Thoughts as they occur to me.
The context window on Gemini is really amazing. Both for the main app as well as notebookllm. I've been pulling entire books down from annas-archive and throwing them in, asking a bunch of questions, pulling out quotes, having it do further searches, etc.
Truely amazing how deep it will let you get into a topic so quickly.
I work with AI professionally and as an individual that thinks things are cool. Actually getting the overview and understanding the big picture vs figuring out practical things that work is very different, and depending on what I'm doing different things catch my attention.
Ksenia Se and I see a lot of the same stuff, but the things that catch my attention are quite different. Things like OpenAI releases, new mid-journey models, the dramas of OpenAI's board room or interesting Inflection "acuihires" are generally intersecting and gives you a sense of where the industry is going.
If you are in the trenches, as it were, there are slightly different thoughts. I'm very interested in things like Ollama, DeepSeek, and the Vercel AI SDK for example that aren't necessarily interesting from a general perspective but quite fascinating on what they can enable. I'm getting to be more of a claude champion (their artifacts are way better in practice, token caching amazing) which is sort of a gut feeling based on working different APIs.
ollama lets you run models in the privacy of your own computer; easy to download to "regular hardware", no "cost", and playing around with various parameter tweaking is super interest. There's something freeing about this.
DeepSeek is wild both in how cheap it was to train, but it goes back to the idea of how small can a foundation model really be. There's something very exciting with this, it feels transformational that smaller models will be smarter than we can imagine now.
The Vercel AI SDK is like low level code plumbing, that you technically don't actually need and you could probably just rewrite it all but it hits the sweetspot of abstractions and usefulness. Its a breath of fresh air after poking around with LangChain – what it attempts is much less, but the balancing of complexity and functionally is Just Better.
Getting your hands deep into it gives a different perspective and it's hard to generalize.
From Sherlock Holmes:
"His ignorance was as remarkable as his knowledge. Of contemporary literature, philosophy and politics he appeared to know next to nothing. Upon my quoting Thomas Carlyle, he inquired in the naivest way who he might be and what he had done. My surprise reached a climax, however, when I found incidentally that he was ignorant of the Copernican Theory and of the composition of the Solar System. That any civilized human being in this nineteenth century should not be aware that the earth travelled round the sun appeared to be to me such an extraordinary fact that I could hardly realize it.
“You appear to be astonished,” he said, smiling at my expression of surprise. “Now that I do know it I shall do my best to forget it.”
“To forget it!”
“You see,” he explained, “I consider that a man’s brain originally is like a little empty attic, and you have to stock it with such furniture as you choose. A fool takes in all the lumber of every sort that he comes across, so that the knowledge which might be useful to him gets crowded out, or at best is jumbled up with a lot of other things so that he has a difficulty in laying his hands upon it. Now the skillful workman is very careful indeed as to what he takes into his brain-attic. He will have nothing but the tools which may help him in doing his work, but of these he has a large assortment, and all in the most perfect order. It is a mistake to think that that little room has elastic walls and can distend to any extent. Depend upon it there comes a time when for every addition of knowledge you forget something that you knew before. It is of the highest importance, therefore, not to have useless facts elbowing out the useful ones.”
“But the Solar System!” I protested.
“What the deuce is it to me?” he interrupted impatiently; “you say that we go round the sun. If we went round the moon it would not make a pennyworth of difference to me or to my work.”
I’ve recently started up a company called The Focus AI.
I’m going to keep writing up things here as I find them, and then cross posting them as it makes sense. Think of what’s happening over there as part of the new mailing list.
One of the most fundamental unanswered questions that has been bothering mankind during the Anthropocene is whether the use of swearwords in open source code is positively or negatively correlated with source code quality. To investigate this profound matter we crawled and analysed over 3800 C open source code containing English swearwords and over 7600 C open source code not containing swearwords from GitHub. Subsequently, we quantified the adherence of these two distinct sets of source code to coding standards, which we deploy as a proxy for source code quality via the SoftWipe tool developed in our group. We find that open source code containing swearwords exhibit significantly better code quality than those not containing swearwords under several statistical tests. We hypothesise that the use of swearwords constitutes an indicator of a profound emotional involvement of the programmer with the code and its inherent complexities, thus yielding better code based on a thorough, critical, and dialectic code analysis process.
On the web I started with chatgpt, and it turns out that it makes more sense than claude or gemini.
Though gemini is way faster.
In emacs I started with Zephyr, and you know? I think it just does the best with defining words. (Which is what I use the most in emacs)
With coding, I started with claude in cursor, and man it really kills it. I switched to chatgpt and you know it just wasnt the same.
So, basically, you just like the thing you used first and there's no objective measure to anything.
JavaScript has won, so of course I've been moving more into typescript and javascript. And these last few weeks as I've been going deeper into the world of AI and LLM I've been dipping into the python ecosystem. And… I'm not convinved.
Obviously you need to go to where the libraries are but ruby is still the most delightful.
I was poking around the implementation of the obsidian-extract-url plugin, and its written in rust but compiled and run as WASM inside of the obsidian plugin environment.
Novel use case for WASM.
One file that contains design, layout, code, and remote server code.
I'm not totally sure how to debug server actions but its a whole bunch of functionality in one place.
No shifting between different files, just doing what you set out to do in one context.