davidmr

I agree. I meant the / to indicate that either one or a combination would have been helpful in different ways, but I certainly could have phrased it better.

Many years ago in a past life, I worked on the network security team at the University of Chicago. We had a similar policy (and they may still for all I know) of just being able to requisition publicly routable IPs and run whatever you wanted on them with no default firewall rules applied at the border. Not for nothing did we call this a "target rich environment".

For all of the cool things I got to do (troubleshoot a breakin at the South Pole, send the RIAA a DMCA takedown notice when they stole our content (absolutely the highlight of my career), etc.), we spent the vast majority of our time on nonsense. We processed dumb breakins by the hundreds, had to enforce DMCA takedowns, and the like.

I'm also all for innovation and giving people the freedom to deploy services and innovate, but I would have killed to deploy all IPs by default behind NAT/firewalls and work with researchers to help them understand their responsibilities before giving them public IPs.

In my limited experience, their root cause analyses are really impressive as well with lots of internal attention and resources. I'm not allowed to talk about any Intel issues, but we reported a very strange issue to Nvidia, sent a couple of dozen cards back and six months later got a truly fascinating report back we with hundreds of pages of compute test result tables and electron microscope images and chemistry lab reports. Anything that hints of a manufacturing problem is taken incredibly seriously.

This is a great start, and I find your code extremely easy to read.

I might suggest some example code for more advanced methods of data ingestion. All the code I see uses feed_dict, which is certainly the most straightforward, but once you start to try to ingest larger out-of-core datasets or use distributed tf, things get exceptionally complicated (to me at least).

Some simple-ish code with queue runners and the new Dataset API would really be helpful.

Thank you. Even if things are a little different for every person as they naturally are, this is good advice and I genuinely appreciate the positive attitude. As someone who has struggled with depression quite a bit, I generally have a knee-jerk reaction to "just be positive" advice, but this really sits a nice middle ground between "be positive" and "find a way to be as positive as you can".

It's just a press release but I've been using them for a while, and it is quite not very to watch a model training or hyperparameter search scale effortlessly over multiple GPUs. It almost makes up for the royal pain in the arse of having to rewrite your code to fit in the newish APIs required by cloud ML engine. Almost.

Even for training? That would be disappointing.

The potential use of TPUs for training is very exciting. They say that they train floating point, but I don't see any indication to the FP precision they're capable of; perhaps I'm missing it. At any rate, I'm really excited to see resources being piled into their Cloud ML Engine product at this high rate.

I've made this comment in a couple of other threads, which subsequently veered off into other territory, so forgive the repetition, but it's a really interesting topic to me. The open-source distributed tensorflow stuff is pretty nice, but it still requires a huge amount of hand coding and tuning the machinery, reminding me quite a lot of just rolling the damn thing in MPI yourself. I'm very excited to see where distributed tf will be in a year or two, but it's a chore today.

Depending on how much these TPUs and other Cloud ML Engine developments help, I'd gladly abandon the attempt to roll it myself with the distributed tf.

The hope is that using Google's secret sauce to auto-distribute the execution graphs and associated data ingestion makes things "just work". At the moment, the documentation and examples for that are a bit all over the place and require writing models to conform to the newish tf.contrib.learn.Experiment API, which is also a bit underdocumented and underexampled. Using it for very large datasets (say >tens of TB) seems to be pretty challenging at this moment (to me at least). For a lot of use cases, BigTable seems to be the ideal ingestion engine for Cloud ML tf jobs, but there's no C native API. You can use BigTable, but you can only dump complete tables into tensorflow rather than querying for relevant data (since the queries cost money, a 5000-core jobs with just a few queries per core would cost you a fortune, so the ability to query BigTable in the tensorflow reader is disabled.

At any rate, I've been banging around on it for a few weeks and am really hopeful. I will follow Cloud ML Engine's career with considerable interest.

The discussion seems to have pretty well settled on Google's policies around service deprecation, but in case anyone's interested in chatting about the replacement API, I'm excited about the long-term prospects of their ML Engine product.

