4 ms·

I think the current state of GPT-4 is overhyped when it comes to programming. It's novel and useful for steering you in the right direction and can reproduce si

I think the current state of GPT-4 is overhyped when it comes to programming. It's novel and useful for steering you in the right direction and can reproduce simple hello-world like examples (even Pong and other common programs) but in my experience fails when things become even slightly nontrivial. Things might change but it seems like GPT-4 has similar performance to GPT-3.5 so there might be some limit to the scaling ultimately. Regardless, learning a new skill, even an obsolete one, is almost invariably a good and enriching thing.

whatyesaid 4y ago
But the typical program decomposes into a lot of grunt functions/code so it's a productivity multiplier, meaning you can have less developers and do more. You've got to know what you expect as output, it just saves time. It's not just "hello world" but it's not the complete app. I'd be curious where it failed for you.
- AnimalMuppet 4y ago
  Right, it saves time. But you have to know enough to spot when it's wrong, and to be able to figure out what to do when it's wrong. To do that, you still have to know a language.
- literallyroy 4y ago
  I find it constantly fails and hallucinates when working with existing systems that weren’t designed in a standard why. For example, a database that isn’t normalized well and has odd one-to-one mappings tripped up its ability to problem solve around hibernate annotations. So perhaps if it builds the system it can build new few features just fine, but it struggles to problem solve when the original design doesn’t follow its expected pattern.
  - whatyesaid 4y ago
    I would try providing it a schema as the first message, just for its internal memory and see if that works, but yeah, I usually ask it to write domain independent stuff and then adapt that.
- giantg2 4y ago
  At least in my experience, those lower level functions need to integrate well. From what I'm hearing, these AI systems don't play well with integration, especially outside of well defined API specs.
- superdisk 4y ago
  In one case I tried using it to help me find the right functions in `oauthlib` that can validate a two-legged OAuth 1.0 signed message. It couldn't really get it and made up information about how OAuth works, but it did get me on the right track so I'm thankful in that regard. I've also been using it while writing Common Lisp, and I'm not sure if it's just because it's a more obscure language, but it commonly produces nonsensical code, recommends functions that don't exist or work differently than it thinks they do, or produces non-idiomatic weird stuff. It also falls apart when writing z80 assembly, again just generating nonsense. It really is a cool tool though, and I use it to get me steered in the correct direction. I've seen other people mention that it can identify unknown unknowns, and I concur it's an amazing tool for that. We'll see if with more specialized training on programming-related datasets if it can get better, I wouldn't write off the possibility. It won't be stealing your job in its current state, though.
  - dmux 4y ago
    >I've also been using it while writing Common Lisp, and I'm not sure if it's just because it's a more obscure language, but it commonly produces nonsensical code, recommends functions that don't exist or work differently than it thinks they do, or produces non-idiomatic weird stuff. I've been using it for Common Lisp as well and have had the same experience. For example, when asking it to help me generate some simple CRUD operations using the cl-sqlite library, it just comes up with function names that simply do not exist in that library [0]: `sqlite3:execute-command` `sqlite3:with-query-results` `sqlite3:get-result` [0] https://cl-sqlite.common-lisp.dev/ https://cl-sqlite.common-lisp.dev/
  - whatyesaid 4y ago
    It seems to really fall apart in anything niche, even if you're just asking information about a topic. Doesn't have enough training data on that topic.
p1esk 4y ago
In my experience GPT-4 is much better at producing correct code than GPT-3.5. Can you provide an example where it’s not the case?
- psyklic 4y ago
  GPT-4 doesn't seem anywhere close to writing or intelligently modifying medium/large programs. This is what most professional programmers do. LeCun at least believes this is a fundamental limitation of AR-LLMs that can't be overcome (e.g. his "Unpopular Opinion" slide -- https://drive.google.com/file/d/1BU5bV3X5w65DwSMapKcsr0ZvrMRU_Nbi/view https://drive.google.com/file/d/1BU5bV3X5w65DwSMapKcsr0ZvrMR...).

whatyesaid 4y ago

But the typical program decomposes into a lot of grunt functions/code so it's a productivity multiplier, meaning you can have less developers and do more. You've got to know what you expect as output, it just saves time. It's not just "hello world" but it's not the complete app. I'd be curious where it failed for you.

AnimalMuppet 4y ago
Right, it saves time. But you have to know enough to spot when it's wrong, and to be able to figure out what to do when it's wrong. To do that, you still have to know a language.
literallyroy 4y ago
I find it constantly fails and hallucinates when working with existing systems that weren’t designed in a standard why. For example, a database that isn’t normalized well and has odd one-to-one mappings tripped up its ability to problem solve around hibernate annotations. So perhaps if it builds the system it can build new few features just fine, but it struggles to problem solve when the original design doesn’t follow its expected pattern.
- whatyesaid 4y ago
  I would try providing it a schema as the first message, just for its internal memory and see if that works, but yeah, I usually ask it to write domain independent stuff and then adapt that.
giantg2 4y ago
At least in my experience, those lower level functions need to integrate well. From what I'm hearing, these AI systems don't play well with integration, especially outside of well defined API specs.
superdisk 4y ago
In one case I tried using it to help me find the right functions in `oauthlib` that can validate a two-legged OAuth 1.0 signed message. It couldn't really get it and made up information about how OAuth works, but it did get me on the right track so I'm thankful in that regard. I've also been using it while writing Common Lisp, and I'm not sure if it's just because it's a more obscure language, but it commonly produces nonsensical code, recommends functions that don't exist or work differently than it thinks they do, or produces non-idiomatic weird stuff. It also falls apart when writing z80 assembly, again just generating nonsense. It really is a cool tool though, and I use it to get me steered in the correct direction. I've seen other people mention that it can identify unknown unknowns, and I concur it's an amazing tool for that. We'll see if with more specialized training on programming-related datasets if it can get better, I wouldn't write off the possibility. It won't be stealing your job in its current state, though.
- dmux 4y ago
  >I've also been using it while writing Common Lisp, and I'm not sure if it's just because it's a more obscure language, but it commonly produces nonsensical code, recommends functions that don't exist or work differently than it thinks they do, or produces non-idiomatic weird stuff. I've been using it for Common Lisp as well and have had the same experience. For example, when asking it to help me generate some simple CRUD operations using the cl-sqlite library, it just comes up with function names that simply do not exist in that library [0]: `sqlite3:execute-command` `sqlite3:with-query-results` `sqlite3:get-result` [0] https://cl-sqlite.common-lisp.dev/ https://cl-sqlite.common-lisp.dev/
- whatyesaid 4y ago
  It seems to really fall apart in anything niche, even if you're just asking information about a topic. Doesn't have enough training data on that topic.

p1esk 4y ago

In my experience GPT-4 is much better at producing correct code than GPT-3.5. Can you provide an example where it’s not the case?

psyklic 4y ago
GPT-4 doesn't seem anywhere close to writing or intelligently modifying medium/large programs. This is what most professional programmers do. LeCun at least believes this is a fundamental limitation of AR-LLMs that can't be overcome (e.g. his "Unpopular Opinion" slide -- https://drive.google.com/file/d/1BU5bV3X5w65DwSMapKcsr0ZvrMRU_Nbi/view https://drive.google.com/file/d/1BU5bV3X5w65DwSMapKcsr0ZvrMR...).