This web site is protected by copyright law.
Reusing pictures or text requires permission from the author.
For more topics, scroll down, press Ctrl-F to search the page,
or check previous months.
For the latest edition of this page at any time,
create a link to "www.covingtoninnovations.com/michael/blog"
2023 February 19
Chatbots yet again
Note about the term chatbot: In what follows I am talking specifically about
new-style generative chatbots that generate text not controlled by any specific payload of information to
be delivered. The term "chatbot" actually means any computer program that accepts questions
from you in plain English; chatbots in that sense have existed since the 1960s.
As long as they deliver accurate information, they can be handy, and generative techniques
can be helpful on the input side, to recognize different ways of wording questions. The answers, however,
must be controlled; for example, if you ask a chatbot your bank balance, it must actually
look at the bank's records and retrieve the true answer, not make up an answer at random.
We seem to be having a public-opinion disaster, or mass panic, as people read texts
written by chatbots and imagine they are the product of a conscious mind. Here are
some of the moment's news headlines. (This is not a clickable link, and anyhow,
they are changing very fast.)
I think even the people who build and deploy chatbots such as ChatGPT have lost perspective on what they are.
You build a text imitator that can imitate, more or less at random, any kind of text it has seen; you feed it all kinds of English-language texts; and then it starts talking like any and all of them. What did you expect?
And the general public is in a panic because they think the chatbots are actually thinking.
Some people can't read words without believing them.
So if a machine says it's conscious, they believe it's conscious. That's not how this works.
Anyway, I have grave doubts about the usefulness of a gadget that generates text at random — often truthful, because it often imitates reliable sources — but it can't be relied on to do so, or even to indicate whether it is doing so.
Running in the other direction, such a gadget would be excellent as a grammar and style checker, simply judging whether something is good, natural English, not whether it is true. Has anyone thought of using it that way?
An hour of the comet's motion compressed into 25 seconds
Here's a better video of Comet C/2022 E3 (ZTF) moving in front of the stars.
On the evening of the 13th, I took 50 one-minute exposures of it, with short fixed
intervals between them, so that they span about an hour. Here they are as a video:
Same setup as before: Nikon D5500 (H-alpha modified), Askar 200-mm f/4 telephoto lens,
Celestron AVX mount, PEC turned on, no guiding corrections, in my driveway in Athens, Georgia.
Workarounds for the problem described below with
cameras I've tested:
Nikon D5300: The off-axis-color problem is absent when not using an electronic lens.
To avoid it with an electronic lens, cover the contacts carefully with tape and use
the lens manually. To avoid lossy compression, expose the sky background and the flats no higher
than 1/4 of the histogram.
Nikon D5500: The off-axis-color problem is present in the blue channel; if it is
noticeable, then at the last stage of processing the stretched image, clip (threshold)
the blue channel (that is, cut off the left edge of the histogram in blue only).
To avoid lossy compression, expose the sky background and the flats no higher than 1/2 of the
histogram, but also not a lot lower, because generous exposure reduces the off-axis-color
problem, even if requires turning the ISO up higher than I originally recommended
(e.g., 400 or 800 rather than 200).
We've known for some time that the raw images from a Nikon DSLR are not truly raw;
corrections and slightly lossy data compression have been performed on them.
Nonetheless, I regularly get good astrophotos with my Nikon D5500.
However, one of the corrections (for off-axis color sensitivity) can look like a faint
circular reflection in the image, and I was bitten by that the other night.
See the blue ring that encircles the middle of the picture? I had to limit the amount of
contrast stretching
to keep it from showing up. Here, of course, you are looking at an extremely
overstretched rendering of the image
Where does it come from? I knew already that Nikon's raw images do a small amount of image compression
that can affect the application of flat fields (pictures of a featureless white surface, which we take
in order to correct for edge darkening and even dust motes in our optics).
But this isn't that. It turns out this is a strange correction that is applied by Nikon because
the sensor supposedly has slightly different color rendition away from the center.
Lengthy discussion of these Nikon quirks, conducted largely by "Sharkmelley," has taken place
on Cloudy Nights Forum; click here
and here.
Note that in my case, this flaw was brought out by extreme contrast stretching of the image.
Basically, I took a severely underexposed image of a very low-contrast target.
The rings popped out when I increased the contrast by a factor of about 1000.