The open-source distributed tensorflow stuff is pretty nice, but it still requires a huge amount of hand coding and tuning the machinery, reminding me quite a lot of just rolling the damn thing in MPI yourself. I'm very excited to see where distributed tf will be in a year or two, but it's a chore today.

The hope is that using Google's secret sauce to auto-distribute the execution graphs and associated data ingestion makes things "just work". At the moment, the documentation and examples for that are a bit all over the place at the moment, and require writing models to conform to the newish tf.contrib.learn.Experiment API, which is also a bit underdocumented and underexampled. Using it for very large datasets (say >tens of TB) seems to be pretty challenging at this moment (to me at least).

At any rate, I've been banging around on it for a few weeks and am really hopeful. I will follow Cloud ML Engine's career with considerable interest.

They're both magnificent planes, but given the option I will usually take the 380. At least on the upper deck, it's far quieter. Over a 10-15 hour flight, that goes a long way with me.

I did! 15 years ago I went and had custom earplugs made with replaceable attenuators, and they're still in perfect shape. Without question, it was the best $200 I've ever spent. I can't recommend them highly enough to anyone else.

Semantic quibbles aside, water cooling is a godsend for one's ears.

I spent a couple of years running the I/O system on the Blue Gene in that photograph (Intrepid). The pictures really don't do it justice, although all the doors are off in that photo, so that wasn't even really what it looked like. It was a big, beautiful, $100m black box.

ANL's new system, Mira has a big mural painted on it, and I think it completely ruins the aesthetic.

I don't miss working in that room, though. The box is very pretty, but the fans are astoundingly loud. As an aside, it's pretty amazing how much heat they need to dissipate. Once a year we had to do a cooling failure test where we all sat in the room by our respective subsystems while Facilities turned off the A/C just to make sure everything would notice the temperature rise and power itself off. When the A/C got cut, you felt the room start to heat up quite quickly almost instantly, within a few seconds.

Respectfully, this is just plain incorrect. You're describing some situations that do happen and just choosing to ignore what happens the rest of the time. When demand precisely matches supply, you're correct--you need a matching engine to connect buyer and seller and no middle man/market maker need get between them. That also happens too.

Assuming that at some point bids and asks will line up and trades will happen doesn't a market make; that's the exchange's job. Making a market means being able to quote prices on both sides of the book and having inventory to trade at the current market price. Carrying that inventory has actual risk involved, and that's why it's not a free service that the exchange (or anyone else) provides. Market makers also get penalized by the exchanges if they're not making markets for some large percentage of the time that the products are trading.

Maybe a better way to look at it is that the exchange is there to match up buyers and sellers at the current price at a given moment in time. The market matches up buyers and sellers at the current price over a period of time. This market stabilizes the price over time. Demand and supply just don't line up perfectly like it seems like they should. That's the difference between the market and the exchange. If I'm buying a product now, I want an idea of the true value of it, and the less the price is whipping around waiting for demand to match the supply, the more I know what the current market rate is.

I'm not saying they're doing gods work, but to say that there's no risk involved or reason for them to exist is incorrect.

If you don't buy any of the above arguments, then I'm interested in your answer to the question of why the exchanges themselves pay market makers to make markets. If the current technology renders market makers obsolete, surely the exchanges would recognize that and keep the money for themselves, no?

That's extremely well said. While it certainly beats the alternative of rolling something like it yourself, I've struggled a bit trying to get in the flow of tf, pardon the expression.

My impression is that it's an ecosystem that's developing incredibly rapidly with all the associated growing pains. Even for such a simple-on-the-surface task as feeding very large datasets into the graphs, there are many different ways, some deprecated already, some about to be deprecated by functionality not in 1.0, etc.

It's also quite new, but the distributed execution, while awesome, still requires a huge amount of hand coding and tuning the machinery, reminding me quite a lot of just rolling the damn thing in MPI yourself. I'm very excited to see where distributed tf will be in a year or two, but it's a chore today.

I'm really bullish on their Cloud ML product, a tool to auto-distribute the execution graphs and associated data ingestion, but the documentation and examples for that are a bit all over the place at the moment, and require writing models to conform to the newish Experiment API, which is also a bit underdocumented and underexampled.

