I love Claude but this button is great. Turns out it doesn’t feel like a conversation if you can’t interrupt!

Cool example of Claude building a webapp from scratch to generate run routes

Fun fact, in 2018 an insight fellow did almost the exact same project.

Doing it all (including figuring out how to host it) took them 4 weeks. This took minutes

It’s easy to forget how far we’ve come

Prompting base models feels like driving a stick shift. More flexible, more annoying.

I’ve seen this happen in practice.

The carefully designed system that went through 8 design reviews and espouses all best practices.

The hacky code written to address an urgent incident.

They converge to the same unmaintanable mess. Code rots.

We've reached the point where models are becoming as persuasive as human writers.

There's a clear model size x persuasion scaling trend, and it doesn't look like it is plateauing.

Claude as a customer support agent can:

Ask for missing information
Call multiple tools in a row, using one’s output as another one’s input
Take actions such as canceling orders (we definitely recommend introducing a confirmation flow)

So excited to finally see tool use out in the world!

Improving Claude’s ability to use tools was a core focus when building the Claude 3 family.

Looking forward to seeing what you all build with it

In Paris for AWS summit this week.

DM me if you want to meet up

Agreed! This is one of my pet peeves.

Folks also severely underestimate the amount of data humans consume. I did the math for a few modalities here:

mlpowered.com/posts/ml-model…

Cool post on a reproduction of the blog post creation demo.

Some differences on approaches but directionally similar.

Minor point on hallucinations: some are actually errors in the transcript!

[Arena Update]

70K+ new Arena votes🗳️ are in!

Claude-3 Haiku has impressed all, even reaching GPT-4 level by our user preference! Its speed, capabilities & context length are unmatched now in the market🔥

Congrats Anthropic on the incredible Claude-3 launch!

More exciting…

This is super cool, had a smile on my face watching it

I’ve started to come around to the view that fine-tuning will become less and less necessary as models become better at in context learning

I’m in this image and I like it.

There is a learning curve to using LLMs for your work, but once you do it makes you more ambitious, and helps so much with procrastination and dealing with cruft.

A short update on interpretability work.
I recommend reading it if you are curious about how LLMs work under the hood.

I don’t think people have fully internalized yet how cheap and how good Haiku is.

Opus gets all the press for good reason, but Haiku pushes the intelligence / cost boundary even more imo

man the internet is quick at finding easter eggs.

🥹Silly little guy🥹

It's Claudes all the way down

Cool to see how much overlap there is between this advice and the advice I give folks wanting to transition to ML roles:

Just do a project that isn’t related to a course, on a topic that is useful to you, ideally with a smart friend or mentor.

Just ran SWE benchmark on opus-20240229, out of 1129 questions it gets 11.07% correct