5 ms·

Is the AI bubble popping? [video]

13 points by grugagag 2y ago

aurareturn 2y ago
No, it isn't. It's just getting started in my opinion. When GPT5 gets released, it's going to take hype to a whole new level again.
- mouse_ 2y ago
  It's hard to say for sure.
- uejfiweun 2y ago
  OpenAI certainly has a vested interest in making people think so. To be honest, from my end, it seems like the biggest step was already taken with the OG ChatGPT using 3.5. I subscribed to 4 for a bit but it shares a lot of the same limitations as 3.5, it was not nearly as big as the leap from having no good chatbots at all to ChatGPT. I'd love to believe that OpenAI has made a major breakthrough with 5, but it strikes me that they're gonna have to strike out beyond "GPT" territory to make a similarly sized dent to the one they made with ChatGPT. That is to say they're gonna have to move beyond transformer architecture.
  - mewpmewp2 2y ago
    To me the main impressive step was absolutely 4. It has made me wonder about how intelligence works pretty much daily. 3.5 seems to be mainly good at summarizing, labelling, while 4 seems to constantly impress with it's ability to pay attention to details and do dynamic reasoning.
    - uejfiweun 2y ago
      Can you give me a little more specific detail on your use cases that fail with 3.5 and succeed with 4? I've tried to get some projects going using the API to generate some certain types of geometric data and both fail miserably for me. Keep hearing that 4 is so much better but I have yet to see it myself.
      - mewpmewp2 2y ago
        One example is a prompt like: Find inconsistencies with data that should be structured in a certain way (...data...), give me JSON response of actions to fix the data. GPT-3.5 gives pointless things, very rarely any of the important things, or simply makes up things, while GPT-4 seems to spot everything at least on the data I provide. What is the geometric data you generate?
        uejfiweun 2y ago
        Thank you for the examples. I definitely haven't used it much for that type of use case but I'll keep that in mind. I don't want to give up too much info about my project because it would probably dox me, but it essentially involves generating consistent geometric shapes in a 2d array. Like the following: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 1 1 1 0 0 0 1 1 1 0 Both current flavors of GPT are unable to generate spatially consistent shapes according to my specifications. Who knows, maybe they will improve in this realm in GPT-5 with the inclusion of their rumored Q* algorithm.
        mewpmewp2 2y ago
        I imagine that I would want to ask for this output differently, as a function that would output this rather than it to write out those tokens. For example: Give a JavaScript function that will output 0 and 1 tokens to render a rectangle on a 2d array. Example output: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 1 1 1 0 0 0 1 1 1 0 It responded: function generateRectangle(rows, cols, rectStartRow, rectStartCol, rectHeight, rectWidth) { // Create a 2D array filled with 0s let array = Array.from({ length: rows }, () => Array(cols).fill(0)); // Fill the rectangle area with 1s for (let i = rectStartRow; i < rectStartRow + rectHeight; i++) { for (let j = rectStartCol; j < rectStartCol + rectWidth; j++) { array[i][j] = 1; } } return array; } function printArray(array) { array.forEach(row => { console.log(row.join(' ')); }); } // Example usage: let rows = 5; let cols = 6; let rectStartRow = 2; let rectStartCol = 2; let rectHeight = 3; let rectWidth = 3; let rectangleArray = generateRectangle(rows, cols, rectStartRow, rectStartCol, rectHeight, rectWidth); printArray(rectangleArray); ------ And then I asked for a triangle instead, but still with square output example ----- function generateTriangle() { const size = 10; const field = Array.from({ length: size }, () => Array(size).fill(0)); for (let i = 0; i < size; i++) { for (let j = 0; j <= i; j++) { field[size - 1 - i][j] = 1; } } return field; } function printField(field) { for (let row of field) { console.log(row.join(' ')); } } const triangleField = generateTriangle(); printField(triangleField); 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 1 1 1 1 0 0 0 0 0 1 1 1 1 1 0 0 0 0 1 1 1 1 1 1 0 0 0 1 1 1 1 1 1 1 0 0 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ---- Then I would execute the function programmatically to generate these tokens.
- rsynnott 2y ago
  You can only get so much milage out of hype; at a certain point you really need to start seeing actual applications.
lamivkeehba 2y ago
Every interview I've watched with members from one of the major AI labs (Anthropic, Meta, OpenAI, Deepmind, etc) seem to believe that we are only getting started. We have yet to hit the limits of scaling. This [1] is a clip from one of the most insightful interviews on the subject. Researchers from Anthropic and Google talking about their work. Full interview here [2] - it's a long listen but if you want to dive into the technical details of some of this stuff, it's great. [1] - https://www.youtube.com/watch?v=UeI29-AdhQI https://www.youtube.com/watch?v=UeI29-AdhQI [2] - https://www.youtube.com/watch?v=UTuuTTnjxMQ https://www.youtube.com/watch?v=UTuuTTnjxMQ
- blibble 2y ago
  would you expect them to say otherwise? such as "we've got nothing, please take away our funding"?
- staticman2 2y ago
  >>> Every interview I've watched with members from one of the major AI labs (Anthropic, Meta, OpenAI, Deepmind, etc) seem to believe that we are only getting started. If your competitor was going around saying "Soon we will build the greatest AI ever and it will run the world economy!!!" would you do an interview where you say "Soon we will build nothing and make no money!!!!"
huhuhu111 2y ago
[dead]
throwaway5959 2y ago
No but the bubble of YouTube tech doomsayers is as large as ever.
dzonga 2y ago
I don't think AI is a bubble. AI, unlike other tech businesses before it, is similar to heavy industry. You need a lot of CAPEX, Capital Expenditure. No more two people in the garage, since you probably need Nvidia GPU's that cost $30k a pop. End of day, all the money being invested into AI ends up within a few companies - Microsoft, Amazon, Google, and Nvidia. Due to cloud costs, and marketing costs.

