| 01:34 | <Hixie> | can anyone find the enumerated attribute which allows the empty string as a valid keyword? |
| 01:34 | <Hixie> | i've looked all over with no luck |
| 01:34 | <Hixie> | i know there was one once... did we remove it? |
| 01:55 | <othermaciej> | I remember there being a bug about this - did it not mention the affected attribute? |
| 02:06 | <Hixie> | seems not |
| 02:19 | <Hixie> | ah, contenteditable |
| 02:22 | <Hixie> | spellcheck to |
| 02:22 | <Hixie> | o |
| 03:23 | <Hixie> | so you can get and set cookies on the document returned by XHR? |
| 03:24 | <othermaciej> | I think in Firefox you can |
| 03:24 | <othermaciej> | in IE the cookie property is not present |
| 03:24 | <othermaciej> | in WebKit it fails, unless we already changed to match Firefox |
| 03:28 | <Hixie> | what should html5 require? |
| 03:33 | <othermaciej> | I vaguely recall discussing this before |
| 03:33 | <Hixie> | firefox doesn't seem to have a .cookie actually |
| 03:34 | <Hixie> | oh let me try an xhtml doc |
| 03:34 | <othermaciej> | the WebKit bug had a bunch of data |
| 03:35 | <Hixie> | xhtml doesn't seem to have .cookie either |
| 03:35 | <Hixie> | i get an XMLDocument in both cases |
| 03:36 | <othermaciej> | https://bugs.webkit.org/show_bug.cgi?id=32115 |
| 03:36 | <othermaciej> | I'm not sure if all the test cases were posted |
| 03:37 | <othermaciej> | someone claimed the following behavior for FF3.5: |
| 03:37 | <othermaciej> | * deleted iframe: no exception, setting and getting allowed |
| 03:37 | <othermaciej> | * createDocument: no exception, setting and getting allowed |
| 03:37 | <othermaciej> | * XMLHttpRequest: no exception, setting and getting allowed |
| 03:42 | <Hixie> | i commented |
| 03:43 | <MikeSmithXX> | does draggable not allow the empty string as a keyword? |
| 03:44 | <MikeSmithXX> | and/or can't it be specified using empty-attribute syntax |
| 03:56 | <MikeSmithXX> | Hixie: ↑ |
| 03:56 | <Hixie> | is the spec ambiguous? |
| 03:57 | <Hixie> | if it is please file a bug, i'm in the middle of a complicated edit :-) |
| 03:57 | <MikeSmith> | hai |
| 04:34 | <Hixie> | hsivonen: http://www.w3.org/Bugs/Public/show_bug.cgi?id=8837 is a pain |
| 06:48 | <Dashiva> | Massive bug activity lately... |
| 06:56 | <Hixie> | people keep filing new ones |
| 06:56 | <Hixie> | makes it hard to get the count down! |
| 06:56 | <Hixie> | somehow people only file new ones when i'm working on fixing them |
| 07:01 | <Dashiva> | That's not so strange |
| 07:02 | <Dashiva> | It feels more useful then since you see other people get responses |
| 07:02 | <Hixie> | heh |
| 07:45 | <othermaciej> | Current bug counts: |
| 07:45 | <othermaciej> | HTML+RDFa: 24 |
| 07:45 | <othermaciej> | H:TML: 10 |
| 07:46 | <othermaciej> | HTML5 + HTML Microdata + HTML Canvas 2D ContexT: 126 |
| 07:55 | <othermaciej> | Hixie: I'm closing some obvious INVALIDs and DUPLICATEs for you |
| 07:55 | <Hixie> | cool, thanks |
| 07:56 | <Hixie> | i just go through them in order of last modified |
| 07:57 | <othermaciej> | 119 now |
| 08:57 | <Dashiva> | XML syntax for CSS... *shudder* |
| 09:03 | <Hixie> | hah |
| 09:03 | <Hixie> | three browsers |
| 09:04 | <Hixie> | three results |
| 09:04 | <Hixie> | gotta love the web |
| 09:04 | <Hixie> | (http://software.hixie.ch/utilities/js/live-dom-viewer/saved/373) |
| 09:05 | <Hixie> | firefox passes the arguments in the order a,c,b |
| 09:05 | <Hixie> | sorry, firefox passes the arguments in the order a,c,b,type |
| 09:05 | <Hixie> | safari passes the arguments in the order b,c,a,type |
| 09:05 | <Hixie> | and opera does a,b,c,type,pluginspace |
| 09:07 | <Dashiva> | Attribute order is relevant? Wow |
| 09:10 | <Hixie> | http://software.hixie.ch/utilities/js/live-dom-viewer/saved/374 -- gecko: type,e,a,d,b; webkit: b,d,a,e,type; opera: a,b,c,d,e,type,pluginspace |
| 09:10 | <Hixie> | wtf is opera doing |
| 09:10 | Hixie | decides opera is disqualified from this round of testing |
| 09:11 | <Dashiva> | Isn't that what IE would do? |
| 09:12 | <Hixie> | dunno what IE would do, I don't have a way to test IE |
| 09:12 | <Dashiva> | Preserving attribute ordering even when removed for compat with COM references or whatnot |
| 09:12 | <Dashiva> | Not specific to <object>, just in general |
| 09:14 | <Hixie> | http://software.hixie.ch/utilities/js/live-dom-viewer/saved/375 -- webkit: x,b,d,a,e,type; gecko: doesn't launch plugin at all |
| 09:14 | <Hixie> | preserving attribute ordering is one thing. Sending attributes that aren't even on the element is another. |
| 09:15 | <Hixie> | what webkit does makes complete sense |
| 09:15 | <Hixie> | and is self-consistent |
| 09:16 | <Hixie> | i think we'll go with that |
| 10:38 | <Hixie> | can <wbr> be expressed in CSS? |
| 10:41 | <Hixie> | 101 bugs, oldest one was changed less than a week ago |
| 10:41 | <annevk> | wbr{ white-space: nowrap } iirc |
| 10:41 | <Hixie> | um |
| 10:41 | Hixie | passes annevk some coffee |
