Notices by mnl mnl mnl mnl mnl (mnl@hachyderm.io), page 2

I'm seriously impressed by the quality of the MoE and even 7B models I've run on my 4 year old M1 over the weekend. This is absolutely solid for most of the really useful ways of LLMs I actually use them for day to day.

Already most of these tools allow me to use local models or cheaply hosted versions of them. I can see the total power used in daily inference be about 10 minutes of running 4 cores.

These models and community finetunes are already much more fun than GPT or claude.

#llms

1/

Really really curious about where things will be in a few months. I don't think the big players will stay on their pedestal for much longer.

They just can't compete with the myriad of options a community can provide.

#llms

2/2

@lain @sofia what about selling people on the concept that the machine could do it better, when it kind of can't? Or if it was the case that a human + the machine could do way better.

@lain @sofia yeah i meant more like "let's fire research assistants and replace them with AI". Maybe not the best analogy here.

@lain @sofia So much of the use-cases really depend on the context. For example, I build many "automate this job entirely", say, writing SEO titles. Because I work for a family business, that means that Joe can now, instead of grinding through 5000 SEO titles, do more product photography.

But, I could also build "fire all content writers" machine, with the same prompt.

@sofia yes. Calling for a lock down of training data based on copyright law would be a terrible thing for opensource models, imo. There's also a lot of research that goes into generated datasets and just curating leads to much better training results.

@sofia like i'm perfectly fine ethically with training my own model on the entirety of all papers in libgen... As long as I don't create the "replace a scientist machine". And while I take openai's stance with a pretty big grain of salt, they aren't (compared to midjourney, say), pushing for the "let's replace people" model, and provide tools to remove your self from the training corpus.

FWIW, lip service does a fair amount here and is not in the headlines.

@sofia i'm really taken aback by the copyright stan-isation of the countercultural debate here, as if the nytimes was a defender of the rights of individual journalists. No, if they were to train their own model on their journalists output they would go ahead, just like adobe and shutterstock and co.

openai paying nytimes licensing fees won't change anything for the better.

Quite cool this printer. I need to find q workflow to link to archive copies not transient pages, but this rules.

Pretty refreshing after the adhesive hell of the iPhone.

#android #fairphone #righttorepair

I should try to do more with shortcuts.app . Since I'm actually planning to switch to android and getting a refurbed fairphone, is there something similar on android?

I've been a little obsessed with QR code for a long time, and realized that they work well to bridge digital and analog. I can print out this drawing and put it into my sketchbook, yet can easily still embed and retain digital information easily (writing urls by hand is... nope).

#sketchbook

@inthehands having examples of mediocrity is what is often missing in teaching and why I think being amongst other students (who are not perfect by definition) is one of the most valuable parts of learning at a school, imo.

Being able to bounce stuff off an ai when self learning is what I really value, because it teaches me what boring is, and it puts me in the teachers shoes to explain what it did wrong.

I often double down on that concept by having the ai explicitly take on that role.

1/

@ravenonthill @graydon @edmistond how is it clear that he was right?

Even as programmers, we program in natural language all the time, in documentation, comments, variable names, rfcs, GitHub issues, planning meetings, etc…

Dijkstra himself was a great natural language communicator, if it weren’t for his cheap quips.

This is the direct consequence of “I program with vim and 80s unix tooling is the best we’ll ever need” thinking, imo

@joesabin I’m kind of fine with code bloat, in the sense that we can build software that does a lot more. But json parsing is just… complex, feature-poor and json is an infuriating format and not even that readable to boot.

@andrej

Depending on the use case:
- XML
- protobuf/flatbuf/any decent binary serialization (much easier to handle)

For readability and human consumption: YAML

@graydon @edmistond which is why we have json and json schema and json ld and bjson and oO

The fact that 99% of my RPC and event serialization has to be done in json is blowing my mind.

Where did we go wrong??

“It’s so simple” they say, wtf?

The snark at #php in US-developer circles never ceases to amaze me.

It’s like people somehow remember their php3 experiments when they were 12 or something and feel the need to show that they are somehow above that.

Nah dude not only is php a decent programming language, but it’s also a language that allows you to have a working site whatever your experience level, which is why 12 year olds have loved it throughout the ages, and that’s fucking amazing in its own right.

I really wish I understood why bing and google rolled out generative ai as part of their offering, as a psychological study. How do the myriad of human parts that make up these company-machines work so hard to make something happen that is… so detrimental to their core business in the first place.

Like is it at the end of the day “we all just wanted a promotion and not make too many waves and not think hard enough”?

Is it sycophantic fog around decision makers?

Why?