---
title: 'DeepMind''s New AI: A Gift to Humanity'
source: 'https://youtube.com/watch?v=Sk9tvyRSCgY'
video_id: 'Sk9tvyRSCgY'
date: 2026-07-24
duration_sec: 715
channel: 'Two Minute Papers'
---

# DeepMind's New AI: A Gift to Humanity

> Source: [DeepMind's New AI: A Gift to Humanity](https://youtube.com/watch?v=Sk9tvyRSCgY)

## Summary

Google DeepMind released Gemma 4, a free and open family of AI models that can run on devices with minimal hardware, including phones and even a Nintendo Switch. This video explores four surprising aspects of Gemma 4: its performance, agentic capabilities, context window, and permissive license, highlighting how it empowers users to own and run AI locally without reliance on proprietary cloud services.

### Key Points

- **The Problem with Proprietary AI** [00:00] — Users of Claude AI lost access due to 'heavy workloads', highlighting the risk of relying on cloud-based, subscription AI services.
- **Gemma 4: Free and Open AI** [01:06] — Google DeepMind's Gemma 4 is a free, open family of models. The smallest models require only a few GB of memory and run on devices like phones and Nintendo Switch without internet.
- **Surprise 1: Performance of 31B Model** [02:50] — The 31B parameter dense model beats models 10-20 times larger on some benchmarks, despite being dense (all parameters active) rather than mixture-of-experts.
- **Four Key Improvements in Gemma 4** [04:41] — 1) Highly curated training data; 2) Hybrid attention (sliding window + global); 3) Better image understanding (no squishing); 4) Shared KV-cache for efficiency.
- **Surprise 2: Agentic Workflows** [06:52] — Gemma 4 excels at tool use, local coding, and agentic tasks like booking flights or summarizing news, making it a viable free alternative to proprietary assistants.
- **Surprise 3: 256k Context Window** [07:51] — The context window is doubled to 256k tokens, allowing processing of long documents, though it's computationally expensive.
- **Surprise 4: Apache 2.0 License** [08:07] — Gemma 4 uses Apache 2.0 license, removing restrictions present in Gemma 3. It allows modification, commercial use, and derivative models with minimal friction.
- **Limitations of Gemma 4** [09:07] — No live database; can be confidently incorrect; struggles with highly complex tasks and images with fine details like thin structures or grass.

### Conclusion

Gemma 4 is a significant gift to humanity, providing free, open, and locally-run AI that cannot be taken away. It empowers individuals and small developers, reducing dependence on big tech companies.

