You've not seen databricks helper fuck up the way I've seen it fuck up.
We are in no danger of AI becoming the next UNISYS MAPPER (the marketing pitch), because AI is the next UNISYS MAPPER (the actual product): great for the trivial cases, and fucking terrible at anything else.
great for the trivial cases, and fucking terrible at anything else
This is pretty much my take on it. Everyone singing its praises has never built anything of worthwhile complexity. There are a handful of people I trust who endorse it, but when pressed about it you'll discover that their process for extracting useful results from it involves so much fiddling and exacting specification that you may as well just write the code yourself because you pretty much already have and the extra time it'll take to debug the hallucinations makes it a wash or net negative.
Everyone singing its praises has never built anything of worthwhile complexity.
Uhh ... I'm currently using Grok to help me work on and analyze a novel algorithm. ~2k lines of pretty dense C++ code with lots of math, bit twiddling, and custom data structures (think interleaved linked lists stored in a single std::vector).
I'm continually astounded by how deeply it understands the operation of the code. I ask it to generate unit tests for edge cases and it just does it. I point out a bug in the output and ask it for a solution and it nearly always gives me a good starting point to get it fixed, if not an outright fix.
It's really damn impressive. It's definitely not perfect, but if I'd had access to this earlier on the project, I have already identified several months worth of hacking and experimentation that it would have saved because its suggestions are that good.
I've not touched Grok so maybe that's better than the handful of models I've tried. I think there's probably some measure of utility in using it as a summary generator, but I remain steadfast in my wariness of it as a tool to generate code.
As a code generator? No, not there yet. I've had it generate some code and it's not terrible but you really have to prod it a lot to not generate extremely inefficient slop (e.g. doing dumb stuff like auto foo = bar - baz; if (foo < 0) ... instead of if (bar < baz) ... ). In general, writing the code isn't the hard part; it can just be tedious in some situations.
But as an analysis tool, it's absolutely fantastic. And given that, it's only a matter of time until the code comes with it.
I'd recommend giving it a try. Hit it with the trickiest bit of code you've ever written. My jaw was on my desk for the first week I was using it, and I paid for the subscription after I hit the 2-hour-15 question limit because I could immediately see the utility.
You've not seen databricks helper fuck up the way I've seen it fuck up.
We are in no danger of AI becoming the next UNISYS MAPPER (the marketing pitch), because AI is the next UNISYS MAPPER (the actual product): great for the trivial cases, and fucking terrible at anything else.
So it makes pajeet codemonkeys obsolete.
It's the exact same Stack Overflow -> production pipeline, but with fewer typos and cut/paste errors.
Sounds good to me!
What do you think it's trained on?
This is pretty much my take on it. Everyone singing its praises has never built anything of worthwhile complexity. There are a handful of people I trust who endorse it, but when pressed about it you'll discover that their process for extracting useful results from it involves so much fiddling and exacting specification that you may as well just write the code yourself because you pretty much already have and the extra time it'll take to debug the hallucinations makes it a wash or net negative.
Uhh ... I'm currently using Grok to help me work on and analyze a novel algorithm. ~2k lines of pretty dense C++ code with lots of math, bit twiddling, and custom data structures (think interleaved linked lists stored in a single std::vector).
I'm continually astounded by how deeply it understands the operation of the code. I ask it to generate unit tests for edge cases and it just does it. I point out a bug in the output and ask it for a solution and it nearly always gives me a good starting point to get it fixed, if not an outright fix.
It's really damn impressive. It's definitely not perfect, but if I'd had access to this earlier on the project, I have already identified several months worth of hacking and experimentation that it would have saved because its suggestions are that good.
I've not touched Grok so maybe that's better than the handful of models I've tried. I think there's probably some measure of utility in using it as a summary generator, but I remain steadfast in my wariness of it as a tool to generate code.
As a code generator? No, not there yet. I've had it generate some code and it's not terrible but you really have to prod it a lot to not generate extremely inefficient slop (e.g. doing dumb stuff like auto foo = bar - baz; if (foo < 0) ... instead of if (bar < baz) ... ). In general, writing the code isn't the hard part; it can just be tedious in some situations.
But as an analysis tool, it's absolutely fantastic. And given that, it's only a matter of time until the code comes with it.
I'd recommend giving it a try. Hit it with the trickiest bit of code you've ever written. My jaw was on my desk for the first week I was using it, and I paid for the subscription after I hit the 2-hour-15 question limit because I could immediately see the utility.