My frustration is likely only because of how awesome tensorflow actually is. It's new, but so promising that I want to use it for everything, regardless of the development roadmap!

I agree up to a point. Certainly the implication you should not continue doing something if a) you think it's fun and/or b) you think it's important seems wrongheaded to me and not a very fun way to go through life. I'm never going to be as smart as Andrew Ng, but I still have fun building ML systems and pipelines.

That said, there are times when I've given up and felt like it was the right thing to do. I went to a pretty good school for Physics because I was sure that was what I wanted to do with my life. After 2-3 years, through no lack of effort on my part, it was just super clear I wasn't going to make the cut--I just didn't have it. I worked hard and got okay grades, but as the material got harder, it was just plain to see who was separating from the pack and those kids just had something I never would, and hard work wasn't going to do it. If I didn't put in 10k hours, I came damned close.

I dropped Physics as a major, switched to Math, dropped the idea of academia as a career and have never looked back. An honest look at one's talents relative to peers and competitors is still important.

In the mid-90s, I was 16 and had just started my very first job at a local ISP after having dropped out of high school (his needing a job with some urgency). It was right around the time that everyone and their mother started to check out this information superhighway thing, and mom and pop ISPs were sprouting up everywhere.

When I interviewed for a job, the owner, who we'll call Frank, seemed like a pretty decent guy. He was definitely quite smart, and knew what he was doing on the telco side. I got the job after like 5 minutes of interview and started the next day.

What Frank failed to tell me in the interview was that he was a completely batshit crazy sociopath. Within a couple of days, there was a fair bit of yelling and swearing about my mistakes, then progressively more and more, followed by genuine emotional abuse and even flipping lit cigarettes at us and throwing other things when we screwed something up.

I'm only 16 and an idiotic 16 year old at that, so I actually stay there like 6 months before I start to realize this isn't normal. One day, a few of us just get up and walk out. About a month later, I heard (and saw evidence as I was walking by) from a friend who was still there, that after hours, someone had shot up the back windows of the office (4th story office in Chicago, so it's pretty unlikely it was random).

Anyway, I manage to forget about it and move on with my life. 20 years later, I look him up to find he had served 3 years in the federal pokey for threatening a judge who had ruled against him in a tax case. Couldn't have happened to a nicer guy.

Pro tip: you don't get a warning when you threaten to kill a federal judge via email. You go straight to prison. Do. To pass go; do not collect $200.

Completely agreed. The landscape 2-3 years ago looked incredibly different, and I picked Mesos for a rather ambitious project. After it became relatively clear that k8s was going to eclipse Mesos and not by a little bit, trying to unwind that decision basically cost me my job through some political infighting. It's a shame, but such is the price of early adoption, I guess.

Given the same information, I'm really confident that I'd still make both choices the same way.

Try actually reading the article. That's not really what it's about.

As a rule these articles about gentrification almost always irritate the hell out of me (likely because I live in a gentrifying neighborhood in Chicago and honestly just don't feel all that bad about it), but this is a rare exception. I think this article is well worth the long read.

A lot of what makes the "news" in Chicago are plays on the old "this was my neighborhood/new people are allowed to live here too" tropes. All that seems to do is to get both sides to dig in, and I've never seen any useful dialogue come from it. Telling me that I'm an asshole because I came in and bought a house for more than it was worth 20 years ago and I should be ashamed of myself doesn't do any more good than me telling someone that I know their rent goes up as the neighborhood gets more expensive, but they don't get to tell me where I can buy my house and by the way, my neighbors houses are worth more now too.

This article changes the conversation in a way that I hope isn't specific to SF by acknowledging that gentrifiers come into contact with new people with real problems. Appealing to someone's humanity to help their neighbor is a much more effective strategy than verbal class warfare. It also helps inspire people to think of ways to get concretely involved in small and meaningful ways, rather than complaining about problems when there are very few realistic alternatives.

I think we've lost the war. I was even asked to meet with my company's architects and designers about a new space we were building out. I basically spent an hour telling them in every different form I could that what I wanted was a place I could go where people could find me, but they would have to knock on a door and open it to talk to me, and these people looked at me like I was from fucking Mars. They'd never heard anything so absurd in their whole life. I even pleaded for cubicles. I can't even remember what it was like to have dignity.