## Transcript

Google DeepMind gave an amazing gift to humanity.  And it is full of surprises. Here’s why. Today,   we are living in the age of AI where these  smart assistants and agents can do things  
we could only dream of 10 years ago. But.  Many of these solutions are proprietary,   require a subscription, and run in the cloud. And then this happens. Yup, some OpenClaw users  
reported losing access to their Claude  AI subscription citing “heavy workloads”. Now, maybe they did something unsavory, I don’t  know. I also understand you pay a fixed rate,  
you can’t eat all you want.  I respect that. However,   this is the point. We have to rely on the  goodwill of these companies for our workflows. So this is why I keep saying over and over that we  
should always look for options where you  we own these AIs and run them on our own   systems for free, forever. No one can take them away. NVIDIA came out with their Nemotron 3 Super,  
which has super capabilities…but its  hardware requirements are also super.  Not so much with Google DeepMind’s new AI, Gemma  4. This is a free and open family of models,  
and yes, finally, the smallest ones  require only a few gigabytes of memory. No need for an expensive GPU. So much so that  I wanted to wait a bit before publishing this  
video to see how you Fellow Scholars use  it in practice. And…look at that! It runs   on your phone without an internet connection.  And folks are already using it in practice to  
create offline translation and summarization  apps. Also, real time image classification   running in your browser while talking like a  bard? No problem. You can already fine tune it  
with Matt’s work. It is so good, it has a little  ecosystem around it already in just a few days.   Because of the brilliance of  you Fellow Scholars. Nice work.
But it gets better. You see, the smallest Gemma is  so small, it runs on…oh my. Look…I love that. It   runs even on an old beat up nintendo switch,  first generation. Not exactly something with  
a lot of memory or processing power. Still  runs the 2 billion parameter Gemma 4 model. Now that is a gift to humanity. But  it gets really strange from here   on out. Here are 4 things that I found  really surprising. Dear Fellow Scholars,  
One, they also have a bigger, 31B  model which was the #3 best open model,   and now hold on to your papers Fellow  Scholars, because it beat some models  
that are 10 times larger. And still  competitive with some that are 20 times   larger. On some measurements. And it is a  dense model. What? What is going on here?
You see, many of the modern AI systems  you encounter are what they call mixture   of experts models. MoE. These are  huge AI models with many parameters,  
and to make sure we don’t burn down all of our  hardware using them, it splits up this big brain   into many small ones. If you have a biology  question, it chops it up into small parts,   and routes them to the parts of the brain that  it thinks are the best experts at processing it.  
Typically, to the top 2 to 8 experts. Only  ask them. Yes, with that, we only activate one   small part of a brain at a time. It makes  sense, right? It’s not a simple process,  
but it is possible. This enables us to create  huge intelligent models that are still efficient. Dense models, however, just light up every  parameter of the system. These are not new,  
and in some ways, these are very  inefficient. You light up all the   31 billion parameters in the brain all  the time, no matter how simple or complex   the question is. But this one…this  one is somehow magically good. How?
They did four amazing things: one, Google  didn’t just dump half the internet into it   to learn about us. They apply super strict  filters to give it only highly curated  
training data. That is actually good advice  for our thinking too. Don’t let everything in,   curate your information diet. There is lots of  noise out there - ignore it. That is excellent.
Two, they use an interesting attention  mechanism that has a sliding window and   also global attention at the same time. What  does that mean? Well, when you read a book,  
you read it line by line to finish a page. That is  a local sliding window. With that, you get all the   details. But sometimes you want to zoom out and  ask, okay, what book are we reading? Which chapter  
is this? That is global attention. Here, they use  both, and call the mechanism hybrid attention. Three, it is better at understanding images.  You know, Gemma 3 had weird glasses on,  
and its image understanding was kind of  a lie. If you gave it a landscape image,   it squished it back to a square image before  processing it, losing some information. It  
squishes everything into its own preconceived box.  Not good. Gemma 4 understands the image as-is,   and the difference really shows on any  benchmark that has to do with images.
Four, it has a shared KV-cache.  KV-cache is short term memory for   what you are currently talking with it,  documents, questions. Now the layers of   this neural network like to recompute their  fresh memory from scratch. This one doesn’t,  
it essentially borrows the memory already computed  by earlier layers. Less work, nearly the same   result. This is one of those ideas where we are  wondering why we didn’t always do it like this.
Okay, and all this was just part of my first  surprise. Second surprise. It is fantastic at   agentic workflows. This is where we don’t just  have an AI assistant that spits out a bunch of  
text, this is when we give it arms and legs and  ask it to do stuff. Tool use, local coding, and   a ton more. Plug it into OpenClaw and it can book  a plane ticket. Look for news and summarize it  
in a more unbiased way. Or write silly emails to  Károly from Two Minute Papers. That sort of thing.   It is really good at that. So when any company  decides that you can’t use their system anymore,  
that’s alright. Just plug in Gemma 4, and you  are good to go. For free. People find that if   you give it custom instructions, sometimes you  don’t even notice the difference. That is huge.
Surprise number three, the context window was  improved to 256k, twice as big as Gemma 3 had.   This is pretty expensive to compute, so don’t  take it for granted. Here, you are not going  
to chuck gigabytes of movies into it, but for  a few long documents, it is perfectly fine. Four, the license. Oh my, the license. This  one gets overlooked so much. Gemma 3 came with  
a Gemma license. In other words, it comes with  strings attached. The model comes with handcuffs,   if you will. If you use it to create training data  for a derivative model. Yup, that one inherits the  
handcuffs too. But, with Gemma 4, not anymore.  Look, Apache 2.0 license. Now we’re talking,  
yes! This license is true to the open  source spirit. You can modify it,   sell it, deploy it commercially with  almost zero friction. Make derivative   models, do a ton of stuff with far fewer  restrictions. This is huge. Thank you so much!
Now, not even this technique is perfect.  For instance, the model does not have a   live database. Without an agent harness, it  cannot browse or look up stuff. Meaning? Well,  
meaning that it can be confidently incorrect.  The internet special. Also for highly complex,   open-ended tasks - it’s not great at that. Or,  
when you have images with lots of high-frequency  visual details, thin structures, blades of grass,   or a fence from far away. Not great at that,  it’s going to need even better glasses.
But, adding this all up, this is an amazing  gift to humanity, one that cannot be taken from   us. This is not for Mr moneybags, this  is for the little man, and it is free,  
for all of us, forever. Hugely appreciated.  Absolutely loving it. What a time to be alive!   Also, I waited with this video because I did  not just want to take the marketing messaging  
and copy-paste it to you. I wanted to see how  you Fellow Scholars are actually using it in   practice. Read through your experiences  with it. Does it really work in practice?   Super important. That’s what we are here  for, not the copy-pasted media headlines.
That needs time. Trying to explain all this in  simple words also takes time. I don’t have a   team here, I do everything from the writing to  recording, video editing. I am trying my best  
here. But it gives you more accurate information,  and that is the most important for me. So, now, after 10 million downloads in the  first week and more thorough testing. My  
opinion is that yes, this thing rocks. I would  like to send a big thank you to every single   scientist who worked on this! And hold on to  this one for dear life because a frontier model  
just got locked down for a few select clients.  You know, it’s a big club. And we ain’t in it. If you enjoyed this, consider  subscribing and hitting that bell.