Such things happen in astrophotography — but, realistically, this was not a properly
exposed image; I should have set the camera to ISO 640 or even 1600; and I should not have performed such
an extreme stretch. If I had exposed more generously, or even set the ISO higher, I would not have had to.
What I've learned is that my old idea of taking minimally exposed pictures at ISO 200 (for maximum
dynamic range) is not always the best way to use the Nikon D5500 sensor. It might well be if not for this
newly discovered quirk. But in fact we may need to set the ISO higher. And when we do, the difference
between this sensor and earlier sensors, such as my trusty Canon 60Da, is much less appreciable.
More of this is expounded at www.dslrbook.com
in the Updates section.
After further tests, I'm definitely keeping the Nikon — but note that I haven't tried
one of the newer, lower-noise Canons (EOS 80D or later), which are probably the best choice.
I'm still disappointed that Nikon doesn't offer us a "truly raw" option in the firmware,
for scientific work, so that we could use their excellent sensors to the fullest.
The astronomer who figured out that the planets go around the sun is 550 years old today.
For earlier Daily Notebook entries explaining how he made his discovery,
click here and here.
I remember his 500th birthday, 50 years ago today. I had recently done a science fair project
on Copernicus and was well aware of the date. At Brookwood Plaza in Valdosta I happened to run into
a VSC astronomer, undergraduate Jimmy Mitchell, who himself remarked on Copernicus' birthday
before I mentioned it. So at least two of us had noticed it.
The 45th anniversary of a very special Valentine card you sent me at Cambridge.
4500 more years would not be enough.
(Our St. Valentine's Day this year was interrupted by the sudden illness of a dog,
the sudden failure of both a computer and a carpet cleaner, and several other
minor calamities. We're going to celebrate it properly in a few days.)
This is one of two astrophotos that I took on the evening of the 13th.
The other, which I haven't processed yet, is of the comet.
This is a stack of 26 1-minute exposures with an Askar 200-mm f/4 astrographic telephoto lens,
Nikon D5500 (H-alpha modified) at ISO 200, and AVX mount with PEC but without guiding corrections,
all in my driveway. For pictures such as this I often use a nebula filter, but not this time.
I actually took 50 exposures, failing to realize that for the last dozen or more, trees
would be in the way.
One of Yale University's architectural misjudgments will soon be a thing of the past,
and I'm not sad to see it go.
Helen Hadley Hall is the dormitory I lived in for my first two years at Yale (1978-80).
For the first semester, I was one of three linguistics students who formed a straight line
in 3-dimensional space, occupying rooms 303, 404, and 505. (I was 404.) Then I moved to
a corner room, 124, which was much more comfortable, especially after I bought my own bed
and moved it in.
The only luxurious thing about Hadley was that everyone had a private room (actually normal at Yale;
the world's best universities don't put people two-to-a-cell the way American colleges have
traditionally done). But the building showed obvious signs of last-minute budget cuts.
I heard, I don't know how reliably, that it was designed to be air conditioned and then the
air conditioning was cancelled at the last minute. What I know is that it was not ventilated
(except by the windows in the rooms) and was uncomfortably hot long past the end of summer.
And it smelled bad!
The picture in the Yale news story shows the patio behind the building, which, in my time, was hardly used.
Because of the high crime rate, we were told to avoid secluded places, so almost the only time
I ever set foot on the patio was once when a party was being held there.
The front of the building looks very much the same and is very close to the street.
Here's an animation I made of 19 separate 60-second exposures of the comet,
taken at 75-second intervals (that is, with a 15-second pause between exposures)
and then displayed for one second each.
This was taken on the evening of February 6, and for some reason the video was viewed over 500
times yesterday (February 11) within a short period of time. (I don't know who drew a crowd to it.)
Each exposure was 60 seconds at ISO 200 with a Nikon D5500 (H-alpha modified) and Askar 200-mm f/4
astrographic telephoto lens, on an AVX mount with PEC and no guiding corrections.
Here's how I made the video:
(1) In PixInsight, using WeightedBatchPreprocessing, generated calibrated versions of all the images.
(2) In PixInsight, determined the proper histogram adjustment, field flattening, etc., for one image,
and then applied it to all the images using an ImageContainer.
(3) They were still not all equal in brightness because of passing clouds, the rising moon, etc.,
so to even them out, I used the PixelMath expression
$T-med($T)+.25
again applied to all of them in
an ImageContainer.
