Still not paying a cent for any LLM/DL. The day i really need it i will run a open source one locally.

Last edited: Yesterday at 1:21 PM

LLM's are getting heavy and obese. Look at android lightweight and fast but not secure.
Ozembic for everybody.

LLMs are like OSes. You can make them lightweight, or you can make them feature/capability rich. Not both.

And I wouldn't characterize Android as light--especially the Pixel or OEM images. The actually light Android images--people often complain about lacking XYZ.

Why do I get a "Wait another hour and they'll announce something "better than ever" in even more hyperbolic terms" vibe from this?

Oh, it's more AI propaganda.

Not at all to disparage Ryan's write-ups, but you'd think Ars writers would get kind of tired trying to keep up with the "most bestest and greatest thing ever!!!!" hype machines that sparks each AI wannabe to announce a Brand New AI Thing every ten minutes.

Just have the fucking bubble pop. I have a ten pound bag of popcorn all set to go for that show. Again, not to disparage the reporting of it at all, but THIS show is kinda repetitive and boring. Like a conga line, a lot more fun to be in than watch, I imagine, and just as directionless as one, too.

Shit I just want a gemma model that is reliable- gemma4:31b did great on local hardware, but still has the infinite loop bug (not to mention a few other items that made it hard to rely on for local work; not even just coding, I was trying to do a newsletter curator and it would just choke processing text).

Usually when we ship a superseding product only a couple weeks after the last version, somebody has an ostrich-sized egg on their face. Really critical cockup that QA/QC should have caught, but so fundamental and embarrassing that it shouldn't even have gotten as far as QA/QC.

But move fast and break things or whatever.

During their earnings call a couple weeks ago, they said that Gemini is going to move to a monthly release cycle, so I'm assuming this is the first example of that new approach.

Usually when we ship a superseding product only a couple weeks after the last version, somebody has an ostrich-sized egg on their face. Really critical cockup that QA/QC should have caught, but so fundamental and embarrassing that it shouldn't even have gotten as far as QA/QC.

But move fast and break things or whatever.

What exactly is the issue with updating the model here. Does it force everyone to migrate their tools or code?

The article doesn't make it clear what was the intended cadence, whether the previous release was delayed, or this one rushed, or they just release whenever they meet some benchmark. I'm not really seeing an issue unless there's an indication of some kind of massive internal fuckup.

I'll start usung it once it has autonomously hacked another company.

/s

Still not paying a cent for any LLM/DL. The day i really need it i will run a open source one locally.

Okay. Thanks for letting us know.

A 15% increase in benchmark results in 3 weeks seems more like benchmaxing than anything else.

In my instance, I have found LLMs to be time wasters. Having to repeat multiple times, them using old data no longer applicable, have to double check all output on the one thing it does semi well, summarize.
The one thing it is good at is polishing a turd and then telling you its a diamond.

The quality of the prompting can have a mjor impact on the quality of the output

Still not paying a cent for any LLM/DL. The day i really need it i will run a open source one locally.

Siri 2 in macOS dev beta 5 is phenomenal.

A 15% increase in benchmark results in 3 weeks seems more like benchmaxing than anything else.

Either that or someone fixed a flaw/bug that wasn't found before release.

Big problem with LLMs is patching isn't really a thing, because none of the AI vendors can explain what exactly will change. They can point to how it works better on benchmarks but not "your previous workflow did x, now it might do y" and it's on the user to do all the testing. Every change has to be treated like a new release.

Deadcat bounce release from their team or the sign of better leadership?

Team wanted to set a new record for fastest abandoned Google product.

The quality of the prompting can have a mjor impact on the quality of the output

That is correct, but that is not how they sell them to the public.