| 10:43 | <Hixie> | man, some of these bugs are like brain teasers in terms of trying to work out wtf the bug is about |
| 10:44 | <Dashiva> | I think that's <nobr>, annevk |
| 10:45 | <Dashiva> | <wbr> seems more like "End previous [anonymous] inline box. Start a new anonymous inline box." |
| 10:53 | <annevk> | oh lol |
| 10:53 | <annevk> | lots of adobe email on public-html |
| 10:53 | <annevk> | aah, shelley powers |
| 10:54 | <annevk> | of course |
| 10:54 | <Hixie> | it must be march |
| 10:54 | <Hixie> | how time flies! |
| 10:54 | <Dashiva> | Maybe she's a precog, so she knew she would be busy with Adobe posts beforehand |
| 10:55 | <Dashiva> | Hixie: That was an attempted answer earlier, by the way. |
| 10:55 | <Hixie> | the <wbr> thing? |
| 10:56 | <Hixie> | thanks |
| 10:56 | <Hixie> | i went with a somewhat different approach, but also prose |
| 10:56 | <Hixie> | i was hoping there was a pure-css solution |
| 10:56 | <Hixie> | (if anyone finds one, please file a bug) |
| 11:01 | <annevk> | omg, whining over the sotd |
| 11:02 | <annevk> | public-html is such a waste of my time :/ |
| 11:03 | <Hixie> | multiple independent whinings over the sotd |
| 11:03 | <Dashiva> | Hixie: What do you mean by pure-css, by the way? |
| 11:03 | <Hixie> | Dashiva: as in, something you could put in a text/css file |
| 11:03 | <Hixie> | specifically, ua.css |
| 11:03 | <Dashiva> | And the problem is <nobr>x<wbr>x</nobr>, right? |
| 11:05 | <Hixie> | that's one of the things that we would need to resolve, yes |
| 11:05 | <Hixie> | aw man |
| 11:05 | <Hixie> | bug 8911 |
| 11:05 | Hixie | glares at anne |
| 11:07 | <annevk> | oh heh |
| 11:07 | <annevk> | I filed the same problem twice |
| 11:07 | <annevk> | see also http://www.w3.org/Bugs/Public/show_bug.cgi?id=8904 |
| 11:08 | <Hixie> | excellent |
| 11:08 | Hixie | marks 8911 as a dupe of 8904, thus modifying both of them and moving them both off the top of the bug list |
| 11:09 | <Dashiva> | What a copout |
| 11:09 | <Hixie> | :-D |
| 11:09 | <Hixie> | i have to deal with 8904 anyway sometime this weekend |
| 11:09 | <Hixie> | it's the P1 i promised anne i'd fix |
| 11:09 | <annevk> | it's blocking any work on XHR, including AnonXMLHttpRequest |
| 11:09 | <annevk> | I can wait another week |
| 11:09 | <annevk> | if you want |
| 11:10 | <annevk> | i'll work on CSSOM instead |
| 11:13 | <Hixie> | i'll do it sometime this weekend |
| 11:13 | <Hixie> | (which ends on tuesday morning for you, since it's a long weekend for me) |
| 11:18 | <Hixie> | http://www.w3.org/Bugs/Public/show_bug.cgi?id=8924 seems like an interesting idea, though it would cause problems in XHTML |
| 11:19 | <Dashiva> | How so? |
| 11:19 | <Philip`> | node.dataset.doquery.range |
| 11:19 | <annevk> | yeah, what Philip` said would be neat |
| 11:19 | <Philip`> | plus some IDL magic to make it reconstruct it to an attribute name with "-" |
| 11:19 | <Philip`> | The problem is you'd want node.dataset.doquery to also act like a string for the attribute data-doquery |
| 11:20 | <Dashiva> | Yeah, we don't want more magic, really... |
| 11:20 | <Dashiva> | The location object is bad enough |
| 11:20 | Philip` | likes magic |
| 11:22 | <Dashiva> | But how is camelcase accessors problematic in XHTML? |
| 11:23 | <annevk> | maybe the API is case-sensitive in XHTML currently? |
| 11:23 | <annevk> | camelcase wfm too |
| 11:25 | <Hixie> | oh actually there's no xhtml problem |
| 11:25 | <Dashiva> | :) |
| 11:25 | <Hixie> | because data-aBc is explicitly not exposed in dataset |
| 11:25 | <Hixie> | sweet |
| 11:25 | <Hixie> | oh, no, i'm wrong |
| 11:25 | <Hixie> | but i can be right |
| 11:26 | <Hixie> | just have to change the spec a little |
| 11:26 | Dashiva | is still wondering what the (possibly false) problem was |
| 11:26 | <Hixie> | data-aa-bb clashing with data-aaBb |
| 11:28 | <annevk> | though those were disallowed |
| 11:28 | <Hixie> | yeah but they're still exposed in dataset currently |
| 11:29 | <Hixie> | i'll just exclude any with capital letters from dataset and we'll be golden |
| 11:29 | <Dashiva> | A custom data attribute is an attribute in no namespace whose name starts with the string "data-", has at least one character after the hyphen, is XML-compatible, and contains no characters in the range U+0041 to U+005A (LATIN CAPITAL LETTER A to LATIN CAPITAL LETTER Z). |
| 11:29 | <Dashiva> | Doesn't that already exclude it? |
| 11:29 | <Hixie> | that's the authoring conformance requirement |
| 11:29 | <Hixie> | doesn't say jack about what dataset does |
| 11:30 | <GarethAdams|Home> | for example, with dataset properties set programatically |
| 11:31 | <Dashiva> | So make the access algorithm replace capital letter with - followed by lowercase letter? That way it can't map to invalid attributes at all |
| 11:32 | <Dashiva> | Well, no... initial capital letter would give data--... but it wouldn't map to a uppercase letter |
| 11:33 | <GarethAdams|Home> | is data--foo invalid? |
| 11:33 | <Hixie> | Dashiva: so data-aB="" would become dataset['a-b']? Then it would clash with data-a-b="". |
