Same thing just happened with Anthropics numbers. Billions short of their stated numbers.

🤞

Yeah, this seems fine.

• using encoders to create useful sequence data representations in much more information efficient ways
• using transformers to act on sequential data, with attention altering focus dynamically instead odd classical AR style
• training neural networks inside differential equations is very effective
• you can output higher dimensional outputs, including classifications that need less human review

This is machine learning mixed with classical modelling, not “AI” by any definition of AI.

The whole threatening careers thing happened with navier stokes too.

These people are completely rotten to the core.

Pushover licensing?

CRAM is an excellent name

I’ve always hated AI comments. They’re the bastard fusion medium blog spam and data labelling from underpaid and exploited people seeping into the slop.

Comments are for why, edge cases, and caveats. They should help explain unintuitive behaviours that usually must be learned the hard way, they help expose a larger structures, they explain an optimization, a shim, or a complex regex, or they explain where a temporary measure exists. Things of that nature.

Self documenting code is about thinking of the reader. I’ve worked in places that uphold self documenting code as gold and banned code comments at all and I’m convinced that’s wrong. It’s a burden on readers to have to go through history to find the addition, but the file has changed names or moved without a proper git mv and it’s a waste of time.

On that note, learning objective C and Swift did a lot for how I think about what naming could be. I try to focus on clear verbs and nouns so the action and target are always clear, avoid side effects, include examples in my docs, and embrace typing structures like nanedtuples in untyped languages.

I really like some of the direction Canonical is going, but very much do not like how they’re getting there.

Rust core utils sounds great, this sounds great, but practically I feel both need a lot of time in the oven to come out right and that’s okay.

I also i feel like snaps have lost and can be considered a mistake. I don’t use snaps, so I can’t explain why, but everyone is saying it.

It doesn’t really feel like work, I just sort by weight+size in descending order

I worked in grocery stores as a teen and I just want that to remain available.

I also do not like the self checkout machine. They have an unacceptable number of steps to pay. They’re horribly programmed. I don’t like the small baggage areas.

This is why I unload them onto the belt in the order they must be bagged

s/pulling back/slowing expansion of installations/g

These are not the same thing, and if you look at any S curve you can see why.

JavaScript stringify strikes again!

Metrics look useful, I have some very niche but well harnessed use cases that mistral medium works for, but it always spends too long gathering context, hopefully large works a lot better for that.

It’s just 1TA49B

5 years ago I’d have thought ML models that size would be implausible.

I have been waiting months for this early summer release lol

Looking forward to reading about it

Shouldn’t there be a lock free version of this?

What if we try fencing tokens?

I mean, I really like the zig approach. It usually turns into “the caller is responsible for managing the memory”

But then when I’m the caller I sweat a little.

The good news is you don’t have to think about it or hand it over. The AI asks for full disk access and then pulls it all on its own. I hope you don’t have anything private in your machine.

I understand, I am having a good day, except I got a 95,000 loc 2900 file AI generated PR dropped on me today and all the variable are single letters used inconsistently. It’s like reading Malboge.

To be fair, my dad was just trying to stop him from getting in trouble with the cops, who largely turned a blind eye to little things that didn’t hurt anyone.