[00:00] Google DeepMind gave an amazing gift to humanity.&nbsp; And it is full of surprises. Here’s why. Today,&nbsp;&nbsp; we are living in the age of AI where these&nbsp; smart assistants and agents can do things&nbsp;&nbsp;
[00:13] we could only dream of 10 years ago. But.&nbsp; Many of these solutions are proprietary,&nbsp;&nbsp; require a subscription, and run in the cloud. And then this happens. Yup, some OpenClaw users&nbsp;&nbsp;
[00:28] reported losing access to their Claude&nbsp; AI subscription citing “heavy workloads”. Now, maybe they did something unsavory, I don’t&nbsp; know. I also understand you pay a fixed rate,&nbsp;&nbsp;
[00:40] you can’t eat all you want.&nbsp; I respect that. However,&nbsp;&nbsp; this is the point. We have to rely on the&nbsp; goodwill of these companies for our workflows. So this is why I keep saying over and over that we&nbsp;&nbsp;
[00:53] should always look for options where you&nbsp; we own these AIs and run them on our own&nbsp;&nbsp; systems for free, forever. No one can take them away. NVIDIA came out with their Nemotron 3 Super,&nbsp;&nbsp;
[01:06] which has super capabilities…but its&nbsp; hardware requirements are also super.&nbsp; Not so much with Google DeepMind’s new AI, Gemma&nbsp; 4. This is a free and open family of models,&nbsp;&nbsp;
[01:20] and yes, finally, the smallest ones&nbsp; require only a few gigabytes of memory. No need for an expensive GPU. So much so that&nbsp; I wanted to wait a bit before publishing this&nbsp;&nbsp;
[01:33] video to see how you Fellow Scholars use&nbsp; it in practice. And…look at that! It runs&nbsp;&nbsp; on your phone without an internet connection.&nbsp; And folks are already using it in practice to&nbsp;&nbsp;
[01:47] create offline translation and summarization&nbsp; apps. Also, real time image classification&nbsp;&nbsp; running in your browser while talking like a&nbsp; bard? No problem. You can already fine tune it&nbsp;&nbsp;
[02:01] with Matt’s work. It is so good, it has a little&nbsp; ecosystem around it already in just a few days.&nbsp;&nbsp; Because of the brilliance of&nbsp; you Fellow Scholars. Nice work.
[02:16] But it gets better. You see, the smallest Gemma is&nbsp; so small, it runs on…oh my. Look…I love that. It&nbsp;&nbsp; runs even on an old beat up nintendo switch,&nbsp; first generation. Not exactly something with&nbsp;&nbsp;
[02:34] a lot of memory or processing power. Still&nbsp; runs the 2 billion parameter Gemma 4 model. Now that is a gift to humanity. But&nbsp; it gets really strange from here&nbsp;&nbsp; on out. Here are 4 things that I found&nbsp; really surprising. Dear Fellow Scholars,&nbsp;&nbsp;
[02:50] One, they also have a bigger, 31B&nbsp; model which was the #3 best open model,&nbsp;&nbsp; and now hold on to your papers Fellow&nbsp; Scholars, because it beat some models&nbsp;&nbsp;
[03:07] that are 10 times larger. And still&nbsp; competitive with some that are 20 times&nbsp;&nbsp; larger. On some measurements. And it is a&nbsp; dense model. What? What is going on here?
[03:24] You see, many of the modern AI systems&nbsp; you encounter are what they call mixture&nbsp;&nbsp; of experts models. MoE. These are&nbsp; huge AI models with many parameters,&nbsp;&nbsp;
[03:36] and to make sure we don’t burn down all of our&nbsp; hardware using them, it splits up this big brain&nbsp;&nbsp; into many small ones. If you have a biology&nbsp; question, it chops it up into small parts,&nbsp;&nbsp; and routes them to the parts of the brain that&nbsp; it thinks are the best experts at processing it.&nbsp;&nbsp;
[03:53] Typically, to the top 2 to 8 experts. Only&nbsp; ask them. Yes, with that, we only activate one&nbsp;&nbsp; small part of a brain at a time. It makes&nbsp; sense, right? It’s not a simple process,&nbsp;&nbsp;
[04:08] but it is possible. This enables us to create&nbsp; huge intelligent models that are still efficient. Dense models, however, just light up every&nbsp; parameter of the system. These are not new,&nbsp;&nbsp;
[04:23] and in some ways, these are very&nbsp; inefficient. You light up all the&nbsp;&nbsp; 31 billion parameters in the brain all&nbsp; the time, no matter how simple or complex&nbsp;&nbsp; the question is. But this one…this&nbsp; one is somehow magically good. How?
[04:41] They did four amazing things: one, Google&nbsp; didn’t just dump half the internet into it&nbsp;&nbsp; to learn about us. They apply super strict&nbsp; filters to give it only highly curated&nbsp;&nbsp;
[04:53] training data. That is actually good advice&nbsp; for our thinking too. Don’t let everything in,&nbsp;&nbsp; curate your information diet. There is lots of&nbsp; noise out there - ignore it. That is excellent.
[05:09] Two, they use an interesting attention&nbsp; mechanism that has a sliding window and&nbsp;&nbsp; also global attention at the same time. What&nbsp; does that mean? Well, when you read a book,&nbsp;&nbsp;
[05:23] you read it line by line to finish a page. That is&nbsp; a local sliding window. With that, you get all the&nbsp;&nbsp; details. But sometimes you want to zoom out and&nbsp; ask, okay, what book are we reading? Which chapter&nbsp;&nbsp;
[05:39] is this? That is global attention. Here, they use&nbsp; both, and call the mechanism hybrid attention. Three, it is better at understanding images.&nbsp; You know, Gemma 3 had weird glasses on,&nbsp;&nbsp;
[05:54] and its image understanding was kind of&nbsp; a lie. If you gave it a landscape image,&nbsp;&nbsp; it squished it back to a square image before&nbsp; processing it, losing some information. It&nbsp;&nbsp;
[06:06] squishes everything into its own preconceived box.&nbsp; Not good. Gemma 4 understands the image as-is,&nbsp;&nbsp; and the difference really shows on any&nbsp; benchmark that has to do with images.
