I am massively pissed off by Ansible’s appalling slowness, and I think pyinfra’s “up to 10x faster” is unambitious. A better performance target should be more like rsync.
Great piece! Any thoughts on non-LLM generative AI? (I am personally finding the image generation to be of extraordinarily limited utility – suffering from absolutely wild hallucinations, impossible to fine-tune, etc. – but I also haven’t engaged with it very seriously.) Also, any thoughts on retrieval augmented generation? Just as a user, it feels like that has been a big breakthrough with respect to reducing the hallucinations of LLMs; is that correct?
I decided not to write about image generators, mainly so I could get my post out before the end of 2023!
2023 was the year Midjourney went genuinely photorealistic - the whole Balenciaga Pope thing back in March, and they released increasingly capable models since then. Still really hard to get exactly what you want out of them, but you can get a photorealistic image of pretty much anything now.
In Stable Diffusion world the thing where you can generate more than one image per second on a consumer GPU is pretty wild. I’m fascinated by the tools which let you draw a sketch and have it turned into a prompt-driven image “live” as you edit it: https://fedi.simonwillison.net/@simon/111489351875265358
As for RAG: it’s a really big deal, I should have included it in my roundup. It’s still hard to do it really well, but getting a basic version up and running is practically the “hello world” of working with LLMs. My best result so far has been this one, which is a Bash one-liner! https://til.simonwillison.net/llms/embed-paragraphs#user-content-answering-a-question
That’s awesome – thank you, and thank you more generally for all of the exploration and writing you’ve done on LLMs and AI more broadly: it’s a tremendous service for the practitioner looking to get past the hype (and hysteria)!
I should publish my TIL about that! They’re really easy to use with Datasette - you can download the pre-compiled .so file from e.g. https://github.com/turbot/steampipe-plugin-hackernews/releases/tag/v0.8.1 and pass it to datasette --load-extension steampipe_sqlite_hackernews.so and start trying it out - run select * from hackernews_show_hn for example.
I first heard the phrase “eject” used in the context of ejecting a Webpack config from a Create React App application’s autoconfig. It may go back further.
A decent, modern configuration management tool that isn’t written in a dynamic language. Something like ansible written in Go.
Edit: I never understood why Hashicorp didn’t create this. It could have used HCL. Maybe because Mitchell was friends with Adam Jacob?
Perhaps Pulumi is close to what you’re describing? Supports Go, Typescript, C#, Java with fully typed SDKs for every cloud
No. I’m looking for on host configuration management like chef, puppet and ansible.
Like Jetportch, before it was discontinued?
propellor?
It’s still Python but https://pyinfra.com/
I am massively pissed off by Ansible’s appalling slowness, and I think pyinfra’s “up to 10x faster” is unambitious. A better performance target should be more like rsync.
Pairs well with Benchmarking “Hello, World!”
Referenced paper What Goes Around Comes Around by Michael Stonebraker and Joseph M. Hellerstein
Great piece! Any thoughts on non-LLM generative AI? (I am personally finding the image generation to be of extraordinarily limited utility – suffering from absolutely wild hallucinations, impossible to fine-tune, etc. – but I also haven’t engaged with it very seriously.) Also, any thoughts on retrieval augmented generation? Just as a user, it feels like that has been a big breakthrough with respect to reducing the hallucinations of LLMs; is that correct?
I decided not to write about image generators, mainly so I could get my post out before the end of 2023!
2023 was the year Midjourney went genuinely photorealistic - the whole Balenciaga Pope thing back in March, and they released increasingly capable models since then. Still really hard to get exactly what you want out of them, but you can get a photorealistic image of pretty much anything now.
In Stable Diffusion world the thing where you can generate more than one image per second on a consumer GPU is pretty wild. I’m fascinated by the tools which let you draw a sketch and have it turned into a prompt-driven image “live” as you edit it: https://fedi.simonwillison.net/@simon/111489351875265358
As for RAG: it’s a really big deal, I should have included it in my roundup. It’s still hard to do it really well, but getting a basic version up and running is practically the “hello world” of working with LLMs. My best result so far has been this one, which is a Bash one-liner! https://til.simonwillison.net/llms/embed-paragraphs#user-content-answering-a-question
That’s awesome – thank you, and thank you more generally for all of the exploration and writing you’ve done on LLMs and AI more broadly: it’s a tremendous service for the practitioner looking to get past the hype (and hysteria)!
Oxide and Friends Topic/Guest?
Absolutely! If @simonw is up for it, we would love to have him!
Closing the loop: Open Source LLMs with Simon Willison (on Oxide and Friends)
I am really looking forward to mixing this with datasette from @simonw
I should publish my TIL about that! They’re really easy to use with Datasette - you can download the pre-compiled .so file from e.g. https://github.com/turbot/steampipe-plugin-hackernews/releases/tag/v0.8.1 and pass it to
datasette --load-extension steampipe_sqlite_hackernews.soand start trying it out - runselect * from hackernews_show_hnfor example.Notes on using these with Datasette and sqlite-utils: https://til.simonwillison.net/sqlite/steampipe
Brendan Gregg weighed in on that other site with what he was seeing at Netflix at the time:
The supplement[1] has the paper and some other papers done with the tools described in the paper.
[1] https://uwdata.github.io/living-papers-paper/living-papers/
I like the “eject” idea, it’s a succinct way of phrasing an friction I often have with tools that wrap something more flexible.
I first heard the phrase “eject” used in the context of ejecting a Webpack config from a Create React App application’s autoconfig. It may go back further.
yeah that IS neat
Has anyone ever heard of something like this - SQL for the OS, but run against a sos report[1]?
I always seem to be using the awk/grep/etc… to explore things ad-hoc.
I did find jc[2] + sqlite-utils[3] to be interesting, but seemed a waste to not leverage all the table definitions from OSQuery.
[1] https://github.com/sosreport/sos [2] https://kellyjonbrazil.github.io/jc/ [3] https://sqlite-utils.datasette.io/en/stable/