Posts by that one in the corner 5493 publicly visible posts • joined 9 Nov 2021
> Systems that use bases other than 10 evolved for sound practical reasons seem to be regarded as superstitions by those who worship the arbitrary decimal systems. Arguably it is the decimal system which is the superstition.
God grief, we seem to have reached the point that you agree totally with my original thesis but we communicated that so badly that both insulted the other's intelligence!
What a marvel the Internet is.
Really can not agree with that.
Given that we had people who were *really* "bad at maths" but were still very happily using old money and were (are) flummoxed by decimal coinage simply not being as easy to use. Many were average humans who just happened to be very young and learning to be numerate, a few were adult and just not numerate for various reasons.
Because when you really are "bad at maths" - which really means "bad at the written down symbolic stuff" - they could, still do, go back to *actual* genuine maths and counting theory: correspondence.
E.g. they had a pound note and wanted to pay three children equal amounts for pocket money: they'd get specie* for the note, (didn't matter the denomination). Then they just had to make up equal piles, one per child. Literally. Put the recipients in a row and just dish out the coins, creating piles in a 1:1 correspondence to them. Trying to claim that the benefits of multiple divisors, on the one area of calculation that everyone has to encounter, no matter their ability, is of no more value than "aesthetically pleasing to know, easily, how many thruppences were in a Crown" is absurd, short sighted, lacking in historical as well as basic arithmetic understanding just for the sake of some imagined jibe at the older generations.
* and before any of the untrusting people start to claim that everyone who was asked to make change for these people were "obviously going to rip them off", there are ways for the totally innumerate to verify; at worst, they are slow at the bank counter (remember, back in the days of lsd you could go into a bank and get lifelike service).
The only "reason" that we so often multiply or divide by 10 is because that *is* "10" - the base of the counting system that we happen to be using - and that it is easy to do so. It makes no difference what base we decide to use everyday, we'd always end up doing lots and lots of multiply/divide by 10 or 100 or 1000 - or 0.1 or 0.01 or 0.001
Your examples "look different" (and hence makes your point - whatever it is) because you decided to write out the second one by deliberately changing the units used.
If we were actually using as our everyday base "the number you get when you cut something in half, then cut all those in half, then cut all those in half and finally cut all those in half" then your second example would naturally read, sticking to one set of unit names instead of jumping to special subunit names,
"2. You have 1 pound of sugar. Weigh out 0.1 pounds"
And at the same time your first example would be upsetting (to the grade level at which you'd aim your two examples)
"1. You have 64 grams of sugar. Weigh out A grams."
(Of course, this is making the assumption that you would still know what a "gram" is, considering that it is clearly a measurement designed by a madman: you determine a particular atomic transition frequency, the speed of light and the Planck length - fine, no issue there, we are a scientifically literate species - then perform a really weird calculation on them -doesn't matter what base you use, it is going to be weird! - and call what you get a "kilogram". Odd name, doesn't make any sense, but who are we to argue? But then to get this "grams" value you divide your "kilogram" by 3E8 ! Madness!)
> He sent an email at the end of the first week when we scaled up complaining that the failure percentage metric had gone up by 25%.
Isn't the "failure percentage metric" == percent of the product that failed? If a batch of 75 has 25* failures then you have a failure percentage of 33%** If you had the same failure percentage and a batch size of 100, you'd expect 33 failures.
If the failure percentage went up by 25% then that rate is now 41%
If that change is also when your batch size went from 75 to 100, you are now getting 41 failures per batch instead of 25 per batch.
> ... 75 to 100 is 25%
Perhaps the first failure is in thinking that "Failure percentage metric" actually involved percentages?
Maybe the metric really have been called "Absolute failures per batch"?
"What we've got here is... failure to communicate. Some men you just can't reach".
* not saying yours were anything like that bad, just want to play with a few figures here
** figures liable to crude rounding
The Romans, in their neverending quest to subsume every local religion, had a go at mapping the Norse gods onto theirs, and therefore the planets, giving us Woten == Mercury (apparently). Leading eventually to the French getting their spelling of Wednesday completely wrong (I mean, the first letter is close, but it is upside down!).
What is sad is the number of people who defend decimal as simpler than any other system because "to multiply or divide by 10 all you have to do is move the point". And get angry when you point out that that exact statement is literally true for every single base, from binary upwards. Then it devolves into "but 10 (decimal) is obviously a better number than 10 (any other base)" and "but why would you ever want to divide by 10 (any other base); you can see that is plainly silly because just look how often we divide by 10 (decimal), you are going to make that so much harder"!
Which is why it made perfect sense to teach them pounds shillings and pence.
Kids will happily count money and how many gobstoppers you can get for a tanner if they are a ha'penny each. Figuring out that getting fourpence a minute will net them them a quid in a hour just comes naturally to the mercenary brats ("come on mister, thruppence a box means you're gonna need to break change at lunchtime and the banks close early today").
