• Everyday AI
  • Posts
  • Secret knowledge for better ChatGPT outputs 🤫

Secret knowledge for better ChatGPT outputs 🤫

🤖 Understanding LLMs' knowledge cutoff, Sports Illustrated's AI-generated controversy, a custom GPT that verifies emails, and more!

Author

Everyday AI
November 28, 2023

Outsmart The Future

Today in Everyday AI
7 minute read

🎙 Daily Podcast Episode: Why do AI chatbots lie? It starts with knowing the knowledge cutoff date. School’s in session y’all. We’re breaking down how and why LLMs have cutoff dates. Give it a listen.

🕵️‍♂️ Fresh Finds: AI that helps you manage and engage contacts, Forbes 30 under 30 for AI, and Tim Cook and Dua Lipa’s chat about AI. Read on for Fresh Finds.

🗞 Byte Sized Daily AI News: Sports Illustrated's AI-generated controversy, Amazon AWS AI updates, and Pika Labs secures $55 million in funding. For that and more, read on for Byte Sized News.

🚀 AI In 5: Doing some email outreach and not sure if you have a valid email address for that prospect? We found a custom GPT that verifies email addresses for free! See it here.

🧠 Learn & Leveraging AI: Still wondering what a knowledge cutoff date is and why it matters when using an LLM? Keep reading for that!

↩️ Don’t miss out: Did you miss our last newsletter? We talked about the future of AI text-to-speech, ChatGPT voice now free for all users, and how to use multiple LLMs at once. Check it here!

Knowledge Cutoff - What it is and why it matters for large language models 🤔

Wondering why AI chatbots lie to you sometimes?

It usually starts with understanding what a knowledge cutoff date is and why LLMs have them.

So what’s a knowledge cutoff date?

We're cutting through the tech jargon to give you a clear view of how AI thinks and learns.

Also on the pod today:

• How LLMs are trained 🧠
• Why knowledge cutoffs matter ☝️
• Comparing different LLMs and knowledge cutoffs 🥊

It’ll be worth your 42 minutes:

Listen on our site:

Click to listen

Subscribe and listen on your favorite podcast platform

Listen on:

Upcoming Everyday AI Livestreams

Wednesday, November 29th at 7:30 am CST ⬇️

Here’s our favorite AI finds from across the web:

New AI Tool Spotlight – Zixflow helps you manage and engage contacts, ob1 generates APIs, databases, and your backend with a prompt, and Walling AI turns your idea into a complete project in seconds with AI.

Big Tech – A senior Google Executive is calling for innovative AI legal frameworks.

Trending in AI - Forbes has released its 2024 30 under 30 for AI enterprise tech.

AI in Medical – See how AI is transforming healthcare risk adjustments.

Pop Culture – Apple’s Tim Cook and singer Dua Lipa sat down and talked AI. Here’s what they said.

1. Sports Illustrated's AI-Generated Writers Stir Controversy 📰

Sports Illustrated recently faced scrutiny for publishing articles under seemingly fake bylines, later found to be AI-generated. These 'authors' had detailed biographies but no real-world presence, with their profile photos available on AI-generated headshot websites. Investigations revealed that not only the bylines but some of the content itself was AI-generated, leading to questions about journalistic ethics.

2. Google and Symphony's AI Venture for Banking Compliance 🏦

Google and Symphony, a tech firm, are collaborating to enhance voice analytics for banks and investment firms. This partnership aims to address the growing regulatory demands in communications compliance, especially in tracking and recording voice and video calls. Utilizing Google Cloud's generative AI platform, Vertex AI, Symphony's Cloud9 voice product will now feature advanced speech-to-text and natural language processing capabilities.

3. Amazon's AI Updates at AWS Re:Invent 🚀

Amazon AWS is set to unveil its bold vision for generative AI at the AWS Re:Invent conference, challenging rivals like Microsoft Azure and Google Cloud. Key highlights include the Bedrock service for easy access to a range of AI models, advancements in vector databases for enhanced semantic search, and the integration of generative AI in applications like Amazon QuickSite and HealthScribe.

4. Pika Labs' AI Video $55M Boost 🎥 

Pika Labs, an innovative startup, is revolutionizing video creation with its AI-powered platform, having secured $55 million in funding. The platform, Pika 1.0, enables users to generate and edit videos from captions and still images, offering styles like 3D animation, anime, and cinematic. Its unique features include extending video lengths, transforming styles, and even adding characters using AI.

5. Two-thirds of Americans Think AI Can Do Their Job 💼

A recent survey reveals that two-thirds of American workers believe AI could perform their jobs. Despite these fears, 78.1% of respondents see AI as a stress reducer at work, and 76.7% think it could shorten the workweek. The survey, conducted by Spokeo, also found that 79.1% of participants advocate for employer-provided AI training.