"What about a health clinic? Or a coffee bar instead?"

I found as many studies as I could about how awful these open office plans are, and printed them all out, and left them there. At the end of the day, on a whim I checked the recycle bin in the conference room they were in. Anyone want to guess what I found?

When I die and get to hell, there will be a Hermann Miller chair in an open office waiting for me. With free snacks and drinks in the kitchen.

I don't yet know if I'm leaving, but I am extraordinarily disappointed. I sat down and did the math a while ago and came to the conclusion that, with the benefit of the taxable/tax-deferred tax management and tax loss harvesting, 15% was a reasonable amount to pay. After I do it again, it may turn out that .25% still puts me ahead, but this leaves a bad taste in my mouth. How long before they eliminate the digital plan and put me in the .4% plan?

I've been spending several hours a day on it over the last couple weeks, and I'm really frustrated with their Cloud ML beta.

I realize that beta means beta, but there's very little working code out there that runs properly (especially Google's sample code), and in many cases their documentation is either outdated or just plain missing.

It just doesn't feel like it's baked, even for one of their betas.

I'm confused. How did the USDS conduct surveillance on basically every citizen in the nation, or are you painting every US gov't employee with a very, very broad brush?

I spent 15 years in academia and now work with dozens of former academics. In all my time, I don't think I've ever met a single person who didn't loathe the journal system from top to bottom, and that includes at least a couple journal editors.

I seethe in anger every time something I'm looking into ends in a IEEE or other journal who militantly police arXiv preprints.

The day we bury the last for-profit closed access journal will be a momentous day in the history of science.

Maybe you and I have different interpretations of what affects us, but I'm still affected if I live in a city with a dramatically increasing murder rate, even if I'm not likely to be murdered. I have friends working in trauma care at hospitals, for the police and fire department, I volunteer my time in neighborhoods I don't live in, etc. More than that, I have a hard time saying that the problems of people just a few blocks west of me aren't a) my problems already, and b) likely to become bigger problems for me in the future. Maybe it's all just misplaced liberal guilt, but I find it hard to ignore.

I'm genuinely surprised that this post is up to 214 comments and nobody has mentioned Paris yet.

I have no idea what the practicalities of a life in Paris are, but it's still at the top of my list. In maybe half a dozen visits, I've found Parisians to be quite friendly for city dwellers (very much counter to the stereotype). If I approach situations in my severely broken French, people figure out quite quickly that their English is better than my French, and like most everywhere else in Europe, nearly everyone under 40 speaks excellent English.

Maybe it's one of those things that looks better to an outsider than it is, but the pace of life is very attractive to me.

I'm far less surprised that nobody has mentioned Chicago, but it's not bad here. For a large city, it has most everything I want: great art, great restaurants, etc. The cost of living here is obscenely low relative to NYC or SF. For far less (monthly rent vs. mortgage payment) than friends who rent one bedroom apartments in either city, we own a modest, but very nice detached house in a nice neighborhood smack in the middle of the city.

Between the recent dramatic uptick in shootings and the truly shitty weather in the winter (though spring and summer are glorious), it's pretty hard for it to bubble up to the top of anyone's list.

Well said. Domination isn't the right word; perhaps "importance to" rather than "domination of".

I realize this is a stupid argument, but as long as someone else is reading it, I don't feel like I have to. I'm not recommending everyone stop reading the news, just me.

I'm conflicted about this.

On the one hand I have also decided to become as ignorant as possible about the news because (at the risk of overusing the phrase), I think we have firmly moved into a post-truth American society. If election outcomes aren't based on facts, what's the point of knowing them?

On the other hand, many other people in this thread are saying that there's never been a more important time to know the news. I understand the sentiment, but I would phrase it differently: there's never been a more important time for good journalists to be doing good journalism. Except for the tricky issue of how to keep good journalists employed without people reading them, I don't feel any particular obligation to be informed. I decided to donate money to ProPublica in the hope that they'll keep the government as honest as possible without me having to deal with it.