> customers using Rovo grow their annual recurring revenue at roughly twice the rate of those who do not
Is "their" disambiguating to "the customers" - or perhaps, just perhaps, he is referring to Atlassian, whose revenue from those customers who, willingly or not, are using the Atlassian AI and providing loads of extra recurring revenue to swell Atlassian's coffers.
Always check claims carefully and be sure the sentence really says what they hope you read it as saying.
Typewriters caused the problem first IMO: you did get them with dipthong keys but it was cheaper to do without. So that started the acceptance of something that looked better than handwritten (for certain values of "better" and ignoring certain superb letterists) but which lacked all the nuance of proper English.
Computerised typesetting, on the other hand, has always had the ability to cope with dipthongs, that odd accent on "naive", the capital S written as two lowercase f's and every other weird and wonderful ligature that a literate publication could ever need: Don Knuth made dang sure of this when TeX was released around 1978, when suddenly far more lovely manuscripts could be produced than there were suitable computer-driven typesetters that could actually realise them! Heck, there weren't even enough graphics displays to give everyone a chance at fully enjoying the fruits of their labours.
But, yes, of course, even as the hardware became available, lesser systems were created and used. With the introduction of desktop publishing, lazy/ignorant journos who couldn't be fagged using the full-size typesetter's keyboard or even look for the ligature in the list became the norm. Select corners of science and maths textbooks, the ones whose authors generated their own galleys and provided camera-ready sheets to the publisher, seemed to be the bastion of beautiful typography. And the glaringly awful in equal measure (luckily a nicely-laid out stats book will be able to illustrate the curve that explains this phenomenon).
Nowadays, it seems to be pot luck whether the combination of author, editor (if they have one) and book designer (if they have one) will result in correctly expressed English or not.
All of which ignores the pressure to create English written in a way that the American market will accept, as the creation of translations for use across The Pond is increasingly deemed an excess expense (as well as a risk to the poor defenceless sentences), which means the use of spellings that have viciously cleaved the unwary dipthong and crushed one half beneath the bootheel of "simplified" spelling (and the resultant jarring pronunciation as the delicate gliding vowels are forgotten: now out of sight and hence out of mind).
Going offtopic (What? In an El Reg comment? Unheard of!) but
> there are things like Lovable
Not knowing what "Lovable" is, did the obvious thing and DuckDuckGo'ed it. Nothing surprising in the results (oh, it is for "apps" and "web apps", ok, that will help get the randos trying it and fulfilling Kerfufflinator's observation).
Then, still on the first page of search results, the website that offers to
>> Create optimal Lovable prompts instantly. Describe your application idea, and our free generator will craft the perfect prompt for deployment-ready full-stack apps.
Something is going wrong here, can't quite decide how to express it...
Collaborative real-time editing of code, where your typing shows up on both screens at once (e.g. ref CodeWright from approximately 1990) requires that the feature be present with Zed.
But making a voice or video call - there are plenty of options for that, it does not require Zed to create yet another! Provide the option to invoke whatever your (IT group's) preferred method is, certainly, but to open the microphone - can you spell "bloat"?
You completely missed the point.
As you say, a literature review is a PART of academic work - e.g. to get a handle on the current state of the field and to check that your "brilliant new idea" wasn't done to death back in 2017.
But the character Asimov described was ONLY capable of literature review and actively shunned doing any new work in the field. For archaeology, classically so, compared to other endeavours (bar its cousin, paleontology) literal field work, brush in hand, is key.
Even if one made an entire career out of performing meta-analyses (and, if you read the story - or consider the comparison being drawn to the real-life pronouncements by Yegge - calling their behaviour "meta-review" is stretching a definition to breaking point) one may be a useful adjunct to the field but not at its peak: running the stats is important but not the be-all and end-all.
If I ask a toddler to "write code" and leave them with the keyboard, my time spent is zero and I get thousands of line of code (big chunky old-fashioned CR key, the size of a little fist, makes a nice thud and is fun to hit).
Now, the time required to make that "code" do anything (so the proud parents can post on FB), let alone do it correctly, scores rather differently.
Sigh.
To own a proper pen that also digitises has been a dream of mine for decades: scribbling notes in the lecture series we still attend and having them available for sharing (you'll have forgotten the subject matter before I get around to inputting it myself - lawd, I hate copy typing).
Many, many such devices have been put onto the market since the 1990s, often using special pads printed with locating marks, sometimes with other, more complicated, methods. But all falling short of being, well, useful. Practical. Not tied into the latest proprietary scheme. Including nowadays, of course, the obligatory 'Phone App (not guaranteed to work on your 'phone for more than three months).
One could think that a company the size of Meta could actually manage the engineering to create such a peripheral, even if only to spy on their own staff. But considering how much they spaffed on all the Meta-crap you know they can't.