| 11:33 | <Hixie> | GarethAdams|Home: the definition of what's valid is what Dashiva quoted above |
| 11:33 | <Dashiva> | Hixie: data-aB wouldn't become anything, it would just be an attribute that's never used |
| 11:34 | <Hixie> | i'm confused as to what you're suggesting |
| 11:34 | <Hixie> | the solution i'm following is trivial |
| 11:34 | <Hixie> | just exclude content attributes with a capital letter from the API |
| 11:34 | <Hixie> | and then do the adjustments |
| 11:35 | <Dashiva> | I'm suggesting you change the "Let name be the concatenation" steps to also replace uppercase letters with hyphen followed by lowercase equivalent |
| 11:35 | <Dashiva> | datalist.doQuery would map to data-do-query |
| 11:36 | <GarethAdams|Home> | Dashiva is saying that dataset['aB'] would be identical to dataset['a-b'] - with the data-aB attribute still being invalid |
| 11:36 | <Hixie> | yes, that has to be part of the solution also |
| 11:36 | <Hixie> | except dataset['a-b'] wouldn't be there |
| 11:37 | <Hixie> | not sure whether to make setting that throw an exception, silently fail, or silently passthrough |
| 11:37 | <Hixie> | i'm leaning towards exception |
| 11:37 | <Dashiva> | Why not allow it? |
| 11:37 | <annevk> | yeah, SYNTAX_ERR |
| 11:38 | <annevk> | Dashiva, enforce some consistency |
| 11:38 | <Hixie> | allowing it leads to a confusing situation where you can compare two strings, be sure they're not the same, then set both and trample each other |
| 11:38 | <Hixie> | very bad |
| 11:40 | <Dashiva> | Although WebIDL would make it fail anyhow, since a-b wouldn't be in the supported property name list |
| 11:40 | <Philip`> | Could make dataset['a-b'] map onto attribute data-a--b |
| 11:40 | <Philip`> | because that should avoid conflicts and avoid forbidding certain strings |
| 11:41 | <GarethAdams|Home> | I was about to say, transparently switching a-b and aB wouldn't help with multiple consecutive - |
| 11:41 | <Hixie> | Dashiva: i use the same algorithm for creation |
| 11:41 | <Philip`> | although actually that probably wouldn't work |
| 11:42 | <Hixie> | Philip`: and map ---x to --x when reading? that way leads madness |
| 11:42 | <Philip`> | because you'd need to do something with data-a-0 attributes, which presumably can only can be dataset["a-0"], and that needs to be reversible |
| 11:43 | <Dashiva> | Map - followed by non-uppercase to _? :) |
| 11:43 | <Dashiva> | Um, non-lowercase |
| 11:43 | <Philip`> | Does IE have some rules for mapping '-'-separated strings onto camelCase strings? |
| 11:43 | <GarethAdams|Home> | well not all dataset keys are mappable to DOM properties, so you don't need to provide for all possibilities |
| 11:44 | <Philip`> | for expandos or whatever they are |
| 11:44 | Philip` | wonders what the exact rules are |
| 11:46 | <GarethAdams|Home> | for most CSS properties, the rule is -[a-z] maps to [A-Z] - that's all. Trying to cover multiple hyphens, numbers etc isn't going to be possible because NameChar in HTML isn't the same as e.g. Javascript |
| 11:50 | <annevk> | Hixie, by the way, in XHR I said that .cookie returns the empty string and does nothing on setting |
| 11:51 | <annevk> | Hixie, but it would prolly be slightly better if HTML5 defined that |
| 11:51 | <Hixie> | cool, i can do that |
| 11:51 | <annevk> | Hixie, or gave me a way to mark the document in a certain way |
| 11:51 | <Hixie> | can you comment on the bug to that effect? |
| 11:53 | <GarethAdams|Home> | maybe of interest? http://www.webdevout.net/test?0a&raw |
| 11:53 | <annevk> | oh actually, it seems it just does it for non same-origin, that seems broken because we disallow access to cookie headers |
| 11:53 | <annevk> | i'll comment on the bug |
| 11:54 | <Hixie> | is Csaba a common name in some locale? |
| 11:54 | <Hixie> | we have two people named that in the spec's acks |
| 11:54 | <Hixie> | and i've never met anyone with that name as far as i recall |
| 11:54 | <Hixie> | but i've only lived in western locales |
| 11:54 | <Philip`> | Wikipedia says Hungarian |
| 11:55 | <annevk> | oh, you just closed the bug |
| 11:55 | <annevk> | was it not about http://www.w3.org/Bugs/Public/show_bug.cgi?id=8924 ? |
| 11:56 | <annevk> | oops |
| 11:56 | <annevk> | left a comment in the wrong bug |
| 11:56 | annevk | wonders how that happened |
| 11:57 | <Hixie> | heh |
| 12:03 | <annevk> | i wonder in that caching thread if there is any observable difference |
| 12:09 | <annevk> | Hixie, Romania maybe? |
| 13:13 | <annevk> | VLC fails |
| 13:13 | <annevk> | F11 is for fullscreen, but it doesn't actually work if you clicked somewhere on the video because that apparently eats keyboard input |
| 13:47 | <Hixie> | nn |
| 13:57 | <Dashiva> | "It would need to be stated that if autoplay is used, loadhint is automatically set to autobuffer." |
| 13:57 | <Dashiva> | Isn't buffering at all rather redundant when you start playing immediately? |
| 14:36 | <annevk> | this atom:id concept utterly failed in practice |
