#Sometimes a great notion text software#
#Sometimes a great notion text code#
#Sometimes a great notion text download#

Today, you can just essentially say, 'Hello, omniscient autocomplete. That the next generation of spammers didn't find a way of getting around it. Then you would have hoped that that model was robust, it didn't drift.

You would have artisanally crafted a model and labeled a lot of data yourself. You would have meticulously engineered features out of the text, out of the metadata about the email itself. If you were getting your start in ML, like 10 years ago, like I did, or 20 years ago, the way that you would have built that first spam classifier problem is you would have downloaded hundreds of thousands of emails. For example, you can use them to approximate and build what 10 or 20 years ago was a mind-numbing process. One of the first things that came out, early research when these models started getting big enough, large enough, where they started having some emergent properties, is that they can solve a variety of other tasks that just aren't autocomplete. It turns out that not only are LLMs good at finishing our sentences, they're very good at solving a variety of other problems as well. It knows what token comes next, because it's read the entire internet, and it does a good job because we gave it treats whenever it gave us a good answer. Given a collection of tokens, it's really just pretrained to understand what token is most likely to come next. This isn't really that much of an aspersion so much as a reminder of what's actually going on underneath the hood. I think for the purposes of this talk, it's fine to have the most antagonistic, most basic, most aspersive description you can possibly have of LLMs, which is that they're omniscient autocomplete with a really bad attention span. I'm not going to go into particularly deep detail about LLMs.

#Sometimes a great notion text code#

This talk is mostly meme and code driven. One of the reasons why these talks are so hard is not only there's a spread of experience, but there's a spread in age.

#Sometimes a great notion text software#

We orchestrate a lot of LLMs to bring personalized Q&A, and personalized software documentation to folks onboarding on to Prefect. Being able to smooth out that onboarding experience by bringing personalized documentation and personalized Q&A to somebody who's stuck on something that you did not anticipate somebody would get stuck on. For many folks, they hit, they smack their face right into that first step, they leave your product and they don't come back. If you've ever built a product yourself, you know that one of the hardest parts is creating personalized documentation and onboarding to folks, especially if your onboarding experience is a bit of like a step function. Really, if Prefect is the backend for data engineering, then Slack is how those 25,000 engineers all come and ask us questions all the time. This is Jeremiah, the founder of Prefect, and Nate Nowack, a senior engineer at Prefect as well. I'm going to use we very liberally when it was mostly these two folks. A lot of what motivated what we built today is one predates me. If you want to get your workflows, pipelines, and agents to actually work, you can use Prefect today to orchestrate and observe your Python code.

It's a product of love from Prefect's backend for data engineering.

#Sometimes a great notion text download#

Of course, Marvin is an open source, free-to-use, you can download it today. You can just drop it in an invisible way among your code that actually works, so it's a tool that you can reach for when it's the right tool to use.

I stole that moniker from Django, which is really, how do you make LLMs an ambient experience so that you don't really have to spend most of your time begging something to work. Marvin is a high-level Python library for building with LLMs quickly and with less code. This is all through the lens of an open source framework that we built called Marvin. I'm going to talk about a bunch of things like this. I'm going to talk a little bit about how folks are using LLMs in the real world, specifically about building NLP pipelines declaratively. I apologize in advance for, if you've ever grown through a math lecture, I'm going to bring a lot of those vibes. Then, asides that, I'm a recovering academic and a recovering mathematician. Ran Insight Data Science here in New York for a number of years. Before that, I was a CTO of a company called Openrole, which used small, medium, and large language models to conduct job searches on behalf of folks. I'm the AI product lead at a company called Prefect.