The best alternative so far has been the Palm Pilot - before HP stuck their oar in and buggered up Graffiti. Sigh.
Today I Learned...
that the US describes the thickness of leather in ounces, where 1 oz equals one sixty-fourth on inch; at least now I can convert the inches to mm (turns out, divide "ounces" by 2.5 to get mm, near as makes no difference for a material that varies across its area).
That is it. None of the sites I double-checked this on had any clue to shed on *why* ounces equate to inches this way. Probably due to some mythic piece of leather of fixed size, one link down the side, that weighed one Troy ounce and measured to be twenty-four sheets of leather to the barleycorn. Then it was rationalised when US units broke free from Imperial.
> The passkey is generated on your device! The only thing shared with the service you are using is the public part of the key! The private key is never stored anywhere except on your device!
Done properly, as with any public/private key system, the private key key should be generated and stored solely locally - and should from that point on be easily and clearly located and managed locally.
SSH being a good example of that - and you'll find plenty of people here who understand SSH (as well as, IMO, a growing proportion who don't understand SSH, even amongst those who actually use it). Similarly PGP, although (again, IMO as we don't have figures) there are fewer who know PGP than SSH.
BUT as with everything, these things can be done improperly: it is perfectly simple to set up a web site that will offer to generate SSH and/or PGP keys for you. Of course, you would have to be insane to go to a website and let it generate private keys for you! Who knows what a random web page is doing?! It *might* be running entirely local JavaScript and not doing any more comms to the outside world after the page load is done, but can you be sure of that? And can Joe Bloggs?
Now look at your web browser as it offers to create passkeys for you.
Gosh, doesn't that look just a web page that is offering to generate private keys for you?
Now, one or two of those offers *do* look sort of like part of the browser's own UI, very similar to the one it uses when offering to save your password for this site (only used for low to zero value sites, of course). Note the "looks". Others appear in the middle of the page, with varying appearance - now, are those happening purely locally? Even the ones that are branded?
> The ignorance on this site is astounding!
And do *any* of those popups provide any guidance whatsoever on what passkeys are, how they work, where you can find any reliable information on all this? Nope. Not that I've been offered[; have you fared better?[1] OTOH if you feel like getting into PGP you can easily (well, it used to be easier, but "modern web design" demands all content be hidden) find out how it works, the web of trust, how it is definitely *you* who own and manage your keychain file. Actually, let me modify that: GPG is all in the open, PGP less so (getting corporate) and anything about digital signatures from Certain Big Corporations is - oh, look, pushing you towards keeping your keychain (not that they'll call it that) on their Cloud "for your convenience"...
> The passkey is generated on your device!
Okay. Let's assume that all these players are genuinely doing it properly (citation!). And the lack of trivially accessible documentation is just forgetfulness on their part (unlike, oooh, ssh -h). That all of these apparent webpages are totally benign as they create your private keys.Then:
> The ignorance on this site is astounding
How about you provide the references, at least to get us started.
* If they are locally generated, how do I generate one? Can they be securely generated by me, using demonstrably local commands?
* Where are they stored? How do I access my own set of private keys (e.g. keychain file)?
Etc etc
[1] Yes, yes, you'd hope that anyone on a techie site like El Reg would have looked up all this stuff, as something to fill the long Winter evenings, but - well, ref my opinion on how many here grok SSH.
> you can buy a USB fob to store all your passkeys
Assuming this isn't just a memory stick (otherwise you'd've called it that)...
Now to identify a genuine and reliable source for one of these USB fobs, verify it works (without giving it access to any real passkeys, just test data - ugh, how does one arrange a test setup for this stuff?), verify that I can copy the data to another one if these fobs (one day, it'll die, at best); what about when I change Wonder if I can issue a "destruct" or "disallow" whe the fob is nicked or otherwise lost?
Not as bad an issue as the sudden reliance on Bluetooth (at least this fob is an "optional" extra).
Although you do seem strangely committed to the idea that, if someone, say a bank, suddenly requires me to use Bluetooth it is entirely my fault that that banking service is now out of my reach.
> If you are testing a tool that is supposed to be able to do a particular task, you don't manually do that task before you put the tool to the test.
If you are testing a tool against a particular task, *before* you run it you need to know what the success criteria are. Otherwise you have no means of knowing if the tool was a success.
To determine those criteria, it may well be sensible (or even absolutely necessary) to manually (or whatever "the old way" is, e.g. last year's milling machine) do the task first on one or more data sets, simply in order to have answer sets to compare against. Especially if you are claiming to measure the new tool against the manual/previous method.
> If running an objective test, you would surely run mythos and then manually look for any flaws that it may have missed ...
Not only for missed flaws *but* but also that the claimed "flaws" are, in fact, flaws.