| 17:21 | <Philip`> | Someone at Google really needs to sort out the problem where emails from Google employees to public mailing lists are very commonly marked by Gmail as "Due to a filter you created, this message was not sent to Spam." and often "Warning: This message may not be from whom it claims to be." |
| 17:35 | <TabAtkins> | Philip`: Man, I know. I get bit by that all the time. |
| 17:36 | <TabAtkins> | Or rather, it did, until I set up my filter to never send them to spam. |
| 19:30 | <wycats> | Hixie: reviewing your comments |
| 19:45 | <wycats> | Hixie: I think what I was really asking for wrt same-origin-policy was the use of "same-origin", not "same-origin-policy" |
| 20:09 | <virtuelv> | othermaciej: if your findings on issue-30 are relevant, I think posting them to the list would be helpful |
| 20:11 | <jgraham> | More than that I would be interesting to see your analysis |
| 20:11 | <jgraham> | Maths is fun :) |
| 20:11 | <wycats> | is http://www.w3.org/Bugs/Public/ down? |
| 20:11 | <wycats> | "Error: Access denied for user 'bugs'@'128.30.52.43' (using password: YES)" |
| 20:11 | <jgraham> | wycats: Yes |
| 20:11 | <virtuelv> | jgraham: that too |
| 20:12 | jgraham | notices his total faliure to make a grammatical sentence |
| 20:13 | <jgraham> | Seriously I've strted sounding like I have been through Google translate a few times |
| 20:13 | <othermaciej> | jgraham: I took the relevant variables to be "longdescs attributes per img" and "valid longdesc values per longdesc attribute", assumed values were only 0 or 1, and reverse-engineered the mean, standard deviation, and standard error |
| 20:13 | <jgraham> | But with more typos |
| 20:13 | <wycats> | jgraham: is that possible? |
| 20:13 | <othermaciej> | writing it out in enough detail to stand up to scrutiny might be more work than the value we'd get out of it |
| 20:14 | <jgraham> | othermaciej: Sure, if ou don't want the bother, don't do it on my account |
| 20:15 | <virtuelv> | the question asked should be a different one, though |
| 20:16 | <othermaciej> | my most interesting conclusion is that if in fact 75% of longdesc values are useful, then Ian would have had to oversample bad longdesc values by a factor of around 400 relative to good longdesc values to get his result |
| 20:16 | <virtuelv> | even with longdesc being used in a useful manner in 75% of the cases used, why is longdesc used on less than 1% of pages? |
| 20:16 | <othermaciej> | 400x oversampling seems like it would be unlikely short of outright research fraud, given the reported methodology |
| 20:17 | <othermaciej> | Ian's study concluded that of the longdesc values that are provided, less than 1% are useful |
| 20:18 | <othermaciej> | I think that's potentially the more relevant point - if longdesc were rare, but very often useful rather than bogus when provided, then one could argue it does more good than harm, despite being obscure |
| 20:19 | <othermaciej> | I should also mention that because of the very large sample sizes, the standard error is ridiculously small, so some form of systematic error is the only way the study's conclusion could be wrong |
| 20:20 | <wycats> | what's the background :/ |
| 20:21 | <wycats> | sorry for asking that in the middle of an existing discussion :/ |
| 20:30 | <othermaciej> | Shelley's claim that Ian's research study on longdesc suffers from some sort of bias that makes it invalid |
| 20:30 | <othermaciej> | but apparently sample bias wasn't what she had in mind |
| 20:33 | <Dashiva> | The data for the second analysis was made public, so anyone could verify the results if they don't believe there's sample bias |
| 20:35 | <othermaciej> | I think Shelley did claim that one had a sample bias (she said it is "not representative of the web, at large") |
| 20:36 | <Dashiva> | That would be bias _against_ longdesc, though |
| 20:36 | <Dashiva> | Since most the longdesc advocacy I've seen has been "Yes, it isn't used normally, but professionals on professional sites use it, and that's enough to make it useful" |
| 20:38 | <TabAtkins> | Shelley did claim that the dmoz data wasn't representative. |
| 20:39 | <TabAtkins> | If I believed she was actually honestly engaging the list, I'd argue that it's breadth means it's still likely reasonably representative, and suggest comparing results from it to results from a random web sampling to verify. |
| 20:39 | <AryehGregor> | You can't do a "random web sampling". Random how? |
| 20:40 | <othermaciej> | she also said that the dotnetdotcom.org data was "anecdotal" because "we don't have access to the methodology determining the web bots path, we have no idea how often its blocked, it doesn't take into count the use of intranet data" |
| 20:42 | <TabAtkins> | Indeed. And again, if I felt she was actually being honest in bringing that up, I'd ask why she believed that the bot, when it is blocked, would bias the result in a particular direction (and again, one could compare its results to other studies that are more reliably representative to look for bias). |
