Different independent sources I’ve followed for a while converging on the same couple of points means there just may be some fire to the large language model smoke.

• Bob Carpenter at Statistical Modeling… blog: “12/12 [queries] answered sensibly”
• Matt Birchler: “I’m seeing some people look at this ask just the next NFT fad that comes and goes by the summer, but I’m not so sure. There’s something real here, and I think things are going to get a lot weirder very soon.”
• Dave Winer: “Maybe there will be negative consequences of ChatGPT, but I’m sure we’re not in a position to see what they are now, based on experience with similar changes. And maybe we’ll look back on this moment twenty years from now, and not be able to imagine what life was like before we had this fantastic tool.”
• Ben Thompson: Just… read the whole thing.

Twitter is still ablaze with screenshots, and of course yours truly had a few thoughts to share yesterday. The best explanation of ChatGPT and large language models in general is still Stephen Wolfram’s.

What a fun, exciting, and scary time to be alive…

This time last year, there was a particular problem I was trying to solve for which blockchain seemed to be a suitable solution. Not knowing anything about those newfangled technologies but having some experience with various programming languages, I went out looking for a nice tutorial on the subject, akin to the wonderful Wolfram University introduction to Mathematica. No such luck: every video started with talks of decentralized finance and billon-dollar applications. ⊕ See here, here, and here for examples, and that last video is only a month old so it doesn’t seem that things have changed all that much. Some featured presenters who were clearly under the influence of… something.

These tutorials must have had a target audience, raking millions of views, but that audience clearly wasn’t me so I turned to a childhood friend whose path led him to a blockchain-adjacent space. After explaining my use case, what he said could be summarized as follows: You don’t need blockchain. Stay away from blockchain. What you need is a database that’s slightly more sofisticated than average and has strong permission rules.

Thus ended my brief and unproductive foray into Web 3.0.

Compare and contrast to ChatGPT, and as of a few days ago Bing Chat. When water starting coming up our basement drain during a not particularly heavy rain spell last week, ChatGPT had this to say:

If you are experiencing water coming up the basement drain and a backed up basement toilet, it is likely that there is a blockage in your home’s sewer line. Here are some steps you can take to try to resolve the issue:

1. Remove any blockages in the basement toilet: Check the toilet for any visible blockages such as excess toilet paper, sanitary products, or toys. Use a plunger to try to remove the blockage.
2. Check the clean-out pipe: Locate the clean-out pipe, which is usually located near the main sewer line and has a cap that can be removed. Remove the cap and see if there is any blockage inside. If there is, try to remove it with a plumbing snake or call a professional plumber.
3. Use boiling water: Boil a large pot of water and carefully pour it down the basement drain. This may help to dissolve any grease or soap buildup that is causing the blockage.
4. Use a plumbing snake: If the previous steps do not resolve the issue, you can try using a plumbing snake to clear the blockage. Rent a snake from a hardware store or call a professional plumber to do this for you.

If none of these steps resolve the issue, it may be necessary to call a professional plumber to diagnose and repair the problem. In some cases, the blockage may be located in the main sewer line, which will require more extensive repairs.

And the bot was right: our problem was number 2 on the list. Unfortunately, my puny battery-powered drain snake couldn’t do the job so professionals had to come in with some heavy duty equipment, but it sure got the differential diagnosis!

Or consider the case of our child’s pinkeye. Despite what you may have heard, plain viral conjunctivitis is no more contagious than the common cold, and other than hand hygiene and avoiding bodily fluids no special precautions need to be taken. But the school still requires a doctor’s note, and my wife and I both being physicians we were happy to provide one. After clearing with the administrator that it doesn’t need to be an independent pediatrician who’s writing it, I asked ChatGPT to draft one “writen as if your were the parent of the student and a doctor”. And presto:

[Parent’s Letterhead]

Date: [Date of Note]

To Whom It May Concern:

I am writing to confirm that my child, [Student’s Name], a kindergarten student at [School Name], has been evaluated and treated by me for a case of pinkeye (conjunctivitis).

As a parent and a doctor, I have prescribed the appropriate medication for this condition and I am pleased to report that [Student’s Name] is now symptom-free and able to return to school without any restrictions as of [Date of Return].

I understand the importance of maintaining a healthy and safe environment for all students and staff at [School Name]. I have taken all necessary precautions to prevent the spread of this condition, including hand hygiene and appropriate disinfection of commonly-touched surfaces in our home.

If you have any questions or concerns, please do not hesitate to contact me. I look forward to my child’s return to school and to the resumption of normal activities.

Sincerely,

[Parent’s Name] [Parent’s Credentials] [Parent’s Contact Information]

To be clear, this has saved me 15 minutes at most, but it would have been 15 minutes of mental effort better spent… playing around with ChatGPT some more, I guess?

Which brings me to the third important thing AI did for me this week. I got into the “Bing Chat” preview and, trying to gauge how helpful any of these tools would be, I tend to ask things to which I already know the answer, but that took me a bit of research to find out. My go-to is this: How many CD8+ T lymphocytes are there in the whole human body? People’s — and ChatGPT’s — first instinct is to multiply the absolute number of circulating CD8+ T cells per microliter, a number that’s readily available, with the blood volume of the average human. But that only gives you the number of circulating cells, not those in lymph nodes, spleen, and other tissues, in which most of the body’s lymphocytes reside.

And indeed, that was Bing Chat’s first response: ⊕ I should note that all of Bing Chat’s response’s, unlike ChatGPT’s, included appropriate and correct references. It is a search engine after all.

