5 ms·

Is anyone else also getting more worried about hard takeoff AGI scenarios?

3 points by upwardbound 4y ago

ggm 4y ago
No. It's massively overhyped. This isn't machine intelligence in the AGI sense it's better complex statistical/genetic algorithms but they still demand training. As I see it, AGI demands self training, or semi undirected training with guided inference. As long as it's 60 billion inputs and mechanical Turk correction feedback it's not AGI. So I see no sign of either aggressive or incremental change in what's being done: to the contrary I see a lot of repetitive re application of NLP and image work but with no intelligence in the AGI sense. Read specialist groups in reddit who analyse what GPT says about their subject matter: riddled with false inferences. Good use of the models is experts checking things and exploring. High recognition rates in image processing and diagnostic inferences are really very cool. I don't think radiographers or specialists or GP doctors expect to be replaced by ChatGPT any time soon. Maybe it helps reduce triage burdens but maybe it also feeds medical insurance nightmares.
- upwardbound 4y ago
  Agreed for sure. That said though, for which time horizon are your predictions going to remain valid? It's an interesting perspective I think to look at AGI risk in a 100 year time horizon because imho I think that's a conservative lower bound estimate for the lifespan of babies born today, if their families can afford access for them to education and good medical care. A lot can change in 100 years. Here's a sampling of current events from 100 years ago, in 1923: https://en.wikipedia.org/wiki/1923#Events https://en.wikipedia.org/wiki/1923#Events March 9 – Vladimir Lenin suffers his third stroke, which renders him bedridden and unable to speak; consequently he retires from his position as Chairman of the Soviet government. March 28 – Regia Aeronautica, the air force of Fascist Italy, is founded. May 27 – The Ku Klux Klan in the United States defies a law requiring publication of its membership. September 4 – The United States Navy's first home-built rigid airship USS Shenandoah (ZR-1) makes her first flight at Naval Air Station Lakehurst (New Jersey); she contains most of the world's extracted reserves of helium at this time. October 29 – Turkey becomes a republic, following the dissolution of the Ottoman Empire; Kemal Atatürk is elected as first president.
  - ggm 4y ago
    Apply the same question to fusion. Or perpetual motion machines. One is making inchingly slow progress across many decades, the other is definitionally impossible. Now, ask yourself where AGI is on this pole: is is merely slow to emerge or is it impossible? What fundamental theory supports it's emergence? What theory with a note to 1950s roots?
    - upwardbound 4y ago
      do you refer to Alan Turing's 1950's publication of a response to 9 common objections to the possibility of AGI? This one https://en.wikipedia.org/wiki/Computing_Machinery_and_Intelligence#Nine_common_objections https://en.wikipedia.org/wiki/Computing_Machinery_and_Intell...
      - ggm 4y ago
        Not that specific. More that McCarthy coined AI as a concept in 55. The perceptron was mid 1940s. So "since 1950" with error bars kind of covers it.

3 points by upwardbound 4y ago

ggm 4y ago

No. It's massively overhyped. This isn't machine intelligence in the AGI sense it's better complex statistical/genetic algorithms but they still demand training. As I see it, AGI demands self training, or semi undirected training with guided inference. As long as it's 60 billion inputs and mechanical Turk correction feedback it's not AGI. So I see no sign of either aggressive or incremental change in what's being done: to the contrary I see a lot of repetitive re application of NLP and image work but with no intelligence in the AGI sense. Read specialist groups in reddit who analyse what GPT says about their subject matter: riddled with false inferences. Good use of the models is experts checking things and exploring. High recognition rates in image processing and diagnostic inferences are really very cool. I don't think radiographers or specialists or GP doctors expect to be replaced by ChatGPT any time soon. Maybe it helps reduce triage burdens but maybe it also feeds medical insurance nightmares.

upwardbound 4y ago
Agreed for sure. That said though, for which time horizon are your predictions going to remain valid? It's an interesting perspective I think to look at AGI risk in a 100 year time horizon because imho I think that's a conservative lower bound estimate for the lifespan of babies born today, if their families can afford access for them to education and good medical care. A lot can change in 100 years. Here's a sampling of current events from 100 years ago, in 1923: https://en.wikipedia.org/wiki/1923#Events https://en.wikipedia.org/wiki/1923#Events March 9 – Vladimir Lenin suffers his third stroke, which renders him bedridden and unable to speak; consequently he retires from his position as Chairman of the Soviet government. March 28 – Regia Aeronautica, the air force of Fascist Italy, is founded. May 27 – The Ku Klux Klan in the United States defies a law requiring publication of its membership. September 4 – The United States Navy's first home-built rigid airship USS Shenandoah (ZR-1) makes her first flight at Naval Air Station Lakehurst (New Jersey); she contains most of the world's extracted reserves of helium at this time. October 29 – Turkey becomes a republic, following the dissolution of the Ottoman Empire; Kemal Atatürk is elected as first president.
- ggm 4y ago
  Apply the same question to fusion. Or perpetual motion machines. One is making inchingly slow progress across many decades, the other is definitionally impossible. Now, ask yourself where AGI is on this pole: is is merely slow to emerge or is it impossible? What fundamental theory supports it's emergence? What theory with a note to 1950s roots?
  - upwardbound 4y ago
    do you refer to Alan Turing's 1950's publication of a response to 9 common objections to the possibility of AGI? This one https://en.wikipedia.org/wiki/Computing_Machinery_and_Intelligence#Nine_common_objections https://en.wikipedia.org/wiki/Computing_Machinery_and_Intell...
    - ggm 4y ago
      Not that specific. More that McCarthy coined AI as a concept in 55. The perceptron was mid 1940s. So "since 1950" with error bars kind of covers it.