(4) Because of an image compression problem, I was not able to make the video in the Blink process
in PixInsight. Instead, I imported all the pictures into Adobe Premiere Elements and showed them
in succession there. Unfortunately, in Premiere Elements, it is not possible to set a picture length
less than one second.
Now for the surprise. My images of Comet ZTF on the evening of February 6 (February 7 Greenwich time)
also picked up another 16th-magnitude comet. Here they both are.
Follow the red marks for the fainter one.
This is a single 1-minute exposure, one of the ones that went into the video above.
The three bright stars are "The Kids" next to Capella in Auriga.
In case the fainter comet isn't prominent enough, here's a very
stretched, grainy rendering in which you can't miss it:
And finally, here's a stack of 20 exposures (all the ones in the video, plus one more), aligned only
on fast-moving Comet ZTF, turning the stars into other streaks and more or less losing the other comet.
I photograph the Orion Nebula often, and had not photographed it this year yet
(my astrophotography was almost shut down during the remodeling),
so I decided to take a shot at it on February 6, even though the moon was
rising. Stack of 20 exposures, Nikon D5500 (H-alpha), ISO 200, Askar 200-mm f/4
lens, AVX mount, PEC, no guiding corrections, some HDR processing in PixInsight.
The most important piece of science news this week, the thing we need to get out to the confused and misled public, is:
ChatGPT is not in any way a step toward making computers think consciously for themselves.
It's good at imitating a conscious person's behavior by paraphrasing English-language texts in a humanlike way.
But it literally does not know what it is saying (there is no knowledge representation in it, except for data about how words are used).
Its real technical triumph is that it demonstrates that grammatical English can be generated from massive
data about how individual words are used, without an explicit representation of the grammar.
That doesn't mean there is no grammar, only that it is implicitly present in all actual use of the language,
which, if you think about it, is no surprise, though it is technically important.
But that is not the same as conscious thinking.
It isn't anywhere near it.
ChatGPT is not a breakthrough toward "artificial general intelligence."
There is still no known technical path toward conscious machines.
They are as fictitious as unicorns or genies.
And then... In the past several days, hundreds if not thousands of people have asked me if I've heard of ChatGPT and
what I think about it. Answer: Yes, I've heard of it; I don't want to read 100 more articles by people who
think it's amazing but don't really know anything about it; and if you want me to do a detailed evaluation,
engage me as a consultant at standard rates.
And a bizarre thing has happened. Some of them claim that because I'm a computational
linguist and know a good bit about how ChatGPT works, I am uniquely UNqualified to comment on it!
I am supposedly blind to how it is "going to put 80% of knowledge workers out of business" and so forth.
Do we have a mass popular delusion going on?
Given our dodgy weather, this may be the best picture of Comet ZTF that I get.
(The comet is now receding.) This is a stack of twenty 1-minute exposures, aligned
on the comet, which was moving, so the stars are streaks.
Nikon D5500 (H-alpha modified, not that it matters) at ISO 200,
Askar 200-mm f/4 lens wide open,
AVX mount with PEC, no guiding corrections.
Is this a "rare green comet streaking across the sky"? You tell me.
I think it's the same color as most comets.
But the press has gone wild with it.
Their latest conceit is that the comet is "over" particular cities at particular
times ("Comet to pass over Houston on February 4" or the like).
That's not how any of this works...
This is a private web page,
not hosted or sponsored by the University of Georgia.
Copyright 2023 Michael A. Covington.
Caching by search engines is permitted.
To go to the latest entry every day, bookmark https://www.covingtoninnovations.com/michael/blog/Default.asp
and if you get the previous month, tell your browser to refresh.
Portrait at top of page by Sharon Covington.
This web site has never collected personal information
and is not affected by GDPR.
Google Ads may use cookies to manage the rotation of ads,
but those cookies are not made available to Covington Innovations.
No personal information is collected or stored by Covington Innovations, and never has been.
This web site is based and served entirely in the United States.
In compliance with U.S. FTC guidelines,
I am glad to point out that unless explicitly
indicated, I do not receive substantial payments, free merchandise, or other remuneration
for reviewing or mentioning products on this web site.
Any remuneration valued at more than about $10 will always be mentioned here,
and in any case my writing about products and dealers is always truthful.
Reviewed
products are usually things I purchased for my own use, or occasionally items
lent to me briefly by manufacturers and described as such.
I am no longer an Amazon Associate, and links to Amazon
no longer pay me a commission for purchases,
even if they still have my code in them.
