AI News: ChatGPT Levels up AGAIN... twice! Code Interpreter, AI job losses, Bing Chat Plugins & AGI

This has been another massive week in AI, with what is being called “the most significant update from OpenAI yet” (quite some statement, given the last 6 months), a new chatbot from a company with strong AI pedigree, and real economic impact being felt.

Let’s go!

BTW you can watch the video of this week’s update over on our YouTube channel here:

OpenAI’s new Code Interpreter shocks data analysts

OpenAI released its homegrown ChatGPT plugin, Code Interpreter. This gives ChatGPT some new abilities:

• Upload files

• Download files

• Run python script

Whilst that might sound pretty unexceptional, it opens up some stunning new capabilities. The most impressive demos so far have been uploading large datasets and having ChatGPT analyse these for trends, create graphs and charts to visualise the data, and even develop and test hypotheses before turning all of this into an academic paper.

Whilst we marketers have been used to being blown away by ChatGPT’s potential, it’s now the data analysts who are realising that the future of their industry looks very different.

Ethan Mollick is a professor at Wharton, and has written a great blog on his experiments with it so far.

Kahn Academy reveals some of the fine tuning it did to make its GPT-4 bot so good

Online education platform Kahn Academy was one of the headline GPT-4 partners when OpenAI launched it. This week founder Sal Kahn’s Ted Talk gave a tour of some of the functionality. The most interesting takeaways weren’t actually in the demo though 🕵️

1. They are using a second AI to moderate each conversation.

2. They taught GPT-4 to ‘think before it speaks’, and showed an example of one of its ‘thoughts’. They found this helps to improve the output.

This is another example of ‘multi-layered AI’ being used to improve functionality and adherence to rules.

We saw how AutoGPT and BabyAGI used a sort of ‘master’ layer to take big picture instruction, chunk this down into tasks assigned to ‘worker’ AIs, receive and process their feedback, and decide what to do next. This process gave it the ability to perform more complex tasks than GPT-4 alone.

Teaching GPT-4 to think before it speaks looks like this, where the ‘thought’ is moderated and communicated to the user by a second layer of AI.

In OpenAI’s Greg Brockman’s Ted talk, he demonstrated how ChatGPT with internet access could be used to ‘fact check itself’.

It’ll be interesting to see how this idea of multi-layered AI develops, particularly as tools like ChatGPT get access to tools, the web, and now the ability to write and execute code. Judging by the output we’re seeing so far, it seems to be a levelling up.

The ‘AI Meteorite’ starts hitting companies. Economic shifts result…

Loser #1: Chegg & Edtech

The first victim this week was homework helping website Chegg, who saw nearly half its value wiped as investors processed news from CEO and experienced tech veteran Dan Rosenweig, that ChatGPT was impacting student sign up numbers.

Loser #2: Walmart Suppliers

News that Walmart has been using AI to negotiate with some of its suppliers came out this week. Whilst this is not a new experiment (it has been running since 2021), it threw up some interesting stats:

• 68% of suppliers approached successfully reached deals

• Average cost savings of 3% negotiated!

• Despite this, 3 out of 4 suppliers actually preferred to negotiate with the bot.

Full unpaywalled version of the Bloomberg article is here.

Loser #3: 7,800 IBM back office staff

This is the first blue chip CEO I’ve seen openly declaring that they were planning mass changes to their staff due to AI.

IBM’s CEO Arvind Krishna said about their HR and back office staff: “I could easily see 30% of that getting replaced by AI and automation over a five-year period.”

That would mean roughly 7,800 jobs lost. Part of any reduction would include not replacing roles vacated by attrition, an IBM spokesperson in full damage-limitation said.

It’s worth saying that this approach is in stark contrast to the one announced by PwC last week (unpaywalled here) who, with their $1B Generative AI investment plans seemed to rule out AI-induced job cuts by claiming they “would leave no one behind”.

Interesting to see such radically different approaches happening, and whether IBM’s bold early allegiance to the ‘job cuts’ side will prove to be the Elon Musk/Twitter moment for large consultancy firms.

Midjourney 5.1 is out! People are happy

The latest release from arguably the leading name in text-to-image AI is out, and so far it looks to be a fairly big leap.

Generating images with text in is significantly better (though still nowhere near perfect). Here’s a good Twitter thread that compares 5 to 5.1

The quality improvement trajectory on text to image is quite stunning. Midjourney IMO is still waiting for its ‘ChatGPT moment’ - if someone can build an easy to use web interface for it and get rapid mass adoption, this will be profound for companies that need creative visual assets very soon.

LinkedIn founder & friends launch a new AI chatbot, PI

PI stands for “Personal Intelligence” and is designed to be a new type of chatbot. Gone are the useful tasks like essay writing, helping with lists, coding and all those other useful applications. Instead, PI wants to have mundane, trivial and banal” conversations with you (yup, that’s a quote from CEO of Inflection AI that built it).

The ultimate vision seems to be to build a helpful and supportive personal chatbot that encourages you to grow and learn.

In practice, PI feels slow and not particularly profound. My personal experience was hitting that ‘boredom wall’ where I just couldn’t go on with it. It tries to be quirky and ‘human’, but comes across as fake (because it is) and very limited if you’re used to using GPT-4.

If talking to ChatGPT is like talking to a supremely well-informed, helpful hyper intellectual, talking to PI is like being stuck in a lift with a stoner who seems to want to discuss mindfulness and energy flows.

Based on the people behind this project, I don’t think it’s the last we’ll see of Inflection AI. But this one feels like a no.

And Finally…

Deepmind CEO Demis Hassabis (who is now in charge of Google’s Brain division as well) thinks that the pace of AI innovation could accelerate. This is a widely-held belief in the AI industry, but in practice is difficult to comprehend.

In this WSJ article (unpaywalled here), he predicts that some form of AGI (AI that is generally better than humans at most economically important tasks) could be here in 3 years.

In somewhat related news, the ‘Godfather of AI’ Geoffrey Hinton has announced his retirement from Google and has instead taken to warning the world about the danger of the tech he has in part helped develop.

I’m usually pretty optimistic, but when the inventor of the monster downs tools and runs screaming from the lab…

If you enjoyed this newsletter…

Please consider forwarding it on to a contact who you think might enjoy it too!

You can also follow Powered By AI on our social channels: