Wilson Bennett LLC的动态

Quick update on the AAA services LLM Agent... Making good progress despite my rusty Python skills. I've gotten to the point of what Pydantic AI refers to as "Reflection and self-correction". I suppose it's just a fancy way of telling the Agent to retry the prompt more than the default of 1. ?? This is a big step towards ensuring deterministic results for the Playbook calling the Agent. As has been the case since the beginning, I can get really close to deterministic output, but at some point, the LLM varies the response which throws the rest of the Playbook off. With Pydantic's validation of the LLM's response, and, telling the Agent to simply retry the prompt, I'm actually seeing self-correcting behavior from the LLM. Pydantic will throw an exception, issue a retry, and then the LLM returns a response that passes validation. My hope is that this is the last of dealing with subtle changes in the LLM's response to the same prompt. Fingers crossed! ??

Researchers who do linguistic analysis on text from documents should consider using riveter, a Python ?? package, that builds on connotation frames of verb predicates to analyze for dinensions of power, agency, etc. I have used it on Reddit posts from several threads, adoption reports, and in my current work, user agreements documents: “You should not use our services if you do not agree with our conditions” Who has the power here? The service provider? Or the user? You get the idea of how useful connotation frames can be. Here’s the paper on the Python package from Maria Antoniak (whom I am a fan of her work). https://lnkd.in/dzwZd883

Build a ChatGPT Clone with Llama 3.1 running locally on your computer (100% FREE and Offline) In this tutorial, we'll walk you through the process of creating a ChatGPT clone by just using Llama-3.1 model, LM Studio, Streamlit and Python. Step-by-step tutorial with Python code: https://lnkd.in/dsEUUQzi If you find this useful, Repost to share it with your friends.

Honest question: why is this project understood to be satire, but the exact same idea in the context of conversational AI is so hyped? https://lnkd.in/eqECGBAD Redditor JirkaKlimes shared a python library called jit_implementation. How it works: 1 You write a function or class signature and a docstring. 2 You slap the @implement decorator on it. 3 The implementation is generated on-demand when you call the function or instantiate the class. Lazy coding at its finest! It's funny, clever, and obviously not intended to be used in production apps. But I hear people raving about ReAct agents and insisting that it's a *good thing* that your business logic is newly invented on-the-fly every time a user interacts with your assistant. If an LLM can do a good job of writing this code, why not generate it once, and then have code you can check, version, modify, and re-use? Here's an example of how you can generate code for your CALM bot. Do it once, and now you have a declarative, reliable bit of business logic https://lnkd.in/ecSphCdK

Nice, and further confirmation of the concept of Pre-Generative Conversational AI proposed in this paper (https://lnkd.in/dxqPGuFc) and implemented in Talkamatic Studio for FAQ and Educational dialogue (Instructional and Search/Sales dialogue are in the pipeline). Just provide some content, generate a dialogue, and edit (no dialogue expertise required). Check it out on https://talkamatic.se/en/ and sign up to try it out!

Let me share a story. I asked Claude about performance comparison between using multiple small regex patterns versus one regex pattern with multiple strings in TypeScript, asking which approach would be faster. Claude performed well by immediately creating JavaScript code and running benchmarks for me instead of providing explanations. I did the same test with ChatGPT, but it first gave lengthy explanations and then asked if I wanted to run benchmarks. When I said yes, it ran the tests in Python. This confused me, so I asked it to run the tests in JavaScript, but it said it couldn't. What I learned is: If you're using Python, ChatGPT might be more suitable. If JavaScript is your main language, Claude might be the better choice.

To effectively scale code and implement foolproof data solutions, you need the best Python Developer in the market to fit your business needs. WeCP's Python Assessment Test is the game-changer you need! ?? ? Evaluate candidates' Python skills through coding exercises, MCQs, and more ? Get detailed insights on performance, plagiarism checks, and proctoring flags ? Access a premium question bank curated by experts Don't settle for less - identify top talent effortlessly! ?? Create your own personalized test with just a prompt using WeCP AI or learn more about this test: https://hubs.li/Q02P2qZS0 #WeCP #Createatest #PythonAssessmentTest

Why did Python switch from an LL(1) parser to a PEG parser? From what I understand, PEG (Parsing Expression Grammar) is clearer than CFG (Context-Free Grammar). This means that a PEG parser always gives one correct parse tree, while LL and LR parsers can struggle with ambiguity. However, PEG does have a downside. Since PEG solves ambiguity by ordering the rules, it won’t warn you if your grammar is still ambiguous, unlike LL or LR parsers, which give errors like "shift/reduce" or "reduce/reduce" conflicts. In software engineering, it’s often better for a bug to crash the program rather than go unnoticed. Anyway, I still think LR parsers are the best.

Sharing a pro tip today! Here's a cheat code for LLMs. Recently I was working with Chat GPT. I was working on a project. This project required python code. I'm not a programmer or engineer. I had Chat GPT create some python code. There was just one problem. I kept getting an error. I input the error into Chat GPT. Same error. I could not figure out why the error was happening. I took the code. I took the error. I input them into Claude 3.5. I had Claude peer review Chat GPT's code. It found a few errors. It provided updated code. I put that code back into Chat GPT. It understood where the error was. It took getting out of it's own "head" to work. Next time you're struggling with an LLM, get it out of its own head. You'll be surprised the outputs you can doing this. P.S. What are some tips you'd suggest with LLMs?

Watching a little Monty Python today, a xmas habit. Reminded me of this... ?? "If you're very very stupid, how can you possibly realize that you're very very stupid? You would have to be relatively intelligent to realize how stupid you are... In order to know how good you are at something requires exactly the same skills as it does to be good at that thing in the first place. If you are absolutely no good at something at all, you lack exactly the skills that you need to know that you're absolutely no good at it." ?? Explains so much, doesn’t it? Sir John Cleese nails the dilemma of ignorance and self-awareness: the worse you are, the less likely you are to know it. It's painfully funny but also painfully true. ?? Surround yourself with people who challenge you and let you know when you're off track. Stop assuming you're an expert just because no one’s told you otherwise.