> "I've seen an uptick in vague ominous attacks on Marlboro saying that our products kill people.
Then don’t vaguepost.
> Trust me when I say you don't want to live in a world where they win the AI race (if that wasn't already blatantly obvious).
They’re not perfect, but they’re the best organization in the running (with the second best person at the helm after Demis). The rest of the field isn’t close.
I’m an American and I just had the thought - if I was working in Japan at a Japanese company and I had the opportunity to hire, would I have a bias to hire other Americans?
Honestly probably, since I understand them the best.
I have fond memories of cs224d [1] taught by richardsocher. It’s a bit dated at this point as it was created in the pre-transformer era, but it was a very cool introduction to applying deep learning to nlp at the time.
I’m thankful for Racket - it got me regularly programming in lisp by virtue of LeetCode accepting it as one of its languages.
I did start to feel Racket’s “wordiness” towards the end - it started to feel a bit like COBOL. I’ve since moved onto Clojure and really appreciate the shorter keywords/function names/fewer parenthesis.
I still miss for/fold though - that thing is an absolute machine.
I had a little excursion into Dyalog APL recently and wound up writing an emacs mode to evaluate Dyalog APL [1]. It was a pretty nice experience using Claude to extract the small subset of features I wanted from gnu-apl-mode [2] to work with Dyalog APL.
I’d really like to properly get into APL though. My plan is to solve a bunch of problems on Kattis [3].
I'm really enjoying this way of learning a new language in the age of LLMs - starting with easy problems on an online code judge website and work with an LLM to come up with/explain simple solutions. It gives me dopamine hits, lots of reps, allows me to start coding right away, and is a nice way to slowly ramp up difficulty and get practice with different features of the language.
Love the story and the article. The only nit I have with it:
> “His answers are… understandable, and maybe in some ways more digestible than we would get from an expert,” he said.
This does not reflect his actual responses? The interviewer keys off his most emphatic sounding words to keep the conversation flowing, but his answers are generally inscrutable.
He did a great job given the cards he was dealt though.
I was curious to get a sense for the overall "success rate" at a glance, so I uploaded the author's data as a spreadsheet and color-coded the conversations based on length (short=red, medium=yellow, long=green) with the help of Claude:
> We retired the “Nerdy” personality in March after launching GPT‑5.4. In training, we removed the goblin-affine reward signal and filtered training data containing creature-words, making goblins less likely to over-appear or show up in inappropriate contexts. Unfortunately, GPT‑5.5 started training before we found the root cause of the goblins.
The prompt is just a short term hotfix/hack because they couldn’t get the proper fix in in time.