Improved red bean paste buns

I’ve posted about red-bean-paste buns a few times before: the recipe I used for making the koshi-an (sweet red-bean paste), a recipe for the bao dough,  an updated recipe for the dough, and some photos of one of my attempts). I’ve not been happy with the dough—it was always too dry and didn’t seal well, so I tried modifying it today, as I had some koshi-an in my refrigerator that I had to use up (from a rather unsuccessful batch of buns a couple of weeks ago).  This recipe takes about 6½ hours from start to finish (if you already have koshi-an, which may take longer to cook, depending how dry your azuki beans are).

Step 1:

1 tsp yeast
½ cup lukewarm water  (this is an increase over previous batches)
½ cup sifted all-purpose flour

Mix together and proof for fifteen minutes.

Step 2:

2 cups sifted all-purpose flour
½ cup powdered sugar
½ cup warm water

Mix with the sponge from step 1 and let rise for 2 hours (until doubled).

Step 3:

1 cup sifted all-purpose flour
2 teaspoons baking powder
1 tablespoon sunflower oil (a more liquid fat than I’ve previously used)
¼ cup slightly warm water

In a large mixing bowl, mix the flour and baking powder, then knead with the dough from Step 2, gradually adding the oil and water. The dough was much more manageable than previous batches—the extra ¼ cup of water seems to make a big difference.

Roll the dough into a sausage shape about 1½ inches in diameter, cut into 12 equal pieces and roll each one into a ball. Place them well apart on a lightly floured cookie sheet and cover with a damp towel. Let rise until doubled (about 3 hours).

Cut 12 3-inch squares of waxed paper.  Roll each ball into a circle about 2½ inches in diameter, but no thinner than ¼” thick. (Using my cookie sticks!) Pinch the outer edges to be a little thinner.

Put about 1 tablespoon of filling in the center of the circle, then pull up the edges and pinch and twist them to seal. Put the bun flat-side-down on a waxed paper square. Let the buns rise for 45 minutes before steaming.

Keep the buns at least ½ inch apart in the steamer, and steam for 20 minutes (somewhat longer than before). Serve hot.

Steamed buns can be frozen and thawed, then re-steamed.

The steam buns stuck a bit to the waxed paper (which is fairly normal), and only a few of them popped open, so this was probably the most successful batch so far, though I did not bother taking pictures this time.

ChatGPT

Like many other people (particularly in academia), I’ve been playing a bit with the new large language model from OpenAI, ChatGPT:  https://chat.openai.com/chat.  This artificial-intelligence program is going to change the game substantially in college courses, as it writes better than many college students, though it tends to write a lot of bullshit and fabulates citations.

Most of my experimenting was triggered by posts or comments on the r/Professors subreddit, where professors have been expressing concern about how this new tool will be used for cheating and how it will affect their courses.

My experience is that ChatGPT writes grammatical prose that sounds good, but doesn’t really understand what it is saying.  When I gave it quiz questions from my electronics course, it provided answers that would sound good to a non-expert, but which are clearly wrong, in the same ways that failing students in the course would be wrong. For example, the prompt “What is the impedance of 20kΩ resistor in parallel with a 50pF capacitor at 1kHz?” results in

The impedance of a resistor is simply its resistance, so the impedance of a 20kΩ resistor at any frequency is 20kΩ. The impedance of a capacitor, on the other hand, is determined by its capacitance and the frequency of the current passing through it. The impedance of a capacitor is given by the following formula:

Z = 1 / (2 * pi * f * C)

where Z is the impedance, f is the frequency of the current, and C is the capacitance of the capacitor. In this case, the impedance of the 50pF capacitor at 1kHz would be:

Z = 1 / (2 * pi * 1000 * 50 * 10^-12) = 1.59e5Ω

To find the total impedance of the 20kΩ resistor and the 50pF capacitor in parallel, we need to use the formula for the parallel combination of impedances:

1 / Ztotal = 1 / Z1 + 1 / Z2

where Z1 is the impedance of the first component (in this case, the 20kΩ resistor), Z2 is the impedance of the second component (in this case, the 50pF capacitor), and Ztotal is the total impedance of the two components in parallel. Plugging in the values from above, we get:

1 / Ztotal = 1 / 20,000 + 1 / 1.59e5

Solving for Ztotal, we get:

Ztotal = 1.19e5Ω

So the total impedance of the 20kΩ resistor and the 50pF capacitor in parallel at 1kHz is approximately 119,000Ω.

