n.b. Almost every time I publicly describe discrete capabilities like these a thought I have is "I want to remember in the future what I previously thought about the future shape of the curve to know if I need to update radically more than I have updated."
@patio11
-
MS Excel’s Value and Limitations in Task Automation
By
–
That might very well have value! MS Excel has a stupid amount of value. But "MS Excel did my taxes for me" is obviously risibly false, right.
-
LLM Limitations: Human Scaffolding Requirements for Output Convergence
By
–
I think this distinction sometimes elides people who are quite steeped in day to day LLM usage, because you can imagine a system that you can cobble together which would converge on correct output after you have spent hundreds of hours on being human scaffolding for it.
-
Engineering Rigor vs. Production Costs in Functional Art
By
–
Right, this is the point: when we think maximum engineering rigor would push costs for a functional art piece to incredible levels we are forgetting what drives teh cost of maximum engineering rigor, which is not "produce a thing that mostly works."
-
LLMs Image Generation Enables Quick Concept Feasibility Communication
By
–
(I think LLMs doing image generation is a bit of an unlock here because it can more quickly communicate “OK so is something in this direction possible enough for you to draw sketches or am I asking for a functioning fusion reactor with inlaid Japanese lacquerwork.”)
-
Scaling Production: Technology Feasibility Beyond Imagination
By
–
That art piece doesn’t feel physically impossible. It feels like something which nobody has scaled production of. But the non-existence of it can be solved with more time than one would expect and probably less money than one would expect.
-
Personal search engine as external memory for professional knowledge
By
–
I have a pseudo-search engine for my own stuff which I use as an external memory aide, because a large portion of professionally relevant memory I want to keep is publicly available, as opposed to in a carefully maintained private notebook.
-
o3 Model Image Critique Test After Three Years
By
–
Tried this on a lark 3 years later: Textual discussion is just o3 given the original image in same thread and told “Critique it.”
-
Who Builds AI Infrastructure Above Frontier Models?
By
–
So: who is building all of that? Clearly every company is not going to do it themselves on top of a frontier model on the backend, right?
-
Social Element: Learning from Previous Tool Usage Patterns
By
–
And there was a minor but important social element, where if e.g. you understood that I had previously used the tool for a particular area, and you were approaching that area yourself, in lieu of asking me what to do you could just sub search my previous usages and then guess.