Forget counting the Rs in strawberry, biggest challenge to LLMs is not making up bullshit about recent events not in their training data

The whisper model has always been pretty crappy at these things: I use a speech to text system as an assistive input method when my RSI gets bad and it has support for whisper (because that supports more languages than the developer could train on their own infrastructure/time) since maybe 2022 or so: every time someone tries to use it, they run into hallucinated inputs in pauses - even with very good silence detection and noise filtering.

This is just not a use case of interest to the people making whisper, imagine that.

Geordi, disgusted: being a _con_tent creator

Geordi, interested: being a con_tent_ creator

Just a couple more GPUs bro, I’m sure it’s around the corner

They added sleeps to training jobs? Sounds like they deserve a raise for improving energy efficiency instead…

I live like 15mi from there, I would prefer the containment bubble to stay intact. But the tech bubble is welcome to go blow up any moment

Some cool stuff: I went to a cs high school kind of deal and we learned some pretty fun stuff, then were encouraged to experiment:

• teacher showed us to render the Mandelbrot set, and when I had that written I wrote a little thing that lets you zoom into it (it was buggy, but who cares! Infinite fractals!)

• in school, I installed Linux on my pc, played around and then fucked up my primary disk’s boot sector and partition table. That meant my DOS install would be fucked up too (and I needed that for homework!), so I read up on MBRs and realized that if I remember the size of the partition, I can restore it to how it was.… and ext2 gives you a correct size even if the partition table is screwy, meaning that with the first primary partition activated I could find the correct size and then make an entry for the second one, and so on, until everything was working again. Data loss prevented!

• I’d written a pretty good solution to a homework problem in C, but typo’d a cleanup command and deleted the source file. The previous data loss scenario had prepared me: I rebooted into a rescue disk, grepped the disk drive for parts of the source and got the file back! (Got multiple copies of the file back, in fact. That’s when I started using version control, hah)

It’s called “safe superintelligence” because they want investors to sign SAFE agreements and feel very smart about it.

Easy to arrive at that being the ideal profile:

• Ufos are a psyop
• Ufos fly
• cruise missiles fly
• hence, cruise missiles are a psyop.

Ehhhh. This is the identical PR they ended up accepting: https://github.com/SerenityOS/serenity/pull/24648

I have feels about the implied “the author behaves like a shithead because he’s ESL” but eh. If it works.

Of course - it’s all part of the cancel culture grift economy. I look forward to us being able to hate-read Jon’s self-published autobiography in a few months.

With just a little bit of creative bin-packing they can share a core with the processes that simulate the smell of sewage and leafblower noise.