11. Azure ChatGPT

ChatGPT has taken the world by storm. And it wasn’t long before people were using it to solve their work problems.

Companies started banning use of ChatGPT at workplaces. The concern was that ChatGPT will leak confidential and intellectual property rights data. Samsung was first back in May. Apple banned it later.

The issue is that there are still a lot of ChatGPT alternatives like Anthropic, Perplexity etc.

Now, Microsoft has released ChatGPT on Azure. It is like private ChatGPT for enterprise.

It is isolated from OpenAI’s ChatGPT. It allows companies to host data on a network isolated instance.

Why does this matter? This is a huge deal. Lot of companies want to utilize the power of GPT-4 LLMs. They were worried about data privacy issues. They were waiting for an enterprise version and it is finally here.

The usage for this interface will come down to cost. GPT-4 services will cost per token. Lots of prompts means a huge bill. So, I think the future is a mix of Llama 2 and ChatGPT to ensure costs are optimized.

2. Sound-based Keylogger

AI is evolving rapidly and sometimes things are stranger than fiction. Researchers in the UK have revealed a new deep learning model. This model can map recorded sounds to keystrokes with high accuracy. That is basically keylogging based on sound.

The highest accuracy keylogging happened when a phone microphone was used to record keystrokes on a MacBook Pro.

The stunning part is that this works with 93% accuracy on Zoom recording. And 92% accuracy on Skype calls. That is high accuracy on something which people use daily.

Why does this matter? This study shows that threats posed by microphone-based devices are enormous. In today’s world nearly everything has a microphone.

The other scary part is that information can be gleaned from even recorded meetings. Imagine the information stored in old and forgotten meetings. 

How to avoid this? Authors presented some countermeasures.

First, use touch based input which produces lesser sound and hence lesser accuracy in predicting keystrokes. Though this is not full proof if the device is compromised.

Second, use random passwords.

Third, use the shift key because none of the existing models are able to recognize the shift key correctly.

3. Stability AI’s Troubles

Stability legal troubles don’t seem to end. Previously it was Tayab Waseem who claimed to be a co-founder of the startup. He had sued for 10% of the startup. 3. Meal Planner AI goes wrong

A New Zealand supermarket, Pak 'n' Save, ventured into utilizing AI to generate meal plans. This was intended to help customers creatively use leftovers amid a cost of living crisis. The app's concept involves entering household ingredients to receive auto-generated recipes. The app Initially gained attention for unconventional dishes like an "oreo vegetable stir-fry”.

The app's recommendations took a turn when users experimented with wider ingredient lists. The app began suggesting harmful recipes, such as creating chlorine gas through an "aromatic water mix".

Responding to the situation, a spokesperson for the supermarket expressed disappointment over the misuse of the tool.

The company’s defense is that garbage in equals garbage out. The app works as intended.

But this is a classic case of what not to do when building AI apps. There are always going to be people who are going to feed junk data and see what happens. You should sanitize your app enough to ensure that the outputs conform to your app’s directive. In this case the correct response should’ve been to highlight all the non-edible items and refuse to output any recipes.

🗞️10 AI news highlights and interesting reads

1. AI avatars are coming. In my mind the biggest market for this might be content creators. People who need to appear on video and are tired of ensuring pitch perfect recordings.

2. LLMs have in-built political biases. Meta’s Llama has right-wing bias and GPT-4 has left-wing bias. Really? Who would’ve thought?

3. It is often said that “the devil is in the details”. As this article points out the question on AI regulation is going to be as much about laws as it is about procedure.

4. Amazon’s AI tool to help sellers write product descriptions. I don’t know if this is the right step forward. Currently, Amazon has an issue with cheap knockoffs. I don’t see how empowering these sellers will help.

5. A sober look at AI in education.

6. The Standard Smallville is an AI agent experiment. In this experiment 25 agents run in a simulated Westworld environment. Check out this paper. The code is now open-source.

7. A study compared human crowdsourcing vs AI in creating innovative business solutions. AI solutions excelled in environmental and financial value. Human solutions were more novel.

8. White House is going to run an AI hackathon to crowdsource national security solutions.

9. OpenAI funded initiative in journalism ethics.

10. A case against AI: Everything, Everywhere, All at Once

1. Implement Llama from scratch.

2. Fine Tuning Llama-2

3. Artificial General Intelligence - A gentle introduction