nostr:npub16sevfkt59gfel9e2hm3c4t2vtfkc2hgzus6xlnp4a5hlndku4fpsjlwex2 Yes! I think you're my first user (or at least the first person who's told me you're using it)

New trick for my TIL website: the "related content" at the bottom of each TIL is now powered by OpenAI embeddings. I wrote a new TIL describing how that works (using new features of my openai-to-sqlite tool): https://til.simonwillison.net/llms/openai-embeddings-related-content

nostr:npub13vj2zrsaas2gcngctp6766vlnfaxhaun6pzh7v5xnnl4ka48k9ss7a9alu Right - a fascinating thing about code is that it ends up being quite a solid workaround for the hallucination problem - because if the LLM hallucinates a solution it usually won't execute or compile!

nostr:npub15ppsm6mqc2ygywe2yy5hwe9drr47gaw6az2gda5hjzzpnae64uzs3n4jdw it's mostly GPT-4, either through the ChatGPT web interface or using my own https://llm.datasette.io CLI tool - I use Copilot mainly as a typing assistant and occasionally have it write me a test or two

There are SO MANY valid criticisms of LLMs that I can respect, but denying how much they can accelerate the act of throwing together useful code isn't one of them

Answered a question on Hacker News about using ChatGPT for programming tasks with a bunch of my own examples... https://news.ycombinator.com/item?id=37126182#37127206

Comments so far include "If someone on my team was doing things this unusually they’d probably be let go" and "sorry but these examples are not impressive at all and by no means a representative of any serious programmer's workload"

A lot of people out there are absolutely determined not to take the practical utility of this stuff seriously!

nostr:npub153xffgkega29h262u5gtnh4gpd8kwecmglxgtt3x3w4edddhzxrqh4wc42 wow, those really are excellent

Anyone got thoughts on good backup strategies for data where the only copy is in an S3 bucket?

I worry about things like catastrophic billing failures or some weird ML algorithm deciding to ban an AWS account for some reason

Backup to a separate account in another region? Mirror bucket contents to a server on another hosting provider?

New release of my llm-mlc plugin for https://llm.datasette.io/ - this plugin allows you to run large language models like Llama 2 using the MLC Python library, which means you get GPU acceleration on M1/M2 Macs! https://github.com/simonw/llm-mlc/releases/tag/0.4

nostr:npub1f00qgrs88cd8rhq5vqeh5s73gh6qtqe3qadw6l909j8s92qhg02s9dr2vn I've been trying that for a while with my annotated release notes series - https://simonwillison.net/series/datasette-release-notes/ - but it's begun to feel like maybe I should put all of that effort into just the release notes themselves rather than splitting it into two.

nostr:npub1qcy4xnxgwr75mlugku3qc2aqskgsymy847helej72ke2y823r7uqesqpzu oh interesting! I don't think I've seen pushback about any of my projects releasing too often, but that's probably because they either don't have huge audiences or don't ship more than once a month or so

Maybe a better way to think about release notes is that they should read (and be written) like blog entries - each one is a chance to both update the users on the latest news about the project and promote that project to people who haven't discovered it yet

New release of LLM, my CLI tool and Python library for interacting with Large Language Models

The main new feature is the ability to assign new aliases to models, useful for dealing with things like "mlc-chat-Llama-2-7b-chat-hf-q4f16_1"

https://llm.datasette.io/en/stable/changelog.html#v0-7

nostr:npub14p3antm8cnvlqx3km7fp4ywyyxq7przhxay0d9gaf6hxkw984cxsm8x6r5 generating little jq and bash scripts is an ideal application for untrustworthy LLMs because hallucinated code won't work, so you can spot any hallucination problems pretty fast!

nostr:npub14p3antm8cnvlqx3km7fp4ywyyxq7przhxay0d9gaf6hxkw984cxsm8x6r5 I found that particular paper very unconvincing once I started digging into the data behind it, they were marking answers as "incorrect" for pretty weak reasons in my opinion

nostr:npub14p3antm8cnvlqx3km7fp4ywyyxq7przhxay0d9gaf6hxkw984cxsm8x6r5 Practice! The more time I spend with different models the better my intuition for if they're going to give me a good answer or not

GPT-4 and ChatGPT are far, far more reliable than the models that I can run locally on my laptop

I'm only just beginning to build that intuition for Llama 2, it'll take a while

I spoke about that a bit in https://simonwillison.net/2023/Aug/3/weird-world-of-llms/#tips-for-using-them

nostr:npub1y4s7hjcxn9e2dxakrl98taqevp8vk9fn4axpkejqcuvackcg7f6sgnvgt8 Yeah this is Llama 2 7B (the weakest version of that model) running locally using my new https://github.com/simonw/llm-mlc plugin

It's very quick to take ethical objections to things that it shouldn't, but you can tamp that down a fair amount by giving it a less restrictive system prompt

... my mistake, that was a bug: It didn't run against the release notes, just against the single prompt "how do I install it" - so this is my bug, not Llama 2 refusing to give me installation instructions for my software

(Still funny though)

Oh this is excellent... I tried running that request for installation instructions through a local Llama 2 7B model and triggered the ethics filter!

Once you have a CLI tool for pulling release notes for a project, you can pipe them into LLM to use them to answer questions:

./combined-release-notes.sh simonw/llm | llm -s 'how do I install it'

https://til.simonwillison.net/jq/combined-github-release-notes#user-content-using-that-with-llm