This Week's Insane AI Breakthroughs: Math, Enzymes, World Generators & More

AI never sleeps and this week has been absolutely insane. Google DeepMind drops an AI that can design new enzymes from

scratch based on your specified purpose. They also released another AI model that can figure out the effects of DNA

mutations in plants. OpenAI drops not one, not two, but over 700 breakthroughs on math problems that we haven't been

able to solve for decades. This is like the biggest day in the history of mathematics. AI was also able to find

some new materials for room temperature magnetic semiconductors. We have a top open-source world generator which you

can interact with in real time. This has memory and can stay consistent for a long time. This AI is even crazier. It

can create simulated movements of human bodies, including the exact muscle and skeletal movements. Chat GPT can now

produce a lot more visual and interactive responses in its answers. We have a new open-source image generator

as well as a new video model and upscaler. We also have some new 3D model generators and a lot more. So, let's

jump right in. First up, we have a new world generator called World Play 2. Now, similar to Genie 3 or Hunyen World,

this also generates a world which you can interact with in real time. So, here are some examples. Now, the really cool

thing is you can do a lot more with this than just walking around. You can interact with different characters or

objects, as you can see here. And you can also prompt it to add additional effects or animations. It also

compresses what happened earlier into its memory so that the world stays consistent even if you walk away and

then you walk back. This is also able to maintain consistency over a really long time. So, here's an example where you

can keep going. And as you can see, there's no like noticeable degradation in quality. So, one of the most flexible

open- source world generators we have so far. Now, at the top, they've released the code to this already. So, in their

GitHub, it contains all the instructions on how to download and run this locally on your computer. Note that this is

based off of Alibaba's 1 2.2. And they also provide a faster turbo model, which can just generate a world in four steps.

Now, for this fast version, it's still pretty huge at 69 GB. This is the full unquantized version of one, which is why

it's so big. If you're interested in reading further, I'll link to this main page in the description below. Also,

this week, we have a new AI called motion space flow. This is a new system for generating human motion from just a

text prompt. For example, here is a person making a high kick with the left leg, or here's a person getting on their

hands and knees and crawling forward. You can also manually constrain any joint to be at a certain position. For

example, you could tell it where a hand or foot needs to be at any point in time and then generate the rest of the

animation around that constraint. So, a pretty useful AI if you want to generate some specific human motion from just a

text prompt. Now, at the top, they released the code to this already. So, if you click on this button and you

scroll down a bit here, it contains the instructions on how to download and run this locally on your computer. They've

also released the training and evaluation scripts for this as well, which is fantastic. If you click on

their hugging face folder, note that the total size of everything is only like 1.1 GB in size, so this should be able

to fit on most consumer GPUs. And this is under the Apache 2 license, which is very permissive. If you're interested in

reading further, I'll link to this main page in the description below. Also, this week, we have a new open-source

video generator called Kandinski 6.0. So, here are some examples of its generations. It's pretty flexible and it

can even do highaction scenes and other complicated movements. Now, they've released two different models. There's a

smaller light variant which is only 3.8 billion parameters and then there's a pro variant which is 29 billion

parameters. Now, I think the main drawback to this is it can only generate 5second clips whereas the other leading

open models like Miniax can do much longer. Also, this currently only supports text to video as well as image

to video, but not reference to video, which I think is now the gold standard. But another cool thing that they

released is a super resolution model, which is called Kandinsky 6 super resolution. This is basically a latent

upscaler. So, you would add this to your pipeline to upscale the latent up to four times. Now, if you're interested in

trying this out, this is already available in diffusers as well as comfy UI and other platforms. Now, for the

light model, it's fairly tiny at only 6.38 GB in size. And then for the full pro model, which is 29 billion

parameters, this is much larger at 60 GB in size. I would say the quality and flexibility isn't as good as Miniax H3,

but this is another decent open- source alternative. If you're interested in reading further, I'll link to this main

page in the description below. Also this week, OpenAI releases a pretty cool feature called intelligent UI. Now, in

the past, when you ask ChatGpt a question, it responds mostly with text, right? Sometimes it might be accompanied

by an image or a chart, but most of it is still text. Now, with intelligent UI, Chat GPT can generate an entire

interactive interface to help answer your question. For example, if you ask it to explain how a bike works, instead

of just giving you several paragraphs, it can create an interactive diagram where you can select different parts of

the bike to learn how they work. Or here's another example. If you ask it to plan for a Sunday lamb roast, you can

see the previous GPT just gives you all text, but for this one, it's able to make it look a lot nicer with images

like this and a breakdown of the cost. And you can also, you know, play around with this to adjust the numbers. And

then it also gives you an interactive checklist like this. Here's another important improvement. GPT6 can now

begin answering while it's still thinking rather than making you wait until all its reasoning is complete. So

if you compare its output speed with the previous GPT, then you can see it's a lot faster. Now here it says they are

rolling this out globally to paid plans. And then starting the next day, the rollout also expands to the free plan.

So, at the time of this recording, you should be able to see this more interactive interface in your chat GPT

responses, even if you're on the free plan. If you're interested in reading further, I'll link to this main page in

the description below. Also, this week, Amazon releases a really cool architecture called a looped diffusion

language model. It's like combining the abilities of diffusion models from image and video generators with large language

models, plus they're also adding a looped behavior on top. So, it's like a Frankenstein of a model. Here's how it

works. You see, instead of a regular language model which generates text strictly from left to right, this model

can predict multiple parts of an answer at the same time and then gradually refine those predictions. That's

basically the diffusion part of the model. And in order to refine its answer, it uses multiple recurrent

loops. As you can see from this diagram, you can think of each loop as like giving it a bit more time to think

through and revise its answer. And the results are pretty impressive. So if you look at its performance on various

reasoning, math and science, and coding benchmarks, you can see that this new model scores the highest for most of

these benchmarks compared to other similarsized models. And because it has that diffusion component, it can

generate multiple tokens in parallel. It's also way faster. So here it says this can generate like 8.5% more

throughput compared to another competitor. Plus, it uses 13.6% less compute. The nice thing is they've

released the model and code to this already. So if you click on this code button and you scroll down a bit here,

it contains all the instructions on how to get started. Note that they've released two different variants. There's

a smaller 1.7 billion parameter model which is around 4 GB in size. And then there's also a larger 8 billion

parameter variant which is around 16 GB. And note that this is under the Creative Commons non-commercial license. If

you're interested in reading further, I'll link to this main page in the description below. Also this week,

Google DeepMind continues to publish some incredible stuff. So they published something called Alpha Protein Novo. And

this is an AI that can basically design entirely new enzymes from scratch. This is like the gold standard. So first of

all, you need to understand what enzymes are. These are basically proteins that help with chemical reactions in your

body. For example, we have enzymes that digest food, enzymes that help you heal, and we also use a ton of different

enzymes in various industries to manufacture chemicals and medicines. The problem is that when we want to get an

enzyme to perform a completely new chemical reaction, we usually have to start with an existing enzyme found in

nature and then kind of tweak it and modify it through experiments. But what if we could just get AI to design an

entirely new enzyme from scratch for any chemical reaction you want? Well, that's exactly what Alpha Protein Novo is

trying to accomplish. It uses AI to design both the 3D structure of the enzyme and also the sequence of amino

acids that make it up. It then uses tools like AlphaFold 3 to predict how this enzyme would fold and interact with

your target. Now, the researchers tested their designs in actual lab experiments and the results were crazy. For example,

they designed an enzyme that can help produce piperodine. And this is a chemical that's commonly found in

pharmaceutical drugs. In another experiment, they also created an enzyme that can break down DEHP. This is an

environmental pollutant that's used as a plasticizer. And this is a pollutant that's really hard to get rid of. But

they were able to use alpha protein novo to basically create an enzyme that can break this down. Now, this is just early

stage research so far. It hasn't been peer- reviewed, but this is an important step toward a future where instead of

searching nature for a particular enzyme that can help with a chemical reaction, we can just get AI to design a

completely new enzyme from scratch based on whichever chemical reaction we want to speed up. So, this could include like

breaking down pollutants or producing chemicals more efficiently. Now, at the top, they released the GitHub to this.

So, if you scroll down a bit here, it contains all the instructions on how to download and run this yourself. Super

exciting times. I think this is like the most exciting breakthrough for me this week. If you're interested, I will link

to this paper in the description below. Now, that wasn't the only release from Google DeepMind. So, they also released

another model called Botanic 1. And this is used to understand plant DNA. Now, here's the problem they're trying to

solve. Imagine we have a melon plant and changing one tiny part of its DNA would make its flower different. and we want

to figure out exactly which part of the DNA is responsible for this. The problem is that DNA is huge and each individual

melon plant contains DNA that's slightly different in different places. So, it's extremely hard to find exactly what part

of the DNA led to this change in the flower. It's really hard to narrow down the possibilities and it takes a ton of

time. But here, Google was able to use AI to make this way faster. So, they used Gemma 4, which acts like a

scientist. So it would plan the hypothesis, write the pipeline and do the analysis. And the second model is

called botanic 1. This is a specialized model trained on the DNA sequences of 320 plant species. So think of this as

an AI that has learned which parts of plant DNA tend to stay the same throughout evolution. Here, Google gave

this AI system almost 2,500 candidates. In other words, possible places where the DNA mutation could

occur. And the system was able to analyze everything and output a ranked list of places worth investigating. And

the result was really good. So this AI was able to rank the correct mutation in first place among the 2500 candidates

with a 90% success rate. Meaning it's really good at pinpointing the exact location in the DNA that would cause

this flower change. This is just one example, but botanic AI can potentially give us further insights into plant DNA

and how certain sections affect plant traits. We could potentially use this to create better genetically modified crops

that are more resistant or produce higher yield. If you're interested in reading further, I'll link to this main

page in the description below. Also, this week we had probably the biggest day in the history of mathematics. So,

OpenAI just released not one, not two, not even a dozen, but hundreds of new math breakthroughs. These are solutions

to problems that humans have not been able to solve for decades. So, get this. They posted 719 manuscripts grouped into

372 families, and they were all generated from an unreleased internal model, which is supposed to be way

better than GPT6 Astra. Now, these span some of the hardest unsolved problems in mathematics, including number theory,

geometry, theoretical computer science. Some results tackle famous problems that mathematicians have struggled with for

decades. And the scale is mind-blowing. So, the model attempted roughly 4,000 problems. And you know what's even

crazier is that the model only thought for roughly 3 hours per problem. And for those of you who are skeptical of the

legitimacy of this, at least at the time of this recording, around a quarter or 162 of these manuscripts are already

leanverified. So this means they've been formally proven to be correct. Now, the remaining like 77% is not necessarily

wrong. They just haven't been rigorously verified yet. And for your information, even this is already remarkable. Like

most of these results are really exceptional breakthroughs that can change a whole field or it's even worthy

of like a top journal publication. And for your reference, even like the best mathematicians in the world can only

solve one or two of these in their lifetimes and most of them solve zero. But here, OpenAI just dropped like over

700 of these in one day. This amount far exceeds all human math discoveries combined in the past few years. That's

how crazy the scale of this is. Now, going over all like 700 of these manuscripts would be overwhelming. So,

at least for this video, let's just go over a handful of the top most shocking breakthroughs in simple terms. Probably

the biggest breakthrough is this quasi reman hypothesis. To understand this, we first need to understand the reman zeta

function. This is basically a formula that can output information about how prime numbers are distributed. Just a

quick refresher, a prime number is basically any number that can only be divided evenly by one and itself, but no

other number. For example, 2 3 5 7 11 and so on. These numbers cannot be divided by any other number other than 1

and itself. Now, what's interesting here is that these prime numbers don't seem to follow any simple pattern. As we

count higher and higher, it becomes difficult to predict exactly where the next prime number will appear. But

mathematicians discovered that instead of studying these prime numbers directly, they could study something

called the reman zeta function. This formula can reveal information about how prime numbers are distributed. It turns

out that this formula, if we plug in certain numbers, it can output zero. So some of these are pretty

straightforward. For example, -2 or -4, -6. But there are also more complex inputs that also output zero. These

inputs aren't ordinary numbers. They're called complex numbers which actually have two parts. But anyway, when we plug

these complex numbers into the formula, they also output zero. And these zeros are called important or trivial zeros.

And they're important because they reveal information about the pattern of prime numbers. It turns out if we plot

all these important zeros on a graph, they all lie along a vertical line. And the Reman hypothesis proposed that all

these important zeros should occur along a perfectly straight vertical line right down the middle of this potential area.

So that's the famous Reman hypothesis. Over the years, mathematicians have checked enormous amounts of important

zeros and they've all been consistent with this prediction. But checking billions or trillions of examples still

doesn't prove that it works forever. Maybe if we keep going, there is an important zero that would not be on this

line. Well, here's OpenAI's discovery. They didn't solve the full Reman hypothesis, but they did prove that

these mysterious zeros cannot appear within 12.5% from the outer edges of this area. So, essentially, they've

narrowed down the area where these zeros can possibly exist by 25%. And this is arguably the biggest breakthrough on the

list. This is worthy of winning a prestigious Fields medal if a human had ever solved it. Now, you might be

wondering, why does this even matter in the real world? Well, prime numbers actually are super important in cyber

security. For example, it's used in encryption and passwords and credit cards. Now, it doesn't mean that this