13 points by grugagag 2y ago

No, it isn't. It's just getting started in my opinion. When GPT5 gets released, it's going to take hype to a whole new level again.

mouse_ 2y ago
It's hard to say for sure.
uejfiweun 2y ago
OpenAI certainly has a vested interest in making people think so. To be honest, from my end, it seems like the biggest step was already taken with the OG ChatGPT using 3.5. I subscribed to 4 for a bit but it shares a lot of the same limitations as 3.5, it was not nearly as big as the leap from having no good chatbots at all to ChatGPT. I'd love to believe that OpenAI has made a major breakthrough with 5, but it strikes me that they're gonna have to strike out beyond "GPT" territory to make a similarly sized dent to the one they made with ChatGPT. That is to say they're gonna have to move beyond transformer architecture.
- mewpmewp2 2y ago
  To me the main impressive step was absolutely 4. It has made me wonder about how intelligence works pretty much daily. 3.5 seems to be mainly good at summarizing, labelling, while 4 seems to constantly impress with it's ability to pay attention to details and do dynamic reasoning.
  - uejfiweun 2y ago
    Can you give me a little more specific detail on your use cases that fail with 3.5 and succeed with 4? I've tried to get some projects going using the API to generate some certain types of geometric data and both fail miserably for me. Keep hearing that 4 is so much better but I have yet to see it myself.
    - mewpmewp2 2y ago
      One example is a prompt like: Find inconsistencies with data that should be structured in a certain way (...data...), give me JSON response of actions to fix the data. GPT-3.5 gives pointless things, very rarely any of the important things, or simply makes up things, while GPT-4 seems to spot everything at least on the data I provide. What is the geometric data you generate?
      - uejfiweun 2y ago
        Thank you for the examples. I definitely haven't used it much for that type of use case but I'll keep that in mind. I don't want to give up too much info about my project because it would probably dox me, but it essentially involves generating consistent geometric shapes in a 2d array. Like the following: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 1 1 1 0 0 0 1 1 1 0 Both current flavors of GPT are unable to generate spatially consistent shapes according to my specifications. Who knows, maybe they will improve in this realm in GPT-5 with the inclusion of their rumored Q* algorithm.
        mewpmewp2 2y ago
        I imagine that I would want to ask for this output differently, as a function that would output this rather than it to write out those tokens. For example: Give a JavaScript function that will output 0 and 1 tokens to render a rectangle on a 2d array. Example output: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 1 1 1 0 0 0 1 1 1 0 It responded: function generateRectangle(rows, cols, rectStartRow, rectStartCol, rectHeight, rectWidth) { // Create a 2D array filled with 0s let array = Array.from({ length: rows }, () => Array(cols).fill(0)); // Fill the rectangle area with 1s for (let i = rectStartRow; i < rectStartRow + rectHeight; i++) { for (let j = rectStartCol; j < rectStartCol + rectWidth; j++) { array[i][j] = 1; } } return array; } function printArray(array) { array.forEach(row => { console.log(row.join(' ')); }); } // Example usage: let rows = 5; let cols = 6; let rectStartRow = 2; let rectStartCol = 2; let rectHeight = 3; let rectWidth = 3; let rectangleArray = generateRectangle(rows, cols, rectStartRow, rectStartCol, rectHeight, rectWidth); printArray(rectangleArray); ------ And then I asked for a triangle instead, but still with square output example ----- function generateTriangle() { const size = 10; const field = Array.from({ length: size }, () => Array(size).fill(0)); for (let i = 0; i < size; i++) { for (let j = 0; j <= i; j++) { field[size - 1 - i][j] = 1; } } return field; } function printField(field) { for (let row of field) { console.log(row.join(' ')); } } const triangleField = generateTriangle(); printField(triangleField); 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 1 1 1 1 0 0 0 0 0 1 1 1 1 1 0 0 0 0 1 1 1 1 1 1 0 0 0 1 1 1 1 1 1 1 0 0 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ---- Then I would execute the function programmatically to generate these tokens.
rsynnott 2y ago
You can only get so much milage out of hype; at a certain point you really need to start seeing actual applications.

lamivkeehba 2y ago

Every interview I've watched with members from one of the major AI labs (Anthropic, Meta, OpenAI, Deepmind, etc) seem to believe that we are only getting started. We have yet to hit the limits of scaling. This [1] is a clip from one of the most insightful interviews on the subject. Researchers from Anthropic and Google talking about their work. Full interview here [2] - it's a long listen but if you want to dive into the technical details of some of this stuff, it's great. [1] - https://www.youtube.com/watch?v=UeI29-AdhQI https://www.youtube.com/watch?v=UeI29-AdhQI [2] - https://www.youtube.com/watch?v=UTuuTTnjxMQ https://www.youtube.com/watch?v=UTuuTTnjxMQ

blibble 2y ago
would you expect them to say otherwise? such as "we've got nothing, please take away our funding"?
staticman2 2y ago
>>> Every interview I've watched with members from one of the major AI labs (Anthropic, Meta, OpenAI, Deepmind, etc) seem to believe that we are only getting started. If your competitor was going around saying "Soon we will build the greatest AI ever and it will run the world economy!!!" would you do an interview where you say "Soon we will build nothing and make no money!!!!"

huhuhu111 2y ago

[dead]

throwaway5959 2y ago

No but the bubble of YouTube tech doomsayers is as large as ever.

dzonga 2y ago

I don't think AI is a bubble. AI, unlike other tech businesses before it, is similar to heavy industry. You need a lot of CAPEX, Capital Expenditure. No more two people in the garage, since you probably need Nvidia GPU's that cost $30k a pop. End of day, all the money being invested into AI ends up within a few companies - Microsoft, Amazon, Google, and Nvidia. Due to cloud costs, and marketing costs.