ajdecon

Really interesting thread! Worth reading since I’ve seen some folks holding up the NY legislation as a good model

Oh I like this description. Then again I also like linear algebra…

Technology Choices TTS Model: Kokoro-82M via kokoro-onnx Use kokoro-onnx for all text-to-speech synthesis. Do not substitute another model without user approval. Reasons (for context, not for debate):

For ${reasons} I’m having Claude draft a plan for Gemini to implement.

I’m amused that Claude is very clearly not open to Gemini’s opinions on its tech choices.

I continue to say, LLMs are expert power tools that have been marketed as user-friendly magic wands.

If you have a gaming PC with a good GPU, or a recent Mac, the “open” models are good enough these days they can do a decent job.

The annoyance is that all the tooling is very much aimed at software devs, so it’s not exactly easy to use. LM Studio is probably the best one to check out right now.

This seems reasonable to me. There are a number of scenarios where LLM text is useful and reasonable, but they should be clearly labeled and not passed off as human work.

Here's a tiny bit of hopium:

In 2024 we were planning on adding 160 GW of renewable energy to the grid by 2030.

As of the start of this year, we had added 100 GW and are planning on adding 220 GW more by 2030.

(In 2022 we planned for 60 GW in 2030. We hit that by the end of 2023)

I can’t decide if that would feel fun or traumatic. And I worry I would just get angry at inaccuracies 😅

I don’t particularly go looking for them, but I definitely mute or block folks I see just generally being assholes in others’ replies.

the closest thing may be UN peacekeeping missions?

or for “major” conflicts the last one I know of was the Gulf War in the 90s, which was actually UN-authorized

Yep. LLMs do well if the work itself is expensive and time-consuming, but checking correctness is cheap and fast. (And ideally automated.)

Software happens to fit in that sweet spot — if you have very good automated tests anyway.

Not so much if a human needs to manually proofread a bunch of text.

I’ve been very happy to see Anthropic stick to their principles. Hoping this continues despite this pressure from the administration.

I’m a fan of Aaron’s podcast, so this is pretty exciting

SCP:GALLIONIC | Official Trailer YouTube video by soem studios

www.youtube.com/watch?v=RSbP...

Until I read this, it did not even occur to me that priests are using ChatGPT to generate homilies, but of course they are, and my long-lapsed-Catholicism is kicking down doors in my skull yelling “HOW DARE THEY?!?”

📸: @leighaf.bsky.social

A calico cat yelling with her mouth wide open while sitting in front of a TV showing the hockey game

Phryne has OPINIONS about the bronze medal hockey game

HuggingFace’s Smol Training Playbook also explains through essentially the same approach and their deviation from it, buried in their model design chapter (in a section called “scaling laws”!)

oooh, thanks for the reminder this article of yours exists. it’s one of the better short explanations of the concept I’ve found!

Not sure! Does Oxide have a sufficiently chaotic-looking lab that you could pose in? ;-)

@ahl.bsky.social @bcantrill.bsky.social we need a new cover image for Oxide and Friends in this style

The first edition is a book I've recommended So Many Times. Looking forward to the updates in the new edition!

Almost every morning while making coffee, I have two dogs and at least one cat circling me demanding things. It’s my morning agility course!

I think this is a pretty good & quick argument regarding why some researchers believe LLMs will continue to progress from a technical lens

(The 'AI Bubble' title will make you think it's about finances/company specifics— it's not. Definitely more about the technology itself)

Heard someone refer to a (large) datacenter as an "ohio-class datacenter" and now I only want to hear datacenter sizes referred to by reference to submarine classes

fwiw I take that as a call to action and hopeful, not fatalistic! we don’t have to accept bad candidates years before an election, we can actually influence outcomes.

but we have to start early, and not just engage every N years (depending on office)

stated another way, leverage is higher the earlier in the process you engage!

donating, lobbying, voting for primary candidates gives you a higher ability to influence the outcome.

by the time you hit the general, it’s a lot lower: one of the people on *that ballot* will take office.

Unironically yes, LOWER DECKS is about how the response to the disappointments of the post-Cold War era and the complexity of the modern world should not be despair or cynicism but instead a recommitment to the original ideals of liberalism, thank you for coming to my TED talk.

yep I’ve lived this experience

Look it’s a lower bound ok?