8 ms·

It's a shame that this has negative connotations. There is a case that this is essential. It seems like we are overcoming an AI winter, using the same technolo

It's a shame that this has negative connotations. There is a case that this is essential. It seems like we are overcoming an AI winter, using the same technologies that was deemed infeasible 20 years ago. When I did competitive programming 20 years ago. I would somewhat jokingly say that I would want to write and AI to solve all the problems. Our mentor laughed it of clearly not considering it even feasible – today ChatGPT would like outperform a highschool competitive programming competition. We probably needed to look away from that generations presumptions in order to leap frog on some of these technologies.

yrgulation 4y ago
“ today ChatGPT would like outperform a highschool competitive programming competition.” Google did it first. You input a search string, find results on first or second attempt and then copy and paste the code.
- ben_w 4y ago
  Not even a close comparison. At best, Googling StackOverflow gets me one function, but only if I know the right keywords and it is short. ChatGPT, I can (and have!) given it example data and said "make a python script to parse this", and it not only gives me code, not only does it work, but when I ask it to extend this to do more or differently, the changes also work coherently with the previous results. For example: https://github.com/BenWheatley/Studies-of-AI/blob/main/coding-assistant/daily-plan-parser-conversation.md https://github.com/BenWheatley/Studies-of-AI/blob/main/codin...
- qu4z-2 4y ago
  Your highschool must have had pretty terrible programmers...
  - yrgulation 4y ago
    [dead]
ibizaman 4y ago
A contrary anectdotal exemple. We use MongoDB where I work and the only justification is it was the new fad at the time. Now it bites us because we use it as a relational DB. PostgreSQL would’ve been much more suitable in our case but it wasn’t in vogue.
- Traubenfuchs 4y ago
  Is there any regret in technology selection more widespread and common than having picked MongoDB during its bizarre new hotness phase? That phase where SQL database administrators where crying, thinking they'd lose their jobs tomorrow and many people claimed SQL was now absolutely deprecated.
  - hot_gril 4y ago
    This was probably the worst. Fads in business logic code or even APIs don't have anywhere near the lasting impact of a fad in screwing up the least easily replaceable part of your system, the DB.
  - edgyquant 4y ago
    In my experience the sql guys were laughing because they knew mongo was a fad and not an actual replacement.
  - CSMastermind 4y ago
    In the last 20 or so years I can't think of one. I've migrated 4 different companies off of Mongo to Postgres and it's a huge project each time.
    - hot_gril 4y ago
      Probably nearly a full code rewrite, right? My job the past 4 years has essentially been to clean up from bad DB decisions.
  - CoolGuySteve 4y ago
    I think using language-provided async calls everywhere will be the next mistake. Instead of callback hell, you get callback hell along with weird async race conditions and a contagion of async functions.
rpdillon 4y ago
I don't think the notion is that we should be relying entirely on cached results from prior generations. Rather, we need to understand what was tried and what worked and what didn't, and why. With this knowledge, we can see when innovation has invalidated prior assumptions, and we can try again.
- madsbuch 4y ago
  Yep, as nobody argues we should completely forget previous generations. The question is where to set the needle. The author definitely thinks we do not dwell enough on the past and should spend more time studying it.
k__ 4y ago
I know it's a hype right now, but I didn't get good results with any AI, be it text or image generation. It all looked impressive the first few days, but the novelty wore off quite quickly.
- poszlem 4y ago
  You are looking at a Nokia 3210 generation of AI.The most primitive version of what is to come. So the fact that the novelty has worn off is not really a good indicator of how much this will transform our everyday lives. Wait for the iPhone generation of AI, always in your pocket, interacting with you multiple times a day, changing how you consume news, how you communicate with the outside world, your friends and family.
  - Existenceblinks 4y ago
    Here we go again "it's early day", "it's like how the internet was born"
    - poszlem 4y ago
      Here we go again. "640K ought to be enough for anybody". "The horse is here to stay but the automobile is only a novelty—a fad."
- ajmurmann 4y ago
  Have you tried midjourney for image generation? Especially for photography style images, I've gotten some really cool results that I go back to occasionally
dgreensp 4y ago
An essay about academia (versus software engineering in industry) would probably make slightly different points. Academics definitely know their history. Also, when an actual disruptive new technology is first invented, jumping on it and trying to do something with it is both trendy and arguably rational, but this is a relatively rare occurrence.
ad404b8a372f2b9 4y ago
The entirety of AI innovation is driven by hardware improvements. Almost all AI research is done at the edge of our compute capabilities, and most of the models which produce results today existed in some form 20 years ago. We didn't leap-frog our predecessors presumptions, we built exactly what many envisioned but were unable to put to fruition with the computers of their time.
- madsbuch 4y ago
  The point is that we retried something that was previously deemed infeasible. Had you been too much aware of the history, you might not have done that.
marcosdumay 4y ago
It was deemed infeasible because it was infeasible. It took 2 generations time for computers to improve enough that it became feasible. And also a huge unexpected discovery about 20 years ago, that you can pre-train a neural network and cost your costs by an order of magnitude.