| 20:42 | <Dashiva> | AryehGregor: Make a statement of what the web is like, then change your sample methodology until it matches |
| 20:42 | <othermaciej> | her use of "anecdotal" to refer to large data sets where she is not fully confident that the methodology provides a representative random sample is an unusual use of the term, I think |
| 20:42 | <TabAtkins> | AryehGregor: Um, why not? First approach: unifoermly sample from Google's cache of pages. |
| 20:43 | <othermaciej> | one could form a hypothesis that certain things missed by the bot would be more likely to contain longdesc attributes, and more likely to contain valid longdesc values in those cases where the attribute is present |
| 20:43 | <TabAtkins> | Second approach: take a large directory, wide enough to remove many forms of systemic bias, and uniformly sample from all linked pages. |
| 20:44 | <AryehGregor> | TabAtkins, "random" isn't well-defined. You have to further specify. See also: Bertrand's paradox. |
| 20:44 | <TabAtkins> | AryehGregor: In the absence of further clarification, lay usage of "random" means "sampled from a uniform distribution". Assuming otherwise is being uselessly pedantic. |
| 20:44 | <AryehGregor> | Sure. Uniform over what set? |
| 20:44 | <AryehGregor> | The same is true in Betrand's paradox. |
| 20:44 | <TabAtkins> | Over some representatively large set of pages. |
| 20:45 | <AryehGregor> | So you mean uniform over all pages, with no weighting? All pages everyone has viewed in the last X days, maybe? Including if it was only viewed by search spiders, or also humans? Including intranets too, or only the public Internet? How large is X? |
| 20:46 | <AryehGregor> | Some of those considerations will materially affect the resulting set, especially whether you include intranet pages. |
| 20:46 | <TabAtkins> | X is sufficiently large. For most of your questions I would hypothesize that it would not affect the sample for the types of variables we're testing. |
| 20:46 | <othermaciej> | hypotheses about intranet contents tend not to be falsifiable |
| 20:47 | <TabAtkins> | There are well-established methods of figuring out how large is "sufficient" based on the desired strength of the test you're performing. |
| 20:47 | <AryehGregor> | Not if you don't define the set in the first place. |
| 20:47 | <TabAtkins> | Though, "a billion" is pretty much always large enough. |
| 20:47 | <AryehGregor> | Not if it's nowhere close to representative. |
| 20:47 | <TabAtkins> | Do you believe that the sets that have been sampled are non-representative for the relevant variables? |
| 20:48 | <AryehGregor> | I don't believe they're non-representative enough to create such a skewed result incorrectly, no. But if you're trying to convince me, you're wasting your time, because I already agree with you. |
| 20:48 | <TabAtkins> | I'm not trying to establish a new test here, just defend the existing tests as being acceptable in their methodology. |
| 20:48 | <othermaciej> | it's defintiely possible for a sample to be very large, but still not random with respect to the variable under study |
| 20:48 | <AryehGregor> | Anyway, I was objecting to your use of the term "random" as though there were only one way to sample a "random" set of pages. |
| 20:48 | <othermaciej> | there is a limit to how much you can control for that |
| 20:49 | <othermaciej> | one possibility is to guess what variables may be correlated, and try to sample from subpopulations to preserve the right ratio |
| 20:49 | <TabAtkins> | AryehGregor: You were being pedantic. ^_^ Like I said, you *know* what "random" means when used like that, and you also know roughly what sort of page distribution we're talking about, and have at least some idea of how to construct a reasonably "representative" set of pages. |
| 20:50 | <othermaciej> | calling something a "random sample" doesn't imply that there is only one way to do it, at least not in any statistics I've heard of |
| 20:50 | <TabAtkins> | othermaciej: Indeed. |
| 20:51 | <TabAtkins> | But still, unless one has reason to believe that the used population *is* biased in that variable, one cannot reasonably suggest that the study sampling from that population is biased. |
| 20:51 | <AryehGregor> | For some definitions of "random" there's no feasible way to construct a random selection of pages, particularly not if you include intranets. |
| 20:52 | <othermaciej> | intranets are for practical purposes outside the realm of science |
| 20:52 | <TabAtkins> | We typically don't include intranets for that reason. |
| 20:52 | <AryehGregor> | That's my point. :) |
| 20:52 | <othermaciej> | because general hypotheses about intranets are for practical purposes not falsifiable |
| 20:52 | <TabAtkins> | There may be a bias introduced through that, but one must also consider the population of disabled web users as a whole and ask if this bias is likely to affect them. |
