Stay calm & keep learning

Hello dear Convofriends! After the Easter break, I'm back with renewed energy and a fresh newsletter. This week, I've got three topics on offer:

🕺 Panic at the disco: rather than freaking out about imminent AGI, most of my news sources seems to have other worries on their mind this week: things are happening so fast that they can 't keep up.

🗣 Vox humana - human voices: on the other side of the panic scale, there's my favorite ML podcasters bringing a sense of reason to the scene. Still a lot of unadressed issues, but at least on a scale that we can all follow.

👩🎓 Sysiphus & co: I'm sticking to my own learning plan. And to be honest: that feels kind of wonderful - taking baby steps in a world where something new seems to happen every hour. And making slow progress!

🙏 Final note: a few words on the passing of Huub Oosterhuis, a Dutch theologist and musician, who wrote some of the most beautiful Dutch devotional music.

Panic at the disco

AI Explained is by far my favorite Youtube channel to learn about generative AI, and stay up to date about what's happening in the LLM space. Today's video sums up the current state of affairs pretty nicely: each time Philip, its creator, wanted to make a video, yet another Big Thing happened. 7 big developments in 48 hours, are we still able to keep up and make sense of it all?

Here's a high level summary of the most important events.

Towards autonomous task completion: AutoGPT and MemoryGPT

AutoGPT, powered by GPT 4, can prompt itself and autonomously complete a task by chaining and evaluationg prompts. It can now also generate and debug code, which basically means that it can develop apps while you instruct it, without you having to give it a separate prompt each time. Have a look at the video for some examples that integrate this with Whisper (around 2:30 timestamp); it's pretty mindblowing.

Release of MemoryGPT by Blackbearlabs.ai, a long term memory add-on for ChatGPT, which basically gives it the ability to remember previous conversations, even after you start a new session.

Anthropic going XXXLLM, and ChaosGPT

Anthropic announces plans for a model that's ten times more powerful than GPT 4. And boy, does it sound...worrying? Take this quote: "“These models could begin to automate large portions of the economy,” [...] “We believe that companies that train the best 2025/26 models will be too far ahead for anyone to catch up in subsequent cycles.” I don't know about you, but that doesn't sound like a recommendation in my ears.

More worrying stuff: people have trained Chaos GPT, a model that makes Bing's Venom look like an absolute darling with puppy eyes. This is an AI that describes itself as destructive and manipulative and has goals like 'Destroy humanity, Establish global dominance, Cause chaos and destruction and Control humanity through manipulation'.

Friendlier stuff: Poe is back! Create your own LLM bot!

After being offline for a while, Quora's bot collection page now let's you create your own LLM bot! Simply design a prompt, select your preferred LLM (ChatGPT or Claude), et voila! Here's a go at Pebbles, a conversational Catbot :-)

Poe is a fun way of practising prompting with direct results and a very user friendly interface. Do give it a try by all means!

Vox humana - human voices

Apart from Poe, this whole conundrum left me getting slightly depressed with the sheer ungrippability of recent developments, and a looming sense of me, a human, seeming to turn irrelevant in the scope of a few years, months even, being replaced by an AI (oh no, correction, a human that uses AI). Fortunately, enter Gary Marcus who has been voicing concerns about the current developments of 'A(G)I' before it became fashionable, and who brings a refreshingly sobering take on things.

This week's episode of Machine Learning Street Talk is great for anyone in need of a bit general reassurance in times of existential dread: debunking AI myths, explaining why LLM doesn't equal AGI, and basically putting the AI hype into a healthy perspective.

Sysiphus & co: learning new stuff!

One of the reasons why I'm slightly less active on social, is that I'm learning new stuff- a lot. I really feel I want to understand some of the basics of prompting LLMs and using them in an end-to-end use case. In code!

The use case I'm working on, is a popular one: accessing a set of private documents that are stored in a central place (like Google docs), and then accessing and querying them through an LLM.

In my case that means that I'm learning to:

• set up a python project and use VS Code and Git for code management
• set up a Google docs project and access the docs through an API
• indexing the docs with LlamaGPT, which is basically a data connector for your private docs, APIs or other sources. Indexing these sources means that they become queryable by LLMs and that you can use that info in your LLM prompts as an extra context.
• using the OpenAI API to actually query these indexes (and chat with the docs, if you will).

It's an extremely slow process, I'm definitely not the most talented developer, and code is definitely not my preferred language. But I did manage to get this end-to-end use case up and running! With a little help of ChatGPT, who patiently answers even the most n00bie newbie questions:

Regardless, I love learning the basics: it forces me to really sit down and focus, formulate my thoughts and think. Plus, it keeps me from being crippled by FOMO of having to learn everything about everything. It's a real pleasure to hear my brain creak every once in a while, and have this satisfying sense of suddenly realising 'hey, I understand!'

Final note

Huub Oosterhuis, Dutch theologist and musician, passed away at the age of 89 on Easter Sunday. As I child, I grew up with many of his Dutch choral works and was always struck by his beautiful lyrics.

One song in particular touches me every time I hear it. 'Ken je mij' has its lyrical roots in Psalm 139. In a week where I've had conversations with four brave young people who are becoming who they've always been, and who face considerable challenges in doing so, it's a song that reminds me of our primary need: to be known and seen for the humans we are. Preferrably not by AI.

Do you know me? Then who do you know?

Do you know me better than I

Do you know me? Then who am I?

Do you know me better than I?

Eyes that look through the sun

Searching for the place where I live

Are you imagery for someone

Who is nice, or immeasurably distant,

who doesn't stand and doesn't fall

And doesn't feel as I do,

not cold and haughty.

About Maaike Groenewege

Maaike is a conversation design lead, linguist and DesignOps coach with Convocat conversational expertise. She helps both starting and more experienced conversational teams in optimising their conversation design practise, NLU analyses and team communication. Her main focus right now is on how LLMs can benefit enterprise conversational AI.

• Join Convoclub, my boutique community for conversation designers, linguists and language lovers. We meet every other Monday at 18:00 CET, find the invites in the Convoclub Events section, or keep an eye out on LinkedIn!
• Find your next book on conversational AI in my Convocat library