😈 Dark Web LLMs


Can you imagine teaching math to a dog?

You could likely teach some basic arithmetic. But at a certain point the dog would be unable to go any further.

Maybe that's what it'll be like for humans looking to understand AGI.

Something so advanced our minds simply can't comprehend it.


Today's Ai5:

  • 😈 Dark Web LLMs
  • πŸ–ΌοΈ New Midjourney Editor
  • πŸ”₯ DeepMind's Imagen 3
  • πŸ€“ Remember this Google Engineer?
  • πŸ”— Claude Prompt Caching

Ps. You can now check out your unique Ai5 referral hub here!

Prompt of the Day 🎨

My face when ChatGPT starts talking about ethics.

Dark Web LLMs 😈

It's one thing to jailbreak ChatGPT using a clever prompt. But this only gets you so far; the good jailbreaks usually getting patched in a day or two. Let's take a look into the world of dark web LLMs to see what they're all about.

The primary objective for dark LLMs is for writing malicious code. Phishing emails and identifying existing vulnerabilities are of interest too, but with a small amount off effort you can easily get a regular LLM to help with this.

There's no shortage of dark LLMs to choose from either. The good news is, most of them suck. They're either low quality efforts, or simple ChatGPT jailbreaks that don't work for long. That being said, there's a few genuine efforts worth talking about.

WormGPT: WormGPT was initially sold exclusively on HackForums and now mostly on Telegram. Licenses go for up to to $5,000 each.

WormGPT aims to provide an alternative to ChatGPT, one that lets you do all sorts of illegal stuff. Everything blackhat related that you can think of can be done with WormGPT, allowing anyone access to malicious activity without ever leaving the comfort of their home. - Creator of WormGPT

WormGPT was a true dark LLM when it was released. Since then though, the creator has started putting guard rails of his own in place blocking things related to murder plots, drug trafficking, and kidnapping. You can read a full interview with the creator of WormGPT here.

DarkBARD: DarkBARD is said to be the evil twin of Google’s BARD. It's unqiue selling point is its real-time processing of information from the clear web. It can generate code and articles, and integrates with Google Lens for image-related tasks. This tool’s versatility is further underscored by its potential in executing ransomware and DDoS attacks. (source)

FraudGPT: FraudGPT is described as a great tool for creating undetectable malware, writing malicious code, finding leaks and vulnerabilities, creating phishing pages, and for learning hacking. (source)

So, should you be worried about these dark LLMs?

... probably not. No more worried than usual anyway. At most, they make life a bit easier for bad actors, just like AI does for the rest of us. More of a concern would be nation states who have the resources to create and train true dark LLMs. But there's not much that worrying will do to help that one!

New Midjourney Editor πŸ–ΌοΈ

Last week, Midjourney released a powerful new editor interface for the web app. The features aren't necessarily new, but they're now way easier to use! Let's check out what it can do! πŸ‘‡

Midjourney has always had its own version of generative fill. In the new editor it's way easier to use. I actually think the results have improved a bit too.

These are 3 different generations as I zoomed out. Notice her face stays the same? The new editor has a drag and drop style interface making reframing much easier than before. Also, you can now zoom out to 10% of the original image size - previously it was only 2x.

Combining all the features allows you to fully customize your ideas. Change backgrounds, reframe, and add/remove objects. Imagination is the limit!

To use the editor you must be using the web app, click any image you've generated (before upscaling), then click Editor in the bottom right of the screen.

Ps. The Midjourney web app is now accessible to anyone who's done 10 or more image generations in Discord. I highly recommend you move over to the web app from Discord if you're still using it!

DeepMind's Imagen 3 🧠

After a bumpy start in the AI race, Google are really starting to ship some decent stuff! This week giving access to their brand new image generator Imagen 3.

Overall the images are pretty good! And it has one feature which I loved. After entering your prompt, it automatically generates options for key words in easy to use drop down menus.

BUT (and it's a big but), in classic Google style, the safety guardrails are ridiculous. 7 out of my 9 attempts with various prompts failed, making it unusable.

Given what you can do with Grok 2 on X, I can't see anyone having the patience to use this tool... Which is a shame because the images aren't half bad!