Use AI to...verify emails? Bounceban custom GPT review 📩

Need to send an important email, but not sure if the email address is correct?

Or, maybe you're doing some cold outreach and don't wanna shell out some big bucks for a paid email verification service.

Well, we stumbled on a Custom GPT that's ACTUALLY useful.

🤷‍♂️ What’s Going On and Why It Matters:

LLM 101 class is in session, shorties.

And everyone can learn, even if you work in the GenAI space like Mike and Natalie.

Today, we dove headfirst into cutoff dates in large language models. 

Here’s the skinny y’all — understanding any GenAI model’s knowledge cutoff is the first and most important thing to understand if you wanna improve your outputs. 

So many times, people are rushing to decorate their second floor windows while ignoring their shaky foundation. Same thing with prompting inside a large language model — the best prompting in the world means jack squat if your foundation is shaky.

Today, Jordan went solo and demonstrated live the ins and outs of cutoff dates. In short, all LLMs are trained on data from the internet. And, like a textbook, there's a publication date for each model. (That’s the knowledge cutoff.) 

We touched on: 

  • Overview of large language models and their training process

  • Importance of knowledge cutoff in maintaining accuracy and avoiding false information

  • Emphasis on transparency in working with large language models

  • The different knowledge cutoff dates for different modes inside ChatGPT

  • Explanation of different models of learning: deep learning, machine learning, and reinforcement learning from human feedback

And much, much more.

Ready to talk about what it means?

Let’s get after it. 🤘

🦾How You Can Leverage:

Cliffs Notes from our LLM 101 class: if you don’t understand knowledge cutoffs, your outputs are gonna be garbage. 

Like we talked about on today’s show, it’s important to view LLMs like a hard-copy, printed textbook. ChatGPT and other LLMs use a fixed data set that is sometimes almost two years out of date. 

(That’s your #1 cause for hallucinations!)

So, here’s how to make sense of it all to grow your company and careers. 

1 – Be Mindful of Cutoff Dates 🗓️

As we showed live, even asking an LLM about its cutoff date can provide perplexing results. Remember — all GenAI models act as the world’s most ridiculously advanced autocomplete.

So, even HOW you ask an LLM about its knowledge cutoff date can produce different results.

Using the Chathub tool we reviewed yesterday, here’s a breakdown of what some popular LLMs say when you ask them about their knowledge cutoff. 

And here’s what they say when you ask them about their training data.

Try this:
Let’s cut through the fluff, and get facts out there. 

Cutoff dates (To the best of our knowledge and testing this week.) 

  • GPT4 (paid version) — April 2023

  • GPT4 with plugins (paid version) — January 2022

  • GPT 3.5 (free version) — January 2022

  • Bing Chat (always free) — Unknown. Uses different GPT models for different modes, so likely between January 2022-April 2023, although it is connected to the internet so it can query information after that date. 

  • Google Bard — Unknown, but likely 2023. Like Bing Chat, Bard is also connected to the website and can query recent information. 

  • Anthropic Claude 2.1 — Exact date unknown, but early 2023. 

  • Meta’s Llama 2 — Exact date unknown. Likely sometime between September 2022 and December 2022. 

2 – Prompt around cutoffs 🚏

Simple way to get around knowledge cutoffs? 

Tell ChatGPT (or whatever LLM you’re using) the newest information.

If you’re stuck with too many hallucinations, try just TELLING an AI chat the most up-to-date information for it to retain. 

It’s like printing off new pages and taping them in that textbook. 

Try this: Only downside with this approach? Memory. Even with extended context windows, LLMs can lose their memory pretty quickly. That means that if you give an AI chat new information, like copying/pasting a long article, you’re chipping away faster at its memory.

To make sure AI chats keep their memory, check out our full episode on tokens, or this quick tutorial on counting tokens in ChatGPT. 

3 – Take our free PPP Prompting class 🆓

Best way around knowledge cutoffs in ChatGPT?

Knowing how to actually use it the correct way. 

More than a thousand business leaders, department heads, students and everyday enthusiasts have taken our free Prime, Prompt, Polish course. (No upsell at the end, either.)

We update it every week with the latest info. 

Try this: Why haven’t you taken this course yet? We keep the registration page private. Just reply to this email with PPP and we’ll send you the registration link. 

Bonus tip: Another great way around knowledge cutoffs? Knowing the right plugins to use. Join us tomorrow LIVE to see what’s new with ChatGPT plugins. (There’s a lot!)

Now This …

We wanna hear from you!

Vote to see live results

Which LLM do you use?

Login or Subscribe to participate in polls.

Reply

or to participate.