As always, a great and easy to understand write-up. I am subscribed to a lot of publishers on Substack, but you are definitely one of my favourites! Please keep on writing these amazing posts.
Circling back to this because Jeff Dean basically confirmed everything you laid out here. He said pretty much the same thing on the Latent Space pod recently. Reasoning and tool use are the two components that work together, and neither is sufficient alone. The bit about tool use being limited by reasoning capabilities especially tracks. I covered the practical bits from that conversation here: https://reading.sh/jeff-dean-on-what-actually-makes-ai-agents-work-dced5bb50206?sk=d8b9e7faac0da6011382834459ca4808
Fantastic write up. Super helpful in understanding the current state and the historical context. People forget that this direction was clear after toolformer though I'm sure everyone is surprised how fast it's developed since.
Totally agree! I wrote a series of write ups on tool usage a super long time ago right after the Toolformer paper was published (I linked all of them in this post). At the time, I remember constantly thinking to myself: "This makes so much sense, I can't believe more people are not focusing heavily on tool use!"
These days, this vision has kind of been realized -- tool use (or more generally agents) are probably the hottest applied / engineering research topic in AI. It just took a little bit for engineers to realize how cool this idea is :)
This is a nice engineering strategy diagram, not a philosophical or architectural first-principles system. It's like calling a prebuilt LEGO set a "new theory of architecture."
Hello Cameron - thanks for this. I feel, based on our own experience of deploying advanced Agentic systems that you've missed the following key points:
* Persona and Memory as core components - you focused on reasoning/action but didn't emphasize these as fundamental architectural elements
* Agentic Teams as deliberative committees - you mentioned multi-agent workflows but not the collaborative decision-making model we've pioneered (understandable, since it's super cutting-edge - see here: www.ai-risk.co/insights
* Economic agency - No discussion of agents as assets, NFTs, or autonomous economic actors (we've just started to implement this)
* Decentralized architectures - Missing the blockchain/ownership aspects of agents
* Human-AI hybrid teams - Didn't cover humans as agents within the system (in person or as clones, something else we've done)
As always, a great and easy to understand write-up. I am subscribed to a lot of publishers on Substack, but you are definitely one of my favourites! Please keep on writing these amazing posts.
Thank you so much for the kind words!
This was really long awaited. Tysm ❤️
No prob! Hope it was helpful!
Circling back to this because Jeff Dean basically confirmed everything you laid out here. He said pretty much the same thing on the Latent Space pod recently. Reasoning and tool use are the two components that work together, and neither is sufficient alone. The bit about tool use being limited by reasoning capabilities especially tracks. I covered the practical bits from that conversation here: https://reading.sh/jeff-dean-on-what-actually-makes-ai-agents-work-dced5bb50206?sk=d8b9e7faac0da6011382834459ca4808
This article is insane, what a great piece.Thank you for the share!! ❤️🔥🙌🙏
Fantastic writeup.
Thanks!
Thank You for sharing your knowledge and demystifying many aspects to even beginners in AI. I was hooked up while reading this.
Of course! I'm glad you enjoyed the article, and I hope it helped you learn something new :)
Fantastic write up. Super helpful in understanding the current state and the historical context. People forget that this direction was clear after toolformer though I'm sure everyone is surprised how fast it's developed since.
Totally agree! I wrote a series of write ups on tool usage a super long time ago right after the Toolformer paper was published (I linked all of them in this post). At the time, I remember constantly thinking to myself: "This makes so much sense, I can't believe more people are not focusing heavily on tool use!"
These days, this vision has kind of been realized -- tool use (or more generally agents) are probably the hottest applied / engineering research topic in AI. It just took a little bit for engineers to realize how cool this idea is :)
This is an incredible deep dive. I really learned a lot, thank you!
I’m glad it was helpful! Thank you for reading
These write ups are so, so useful for understanding what's happening with AI. Thanks for sharing them publicly.
Of course! Thanks for reading!
good
good
https://substack.com/@readfuturecore/note/p-166422582?r=5vwhbx&utm_medium=ios&utm_source=notes-share-action
Great right up, thanks for explaining this from first principles. Appreciate it.
This is a nice engineering strategy diagram, not a philosophical or architectural first-principles system. It's like calling a prebuilt LEGO set a "new theory of architecture."
Hello Cameron - thanks for this. I feel, based on our own experience of deploying advanced Agentic systems that you've missed the following key points:
* Persona and Memory as core components - you focused on reasoning/action but didn't emphasize these as fundamental architectural elements
* Agentic Teams as deliberative committees - you mentioned multi-agent workflows but not the collaborative decision-making model we've pioneered (understandable, since it's super cutting-edge - see here: www.ai-risk.co/insights
* Economic agency - No discussion of agents as assets, NFTs, or autonomous economic actors (we've just started to implement this)
* Decentralized architectures - Missing the blockchain/ownership aspects of agents
* Human-AI hybrid teams - Didn't cover humans as agents within the system (in person or as clones, something else we've done)
Fascinating read! I’m Harrison, an ex fine dining line cook. My stack "The Secret Ingredient" adapts hit restaurant recipes for easy home cooking.
check us out:
https://thesecretingredient.substack.com