https://news.ycombinator.com/item?id=48743472
I will need to work on my manifesto when I get home. But :
three hypothetical Ai queries:
questions only of interest to a certain type of compulsive fan ... or a person who noticed something and is curious
how do we solve this?
We do NOT just train an LLM on all the survivor results and hope it can think it out. No. We use a database. The LLM can build the database. The LLM can make the database query. The LLM can design the system to do steps 1 and 2. What it can't do is give the result fully-formed, like Athena jumping out of the head of Zeus.
where is Atacama going?
the various ideas:
π₯ unfortunately none of that seems project-worthy, right now ...
π¬ tell me what you eat, and i tell you what you are.
π₯ the quote was supposed to be "perfection is achieved when there is nothing left to take away".
π‘ for example: the "subject" field. it is, largely, an exercise in digressions. also: the built-in section breaks. just have multiple messages. in a channel.
the nested footnotes βοΈ colortext will have to go. it is nice-to-have for the writer, but too confusing for the reader (man or AI).
what is left?
a thought. some words. and, commentary.
some of the commentary is load-bearing. other parts are for reference for the less-informed reader.
π₯ the commentary that makes it clear left means "remaining" and not "the opposite of right" ... is not load-bearing, normally.
We have the new Atacama app it seems OK not great but OK.
The footnotes don't quite work right the speech to text is mediocre the category selection is bare bones but it is better than nothingβοΈ It might not be better than what we had
These tasks are one end of the spectrum of "language model" tasks.
A "language model" takes language and does operations on it. Not thought. The cross-language translation and "sentence decomposition" rely on a simpler understanding of knowledge. β¨ it is much easier to say "I can think of a castle" than it is to build one
We do need some element of the second type of model.
Which is going to be an LLM/software centaur version.
the LLM here does need to be able to do thought tasks. not just language-model.
π₯ there is a reason we don't put a dictionary in charge of the country.
The sentence decomposition logic is almost done.
The design:
For a sentence, the decomposition identifies:
We do not yet have a full "sentence diagram" out of the words. The traditional Reed-Kellogg sentence diagram is uninteresting -- it only applies to English, is outdated, and remains confusing. However, the data missing to be able to generate this will be relevant later on.
I have been trying to run some very simple API calls using Claude to operate the new Barsukas HTTP API.
The flow is a 20-word prompt, reading 3-4 files from disk, and making a 4-line Python script to call the API.
This costs 22 cents. βοΈ which feels like too much
Why so much? There is something like 8 KB of system prompt and 14 KB of system tools. All of which should be cached/built-in to the model, but is instead just charged to the user. And several rounds of "cached reads" as the tool pauses for inputs.
Fortunately, the $20/month subscription gives something like $300/month of credits. When the cost is inflated by a factor of 5, it helps that the currency is also inflated.
On the other hand, the sheer volume of struggles Claude Opus is having with this task (simple directions like "only read the api/ dir" get ignored about 1/4 of the time) does make me want to look for a different harness, which can pivot between model providers. βοΈ it's not as bad as it was six weeks ago, but it is still very disappointing for the "recommended/expensive model" that it can't do this well.
a recent bug in Greenland: I was asking ChatGPT 5.4 mini to do a per-word breakdown of a sentence. But, sometimes it left words out. Other times, it would include the trailing period as a word, despite being instructed not to.
The cause of the bug: the prompt also asked ChatGPT to get the "word count" of the sentence. It is bad at this.
So, if it said The dog is on the log has 4 words, it would stop after 4 words and ignore the log when processing the sentence.
Of course, Python can easily do this. π‘ and, in the future, AI systems will know to just use Python to get this information rather than trying to guess it themselves. But here, it wasn't even used; I don't remember/know why a "word count" got added to the response.
I compared the Cambridge YLE list to the Trakaido wordlists. What I found was that most of the missing words were either deliberate βοΈ modal verbs like "can", as well as pronouns, are handled outside the wordlists, correct π‘ ice skating, jellyfish, or toothache do not need to be on the list, or irrelevant π‘ having "grandfather" instead of "grandpa" is fine ... though it would be good to improve the synonyms.
Click title to read full message...