watson Ξ

Cloud cover map... digital.weather.gov?zoom=7&lat=3... (Note you need to scroll to the right time)

Eclipses visible in San Francisco, California, USA Which upcoming lunar and solar eclipses are visible in San Francisco, California, USA, and what do they look like?

Bunch of eclipse timing info... www.timeanddate.com/eclipse/in/u... (put in your location if out of PDT)

Thursday night 10pm PDT all across the US total lunar eclipse.
Map near the Bay area is the cloud cover expected at this time tomorrow. Fingers crossed it shifts a little! Links to cloud map (whole country) and helpful info on timings wherever you happen to be provided in thread below

I filled in this form just now and was emailed an API key a few seconds later, strongly recommend doing the same

Anthropic CEO Dario Amodei warns: AI will match ‘country of geniuses’ by 2026 Anthropic CEO Dario Amodei calls Paris Summit a "missed opportunity" as U.S. and European leaders clash over AI regulation.

2026 is next year. As I see it, this is an invitation to evolve. Regulation is necessary, but what if AI could also help regulate us? Could it, if shaped with intention, become a true partner in collectively amplifying human wisdom and helping us meet this moment?

venturebeat.com/ai/anthropic...

Seems pretty reasonable to me for there to be extensive trees of selection and mixture. Seems hard to have optimality on the key metrics that matter (speed, cost, quality) otherwise

Also in some sense, all neural architectures by their very nature are mixtures of experts! So... Yes :)

I believe this would have to assume that such a product would be perfect and that workers that use it would be perfect. Otherwise, companies want support, customization, additions etc. The same reasons that complex open source software can still make money.

The immediately adjacent example of open source software would seem to call this statement into question. Where does the "extremely unlikely" part come from with that taken into account?

Datasets & Experiments - LangfuseDiscordGitHubGitHub Use Langfuse Datasets to create structured experiments to test and benchmark LLM applications.

Parea is a solution that is workable but still young. We've been using it with medium success.

Langfuse is something else I have been considering: langfuse.com/docs/dataset...

Yeah it seems to be a weird blind spot for the industry. Weird because it seems like everyone building anything for which they desire defined repeatable behavior with LLMs should need this more than python developers need pytest. Otherwise you wind up with regression whack-a-mole

@simonwillison.net do you have a favorite way to do pytest-like LLM evals? Deterministic functions and/or LLM-as-judge

@5calls.org is making it easy for us to get in touch with reps and push them to stand up

Great tool (and a good account to follow here) to access your reps very quickly and know what to say. You could also paste in the script to Claude or ChatGPT for quick customization.

Let's spur these ppl to action. Complacency from our reps is not acceptable

Is there a reason there's an email wall before I can get any useful info ? I'm sure that turns a lot of ppl around. 5calls.org mentioned above seems lower friction

My housemate is part of a team frantically backing up science data from public servers due to be purged. Let it be known in 2025 these are the times in which we live in the United States

My understanding is you can't. But you can use a pin 📌 emoji apparently in a comment combined with a plugin or add-on of some kind. To much for me

Of course you can always copy out the link

A man dressed in brown and khaki hugging an enormous chinkapin oak, Quercus muehlenbergii. Source: Katherine Shaw

I have a suggestion for you. In times of trouble and strife, there is therapeutic benefit in being among trees and in physical contact with them. So if you can, go hug a tree today, the bigger the better. This huge old chinkapin oak is known as The Great God Tree.

💯

Seems like a huge opportunity for a group to come in and build FB lite on top of the AT protocol, even if they want to keep BSky an X clone

v cool Carson. Congrats. I'll be following this

Truly entertaining to read. R1 is the latest from the leading Chinese AI lab... While o1 is flipping to Chinese to reason, R1 is reasoning like the voice in your native English-speaking head.

From what I understand, more fun to read the reasoning than be impressed by output so far

I'm compiling my own list here, with yours truly included by pure nepotism

bsky.app/profile/did:...

Watch Duty is well built and wonderful for alerting you of fire risk nearby and then providing you with the most detailed up to date information. Kudos to you @lukas.blue

Hey, I (am one of the people who) made this!

And yes, days like today are what it's built for.

Fun fact:
(20 + 25)² = 2025

The next time this will happen will be in 1000 years! 🤯

These are indeed useful, especially for people getting started here...