Today we’re rolling out Gemini 4 Carbon, our strongest model with frontier performance in cybersecurity capabilities and long-horizon agentic workflows.
We’ve poured a lot of love into it, and it’s very good at working through hard problems.
Big news: Gemini Omni 1.2 Flash by @GoogleAI debuts at #1 in both Text-to-Video and Image-to-Video, and #3 in Video Edit!
Gemini Omni 1.2 Flash landed #1 in Text-to-Video with 1541 pts. This is a +27 pt improvement from the previous Gemini Omni 1.1 Flash at #2 with 1514 pts.
With this release, Google’s Omni variants take all 3 top slots in Text-to-Video.
See the Image-to-Video and Video Edit results in thread.
Congrats to the @GoogleAI team on this release!
Introducing Gemini 4 Carbon, it's a big jump over Argon which we released just 2 weeks ago.
Gemini 4 Carbon is SOTA on cyber, long horizon tasks, ranks #1 on Artificial Analysis V5, significantly improves on usability in domains like coding, and is more aligned.
This is Genie 4, our most realistic and interactive world model yet. 🌐
Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
@ahatamiz1 another ad-hominem... none of the statements drop the claims before it. https://t.co/SlCsIa2xl0
and I was working on sparse attn when it was not cool, look up routing transformers. Have fun making personal attacks, I am not gonna engage further.
@ahatamiz1 important to note I am not saying SWA is better or worse than linear. I think they are prob almost equally good if training recipe is correct, and the moat is not this, it is the data flywheel and great RL training recipe.
this is a fall back to ad-hominem...
You said earlier: "SWA is lossless inside the window and zero outside of it and it is exactly why it fails so badly. (https://t.co/WXVTxeTydu) for example shows that larger sliding windows hurts the long-context performance." and I asked a simple question to show the flaw in how you concluded what the paper really says.
In fact, the paper explicitly says: "We establish that windows that are too long actually hinder the linear attention layers from learning to model long-term dependencies during training. We hypothesize that this degradation is due to under-training of linear attention layers on the long context recall tasks."
If SWA layers "zero outside of the window" as you claim, then that should not cause linear attention layers to get under-trained.
@ahatamiz1 Not saying the effective receptive field is sliding window * layers but it is not zero either. In practice it makes a huge difference whether you do the proper overlapping SWA and not the chunked version with no overlap, like what llama 4 did