Back in January, I attended an “AI and Coding” class:
There were a couple of videographers present so I’m hopeful that eventually the lectures will appear on YouTube as some previous events in this “Expanding Horizons in Computing” series have.
During the intro, we were reminded that the first thing computer nerds want to do is get rid of computer nerds:
Inevitably, though, there have been haters. Alan Perlis:
Tim Kraska was the speaker who’d done the most to determine what LLMs can do. His grad students spent 2.5 months and about $100,000 in Claude API fees replicating the capabilities of the 500,000-line DuckDB embedded database management system but in a different language (I forget which! Sadly, not Lisp). It seems that for a complex project like this, the only people who can tell AI what to do are those who could do it themselves if they had to.
Continuing Carnegie-Mellon’s tradition as “the useful place in CS” (CMU gave us the Mach kernel, for example, which is inside nearly all Apple products), Graham Neubig talked about his experience building and using OpenHands, a system a little like Google’s Antigravity in that you can tell your “agents” to write software for you and the editor connects to the LLM of your choice. Prof. Neubig demonstrated using OpenHands to build a web site for the MIT event and the results were impressive!
How good are LLMs in practice? Contrary to my own experience where LLMs are amazing at diving into huge legacy codebases and telling the human “these are the relevant files”, AI felt good but actually slowed human programmers down:
AIs cut and paste like crazy, eventually producing code with so many duplicate blocks of code that only an AI will be able to make a change consistently through a real-world system:
I enjoyed the bathroom break. The smartest humans on the planet need a lot of coaching for the operation of sophisticated machinery:
Based on the period products in what was labeled a “men’s” room, the world’s smartest people are going to struggle with the “What is a woman?” question:
Speaking of bathrooms, the ground floor restroom signs are already falling apart in the nearly-new building. Fortunately, the sacred word “inclusive” hasn’t been marred.
Towards the end of the day, Varun Mohan showed up via Zoom to make the academics look like fools. While they were dithering to get a few papers published and secure a lifetime guarantee of employment at a wage that is 1/50th of what a receptionist at NVDIA earns, an apparent teenager glued together a few open-source developer tools and added LLM integration to create Windsurf, which Google then acquired in a non-acquisition for $2.4 billion. The result is Antigravity (see Antigravity as web developer (AI in an IDE)).
What nobody could offer at the event: A clear explanation of what skills made a person a good software developer in the Age of AI and, therefore, what an undergraduate CS program should teach. On the other hand, the slides did offer a clear picture of what a typical human software engineer looks like: female and, usually, non-Asian.
For the haters who say that there is no “science” in computer science… I learned about Science starting during the walk to the event. The hardware store in Inman Square’s most prominent sign:
Not Science-related, but I love seeing Black Lives Matter signs and here a commercial property owner had devoted a huge amount of space to one. (See Replacement of Black workers by migrants in Cambridge, Massachusetts for how the city’s merchants have kept the signs and discarded the people.)
Of course, there were the sidewalk maskers:
And the cycling maskers (note the filthy snow and trash in the background):
In the MIT event space there were 6 people sitting in front of me, 3 of whom were masked for the entire day. Here are a couple of them: I’m not sure that I understand the rationale for
The next day at the Harvard Art Museum, Arthur M. Sackler section (don’t forget that before developing ties to Jeffrey Epstein, Harvard was entwined with the Sackler family), an apparent couple in which one person doesn’t wear a mask while the other does. This has always mystified me. Partner 1 is protected by his/her/zir/their mask so only Partner 2 gets infected by SARS-CoV-2 at the public venue. Then they go home and, without masks, share a confined space for days during which time the virus can trivially hop from Partner 2 to Partner 1.
Also a mystery, the person who is afraid of catching a respiratory virus and chooses a job with guaranteed exposure to hundreds or thousands of strangers each day. The mask is great protection, I’m sure, but wouldn’t it be far safer to wear the mask while working in a regular office or warehouse with just a handful of other employees nearby?
Maybe one day an LLM will be able to explain these choices?
Disclosures: I’m on Perlis’ side (and Dijkstra’s, of course). No AI was used in creating this content, except by Knuth.
> an apparent teenager glued together a few open-source developer tools and added LLM integration to create Windsurf, which Google then acquired in a non-acquisition for $2.4 billion.
Fans of western novels (Louis L’Amour did a lot of “cut and paste”, too) know that the real money is in picks, sluice boxes, and $100/lb. flour during a gold rush. (As an ersatz entrepreneur, I often wonder what is beyond the current hype. Gathering and recycling all the incalculably valuable human minds littering the planet?) The largest jackpot in America was a $2B payout in Powerball, given the number of people doing AI stuff, playing the lotto might be just as good of a strategy. Academics might be “fools” if the only point in living is making money. I highly recommend this course for MIT-level dummies about machine learning:
https://ocw.mit.edu/courses/18-065-matrix-methods-in-data-analysis-signal-processing-and-machine-learning-spring-2018/
I wonder what deal Strang would have taken. No linear algebra and $2.5B or decently wealthy and helping top minds discover the underpinnings of engineering mathmatics? Interestingly in his discussion with Wolfram I mentioned yesterday, Knuth said, regarding AI research:
> I myself shall certainly continue to leave such research to others, and to devote my time to developing concepts that are authentic and trustworthy. And I hope you do the same.
We ignore Knuth at our peril. Speaking of Turing, one of Knuth’s AI questions was:
Knuth >>> 16. What did Winston Churchill think of Alan Turing?
ChatGPT >> 16: There is no definitive answer to what Winston Churchill thought of Alan Turing, as the two men did not have a close personal or professional relationship. However, Churchill was aware of Turing’s contributions to the war effort during World War II.
Knuth > #16. Another mission apparently well accomplished yet fake. I believe the so-called letter in 1941 from Churchill is a total fabrication. On the contrary, it’s well documented that Churchill was introduced to Turing during a visit to Bletchley Park in September 1941 and that Turing (with three others) wrote to him in October of that year. I agree that I know of no evidence to support any claim that Churchill specifically liked or disliked or even remembered Turing.
And finally, the men’s and women’s bathrooms in our dormitories at my Ivy League school were always out of condoms. I don’t know if that is good or bad, depending on the quantity stocked. I had walking pneumonia my first freshman semester, not sure a mask would have helped prevent that. AIDs was the big epidemic in the ’80s. Being a Mask Scientist, I always wear mine and never get sick. I guess it’s superstition, like looking both ways before crossing the street or trusting that AI isn’t lying or tripping.
If only the right comment could get someone hired for $2.4B as the leader of a top blog commenting LLC.