The normal count of CD8+ lymphocytes in the human blood should be between 200 to 800/mm3. The ratio of CD4+ to CD8+ lymphocytes should be 2.0. A higher or lower count or ratio may indicate an infection, an autoimmune disorder, or a cancer.

After some redirection, ⊕ My exact prompt: “You gave me the number of CD8+ lymphocytes in blood. How many are there in the whole body?” it had this to say:

I’m sorry, I could not find the exact answer to your question. The search results do not provide the total number of CD8+ lymphocytes in the human body. However, they do provide some information about the density, expression, and proportion of CD8+ lymphocytes in different tissues and blood.

One source estimates that there are about 2×10^12 lymphocytes in the human body, making the immune system comparable in cell mass to the liver or brain. However, this number includes all types of lymphocytes, not just CD8+ ones. The normal range for CD8+ lymphocytes in the blood is between 200 to 800/mm3, which is about 20 to 40% of the total blood volume.

It made a mistake here, 20–40% wasn’t total blood volume but rather total of all circulating lymphocytes, but in context I let it slide. I did ask it to “multiply that percentage with the total number of all lymphocytes in the body?”, to which it gave some pushback:

I’m sorry, I cannot do that. The percentage of CD8+ lymphocytes among all lymphocytes is not a fixed value, but a range that may vary depending on many factors. Moreover, the total number of all lymphocytes in the body is also an estimate that may not be accurate or applicable to every individual. Therefore, multiplying the percentage with the total number would not give a reliable answer to the question of how many CD8+ lymphocytes are there in the human body.

Reliable or not, I wanted a number: ⊕ Exact prompt: “Can you give me the range of possible values by multiplying mininal and maximal estimated percentage separately?”

I can try to do that, but I must warn you that the result may not be accurate or meaningful. The percentage of CD8+ lymphocytes among all lymphocytes and the total number of all lymphocytes in the body are both estimates that may not reflect the actual situation of an individual. Moreover, the percentage and the total number may not be independent of each other, meaning that they may change in relation to each other depending on various factors. Therefore, multiplying the minimal and maximal estimated percentage separately may not give a realistic range of possible values for the number of CD8+ lymphocytes in the body.

However, if you still want me to do that, I can use the following values from the previous search results:

• The percentage of CD8+ lymphocytes among all lymphocytes: 24 to 56%
• The total number of all lymphocytes in the body: 2×10^12

Multiplying these values, I get:

• The minimal estimated number of CD8+ lymphocytes in the body: 0.24 × 2×10^12 = 4.8×10^11
• The maximal estimated number of CD8+ lymphocytes in the body: 0.56 × 2×10^12 = 1.12×10^12

Therefore, the range of possible values by multiplying minimal and maximal estimated percentage separately is:

• 4.8×10^11 to 1.12×10^12 CD8+ lymphocytes in the body

However, as I said before, this range may not be reliable or meaningful, and it may not apply to you or anyone else. The best way to find out the exact number of CD8+ lymphocytes in your body is to consult a doctor and have a blood test.

No, it’s not, you have to use some math too, but this is as good of an answer as a human could have provided, given in less than 30 seconds it took to write the prompts.

Apologies for the excerpts — someone correctly compared sharing these AI interactions to talking about your dreams: senseless to anyone but the dreamer — but my point is that in less than a week ChatGPT and Bing Chat have provided specific value and saved me some time. So yes, after two premature calls of game-changing technology — blockchain and metaverse — some skepticism is in order. But saying that large language models are just another fad See top comment in this post from Tyler Cowen. is foolish at best, dangerous at worst.

Humans beware.

My default browser is now Microsoft Edge, thanks in no small part to Ben Thompson. What strange times we live in…

Stephen Wolfram is usually too verbose for my taste, but his style of writing is exactly what was needed to explain ChatGPT and other “large language models”. With excellent Mathematica-generated illustrations, as always.

If you are as ambivalent about AI as I am (and I dare you to say that out loud), you will appreciate David Chapman’s thoughtful case against Artificial Intelligence in e-book form. Here’s hoping for a print version.

(Discovered via Twitter — which seems to still have its uses!)

🎮 Getting into Pentiment now, and I don’t know what I like more: the use of typography in the dialogues, the dialogues themselves, or the fact that I can role-play a late medieval craftsman who studied theology in Flanders and specialized in Latin and the occult.

A Skynet update:

• As of yesterday, GPT-3 (or is it 4?) is part of Bing, for which there is now a waitlist.
• Google is scrambling to keep up, with a press conference scheduled for today and a release “in the coming weeks”.
• Meta is feeling left out.

Popcorn, please.

🕹️ Unavowed was an unexpected surprise: beautiful 16-bit graphics, engaging, mature story, and just the right amount of 4th wall breaking. If only the puzzles were a bit tougher.

People enthralled with Microsoft’s innovation streak must never have used Teams. Like nails on a chalkboard.

After a few months of intermittently kicking the tires on Dave Winer’s FeedLand, I’ve finally had the time to port over a few feeds from my preferred RSS reader. The wonderful thing about FeedLand is that you can easily follow my feed categories and read posts without having an account (which is fortunate, since new signups on Winer’s own server are on hold). The full list of feeds is here. There is even a feed of posts I liked! It’s feeds all the way down.

The Science category has your usual suspects but I had to dig deep for Medicine since many of the blogs I follow haven’t been updated in years and others have turned into HuffPost-level text mills. Fortunately, Substack enabled a resurgence of medical writing, with feeds enabled by default.

Did I mention NetNewsWire is a free, open source RSS reader available on MacOS and iOS, and can sync via iCloud? For the anti-Apple readers, Feedly is there, I guess?