🤖 GPT-4 Is Not Perfect

What's up? You're reading Inclined AI. It’s Wednesday, so treat yourself with a BEC sandwich and a black coffee. It’s the breakfast of champions.

Here are the headlines for said champions:

• Microsoft lambasts GPT-4

• Marc Andreessen rants about regulation

• AI chatbots can find your personal data super fast

• AI could cause a financial crash, apparently

Be a sponsor of our newsletter.

Microsoft showed us GPT-4’s warts, and you won’t be surprised.

It turns out that following instructions like a good AI model should do gets OpenAI’s model into more trouble than its predecessor, GPT-3.5. To phrase it another way: GPT-4 is prone to jailbreaking.

Here’s what the research team at Microsoft found:

Compared to other models, GPT-4 follows commands with the accuracy of a seasoned Marine and comprehends instructions better than other Transformer models available. And that should be a good thing.

However, when you expose a tool like GPT-4 to the entire internet and give it a firehose dosing of terrible things, issues happen. That’s why Bing Chat had some issues in the beginning.

But why is Microsoft telling us this?

Admitting errors in their golden AI model is a bad idea, right? Well, the footnotes help explain the context. They alerted OpenAI to these issues a while back and addressed the problem (we think).

Still, you should see this as a crack in GPT-4’s armor and one that can widen with time, too. This is why those same researchers are encouraging academics to study this same effect and give this study more peer review.

Let’s try to fix this for all models, so AI is safer in the long run.

Life-Changing Concepts → gives you an unfair advantage in business using mental models

Master ChatGPT → check out the ChatGPTricks e-book that takes you from zero to hero

Marc Andreessen is an “e/accel” and wants you to know it.

The tech icon published a blog entitled “The Techno-Optimist Manifesto” on the a16z website this week. It’s an obelisk for Andreessen that shouts to the investment firm and anyone who’ll listen that “tech is better off unregulated by ethics.”

That sounds weird, but wait, there’s more.

The whole piece is a Libertarian rallying cry, punching at a straw man until it’s all out of stuffing. You can read it for yourself because you’ll find interesting points there, too.

I bring it up to highlight a growing movement within the AI space.

Effective Accelerationists are clawing back and pushing their argument with full force now. They want unregulated AI unhampered by fear of the short-term.

AI researchers are splintering and splitting over this core belief. Do we trust AI to develop unhindered, or should we wait for humans to decide what’s best?

I don’t have an answer, but I expect the other side of this argument to keep chiming in. Another manifesto is coming sooner or later. And another one after that.

🪪 AI chatbots can get your personal information with only a few details

🎧 Riffusion goes viral and uses that momentum to raise $4m for their AI-generating music app

🧐 Deep thinkers debate whether AI can become conscious

🐾 Baidu fights to become more relevant by updating Ernie

🛡️ Yoshua Bengio, a Godfather of AI, wants an organization made to defend humanity

🇬🇧 Even Cambridge researchers doubt that the UK can lead the next wave of generative AI

©️ Microsoft and Adobe are trying to promote a new symbol for AI-generated images

📉 AI could cause a financial crash within the next decade

• Softr - create business apps (link)

• Expresso - workplace wellbeing solution (link)

• Streamlabs - fast & efficient way to edit podcasts (link)

• DeepMate - prepare, conduct, get feedback on interviews (link)

• SingleAPI - convert the internet into your own API (link)

• TimeToTok - helps TikTok users maximize growth (link)

• AgentLabs - chat-based UI as a service (link)

• Korl - product presentations tailored to each audience (link)

Channel Crossover

This is what happens when Animal Planet meets Food Network.

_{- That’s it for today. I hope you enjoyed the latest edition of inclined.ai - Davis.}