This “answer” replicates two of the common mistakes of failing students: it neglects that the impedance of a capacitor is a purely imaginary number and it gets the arithmetic of fractions wrong. Until the large language models can be appropriately merged with the math AIs (like Photomath or Wolfram Alpha), engineering faculty do not have much to fear from ChatGPT.  I guess that it will take 5–10 years to do that sort of merging, as the two approaches to AI are not very compatible.

I tried some more “creative” prompts also and tweaking the prompts to get better results. I tried one series of prompts (I won’t show the results for any but the last iteration):

• Write a song to be sung by students with the theme and title “When will I ever need this?”
• Try again but with rhyming verses.
• There is no rhyme there.
• Try again with real rhymes.
• Try again with a blues-song format and general air of depression.
• Try again without the upbeat bridge.
• Remove the current outro and end with singer being lost and forlorn.
• Try again incorporating “1812”, “square root of 1066”, and “the quadratic formula”.

resulting in 

Verse 1: Sittin’ in this classroom
Feelin’ like a waste of space
Teacher’s talkin’ at me
But I can’t hear a word he’s sayin’

Chorus: When will I ever need this?
Will it help me find my way?
Will it make me rich or famous?
When will I ever need this?

Verse 2: I ain’t sayin’ it ain’t important
But it just don’t make no sense
Memorizin’ 1812 and the square root of 1066
And the quadratic formula too

Chorus: When will I ever need this?
Will it help me find my way?
Will it make me rich or famous?
When will I ever need this?

Outro: Feelin’ lost and forlorn
In this never-endin’ maze
I just don’t know where to turn
When will I ever need this?

Chorus: When will I ever need this?
Will it help me find my way?
Will it make me rich or famous?
When will I ever need this?

The main use case for faculty seems to be creating email templates or assignment prompts,  Here is an example of an email template (with the prompt provided from a question asked by a professor on the subreddit). Prompt: “Write an email telling a student that I will not change your grade because that is the grade you earned, and that there are no more opportunities for extra credit as the term is over.”  I had to reprompt with “Try again without inviting the student to reply.”

Dear Student,

I want to let you know that I have carefully reviewed your work and have determined that the grade you received is an accurate reflection of your performance in the course. As such, I will not be changing your grade.

Furthermore, the term is now over, so there are no more opportunities for extra credit. I understand that you may be disappointed with your grade, but please know that it is my belief that you earned the grade you received through your work in the class.

Sincerely,

[Your Name]

The main ways that people have thought of to avoid or detect ChatGPT cheating have included highly technical content that requires math skill; checking citations (ChatGPT likes to fabulate them); making prompts that use a visual image (it will be a while before the large language models are merged with computer vision); requiring web lookup of current events (newer than the model’s training—the models currently do no search); requiring live, in-class writing; and requiring multiple stages of writing process (proposal, outline, draft, edited draft).  I don’t think that the last approach would work, as ChatGPT seems capable of doing outlines and editing drafts.

Someone suggested that real student writing could be detected by the cliched “hooks” that students use, but that is easily faked also: “Write a 200-word essay about cheating using AI, starting with ‘Since the dawn of education, students …’ “

Since the dawn of education, students have always looked for ways to cheat in order to get ahead. With the rise of technology, cheating has become even easier with the use of artificial intelligence (AI). AI can be used to generate answers to tests or homework assignments, or even to write essays.

While using AI to cheat may seem like an easy way to get good grades, it is ultimately a dishonest and unethical way to succeed. Not only does it go against the principles of fair competition and hard work, it also undermines the value of education. By cheating, students are not truly learning and are not gaining the skills and knowledge necessary for future success.

Furthermore, using AI to cheat can also have serious consequences. Students caught cheating can face disciplinary action from their schools, and in some cases, it can even result in expulsion. In addition, using AI to cheat can also have long-term consequences on a student’s future, as it can damage their reputation and make it more difficult to get into college or find a job.

In conclusion, while using AI to cheat may seem like an easy way to get ahead, it is ultimately a dishonest and unethical action that can have serious consequences. It is important for students to understand the importance of hard work and honesty in education, and to resist the temptation to cheat.

I’m convinced that Chegg will be buying a large language model and retraining it on their large database of cheating materials, resulting in an AI that can replace the “experts” that they hire in India. The results from ChatGPT are already about as good as Chegg answers, and training a dedicated CheatGPT bot will undoubtedly increase Chegg’s profit margin (no labor costs for “experts”).