| 20:53 | <AryehGregor> | Anyway. |
| 20:53 | <TabAtkins> | That is, is an average disabled web user going to be affected by a higher or lower prevalance of correct @longdesc usage in intranets? |
| 20:53 | <TabAtkins> | I'd argue not. |
| 20:53 | <Dashiva> | Counterargument: disabled users aren't average |
| 20:53 | <othermaciej> | you don't have to assume anything about intranets really, just accurately state the result as being about publicly available web content |
| 20:54 | <TabAtkins> | Are they non-average in a way that would bias them strongly towards using intranets, and thus being more highly affected by them than a normal web user? |
| 20:54 | <othermaciej> | then anyone who wants to make a separate hypothesis about intranet content has the burden of proof to demonstrate their hypothesis |
| 20:54 | <othermaciej> | if they feel that would provide relevant evidence |
| 20:55 | <Dashiva> | TabAtkins: If the assumption is that intranets are better at longdesc, I'd assume there would be a similar assumption about disabled users |
| 20:55 | <Dashiva> | But nobody is of course providing any data, just complaints about the data that does exist |
| 20:55 | <othermaciej> | that disabled users are more likely to use intranets? |
| 20:56 | <TabAtkins> | Dashiva: You would assume that, given that assumption, disabled users are more affected by intranets than non-disabled people? |
| 20:56 | <othermaciej> | that might be a testable hypothesis |
| 20:56 | <TabAtkins> | That sounds testable, yeah. |
| 20:56 | <TabAtkins> | I doubt it's true, and don't believe it's relevant for our purposes, though. |
| 20:57 | <Dashiva> | I think the strongest argument against sample bias is that nobody has produced a sample biased the opposite direction |
| 20:57 | <TabAtkins> | (Rather, I doubt it's true, and *strongly* doubt it's true *enough* to matter to us.) |
| 20:57 | <TabAtkins> | Dashiva: I'll bet Shelley could produce one. |
| 20:57 | <Dashiva> | And a real sample, not just a dozen sites |
| 20:59 | <othermaciej> | well like I said before - for studies to be getting these results through sample bias, if the true situation is much different, the oversampling factor for bad longdesc values would have to be huge |
| 21:01 | <Dashiva> | There might not be enough longdescs in existence :) |
| 21:01 | <TabAtkins> | Well, 1 in 1000 pages contain @longdesc, apparently. That's a decent population. |
| 21:02 | <TabAtkins> | (Rather, that's the proportion given by one study.) |
| 21:02 | <TabAtkins> | So, given the estimated 1e12 pages on the web, about 1e9 of them contain a @longdesc attribute. |
| 21:03 | <othermaciej> | one study specifically identified 1.3e6 img elements as having a longdesc attribute |
| 21:03 | <TabAtkins> | Yeah, out of roughly 1e9 pages sampled. |
| 21:03 | <othermaciej> | hey, HTML WG made daringfireball |
| 21:04 | <Dashiva> | Yeah, but that's just attribute presence |
| 21:04 | <Dashiva> | Conforming, useful instances are a tiny fraction of that again |
| 21:04 | <othermaciej> | point being - that's a lotta longdesc, most of it apparently bogus |
| 21:06 | <othermaciej> | Ian's study estimates that something like one in a million images have a potentially useful longdesc |
| 21:08 | <TabAtkins> | AryehGregor: Bertrand's paradox doesn't apply here. In there you have a continuous variable that you're attempting to sample, and so the distribution you use to sample it matters quite a lot. A discrete variable, on the other hand, can be given a very simple uniform distribution. The example given by Bertrand matters because there *is* no uniform distribution over all possible chords of a... |
| 21:08 | <TabAtkins> | ...circle. |
| 21:09 | <TabAtkins> | However, I unfortunately will have to leave before I see a rebuttal, as I've got to head out to meet with my parents. Feel free to disagree with me, though, and I'll see it when I get back. ^_^ |
| 21:09 | <AryehGregor> | That's correct, but you have no feasible algorithm for uniformly sampling pages on the web. |
| 21:10 | <TabAtkins> | If I have all the pages, I certainly do. I don't have all the pages, of course, so we can then argue over whether the pages I have access to are representative, but that's a different issue I think. |
| 21:11 | <AryehGregor> | So the same general principle applies: you're suggesting someone sample something at random, but since uniform sampling is impossible (practically in this case, logically in Bertrand's), it's not clear what distribution you're asking for. |
| 21:11 | <AryehGregor> | The obvious algorithm would be to collect all the pages and then sample in a usual fashion, yes. But presumably you were not suggesting that Shelley do that. |
| 21:12 | <othermaciej> | studies do random samples of persons residing in the United States, but I am not sure anyone has an accurate, up-to-date and complete list |
| 21:14 | <othermaciej> | nor is every person in the United States equally convenient to reach |
