9 ms·

GPT is master of "approximate". Automating "correct" is still an unsolved problem. Will it be solved tomorrow? 5 years from now? Never? Who knows? Bicycle

GPT is master of "approximate". Automating "correct" is still an unsolved problem. Will it be solved tomorrow? 5 years from now? Never? Who knows? Bicycles, and their near-optimal designs, are still a $68B industry.

carapace 3y ago
I really didn't expect that comment to go over so poorly. I am trying to be constructive here, not critical. > Automating "correct" is still an unsolved problem. That's our job. Cf. "Augmenting Human Intellect: A Conceptual Framework" SRI Summary Report AFOSR-3223 By Douglas C. Engelbart October 1962 https://dougengelbart.org/pubs/augment-3906.html https://dougengelbart.org/pubs/augment-3906.html And "An Introduction to Cybernetics" by W. Ross Ashby http://pespmc1.vub.ac.be/ASHBBOOK.html http://pespmc1.vub.ac.be/ASHBBOOK.html PDF: http://pespmc1.vub.ac.be/books/IntroCyb.pdf http://pespmc1.vub.ac.be/books/IntroCyb.pdf Specifically "Amplifying intelligence" which comes at the end so you kinda have to read the whole book to get it. It's worth reading anyway. > Bicycles, and their near-optimal designs, are still a $68B industry. What does that have to do with anything?
- hello_computer 3y ago
  FWIW, I don't know why they downed you either. Just thought that your epitaph was premature. Despite recent developments, the symbolic calculator via teletype is still getting me to the point of "correct" faster than ChatGPT. If I owned a search engine company, or wrote clickbait listicles for a living, I'd be worried, but I don't. Natural language is packed with ambiguities and unspoken/unwritten context, and is thus a terrible programming language. People will end up spilling paragraphs of prose into text models, to do poorly and unpredictably, what a couple lines of bash would have done perfectly. For "approximate", natural language is fine, and possibly even optimal, but for precision, for "correct", it is unsuitable. It is why every profession develops its own domain-specific vocabulary and re-definition of phrases: they need more precision. Don't get me wrong. Those paragraphs will most certainly be spilled, but this is not the demise of programming. This is the next degeneration of programming! The Netflixification of programming; the union of hacks and technology, generating infinite low-quality content for insatiable consumers who have lost all faculties of discrimination. Perhaps the "most needful" software has already been written. But how much of it is any good? Last I checked, everyone's bug tracker is still overflowing. Some are easy fixes. Others will be decade-long slogs. I wish we were as close to done as you think, because I think with these new models, we're actually going to start slipping further away. > What does that have to do with anything? If the market size serves an estimate of work the world wants performed, and something as "done" as a bicycle still pulls-in $68B/yr, while software still earns trillions, then by the money-metric we have quite a ways to go before we are anywhere near finished.
  - carapace 3y ago
    Cheers, well met. > thought that your epitaph was premature. It's possible. I gave up prognostication when I didn't predict Twitter. I recently checked Craigslist for jobs in my line and it was empty... > the symbolic calculator via teletype is still getting me to the point of "correct" faster than ChatGPT. I prefer it myself. But we're not the mass audience, eh? My entire career is built on the fact that most people won't even learn Python or JS (let alone bash.) When we connect these models to empirical feedback devices they'll be able to tell us correct things. Schmidhuber says his goal is to "build an automatic scientist and retire." > Natural language is packed with ambiguities and unspoken/unwritten context, and is thus a terrible programming language. People will end up spilling paragraphs of prose into text models, to do poorly and unpredictably, what a couple lines of bash would have done perfectly. But you have to learn bash first. Now we can offload all that overhead to (fast powerful) external processors. > For "approximate", natural language is fine, and possibly even optimal, but for precision, for "correct", it is unsuitable. Nyet. Read the links I gave you. Our task is selection from the choices the machine generates. That's how you build an intelligence amplifier. The machines are capable of arbitrary precision. > consumers who have lost all faculties of discrimination. That's the only faculty the machine can't replace. That's the purpose of humanity: to answer the question "What are people for?" I don't know what's going to happen, but I am pretty sure of the overall envelope of possibility. We might become the slobs in Wall-E, I don't know. - - - - > If the market size serves an estimate of work the world wants performed, and something as "done" as a bicycle still pulls-in $68B/yr, while software still earns trillions, then by the money-metric we have quite a ways to go before we are anywhere near finished. I'm still not sure I'm getting what you mean? What's "finished" mean here? People buy bikes, yes, but I don't see the connection to my thesis that talking computers will rapidly make obsolete things we programmers take for granted now: Programming Languages, Bug Trackers, Bugs, Frameworks, Libraries, Apps, etc. Just as the smartphone subsumes many previously-separate devices, these "Talking UI" systems will subsume most of what we think of as programming.
    - hello_computer 3y ago
      > Nyet. Read the links I gave you. Our task is selection from the choices the machine generates. That's how you build an intelligence amplifier. The machines are capable of arbitrary precision. I'm already familiar with the Engelbart paper, but this is what people have already been doing with Google; with copy-paste from Stack Overflow. More often than not, all of the generated choices are either partially or wholly incorrect, and in need of either a cross-breeding or a euthanization. In this respect, LLMs are no different from the answer generators which preceded them. Only the priesthood of RTFM knows what to breed, what to gene-edit, and what to put down.

