Great read. I have been using Claude to plan a short runway business launch. Definitely some EOS involved. I simply couldn’t optimize my human body and mind sufficiently to “run the tasks” that effectively.
Of course I felt great working on the plan, laying it all out. I never could keep pace—too much extrinsic variability (ie life) despite my own shortcomings in the same (ie flexibility of pace/recognition of need for rest). I changed the plan to reflect stages of development rather than a timeline.
I wonder how much of the initial plan rigidity reflected temporary user state. The existing guardrails are not nuanced—LLMs’ inability to read time is a much more serious issue than currently understood.
Thanks for sharing! Nice bundle of ideas in that piece and nice to hear from a ex-Ramp guy, I have a bunch of friends who worked there in the early days
One piece of bad LLM ergonomics is the emotional rollercoaster of having good session grind to a halt because of some mis-specified intent. Some type of “undo” where you could rewind a few turns and resume from a known good state would be nice.
Love the Cognitive Ergonomics framing! A pattern I've seen a few times with clients is what I've been jokingly calling the Workslop Kessler Syndrome. Say someone in a group is tasked with something they're not entirely comfortable with and uses an LLM to generate some output, but due to their limits (confusion, time, etc.) they cannot quality control the output before passing it on. Others are then confused by the output and use their own LLMs in turn to try and understand it. This is not always a problem, and sometimes the LLMs clear things up, or an outside observer can step in and explain or simplify. But occasionally the cascading confusion, accelerated by Default Skimming + Brandolini's Law, results in a kind of collective DOS attack that reminds me of Kessler Syndrome. I think one possible protocol is to periodically get people to switch to verbal discussion (similar to with AI psychosis), where it might be quicker to untangle. Though in at least one case I've seen it get to the point where a whole avenue of exploration was abandoned in part because the outputs had got so far ahead of the group's thinking.
Great read. I have been using Claude to plan a short runway business launch. Definitely some EOS involved. I simply couldn’t optimize my human body and mind sufficiently to “run the tasks” that effectively.
Thanks! And yeah exactly. LLMs are demanding of their users in hard-to-quantify ways. How’d you feel during / after ?
Of course I felt great working on the plan, laying it all out. I never could keep pace—too much extrinsic variability (ie life) despite my own shortcomings in the same (ie flexibility of pace/recognition of need for rest). I changed the plan to reflect stages of development rather than a timeline.
I wonder how much of the initial plan rigidity reflected temporary user state. The existing guardrails are not nuanced—LLMs’ inability to read time is a much more serious issue than currently understood.
lol love the protocol format + it does sort of feel like we're not putting enough words to previously-unseen phenomena (of which there are many)
also recently wrote some of my anecdata on the topic of "standards" and "variability" in this context here! https://text-incubation.com/ai-agents-make-small-companies-bigger
Thanks for sharing! Nice bundle of ideas in that piece and nice to hear from a ex-Ramp guy, I have a bunch of friends who worked there in the early days
One piece of bad LLM ergonomics is the emotional rollercoaster of having good session grind to a halt because of some mis-specified intent. Some type of “undo” where you could rewind a few turns and resume from a known good state would be nice.
Yea that is a frustrating part of these model. Also makes me think that, ironically, LLM memory is too weak and too strong at the same time
Love the Cognitive Ergonomics framing! A pattern I've seen a few times with clients is what I've been jokingly calling the Workslop Kessler Syndrome. Say someone in a group is tasked with something they're not entirely comfortable with and uses an LLM to generate some output, but due to their limits (confusion, time, etc.) they cannot quality control the output before passing it on. Others are then confused by the output and use their own LLMs in turn to try and understand it. This is not always a problem, and sometimes the LLMs clear things up, or an outside observer can step in and explain or simplify. But occasionally the cascading confusion, accelerated by Default Skimming + Brandolini's Law, results in a kind of collective DOS attack that reminds me of Kessler Syndrome. I think one possible protocol is to periodically get people to switch to verbal discussion (similar to with AI psychosis), where it might be quicker to untangle. Though in at least one case I've seen it get to the point where a whole avenue of exploration was abandoned in part because the outputs had got so far ahead of the group's thinking.
Knight takes h5 pawn...