Hacker News new | ask | show | jobs
by antipaul 69 days ago
Seems like the most basic thing ever.

Can’t even use the new Gemma on device model… no model selection besides fast/normal/thinking.

Also requires Google login

2 comments

Google models are good, their products are all subpar
I recently discovered you can lose chat history by switching between using Gemma web "app" and at least the Android version.

Simply start chat on web, head to couch and continue on phone, later go back to web and, crucially, do not refresh your browser, but resume chatting.

All that lovely stuff you chatted about on your phone will be gone from the chat...

Android app also loses anything you've typed but not submitted if the app is closed while in background. I guess kinda expected but also annoying if you get interrupted.

I also found Gemini oddly less accurate at web searches, and less willing to do them compared to ChatGPT. Working on an electronics project Gemini relied much more on innate, hence sometimes incorrect knowledge, while ChatGPT 4.5 went and verified datasheets and whatnot much more rigorously.

I wouldn't attribute that as much to the model as I would the tools, prompts, and app plumbing around them. This is where I find Google to be really lacking in comparison to Anthropic and my custom stack. I don't use OpenAI so cannot offer a comparison or opinion.
Right I meant the experience from the apps, since that was what OP talked about.

I couldn't even get the Gemini VSCode plugin to work so haven't tried that, just kept asking me to log in and saw others reporting the same.

So yeah...

Do you know anyone who is innovating beyond basic chat?

Yes there's Claude Code, but I just can't stand them with woke politics, limits, poor prompt adherence. ChatGPT has been ok, but mostly same chat window (with terrible model hallucination). At least in google's ai studio you can interrupt using keyboard.

Anthropic is out in front with their Desktop app (imo), I have to evaluate it for rollout at a client, latest update brought significant improvements. I find Claude and the soul thing to be annoying. I won't touch Sam Altman stuff, for similar economic / political reasons in the opposite direction. OpenAI is the most likely to fail too (imo) and rumor has it they have not had a successful pre training run since Ilya left, with the 5.x series being mid/late iterations on the unreleased 4.5

Personally, I built a full custom agent stack so I can have full control. There should be lots of options these days for beyond just chat. Search my HN or GitHub history if you want more details on my take.

Of course it requires a Google login. It’s an app for a Google service. That service is about Gemini, their strongest consumer model. Gemma is not much more than a tech demo.