| 21:15 | <AryehGregor> | They try to find representative cross-sections, but only by carefully controlling certain variables and hoping those are the relevant ones. The people who agree to participate in studies aren't representative, after all. |
| 21:15 | <othermaciej> | usually you try to estimate how much your sampling methodology might correlate with the variable of interest |
| 21:15 | <othermaciej> | and apply that to your error estimates |
| 21:16 | <AryehGregor> | Probably this all sounds more reasonable to a scientist than to a mathematician. |
| 21:16 | <othermaciej> | you don't generally say "I am not sure if my sample is perfectly representative, therefore I assume my error bar is unbounded" |
| 21:16 | <AryehGregor> | (a pure mathematician, specifically) |
| 21:16 | <AryehGregor> | Well, no, that's a silly thing to say. |
| 21:19 | <othermaciej> | perhaps Ian should have reported an estimated error for his study, then one could debate the size of the confidence interval rather than the statistically dubious question of "valid or invalid" |
| 21:20 | <othermaciej> | I believe he studied physics so I assume he knows how to do that sort of thing |
| 21:30 | <jgraham> | TabAtkins: It is worth noting that "all the pages on the internet" is a countably infinite set of pages |
| 21:32 | <AryehGregor> | jgraham, naturally you have to define it as something like "all pages viewed in the last X days". |
| 21:33 | <jgraham> | AryehGregor: Theoretically, yes. |
| 21:33 | <jgraham> | Although that's not a very practical thing to do |
| 21:35 | <jgraham> | Instead you get samples like "all the pages in the Google index" and assume that doesn't contain, say, 50% identical pages from a single calendar site that has an infinite number of possible pages |
| 21:36 | <jgraham> | Which is a reasonable assumption because an index that did have that kind of bias would be utterly useless as the basis for a search engine |
| 21:36 | <Dashiva> | I wonder how Google actually handles that. A maximum number of pages per domain, maybe? |
| 21:36 | <AryehGregor> | I imagine it's complicated. |
| 21:37 | <jgraham> | I guess their bots only go so deep |
| 21:37 | <othermaciej> | is there really such a thing as a server that could produce an infinite number of pages, even in theory? |
| 21:37 | <jgraham> | othermaciej: Well it depends what you mean by "pages" presumably |
| 21:38 | <othermaciej> | surely there are resource limits that fall short of the infinity point |
| 21:38 | <AryehGregor> | othermaciej, if you allow an unlimited amount of time, sure. |
| 21:38 | <AryehGregor> | Assuming there's no reason the server can't exist for an unlimited amount of time. |
| 21:39 | <AryehGregor> | E.g., no heat death of the universe or anything like that. |
| 21:39 | <jgraham> | If you take a HTTP view of the universe where each thing at given URL is the same even if it is actually different then you are limited by the number of unique urls |
| 21:39 | <AryehGregor> | jgraham, what about POST? |
| 21:39 | <othermaciej> | I was going to say at the very least it has to fit the whole URL in memory, thus can serve resources for only a finite number of URLs, but I suppose it could stream the URL |
| 21:39 | <AryehGregor> | Well, also pages can have different representations. |
| 21:39 | <AryehGregor> | I mean: <?php while (true) { echo rand(); } |
| 21:40 | <jgraham> | AryehGregor: In what context? |
| 21:40 | <AryehGregor> | Or for more elegance: <?php while (true) { echo rand(0, 9); } |
| 21:40 | <AryehGregor> | That's a lot of different pages. |
| 21:40 | <AryehGregor> | Clearly we must be talking about the actual streams of bits, not just the URLs. |
| 21:40 | <othermaciej> | but a finite number |
| 21:40 | <AryehGregor> | After all, one document at the URL might contain longdesc, and another not. |
| 21:41 | <AryehGregor> | Why is it a finite number, if there's no buffering and the server is given unlimited time? |
| 21:41 | <AryehGregor> | Heat death of the universe? |
| 21:41 | <Dashiva> | The number of URLs is only limited by server restrictions, though |
| 21:41 | <othermaciej> | ah, I failed to think of the while(true) |
| 21:42 | <othermaciej> | I think to have a definition suitable for a study, you need to sample pages as they are at the time you visit a URL, because even normal pages can change and you don't want to sample over and over til the Web stops changing |
| 21:42 | <othermaciej> | that would likely be a long wait |
| 21:54 | <Philip`> | othermaciej: "the size of the confidence interval" - if I remember correctly that's proportional to 1/sqrt(n), and if n=1e9 then it's going to be pretty close to zero and uninteresting and completely insignificant compared to even tiny systematic biases |
| 21:54 | <othermaciej> | Philip`: yes, the standard error is tiny if you assume no sample bias |
| 21:55 | <othermaciej> | or rather, I should say, the standard error is tiny, and if you don't assume sample bias or measurement error or something of that nature, that's about the only thing that factors into your confidence interval |