> ... afterwards?
You can not "manually look..." AFTER the objective test! If you have not done that comparison then your test is not complete - in fact, the *test* has not even begun, only one part of the preparation for the test proper! That part being gathering one set of the values that will be tested against the "ground truth" of the manual/old-style run. The temporal ordering of those two runs isn't relevant, of course, as they ought to be independent of each other (otherwise you have more weirdities that your protocols need to be modified for and carefully described in the test report).
And the LLMs are tireless, relentless. Even the human intervention during the process ("and about ~20 hours of me unsticking it from dead ends") could be automated away: there are whackily crude search algorithms that you would never apply when you understand the search space, but if you don't happen to have on-hand "anyone skilled enough" to provide a good definition of a "dead end"...
Because, what the heck, you are a Bad Guy looking for The Big Score, the payoff may be huge compared to the cost - and you are using stolen credentials to "pay" for the compute anyway (remember, Bad Guy!) so sod it, let it run.
Based on current marketing guff, that could be anything from one of the usual suspect bitbarns to a Silicon Valley "artisanal ice cream sundae*" supplier
* "Our sprinkles have artificial intelligence but no artificial colors"
> The real question is what
is the actual, day-to-day use of LLMs for answering medical questions?
Oh, that would be the general public, asking whatever models they can get their hands on, by a long, long way. Joe Bloggs asking why he has a headache. And look, the *worst* behaviour is at the *start* of the process, when Joe is the most easily misled.
That sounds like a pretty "real" situation to be concerned about.
> Additionally, the research is now even more obsolete because all of the providers have newer models that perform better
Citations! Do these newer models perform better for this use case? Where is the data to back that up? Or are they just "better" at everything, because of course they are? (And take care answering that, as we have a history of newer models being worse on all sorts of metrics, like sycophancy).
Not that that matters in the case of Joe Public (see above) as he doesn't get access to these fancy-schmancy models, not unless he has paid the subscription fee.
> we can mandate and scale the best AI diagnostic tool
And this tool happens to be - which? Is it the arbitrary LLM that is claimed to be trained more on proper medical texts than the output of Dr Google and Consultant Reddit? Is it "the best" choice from the pile of junk or "the best" because we've taken the project as seriously as we do building a bridge, putting details of every bolt onto the plans?
And after you've mandated its use (oooh, loadsa dosh going to the bitbarn owner who can convince the regulators that their's is the best) what is done to it for the next release? A total rebuild because we found a way to tag Dr Google results as something to deprecate or a new UI with a flat toolbar?
Absolutely.
What you describe is known as "using tools for a job they are good at" (or, at least, "jobs they are our best bet at so far, but that is a low bar, treat their results with care").
Use machine learning to discover/refine/verify patterns. Make it write out what it has found, in excruciating detail, with stats and cross-references back to significant inputs. You can use LLMs, as you indicated, to preprocess plain-texts, but double and triple check what they generate (you can afford to do this, as the preprocessing is done once, we aren't going to use this LLM for every query of the final system).
At the end of it, you get a load of rules, with statistical weightings and references to how everything was generated. You can then use these rules to ask questions to reach a diagnosis. A diagnosis that can be interrogated and explained. Even modified in the light of the explanation (which becomes another datum, with citations, for an iteration of the ruleset).
As you point out with the reference to IBM's older attempts, this is not a new concept - go back to the 1970s and 1980s "Expert Systems", for starters, then work your way along, thinking about what we *could* have built by now, bringing all these things together.
But, of course, all of the above takes time, effort, isn't commercially viable, it can't be done, etc etc.
Meanwhile, talking about tools for the job, here is our expensive chatbot hammer: it isn't very good at screwing, unless we swing it *really* hard, like THIS! Now you are properly screwed - over.
> The tendency ... take this ... come back in a week ... so maybe it is this...
The GP is doing proper practical diagnostics, which includes (but is not limited to) short-term prescribing to:
- alleviate gross symptoms to allow specific differentiators to become visible
- tackle the most likely or most immediate risk
- starting the process of determining how you, as an individual, react to the choice of available treatments and/or the dosage
and, in no small number of cases,
- dealing with hypochondria by providing relief from non-threatening symptoms, if not an outright placebo ("take two aspirin" is sound advice in so many cases - yes, even where long-term use of aspirin is contra-indicated due to slightly raised bleeding risk, the relief from just those two will outweigh); although this is the case where it is likely that the GP will make notes but not actually tell you why this course is indicated...
If, as your comment seems to indicate, you are regularly encountering assorted GPs who don't take the few seconds to tell you which situation applies in your case, then you've been very unlucky with your GPs. Of course, I'm not going to suggest any alternative, such as you've been inattentive or even that you may fall into the "no small number.." category, but if you were including in your "tendency" anecdotes from everyone down the pub, perhaps some of them...
