise? At any rate, having to support a different computing platform / API would be another blocker to general adoption.

[0] https://github.com/tensorflow/tensorflow/issues/22

maldeh··on Caffe2 adds 16 bit floating point training support on the NVIDIA Volta platform
From first looks, there is little doubt that that NVIDIA's Volta architecture is a monster and will revolutionize the AI and HPC market. But the article seems to avoid quantifying how 16-point FP operations are beneficial against 32- or 64-bit FP operations in real-world usecases, or how the Caffe2 / NVIDIA architecture provides any significant boost to FP16 in particular, especially apropos to images (or why FP16 is better for image data in general).

I'm interested more in understanding why Caffe2 would outperform Theano, Tensorflow, MXNet, etc. once Volta chipsets are generally available, beyond early pre-release optimization, particularly when most of the front-runners are already leveraging / taking into account NCCL, CuDNN, NVLink, etc. When the burden of adding support for new NVIDIA primitives is so low, what gives Caffe2 an advantage beyond an ephemeral "we were partners with NVIDIA first" one-up that would last for a couple of months at most?

(Apologies in advance if this post sounds overly negative, but I am constantly evaluating the current crop of frameworks for the trade-offs they enforce on the problem space, and a definitive answer would be very helpful.)

maldeh··on Inside an AI 'brain' – What does machine learning look like?
One really contrived usecase may be "flame graphs" for high level diagnostics of learned models in a production pipeline, as you collect more data or modify features. But I don't imagine it'd be very good.

- the images and false colors need to show some semblance of stability for a given network between epochs; and it needs to be robust against changing data or input structure.

- requiring visual inspection doesn't give you something you can automate with, unlike an evaluation score.

- if there is indeed a significant deviation in "MRI"-like scans between batches, its diagnostic utility ends there - it tells you nothing about what caused a change.

maldeh··on Ask HN: Who really gives your personal info to Intelius, Instant Checkmate, etc?
It's likely the locality / ephemerality / repudiability of those identifiers made them less of a big deal in an era when people couldn't instantly look these details up or build a detailed profile of everybody given the right data sources.
maldeh··on A Formal Security Analysis of the Signal Messaging Protocol
To say "OWS shut down this fork" is a rather drastic distortion of the truth. The project was simply not entitled to running on OWS servers, or of re-using the Signal trademark, both of which are perfectly understandable.

And if you look at the repo that you linked, this is called out right at the top of the README:

In the same thread, Marlinspike said that he would be willing to consider "a clean, well written, and well tested" pull request that would add WebSocket support to the Android version of Signal. This would effectively eliminate Signal's dependency on GCM and thus allow the official Signal app to function on custom Android ROMs that do not include Google Play Services. There is now a bounty on making this pull request, and the person/team making the pull request would also receive whatever the BitHub payout would be at the time.

...

To be clear, Signal relying on GCM is only an issue for people who use a custom Android ROM without Google Play Services. For the vast majority of people who do have Google Play on their phone, this issue is completely irrelevant. Signal is designed so that GCM is only used for a wakeup event and never sees any of the encrypted Signal messages.

maldeh··on Washington Post Is First Paper to Call for Prosecution of Its Own Source
You're right that they don't have a bunch of obligations to stick by Snowden through legal challenges. Yes, it is ethically acceptable to publish public-interest excerpts from a source they believe is intrinsically illegal. They can also loudly have problems with the way he engaged with other media actors (a la publishing overly sensitive information) and still have a relationship with their source. Their contract with him is purely "do your best to mine and publish what you believe to be in the public interest from this trove of (possibly illegal) data".

Simply put, you can say WaPo initially betrayed Snowden by (in their mind) publishing sensitive info (betrayal by incompetence). If their present thesis is correct, then they were intimately involved in increasing public harm from the leaks. At this juncture they had a number of options to make things right by their perception, but the one they chose was to attribute all blame to him (betrayal by deception).

Not that Snowden had some special right not to be betrayed. He engaged with a lot of media actors, laid a lot of trust by their sense of journalistic ethics, and had to have accepted the risk of whatever fallout could happen from some of the outlets not acting in a trustworthy manner (mitigated by how selectively he involved people). That one of them "betrayed" him by not doing their job well, or pointing fingers at him is not really interesting. Again, I find the WaPo's dishonesty about their own role more interesting.

maldeh··on Washington Post Is First Paper to Call for Prosecution of Its Own Source
is there an editorial they could have written that would advance their opinion about clemency and not been treacherous?

I think the nature of the opinion itself makes it an unavoidable binary of betrayal / non-betrayal here, since to say now that the leaks were unjustified, is a complete reversal from their prior implicit stance.

Given that Snowden approached them for editorial oversight, and they made the executive decisions to publish some documents, to declare later that those same reports were illegal would have to be a betrayal of their source's trust by definition, regardless of whether it is justified, moral or correct.

I agree that hypocrisy is an independent dimension. They could do a complete moral cleansing by taking responsibility for their actions, but even that wouldn't negate the betrayal.

As an aside, I am not put off by the treachery or disloyalty as much (since that can be inevitable at times - Snowden himself betrayed his employers out of his moral convictions), just WaPo's intellectual dishonesty as they went about it.

maldeh··on Yahoo probes possible huge data breach
However, attempts to contact more than 100 of the addresses in the sample saw many returned as undeliverable with auto-responses reading: "This account has been disabled or discontinued," which might suggest that the data is old.

Or another sad possibility is that this may be representative of any sample of yahoo email addresses.

maldeh··on Soylent CEO Could Face Criminal Charges for LA Hilltop 'Experiment'
Perhaps this is distracting the conversation then. I just brought up rust because it is a common element of rental inspection (http://www.dca.ca.gov/publications/landlordbook/looking.shtm...), and because inhaling iron particles has been known to cause respiratory issues - probably more relevant in a human enclosure than in a mailbox (albeit not as serious as asbestos or lead). But the point stands, if the container is generally accepted as a public-access dwelling, then it is not unreasonable to subject it to inspection for livability.
maldeh··on Soylent CEO Could Face Criminal Charges for LA Hilltop 'Experiment'
I may be missing something, is this not a thing people say?
← PreviousPage 3 of 5Next →