Hacker News Reader

Comment by j-bu | Hacker News Reader

"The knowledge cutoff date for Gemini 3.8 Flash is March 2026 – users can expect updated information for some domains while in others they may experience the model’s knowledge is limited to January 2025 (in line with the Gemini 3 Model Family)."

Kind of wild that they haven't (successfully) pretrained a base model since Jan-25.

View on HN

I'm curious if the knowledge cutoff is important, when the interface (Gemini app) can search online for recent information. Is there a big advantage to having everything internal?

Not directly - but latest research advancements, cleaner / richer datasets, etc. still require fresh base models. Not everything can be fixed through post training alone (e.g. why GPT-5.5 "Spud" was such a big jump, and also why GPT-6 "Astra" is now supposedly another big leap). Ofc model size etc also plays a role, but my (admittedly limited) understanding is that new base models _can_ also lead to big jumps even keeping parameter counts constant.

You don't need everything internal, but having some idea of recent events is useful. If you ask it to implement some local AI there's a decent chance it will try to use qwen 2.5 without wondering if anything better came out since

Search grounding is expensive, you can't force the model to do it either. I use Gemini a lot and it often replies with out-dated data. The more detailed the information you're asking, the more likely it is to be wrong.

very important actually. just try to generate code for fresher frameworks/libraries. gemini sucks so bad in real work usage, everything it suggests are outdated and mostly useless.

Sometimes, it’s quite presomptuous and doesn’t search when it should. I’ve had to argue too many times with it that, yes, the Nintendo Switch 2 is real.

That extremely likely just means that they're preparing an omega huge Gemini 4 Pro release and that that's what training right now on most of the compute

Reply on news.ycombinator.com