carapace 3y ago

I really didn't expect that comment to go over so poorly. I am trying to be constructive here, not critical. > Automating "correct" is still an unsolved problem. That's our job. Cf. "Augmenting Human Intellect: A Conceptual Framework" SRI Summary Report AFOSR-3223 By Douglas C. Engelbart October 1962 https://dougengelbart.org/pubs/augment-3906.html https://dougengelbart.org/pubs/augment-3906.html And "An Introduction to Cybernetics" by W. Ross Ashby http://pespmc1.vub.ac.be/ASHBBOOK.html http://pespmc1.vub.ac.be/ASHBBOOK.html PDF: http://pespmc1.vub.ac.be/books/IntroCyb.pdf http://pespmc1.vub.ac.be/books/IntroCyb.pdf Specifically "Amplifying intelligence" which comes at the end so you kinda have to read the whole book to get it. It's worth reading anyway. > Bicycles, and their near-optimal designs, are still a $68B industry. What does that have to do with anything?

hello_computer 3y ago
FWIW, I don't know why they downed you either. Just thought that your epitaph was premature. Despite recent developments, the symbolic calculator via teletype is still getting me to the point of "correct" faster than ChatGPT. If I owned a search engine company, or wrote clickbait listicles for a living, I'd be worried, but I don't. Natural language is packed with ambiguities and unspoken/unwritten context, and is thus a terrible programming language. People will end up spilling paragraphs of prose into text models, to do poorly and unpredictably, what a couple lines of bash would have done perfectly. For "approximate", natural language is fine, and possibly even optimal, but for precision, for "correct", it is unsuitable. It is why every profession develops its own domain-specific vocabulary and re-definition of phrases: they need more precision. Don't get me wrong. Those paragraphs will most certainly be spilled, but this is not the demise of programming. This is the next degeneration of programming! The Netflixification of programming; the union of hacks and technology, generating infinite low-quality content for insatiable consumers who have lost all faculties of discrimination. Perhaps the "most needful" software has already been written. But how much of it is any good? Last I checked, everyone's bug tracker is still overflowing. Some are easy fixes. Others will be decade-long slogs. I wish we were as close to done as you think, because I think with these new models, we're actually going to start slipping further away. > What does that have to do with anything? If the market size serves an estimate of work the world wants performed, and something as "done" as a bicycle still pulls-in $68B/yr, while software still earns trillions, then by the money-metric we have quite a ways to go before we are anywhere near finished.
- carapace 3y ago
  Cheers, well met. > thought that your epitaph was premature. It's possible. I gave up prognostication when I didn't predict Twitter. I recently checked Craigslist for jobs in my line and it was empty... > the symbolic calculator via teletype is still getting me to the point of "correct" faster than ChatGPT. I prefer it myself. But we're not the mass audience, eh? My entire career is built on the fact that most people won't even learn Python or JS (let alone bash.) When we connect these models to empirical feedback devices they'll be able to tell us correct things. Schmidhuber says his goal is to "build an automatic scientist and retire." > Natural language is packed with ambiguities and unspoken/unwritten context, and is thus a terrible programming language. People will end up spilling paragraphs of prose into text models, to do poorly and unpredictably, what a couple lines of bash would have done perfectly. But you have to learn bash first. Now we can offload all that overhead to (fast powerful) external processors. > For "approximate", natural language is fine, and possibly even optimal, but for precision, for "correct", it is unsuitable. Nyet. Read the links I gave you. Our task is selection from the choices the machine generates. That's how you build an intelligence amplifier. The machines are capable of arbitrary precision. > consumers who have lost all faculties of discrimination. That's the only faculty the machine can't replace. That's the purpose of humanity: to answer the question "What are people for?" I don't know what's going to happen, but I am pretty sure of the overall envelope of possibility. We might become the slobs in Wall-E, I don't know. - - - - > If the market size serves an estimate of work the world wants performed, and something as "done" as a bicycle still pulls-in $68B/yr, while software still earns trillions, then by the money-metric we have quite a ways to go before we are anywhere near finished. I'm still not sure I'm getting what you mean? What's "finished" mean here? People buy bikes, yes, but I don't see the connection to my thesis that talking computers will rapidly make obsolete things we programmers take for granted now: Programming Languages, Bug Trackers, Bugs, Frameworks, Libraries, Apps, etc. Just as the smartphone subsumes many previously-separate devices, these "Talking UI" systems will subsume most of what we think of as programming.
  - hello_computer 3y ago
    > Nyet. Read the links I gave you. Our task is selection from the choices the machine generates. That's how you build an intelligence amplifier. The machines are capable of arbitrary precision. I'm already familiar with the Engelbart paper, but this is what people have already been doing with Google; with copy-paste from Stack Overflow. More often than not, all of the generated choices are either partially or wholly incorrect, and in need of either a cross-breeding or a euthanization. In this respect, LLMs are no different from the answer generators which preceded them. Only the priesthood of RTFM knows what to breed, what to gene-edit, and what to put down.