yrgulation 4y ago

“ today ChatGPT would like outperform a highschool competitive programming competition.” Google did it first. You input a search string, find results on first or second attempt and then copy and paste the code.

ben_w 4y ago
Not even a close comparison. At best, Googling StackOverflow gets me one function, but only if I know the right keywords and it is short. ChatGPT, I can (and have!) given it example data and said "make a python script to parse this", and it not only gives me code, not only does it work, but when I ask it to extend this to do more or differently, the changes also work coherently with the previous results. For example: https://github.com/BenWheatley/Studies-of-AI/blob/main/coding-assistant/daily-plan-parser-conversation.md https://github.com/BenWheatley/Studies-of-AI/blob/main/codin...
qu4z-2 4y ago
Your highschool must have had pretty terrible programmers...
- yrgulation 4y ago
  [dead]

ibizaman 4y ago

A contrary anectdotal exemple. We use MongoDB where I work and the only justification is it was the new fad at the time. Now it bites us because we use it as a relational DB. PostgreSQL would’ve been much more suitable in our case but it wasn’t in vogue.

Traubenfuchs 4y ago
Is there any regret in technology selection more widespread and common than having picked MongoDB during its bizarre new hotness phase? That phase where SQL database administrators where crying, thinking they'd lose their jobs tomorrow and many people claimed SQL was now absolutely deprecated.
- hot_gril 4y ago
  This was probably the worst. Fads in business logic code or even APIs don't have anywhere near the lasting impact of a fad in screwing up the least easily replaceable part of your system, the DB.
- edgyquant 4y ago
  In my experience the sql guys were laughing because they knew mongo was a fad and not an actual replacement.
- CSMastermind 4y ago
  In the last 20 or so years I can't think of one. I've migrated 4 different companies off of Mongo to Postgres and it's a huge project each time.
  - hot_gril 4y ago
    Probably nearly a full code rewrite, right? My job the past 4 years has essentially been to clean up from bad DB decisions.
- CoolGuySteve 4y ago
  I think using language-provided async calls everywhere will be the next mistake. Instead of callback hell, you get callback hell along with weird async race conditions and a contagion of async functions.

rpdillon 4y ago

I don't think the notion is that we should be relying entirely on cached results from prior generations. Rather, we need to understand what was tried and what worked and what didn't, and why. With this knowledge, we can see when innovation has invalidated prior assumptions, and we can try again.

madsbuch 4y ago
Yep, as nobody argues we should completely forget previous generations. The question is where to set the needle. The author definitely thinks we do not dwell enough on the past and should spend more time studying it.

k__ 4y ago

I know it's a hype right now, but I didn't get good results with any AI, be it text or image generation. It all looked impressive the first few days, but the novelty wore off quite quickly.

poszlem 4y ago
You are looking at a Nokia 3210 generation of AI.The most primitive version of what is to come. So the fact that the novelty has worn off is not really a good indicator of how much this will transform our everyday lives. Wait for the iPhone generation of AI, always in your pocket, interacting with you multiple times a day, changing how you consume news, how you communicate with the outside world, your friends and family.
- Existenceblinks 4y ago
  Here we go again "it's early day", "it's like how the internet was born"
  - poszlem 4y ago
    Here we go again. "640K ought to be enough for anybody". "The horse is here to stay but the automobile is only a novelty—a fad."
ajmurmann 4y ago
Have you tried midjourney for image generation? Especially for photography style images, I've gotten some really cool results that I go back to occasionally

dgreensp 4y ago

An essay about academia (versus software engineering in industry) would probably make slightly different points. Academics definitely know their history. Also, when an actual disruptive new technology is first invented, jumping on it and trying to do something with it is both trendy and arguably rational, but this is a relatively rare occurrence.

ad404b8a372f2b9 4y ago

The entirety of AI innovation is driven by hardware improvements. Almost all AI research is done at the edge of our compute capabilities, and most of the models which produce results today existed in some form 20 years ago. We didn't leap-frog our predecessors presumptions, we built exactly what many envisioned but were unable to put to fruition with the computers of their time.

madsbuch 4y ago
The point is that we retried something that was previously deemed infeasible. Had you been too much aware of the history, you might not have done that.

marcosdumay 4y ago

It was deemed infeasible because it was infeasible. It took 2 generations time for computers to improve enough that it became feasible. And also a huge unexpected discovery about 20 years ago, that you can pre-train a neural network and cost your costs by an order of magnitude.