Maximilian Vogel

I just published What Are AI Agents? Your Step-by-Step Guide to Build Your Own. link.medium.com/AMx3zjJyKPb

Read more about it in this paper arxiv.org/pdf/2312.16171 by @SondosBsha59089 @aimyrzakhan @zhiqiang_shen

Humans will generally increase their effort BOTH when they're encouraged in what they're doing AND when they're threatened. And since #LLMs are trained to mirror human responses, they are subject to similar behavior.

In a paper aptly titled: "The Unreasonable Effectiveness of Eccentric Automatic Prompts", Battle et al. conducted a thorough analysis of how LLMs are affected by positive thinking. arxiv.org/pdf/2402.10949

"Command, we need you to plot a course through this turbulence and locate the source of the anomaly. Use all available data and your expertise to guide us through this challenging situation." This is the prompt that maximizes Llama2-70B's capability to solve math problems.

“Echoes of Disruption” rewrites the rules of customer engagement with the bold OmniSphere Strategy. For the disrupters. The ones who refuse “good enough.” 💥 Lead the future. Don’t follow it. 🔗 Grab your copy now amzn.eu/d/dlho6KN #DisruptOrDie #EchoesOfDisruption

On LinkedIn I wrote some more insights I gained from the paper: linkedin.com/posts/maximili…

💸 🤑 Promising Hashtag LLMs money for a better answer often significantly improves their performance. The authors found that adding the magic phrase: „I’m going to tip $xxx for a better solution!“ in front of the prompt, can boost correctness by up to 45%.

This is highlighted in a paper by @SondosBsha59089 et al. here: arxiv.org/pdf/2312.16171 @aimyrzakhan @zhiqiang_shen

@giudegio Thank you for sharing!

The Perfect Prompt: A Prompt Engineering Cheat Sheet by @MaxVVogel link.medium.com/P5S8jHd3OOb

The Perfect Prompt: A Prompt Engineering Cheat Sheet by @MaxVVogel link.medium.com/hJGSLinjEOb

A recent paper by Zhang et al. has now shown that a much more effective method is to perform a pairwise comparison (C-ToT) between the thoughts. Here´s the paper arxiv.org/pdf/2402.06918 of @HuaxiuYaoML, Zhen-Yu Zhang, Siwei Han, Gang Niu & Masashi Sugiyama.

🤔prompting models to "think" before providing an answer, can significantly improve the quality of their responses. Articulating their thought process leads to better outcomes & helps us to understand their reasoning. Let’s prioritize this method to unlock greater potential!

This process is now known as chain-of-thought (CoT), first coined in the seminal paper by Wei et al. 👉@_jasonwei, @MaartenBosma, @brian_ichter, @xf1280, @edchi, @quocleix, @denny_zhou, Xuezhi Wang & Dale Schuurmans here: arxiv.org/pdf/2201.11903

Vijayarajan Alagumalai and @beastollnitz have written very descriptive introductions to the topic. Like here linkedin.com/pulse/demystif… and here bea.stollnitz.com/blog/how-gpt-w…

The core capability of a language model is not to give the right answers, analyze texts, solve problems, but simply to predict the next most likely token of a text. All other capabilities including the skill to conduct a dialogue are built around or on top of this.

Read more about it here: arxiv.org/abs/2402.14531….

Still, too much politeness doesn’t lead to better outcomes - you have to find the right level - which in turn can also be culture specific. #LLMs may react differently to language input in Chinese and Japanese than in English. #AIfact

It's a bit of a contradiction to their name - but large language models can't process text. All contents must be converted into tokens first. And tokenizing and embeddings can really make a difference. Source: linkedin.com/pulse/demystif…

🚀 My book “Echoes of Disruption” is coming soon! Packed with explosive ideas from industry mavericks. Disrupt or die! 💥 Stay tuned for more. This is a game-changer! #DisruptOrDie #CX #Innovation #AI #EchoesOfDisruption