Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.
- 187
- 1
Jump in the discussion.
No email address required.
Notes -
So I know a lot of posters here are doing some really advanced stuff with AI, but I have a quick little anecdote from my own luddite life that might help explain why most normies don't really see AI as something particularly useful.
I just asked Google's free AI what the current value of 0.035 ounces of gold is and it completely face-planted, ignoring the decimal and the zero and giving me the value for 35 ounces. This isn't a case of overactive guardrails or confusing syntax or hallucinating plausible details into a news story; it's the most basic of transcription errors, the kind that dumb, pre-LLM systems have solved for decades. Old google never would have made such an error.
This is the technology that's going to completely upend the world order? It makes sense for normies to be skeptical. Regardless of what the frontier labs have cooking, the AI most people actually interact with is worse than useless half the time.
Anybody else have one of these "the fuck are we even doing here?" moments with AI recently?
I'm amazed at how many people don't cough up even the $20 required for the full-fat experience. At least OpenAI has recently announced that they're making unlimited usage of a newer model, 5.6 Luna, available to the masses for free. It is still several steps down from Sol, which continues to genuinely impress me.
Google's "free AI" is an incredibly imprecise term. It can mean Gemini 3.5 (Flash? Lite?) when used through the Gemini app. It can be the search-specific Gemini 1.5, IIRC, used if you just "Google" something. As you can imagine, not even Google can afford to use the best of the best when serving so many people at such scale, with the majority being trivial queries.
I would pay a great deal more than $20 for an AI plan, but luckily, I don't have to. I've got tons of Claude Max, because I decided to accept payment in that form instead of figuring out intercontinental financial transfers after my participation in Unslop.
I remember when a Dutch junior of mine was consulting another senior for advice when working on a meta-analysis meant for a big name journal. I could have clapped like a seal when the latter mentioned Claude, in addition to ChatGPT, for help with the stats scutwork. Then I asked, and confirmed, that she only had a free plan. I was so dismayed that I immediately offered to share my Max, and walk her through things. It didn't hurt that she's a pretty girl, and it would have provided an opportunity to go out with her (that ended up happening anyway), but I am beyond annoyed when other doctors use free-tier LLMs for clinical work. They still function adequately, but anyone who can afford better should use better. $20 is really not an onerous ask for a significant amount of intelligence on tap.
Yeah, very recently decided to try the Claude Pro-tier AI for a month, and I'm very impressed. I did have access to Gemini 3.1 Pro through work, but it feels like a night-and-day difference.
We'll see if I still feel this way in a couple weeks, but for now I'm firmly in the "very, very useful" camp.
I'm actually a huge fan or 3.1 pro (high) personally. Opus and 5.x xhigh are definitely better at difficult problem solving but 3.1 pro can handle most things and is much more chill and less anal about many things.
Would be interesting to hear what specific tasks 3.1 pro underperforms on that opus can take care of
I've tried to use Gemini 3.1 pro to handle some basic coding tasks at work, and it does not do particularly well. I don't want to give exact specifics to avoid doxxing myself, but I can give you the shape of it. Skip to the most direct example if you don't feel like reading.
I have a java project. Currently it uses all-java libraries for image processing. Java kind of sucks for image processing because the integrated entry point always loads the entire raster into memory at once. If you have a limited heap, you're going to run into a world of pain trying to work with the standard library.
To get around that problem, I've looked at a native library that has java FFM bindings. Unfortunately, to use FFM, I had to upgrade to JDK 25. We had some test failures, and I figured that Gemini/antigravity could handle that kind of scutwork. It could not. It tried to make massive, architecture-level changes to two different subsystems in our codebase rather than just fix a classloader problem. Eventually I gave up and did it myself. I lost about a day to this.
After upgrading the JDK, I handed off the work to another developer to handle writing a small wrapper around the FFM library to unpack the native libs. She immediately tried to use Gemini, and lost four days to its confabulations. She came back to me repeatedly telling me it was impossible, and that we couldn't possibly do this on windows because Gemini gave her a trivially disprovable assertion. I eventually gave up and handed it to another developer who engaged his brain and had it done in a few hours, on all our supported platforms, with tests.
The most direct example: After that, I started converting one of our image processing routines to use the native lib. I figure that since the problem was easy and both the native library and the FFM bridge are both exquisitely documented, and they're both open source, this should be trivial for Gemini. Well, it turns out that both the native lib and the FFM bridge were both mostly written after January 2025, so neither 3.1 Pro and 3.6 flash consistently had the APIs inside their training window. It also didn't really have many examples to match on because this isn't a basic Python CRUD app.
You would not believe the absolute fever dream of a codebase it tried to cook up. It couldn't get function names right, and when it could, it couldn't consistently distinguish between the native library and the bridge. It frequently failed to even be consistently wrong. Eventually it finished, and it solved the problem by importing the FFM bridge but not actually using any of the underlying native calls.
I was a little disappointed.
RIP in pepeloni. But to be fair most agentic ai loves to do this unless you prompt it correctly.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link