Topic
LLM
Imported from Atacama channel llm
Recent posts
https://mashable.com/tech/anthropic-fable-5-disproves-jacobian-conjecture
https://news.ycombinator.com/item?id=48743472
deliberate pigheadness is not intelligence
three hypothetical Ai queries:
I have been trying to run some very simple API calls using Claude to operate the new Barsukas HTTP API.
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.
One of the under-appreciated benefits of using LLMs for coding is that they are experts on most topics.
New Rule: when running a command, any pipe to "head" or "tail" should pipe to a version that automatically tee-s the output to a temp file as well, or should otherwise explicitly contain the tee in the command prompt.
Some Redditors are going more insane now that GPT-4o has been shut down.
In reply to https://freddiedeboer.substack.com/p/im-offering-scott-alexander-a-wager :
today, in Trakaido/Barsukas development: I want to add 20-30 cities to the word list with translations in 14 languages.
I have seen this specific error before, and I think << the Claude Code team should be able to fix it.
Somehow Barsukas has become the other server as part of the Trakaido project. It contains translation data, sentence data, etc.
thoughts on Trakaido development
The low-hanging fruit is having agents do things like type checking. there is no reason to not have your main codebase include type hints for JavaScript, Python, etc. any more. On the other hand, there is no need to force developers to use them. An agent can modify the pull request to add them.…
the pattern of LLM work systems
Claude seems to be standardizing on a new "agent session" method, and I think I like it.
an LLM-coding post-mortem: PicPickr
https://harshdeepgupta.substack.com/p/6-months-and-485-my-journey-into
I asked << gpt-oss:20b to do some basic Lithuanian translations. Its attempts to do so (or to << think its way out) were flawed ...
the future is today.
the opposite of wrong is still wrong
Seen on the internet: https://a16z.com/geo-over-seo/
beware fools who yell at clouds
Seen on Substack https://substack.com/@tedgioia/note/c-128027920 :
contra Brian Balfour on LLM Platforms
https://blog.brianbalfour.com/p/the-next-great-distribution-shift my verdict: it stinks!
are whales a type of big fish?
The market for learning the Lithuanian language is quite small.
Perhaps the correct question is not << why does the machine need to be polite . Perhaps the correct question is << regardless of the motivation, make it be polite . there is a certain evolutionary need; people tend to like things that are polite more.
greenland, a post-mortem, part 3
JSON output is almost a necessity for an LLM to be usable today. All of the major LLM platforms have it in some form. But, if you are using a model from 2023, it might not support it, or it might not work very well.
greenland, a post-mortem, part 2
That link is https://spaceship.computer/greenland/ .
greenland, a post-mortem, part 1
Other than the << fit and finish , the Greenland project is done.
Trying the << qwen3-1.7b-q6_k.gguf model QWEN3, 1.7B parameters, quantized to 6-bit, there is a clear problem of runaway-thought (or at least looping responses).
more "fake thought" at the margins
QWEN3 is out. I'm starting tests with the 4B model.
To write a few paragraphs on a topic, there is (roughly) a four-step process:
If I had a three-word summary for what I am working on, it is << tools for AI .
https://www.dailykos.com/stories/2025/4/7/2315259/-Please-ban-all-AI-images-in-all-diaries-Add-it-to-the-ROTR this is a top-rated user-generated "diary" on liberal blog DailyKos
In my testing, I am starting to make a distinction between two types of "tests" for LLMs.
https://annelutzfernandez.substack.com/p/resisting-ai-in-education-part-iii
more arguments, for another blog post: https://annelutzfernandez.substack.com/p/resisting-ai-mania-in-schools-part-baf
https://annelutzfernandez.substack.com/p/resisting-ai-mania-in-schools-part
Gemma3 is out: https://blog.google/technology/developers/gemma-3/
https://www.nbcnews.com/tech/tech-news/arizona-moves-ban-ai-use-reviewing-medical-claims-rcna193135
a possibly more-generous interpretation
What will the population of Earth be in 2200? as a reminder, there are no credible theories for more than a de minimis number of humans to live off-Earth And, how does AI affect this?
The short answer here is the very mundane << we don't need software, we need humans .
thoughts on LLMs for Social Networks
in the future, your LLM will talk to my LLM, and neither of us will benefit from the sage information exchanged.
thoughts on the LLM-powered encyclopedia
it would be a competitor to Wikipedia.
https://news.ycombinator.com/item?id=42458752
https://viruscomix.com/page490.html
Here is Deepseek-8b's response to the same question.
The new "deepseek-r1" models are out. Both a large model I can't use, and "distilled" models.
https://andymasley.substack.com/p/individual-ai-use-is-not-bad-for << Using ChatGPT is not bad for the environment
Working with LLMs is a speed-run through 50 years of software engineering.
I asked the machine to rename the blueprint file "blueprints/messages.py" (to avoid naming collisions with the common/messages.py library file).
Initial indications are that phi4:14b is substantially slower than gemma2:9b on a 24GB MacBook Pro, with no noticeable performance difference. part of the slowness is related to the "two phase response" code I added to work around Ollama's inability to do structured JSON responses correctly. Even…
We asked llama3 and phi3 (April release) about various random article-topics from Wikipedia.
battle of the heavyweights (1)
The larger models do not struggle with the types of prompts suited for "middleweight" models. We ask three of them (ChatGPT4, Claude2.1, "Bard") the hallway question, to demonstrate their competence and mastery of simple tasks.
battle of the middleweights (3)
sometimes, less is more
battle of the middleweights (2)
ask a stupid question, get a stupid answer
I asked four LLM models the same question: "Write a 500-word essay on the history of the concept of a "hallway", and how they are used today."
in which I ask the machine for age-appropriate book recommendations
I asked ChatGPT4 to give some suggestions for books for me to recommend to children, based on the Reading List here.
a course in LLM embeddings (1/17)
or, "LLM embeddings and outlier dimensions"
Three threads, twisting and turning.