[06:20] Four, it has a shared KV-cache.&nbsp; KV-cache is short term memory for&nbsp;&nbsp; what you are currently talking with it,&nbsp; documents, questions. Now the layers of&nbsp;&nbsp; this neural network like to recompute their&nbsp; fresh memory from scratch. This one doesn’t,&nbsp;&nbsp;
[06:37] it essentially borrows the memory already computed&nbsp; by earlier layers. Less work, nearly the same&nbsp;&nbsp; result. This is one of those ideas where we are&nbsp; wondering why we didn’t always do it like this.
[06:52] Okay, and all this was just part of my first&nbsp; surprise. Second surprise. It is fantastic at&nbsp;&nbsp; agentic workflows. This is where we don’t just&nbsp; have an AI assistant that spits out a bunch of&nbsp;&nbsp;
[07:06] text, this is when we give it arms and legs and&nbsp; ask it to do stuff. Tool use, local coding, and&nbsp;&nbsp; a ton more. Plug it into OpenClaw and it can book&nbsp; a plane ticket. Look for news and summarize it&nbsp;&nbsp;
[07:21] in a more unbiased way. Or write silly emails to&nbsp; Károly from Two Minute Papers. That sort of thing.&nbsp;&nbsp; It is really good at that. So when any company&nbsp; decides that you can’t use their system anymore,&nbsp;&nbsp;
[07:38] that’s alright. Just plug in Gemma 4, and you&nbsp; are good to go. For free. People find that if&nbsp;&nbsp; you give it custom instructions, sometimes you&nbsp; don’t even notice the difference. That is huge.
[07:51] Surprise number three, the context window was&nbsp; improved to 256k, twice as big as Gemma 3 had.&nbsp;&nbsp; This is pretty expensive to compute, so don’t&nbsp; take it for granted. Here, you are not going&nbsp;&nbsp;
[08:05] to chuck gigabytes of movies into it, but for&nbsp; a few long documents, it is perfectly fine. Four, the license. Oh my, the license. This&nbsp; one gets overlooked so much. Gemma 3 came with&nbsp;&nbsp;
[08:20] a Gemma license. In other words, it comes with&nbsp; strings attached. The model comes with handcuffs,&nbsp;&nbsp; if you will. If you use it to create training data&nbsp; for a derivative model. Yup, that one inherits the&nbsp;&nbsp;
[08:36] handcuffs too. But, with Gemma 4, not anymore.&nbsp; Look, Apache 2.0 license. Now we’re talking,&nbsp;&nbsp;
[08:48] yes! This license is true to the open&nbsp; source spirit. You can modify it,&nbsp;&nbsp; sell it, deploy it commercially with&nbsp; almost zero friction. Make derivative&nbsp;&nbsp; models, do a ton of stuff with far fewer&nbsp; restrictions. This is huge. Thank you so much!
[09:07] Now, not even this technique is perfect.&nbsp; For instance, the model does not have a&nbsp;&nbsp; live database. Without an agent harness, it&nbsp; cannot browse or look up stuff. Meaning? Well,&nbsp;&nbsp;
[09:20] meaning that it can be confidently incorrect.&nbsp; The internet special. Also for highly complex,&nbsp;&nbsp; open-ended tasks - it’s not great at that. Or,&nbsp;&nbsp;
[09:33] when you have images with lots of high-frequency&nbsp; visual details, thin structures, blades of grass,&nbsp;&nbsp; or a fence from far away. Not great at that,&nbsp; it’s going to need even better glasses.
[09:47] But, adding this all up, this is an amazing&nbsp; gift to humanity, one that cannot be taken from&nbsp;&nbsp; us. This is not for Mr moneybags, this&nbsp; is for the little man, and it is free,&nbsp;&nbsp;
[09:59] for all of us, forever. Hugely appreciated.&nbsp; Absolutely loving it. What a time to be alive!&nbsp;&nbsp; Also, I waited with this video because I did&nbsp; not just want to take the marketing messaging&nbsp;&nbsp;
[10:14] and copy-paste it to you. I wanted to see how&nbsp; you Fellow Scholars are actually using it in&nbsp;&nbsp; practice. Read through your experiences&nbsp; with it. Does it really work in practice?&nbsp;&nbsp; Super important. That’s what we are here&nbsp; for, not the copy-pasted media headlines.
[10:32] That needs time. Trying to explain all this in&nbsp; simple words also takes time. I don’t have a&nbsp;&nbsp; team here, I do everything from the writing to&nbsp; recording, video editing. I am trying my best&nbsp;&nbsp;
[10:47] here. But it gives you more accurate information,&nbsp; and that is the most important for me. So, now, after 10 million downloads in the&nbsp; first week and more thorough testing. My&nbsp;&nbsp;
[11:00] opinion is that yes, this thing rocks. I would&nbsp; like to send a big thank you to every single&nbsp;&nbsp; scientist who worked on this! And hold on to&nbsp; this one for dear life because a frontier model&nbsp;&nbsp;
[11:14] just got locked down for a few select clients.&nbsp; You know, it’s a big club. And we ain’t in it. If you enjoyed this, consider&nbsp; subscribing and hitting that bell.