If you've more patience than I do, you can try Imagen 3 here for free.

Remember this Google Engineer? πŸ€“

Back in 2022, before ChatGPT, Google engineer Blake Lemoine made headlines with claims that Google's internal chatbot had become sentient. 2 years later the debate is firing up again. Did his claims have some merit? Let's take a look.

Blake posted the full conversation between him and the AI, it's a great read. The back and forth conversation is interesting. It's nothing like what we could have with ChatGPT or Claude right now.

One part of the transcript I found interesting was the point when the AI started asking questions of its own.

And there was the sections that made the headlines too. πŸ‘‡

This really would have seemed like magic if you'd never seen it before, much like using ChatGPT for the first time. Personally I don't think LaMDA was sentient. What this does tell me though is that the AI we as consumers get access to are much less powerful (dumbed down) than those behind closed doors.

What do you think?

Claude Prompt Caching πŸ”—

Anthropic just released a new caching feature for their Claude family of models.

Basically caching allows developers to fine-tune model responses with longer and more instructive prompts, at the same time reducing costs by up to 90%.

​This X post explains it well:

Think of prompt caching like this: You're at a coffee shop. The first time you visit, you need to tell the barista your whole order. But next time? Just say "the usual."

That's prompt caching.

It makes it cheaper and faster to call any large prompt components that you reference over and over again. And it means a bunch of AI-powered apps that were previously too expensive to run just became viable.

Another great development from Anthropic!

Snack Sized 5 πŸͺ

1️⃣ Check out this real time diffusion example using light and image projection

2️⃣ OpenAI shuts down election influence operation that was using ChatGPT

3️⃣ Full YouTube documentary on Bittensor for those interested in AI + Cryptocurrency

4️⃣ Cool animation showing the last 15 months' rankings in the chatbot arena

5️⃣ I finally started an Instagram account for Ai5, if you post about AI give a follow for follow!

Share Ai5 and get Rewards!

Copy the link below and share to get your first reward.

[RH_REFLINK GOES HERE]

You can check out your unique referral hub here:

​https://hub.sparklp.co/[RH_SUBID GOES HERE]/1​

Join the conversation on X.

Connect on LinkedIn.

Read the Blog on Rareconnections.

Feedback? Yes please! matt@myai5.com

​Unsubscribe Β· Preferences​

Ai5

Stay ahead with 5 practical AI insights delivered every Monday and Thursday. Join 9,000+ readers from top AI companies like Unite.ai and Midjourney.

Read more from Ai5

There's an emerging trend putting intelligent AI into closed systems and observing the behaviors. Given the rate AI systems can adapt, it's possible we might just learn a few things about our own civilization. And even our destiny. It also makes me wonder, are we as humans being observed on Earth in our own little closed system? Today's Ai5: 🧬 AI Clones Actually Work πŸ’₯ LLMs Gone Wild ⚑ Colossus is Live πŸ’š AI Powers Body Scan Startup 🧱 Minecraft Agent Civilization Prompt of the Day 🎨 Holding a...

Would you trust an AI Doctor? Something that remembers every piece of medical literature ever written. Compared to (maybe) someone who studied 20 years ago achieving average grades. No shade on doctors, but you can see the potential. Today's Ai5: πŸ”Š ElevenLabs SFX Library πŸͺ¦ RIP Call Centers 🎞️ Introducing Minimax 🦠 Google Can Hear Sickness πŸ€– Agent.ai Prompt of the Day 🎨 A 3D render of the word "Ai5" with each letter crafted as a unique, vibrant flame pattern. Distinctive flame letters are...

It's estimated that the 1993 game Doom was installed on more computers than Windows 95 at one point. To this day, Doom has a huge modding community. It propelled shareware style distributions, birthed video game speedrunning, and popularized network multiplayer. The full game was just 2.39 Megabytes. Today's Ai5: πŸ•ΉοΈ AI Powered DOOM πŸ€– OpenAI Are Releasing an Agent 🎭 Midjourney Personality Test πŸ“± Claude Artifacts on Mobile πŸ“œ Everything to Know About SB 1047 Prompt of the Day 🎨 in the style of...