| 21:57 | <jgraham> | I think that is almost always true though, in the sense that if you are in a regime where your measurement is only just significant within random errors, it is almost certianly insignificant due to systematic errors you weren't accounting for |
| 21:57 | <othermaciej> | Philip`: I did compute the standard error, and both with the 1 billion sample of pages and the 13 billion subsample of pages containing longdesc it is very small |
| 21:57 | <Philip`> | Is it possible to come up with a confidence interval based on any other kind of error, that is not completely meaningless? |
| 21:57 | <jgraham> | ("90% of 2 sigma results are noise") |
| 21:57 | <othermaciej> | well, you could assume a statistical distribution of sample bias, but I have no idea if that is meaningful |
| 21:58 | <othermaciej> | jgraham: "longdesc value is bogus" seems to be a 6025 sigma result |
| 21:58 | <jgraham> | othermaciej: Sure, I am not suggesting this result is in that regime |
| 21:58 | <othermaciej> | or rather, longdesc value is bogus more often than it is useful |
| 21:59 | <Philip`> | Seems the more fundamental problem is that nobody can even define what population they care about bias from |
| 21:59 | <othermaciej> | sorry, 6525 sigma or so |
| 21:59 | <jgraham> | Philip`: That itself may be telling |
| 22:00 | <AryehGregor> | 6525 sigma? Seriously? |
| 22:00 | <Philip`> | except for some easily definable populations ("all the pages in Google's cache", "all the pages in dmoz.org", etc) where it's trivial (and in some cases you don't even need random sampling because you can test the entire population) |
| 22:00 | <othermaciej> | which by the way, if you compute the confidence level for that sigma level, I think the number by which it is less than 100% probably doesn't have a name |
| 22:00 | <AryehGregor> | Five sigmas is enough |
| 22:00 | <othermaciej> | AryehGregor: that's the number of standard errors out 50% valid londesc would be, given Hixie's study |
| 22:00 | <AryehGregor> | I've heard that five sigmas is enough for particle physicists. |
| 22:01 | <othermaciej> | I am not really sure how you statistically model the possibility of sample bias |
| 22:01 | <Philip`> | s/trivial/trivial to get a uniform random sample/ |
| 22:01 | <jgraham> | AryehGregor: Yeah, that sounds about right |
| 22:02 | <jgraham> | (well people will believe 3 sigma results too but not really be that surprised if they vanish) |
| 22:03 | <jgraham> | (so if you were to propose for funding on the basis "we could measure X to 3 sigma given ideal performance" there might be some skepticism) |
| 22:04 | <jgraham> | In fact I think I might have had applications for telescope timed down on roughly that basis |
| 22:04 | <jgraham> | s/timed/time turned/ |
| 22:05 | <othermaciej> | how do physicists deal with the fact that there may be systematic bias? |
| 22:05 | <othermaciej> | assume irrelevant if you've got enough sigmas? try to eliminate? guess and account for in calculations? |
| 22:06 | <jgraham> | othermaciej: If you know about it, I think it is close to "guess and try to account for it" |
| 22:07 | <AryehGregor> | Bug tracker down? http://www.w3.org/Bugs/Public/show_bug.cgi?id=8784 |
| 22:08 | <jgraham> | But often you simply don't know; I remember being told that if you plot the best measurement of the speed of light over time it is flat wihin error bars for short periods, then jumps by more than the nominal error bars then is flat again, and so on |
| 22:08 | <jgraham> | I seem to remember not being able to find a URL to back that up though |
| 22:08 | <AryehGregor> | I've heard similar things about Millikan's oil-drop experiment. |
| 22:09 | <othermaciej> | this talk of experimental error reminds me of this old joke: http://www.gdargaud.net/Humor/OddPrime.html |
| 22:09 | <Philip`> | jgraham: That sounds like proof that the speed of light is changing |
| 22:09 | <othermaciej> | AryehGregor: I think the measurement of the fundamental charge too a slow crawl from Milikan's measurement to the current value |
| 22:09 | <AryehGregor> | othermaciej, that's the kind of page that should have stopped after about five lines. |
| 22:10 | <othermaciej> | all the versions of the joke I have heard stop at the engineer |
| 22:11 | jgraham | is ashamed he knew waht the joke was just by reading the URL |
| 22:11 | AryehGregor | isn't |
| 22:11 | <AryehGregor> | othermaciej, me too. |
| 22:11 | <othermaciej> | although I like the Programmer entry on that page |
| 22:11 | <jgraham> | And you heard that URLs are just opaque strings... |
| 22:12 | <jgraham> | (I had also only seen the first three) |
| 22:35 | <hober> | I've made a bunch of progress on my zero-edit counter proposal for ISSUE-95 (hidden=""). |
| 22:35 | <hober> | If anyone's interested in reviewing / commenting, I'd really appreciate it. |
| 22:35 | <hober> | http://hober.jottit.com/ISSUE-95_Change_Proposal |