d3a2d431-1d2e-4be6-b599-e8fa289f2a9a

[giscus[bot]]

d3a2d431-1d2e-4be6-b599-e8fa289f2a9a

Learn when to use fast System 1 decision models like Jev to classify, route and score, and when to use System 2 LLMs like Claude and ChatGPT to write and reason.

https://www.ssw.com.au/rules/system-one-vs-system-two-ai

[jeoffreyfischer]

Great topic - I already knew about System 1 and 2 thanks to this great video from Veritasium: https://www.youtube.com/watch?v=UBVV8pch1dM

[AntPolkanov]

Interesting!
It's funny how LLMs resemble compilers. Soon there will be debates like "Strongly-typed LLMs are much better than JavaScript-like dynamic LLMs"😄 But we all know that strong types always win.

[AntPolkanov]

That's a huge win for classification tasks!
I think we can expect a Visual Jev soon 👁️

[TobyChurches]

Really cool topic. These System 1 models really seem like they would excel in their very specific classification / closed answer use cases. But at the moment I still see System 2 models being being much more popular, due to their flexibility and ease of use. I think I combination of the 2 is where the AI will really shine.

[18-th]

If it works well, this should be great for "vibe checks" that are hard to express as code. Looking forward to playing with this more.

[vladislav-kir]

Sounds like a nice tool to test stuff with AI. Keen to try it.

[tomek-i]

time to think about rewriting some old projects :-D

[RobJT-wq]

Using it for triaging will be the best use case for the SysAdmin team

[ricksu978]

Jev can currently handle about 32K to 64K tokens of context, which is limited compared to System 2 reasoning models.
When writing questions, keep them concise and avoid including large amounts of background text.

[isaaclombardssw]

Not significantly affecting workflows, but to police LLM usage (routing to different model classes) it looks useful.

For most decisions that matter, I appreciate the ability to research, source, and further confirm that LLMs provide, rather than a mystery number. In applications (such as YakShaver) I can see a use case for cheap classification.

The pitch and website for Jev are compelling, the marketing is great. Curious to see where it goes. 👀

[suiyangqiu]

Very interesting and keen to try it, I think the future of these are not using either System 1 or System 2 AIs, instead, it would involve integrating the use of the both of them to work alongside one another.

[ianhoffmanx]

I think this could be really cool for dynamic branching and stuff for leadgen forms.

[chrisschultzssw]

Sounds great for triaging tickets in systems like Zendesk.

[kulesy]

Very important to recognize the distinction between System 1 and 2. Performance benefits can be huge!

[bahjatalaadelssw]

Would love to see how System 2 can tackle story narratives and keeping context through dialogue.

[Berber-Mama]

I’d like to try this for sorting customer enquiries, with a person checking the ones the AI is unsure about.