New o1 model can solve complex tasks due to a new processing step before answering.
See full article...
See full article...
OpenAI’s new “reasoning” AI models are here: o1-preview and o1-mini
- Thread starter JournalBot
- Start date
I recently watched the Star Trek: TNG, S01E15 episode again where the Binars create a very lifelike woman in the holodeck for Riker.
Riker: "I mean, how real are you?"
Minuet: "As real as you need me to be."
Previously, those and other lines in this scene never bothered me. This time, they gave me chills. Using an AI chatbot is partially a self-reflection and therefore can feel pretty weird.
No solid conclusions or judgment here, just something that really struck me.
Riker: "I mean, how real are you?"
Minuet: "As real as you need me to be."
Previously, those and other lines in this scene never bothered me. This time, they gave me chills. Using an AI chatbot is partially a self-reflection and therefore can feel pretty weird.
No solid conclusions or judgment here, just something that really struck me.
Upvote
5
(6
/
-1)
That sort of comparison is something humans do all the time. Sometimes, it's apples vs oranges, but sometimes it's "if it gets this simple thing wrong, how am I supposed to trust anything else it says?"
Which isn't unreasonable.
Yes, I wouldn't expect a spell checker to do math, but the generalized nature of an LLM makes it very difficult for people to discriminate.
Last edited:
Upvote
16
(16
/
0)
The comparison with an image generator is insightful. Thanks for that.
There is indeed something different about language. I think that's why using an AI chatbot feels weird to me; it's like a talented BS artist with good diction and grammar, but doesn't actually know anything. Every time I encounter that in a chat session, it's rather disorienting.
Upvote
11
(12
/
-1)
A machine told me that bears like false equivalencies.
Upvote
9
(10
/
-1)
Upvote
0
(0
/
0)
Are they using it because it's quick and convenient or because the quality of the end result is better? We don't have that data.
Easier lines up with convenience more often than quality.
Now, doing things for convenience is perfectly fine, but then one needs to be clear about which dimension is being valued.
Upvote
11
(13
/
-2)
Perhaps you could expand on what you mean by production level coding.
For example, when you code for "non production", do you have a different standard for accuracy?
Upvote
0
(0
/
0)
If you haven't read the referenced "Thinking, Fast and Slow" by Daniel Kahneman, I highly recommend it. Your comments/questions are well addressed there.
Really.
Upvote
4
(4
/
0)
I don't do production level coding so I don't really even speak in those terms. I thought I could simply exclude that and people would know that I'm only referring to hobby/prototyping/testing/etc.
Lots of people need to leverage coding to play around with some idea where the code doesn't have to match the same standards as production level coding.
As a non-SW developer I feel like different levels of accuracy/reliability are needed depending on what the code is for. Production level code still covers UI code, gaming code, and low level firmware level code which all has different levels of accuracy and reliability needed.
The coding I do is infrequent coding when I want to test out an idea or to trouble some hardware. I don't need any have the same kinds of bullet proofing or security concerns that production level code needs.
I'm certainly going to do sanity checks on the code with test cases, but the code doesn't have to be perfect. It's probably better if it has some glitch to prevent me from being overcomplacent. Haha.
So the code AI produces works generally well for my use case especially when there is a large representation of what I'm doing in the training data.
But, I do feel like there is a lot more they could do that I felt was low hanging fruit. Like it should run the python code itself. Most of the time I simply feed it back the error message, and it fixes itself.
So in terms of actually making money I wish they'd focus on improving the models for things they are good at, and to stop trying to make them general.
Now that being said I have been impressed with o1-preview in answering the dumb questions I sometimes ponder.
Hmm. There's a bit of conceptual barrier here (I am a SW dev), but let's see how we do.
Outside of medical applications which require extra rigor, UI code, gaming code, and low level firmware code are all basically the same level of accuracy and reliability. It all has to work.
The analogy that suddenly occurs to me is weekend "Learn to be a coder!" sessions. These are useful as an intro, but don't make anyone a "coder". Similarly, if hobby/prototyping/testing/etc equates to "I just need to slap a few lines of code together to process this CSV file (or whatever), then AI code generation works reasonably well.
Anything much beyond that ventures very close to "production" quality code that has to handle edge cases, corner cases and all the other fun stuff. To put it a bit differently, there's a whole lot of non-production code that still requires production features in order to be useful. In that context, close doesn't count, which I think is what the other poster was driving at.
Last edited:
Upvote
4
(4
/
0)
I assure you that firmware does not get the highest level of rigor and testing. I can see why you would think that, but it's not true. Also, game developers have large testing/QA departments, but yet lots of bugs remain.
So why is that? The discrepancy in your experience is likely due to size of code base and functionality, your level of interaction with the software and how close that is to your professional expertise, experience and intuition. The closer it is, the more likely you are to notice bugs.
Having been doing software dev for 20 years, there is buggy code everywhere. It's more obvious in the case of games and your Rivian, but it's there even in more critical stuff. Intel's recent blunder with code requesting the wrong voltage levels is far from unique: reading errata on firmware and other foundational stuff can be pretty appalling. That's why I said it's basically the same level of accuracy and reliability. However, it mostly works in each domain and that seems to be good enough for the majority of organizations.
What's interesting and frustrating is that the problem isn't in the code, it's in the developers and organizations. It took me years to see this. AI isn't going to fix that.
Upvote
2
(2
/
0)
I'm glad it worked for you.
Refactoring is often undertaken for code modernization (i.e. keeping current), but is also a great opportunity to evaluate current requirements. I've removed lots of code during refactoring because it turned out it mapped to a dead business requirement.
Upvote
1
(1
/
0)
- Status