For my Spring quarter class, I will have the students use ChatGPT (assuming it is still free then) to create the summaries that they need to write for the class—they will have to provide ChatGPT with the info and have it turn it into readable prose of about the right length.  I think that this approach will result in more readable summaries than in the past, while still challenging students to come up with the key points that need to be communicated. I suspect that a lot of writing in future will be done by a combination of human and artificial intelligence, and I think that this 2-unit, non-required, P/NP course will be a good way for students to experiment with that approach to writing (with no academic-integrity issues interfering).

Empty office!

I finished emptying my office today, just a little shy of 18 months after retiring and six months after last year’s recall-to-service ended.  I will be teaching a course in Spring 2023 again, but without the benefit of a physical office on campus. 

I had about 40 gallons (150 L) of paper to recycle, a similar amount to take home to sort out and refile, and a somewhat larger amount to shred (old student work, old grade records, old papers I’d refereed).  I had about 30 shelf feet of books to get rid of, but I managed to give most of them away by putting a book cart in “Jack’s Lounge”, the study area in the Baskin Engineering Building, during the last few days before finals. Several of the books were ones I had used as a grad student at Stanford in the late 70s and early 80s and might have been worth something as collector’s items, but I decided that I was not up to putting them on Amazon or e-bay. I discarded a lot of research notes from old projects that were no longer interpretable (or no longer interesting, or both). I also discarded a lot of handouts for my old courses—most of that course design is now lost (though some will undoubtedly still be on the computer as LaTeX or .html files).

While clearing out the bookshelves and filing cabinets, I also found some interesting things, like the student evaluations from the first course I ever taught (Summer of 1978)—I haven’t gotten up the courage to look at them again. I also found a lot of old posters, some of which may be worth framing and hanging in the little bathroom, to replace the current very faded posters.

I’m probably going to have to buy a new filing cabinet to store stuff I saved in, though I can probably thin what I brought home by another 10–20%. It will probably take me several months to get stuff put away. It was enough of a problem parting with so many of my “treasures”—going through it again so soon would be too difficult.

I’ve announced to the department chair and department manager that the office is available for the next occupant, other than moving my University-owned Linux box across the hall to the grad-student bullpen.  I would have moved the computer myself, but IT has to bless the Ethernet lines. If you plug the wrong computer into one of the ports, the switch detects the foreign computer in a couple of minutes and shuts down the port.

 

 

End of Project Baseline

Yesterday was my final in-person visit to Project Baseline, just over 4 years since my first visit.  They have really reduced the ambition of the project since then—they took the same urine sample, 25 vials of blood (about 200 ml), and swabs for oral, nasal, and skin microbiomes, but the rest of the assessment was just vital signs and a few low-cost fitness tests (balance, grip strength, sit-stand, walking test).  All the expensive stuff they started with (stress echocardiogram, chest CAT scan, ankle-brachial index, pulmonary function) has been discontinued—I think that the heart (pun intended) went out of the project when the original PI (Dr. Sanjiv Sam Gambhir) died in July 2020.  Also COVID messed up their ability to do detailed in-person studies for a couple of years, and I think that Verily lost interest in doing the project, so has just been keeping it ticking at minimal levels.

The one interesting result from this visit was that my 6-minute fast walk was 729 meters, substantially more than the 630 m of my first visit and something like 660 m on my second visit. I did push myself a bit more this time, getting a little out of breath by the end of 6 minutes. The 2.025 m/s pace is about 4.53 mph or 13:17 per mile. That is a pretty good pace for someone of my age, though I had difficulty finding any large studies that establish reasonable guidelines for healthy adults. I found one small study for people 20–50 [https://doi.org/10.1016/j.rmed.2006.01.001] and lots of studies for people who were unhealthy in various ways. The linear regression from that study suggests that “normal” for men of my height and age is around 551m, but that is extrapolating the linear regression well out of the age range. A somewhat larger study [https://doi.org/10.1183/09031936.00194909] set age- and sex-based standards for an older population—but even my original distance was above the 75%ile for my age, and the latest measurement was above the 75%ile even for men 20–30 years younger than me.

I’ve returned the Verily Study Watch and the hub for uploading data from it—I don’t think I’ll miss the watch, as it was not a very useful piece of equipment from a consumer standpoint—all the processing was done in the cloud after data was uploaded, and almost nothing was shared with the person wearing the watch.

I might get myself a consumer-level fitness monitor—they range in price from $40 to $400 (though I’m sure you can find more expensive ones). I’ll probably want one that I can download data from without needing an expensive subscription plan.  It might be nice to have one that I can wear on my ankle for tracking bicycling as well as walking, but I don’t know whether any are designed for that obvious use (nor whether pulse monitoring at the ankle works as well as it does at the wrist).  I’ll have to look into what’s available.