The part about AI helping to build up tacit knowledge much more rapidly than before doesn't sound right to me though.
I guess it could definitely be used to capture/document tacit knowledge that already exists.
But for knowledge that doesn't exist yet, even if an AI were to suggest a million different experiments, the bottleneck here will be running all those experiments, waiting for the results and making changes accordingly. This loop would inevitable be slow and could even span decades (if that was how long it took to build the tacit knowledge in the first place).
I've been looking at it as a bus factor reduction. The local LLM allows me to effortlessly add documentation, comments, decision points and tests. The quality of these dont particularly matter, because I'm using a local LLM and my equipment is owned by my employer so they'll get the same model I used who can read all the stuff.
These are things I'd never do; I'm not a professional programmer; I get paid for a distinctly different capabilty. Being about to write software to complete my job means I've got more time to do the critical things, like evaluate data, etc.
So before LLMs, I've architect hundreds of small programs for analytics and system admin. Anyone coming after meet would degrade 20 years having to do things by hand or buy a bunch of subscripts.
Now I can set the LLM on a schedule to document how everything works. Whoever replaces me then just needs to have a desire to carry on and it's near seamless.
Obviously, you have to care about continuity, and not treat staff like they're interchangeable, so this does not align with the current capitalist "Stop hiring humans"; but if we enjoin a better socialism, we reward ourselves.
Of course, some people are assholes and 30% will assume everyone else will be an asshole, so we'll get the worst outcomes.
Very interesting.
The part about AI helping to build up tacit knowledge much more rapidly than before doesn't sound right to me though.
I guess it could definitely be used to capture/document tacit knowledge that already exists.
But for knowledge that doesn't exist yet, even if an AI were to suggest a million different experiments, the bottleneck here will be running all those experiments, waiting for the results and making changes accordingly. This loop would inevitable be slow and could even span decades (if that was how long it took to build the tacit knowledge in the first place).
I've been looking at it as a bus factor reduction. The local LLM allows me to effortlessly add documentation, comments, decision points and tests. The quality of these dont particularly matter, because I'm using a local LLM and my equipment is owned by my employer so they'll get the same model I used who can read all the stuff.
These are things I'd never do; I'm not a professional programmer; I get paid for a distinctly different capabilty. Being about to write software to complete my job means I've got more time to do the critical things, like evaluate data, etc.
So before LLMs, I've architect hundreds of small programs for analytics and system admin. Anyone coming after meet would degrade 20 years having to do things by hand or buy a bunch of subscripts.
Now I can set the LLM on a schedule to document how everything works. Whoever replaces me then just needs to have a desire to carry on and it's near seamless.
Obviously, you have to care about continuity, and not treat staff like they're interchangeable, so this does not align with the current capitalist "Stop hiring humans"; but if we enjoin a better socialism, we reward ourselves.
Of course, some people are assholes and 30% will assume everyone else will be an asshole, so we'll get the worst outcomes.
Just like nuclear energy.