discovery can allow someone to suddenly hack your credit card or break encryption, but this discovery could

help mathematicians understand the pattern behind prime numbers a bit better. Keep in mind, this is a problem

that people have struggled to solve for over a century. If you want to create your own AI influencer and turn it into

basically unlimited content, definitely check out Higsfield, the sponsor of this video. Higsfield is one of the fastest

growing generative AI platforms, and they just launched AI influencer 2.0. This gives you an entire workflow for

creating a digital character and turning one idea into tons of different videos. You can start by building your

influencer completely from scratch. You can customize things like age, facial features, hair, body type, clothing,

accessories, and a lot more. Or you can simply upload a photo of yourself and create a character based on your own

appearance. What I especially like is that once you create your character, Higsfield saves its identity. So instead

of generating a slightly different person every time, you can keep using the same recognizable character across

different outfits, locations, lighting, and videos. But the really impressive part is Genjutsu. You can choose one of

their built-in trending videos or upload your own reference clip. And Higsfield recreates that entire performance using

your character that includes the body movement, timing, camera movement, and even the actions happening inside the

scene. And once you've generated the video, you can keep remixing it. You can swap the outfit, product, background, or

other elements while preserving the original motion. So, one performance can quickly turn into dozens of different

creative variations. If you want to create a ton of influencer content at scale, Higsfield's AI Influencer 2.0

gives you the entire workflow in one place. Check it out using the link in the description below. All right, moving

on to another breakthrough, which is number four, Hilbert's 10th problem over rational numbers. Now, to understand

this, we first need to understand what rational numbers are. These are basically any numbers that can be

expressed as fractions. for example, a half or a quarter or even regular whole numbers like 2 3 and four because you

can express them as fractions. It's just divided by 1. Now imagine I gave you an algebra equation where you need to

figure out the value. Now in this case the answer would be pretty obvious. It's either three org -3. But what if I gave

you a much more complicated equation involving dozens or even hundreds of unknown numbers? Well, back in 1900, a

mathematician called David Hilbert proposed a famous question. Is there a method which can take any equation like

this and determine whether it has a solution using only whole numbers? Now, here's what's interesting. In 1970, we

already proved that this was impossible. In other words, there's no universal method that can always determine whether

any given equation has a solution using only whole numbers. So, this has been disproven. But what if we also allowed

fractions instead of whole numbers? This would make things easier, right? Because now we have a lot more possible answers

to choose from. Well, we couldn't actually prove this. And this has remained unsolved for decades, at least

until now. So, OpenAI has published a proof claiming that even if we allow fractions, it's still impossible to

create a universal formula that can always determine whether these equations have a solution. Now, why does this

matter in the real world? Well, this discovery actually establishes a limit on what computers and AI can do. Even if

we had computers that were millions of times more powerful than what we have today, we still couldn't create an

algorithm that's guaranteed to solve this particular problem for every possible equation. Now, that doesn't

mean AI can't solve complicated equations. It just means there can never be one universal method that always

gives us a correct yes or no answer for every equation in this category. So, it kind of tells us that certain problems

are just impossible to solve with one universal algorithm, regardless of how powerful our computers become. All

right, moving on to the next major breakthrough. This one involves matrix multiplication. A matrix is basically

just a grid of numbers. And it turns out that multiplying these grids together is actually one of the most important

calculations in modern computing. It's used in all sorts of technologies, including computer graphics,

simulations, and especially AI. In fact, when you use an AI model, a huge amount of compute goes into just multiplying

these matrices together. But here's the problem. As these matrices get bigger and bigger, the amount of calculations

required to multiply them also grows really quickly. So, mathematicians have spent decades trying to figure out how

to multiply these matrices more efficiently. Well, back in 1969, a mathematician discovered something

remarkable. He found a clever way to reorganize the calculation so that multiplying matrices required fewer

operations than previously thought. And ever since then, researchers have continued improving these algorithms and

gradually reducing the amount of work required. And we measure this using something called the matrix

multiplication exponent. Basically, this is a number that tells us how quickly the computational work grows as the

matrices get larger. Now, the traditional method has an exponent of three, but over the decades, we've

managed to push this number down to around 2.37. The ultimate goal is to get as close to 2 as possible because that's

the theoretical minimum. Well, here's OpenAI's discovery. They published a new math proof claiming to reduce this

number all the way down to 2.25, which is a huge improvement over the previous record. This might not sound like much,

but again, keep in mind that when we're running AI models as well as other technologies, an insane amount of

compute goes into just multiplying these matrices. So like one round of multiplication might not seem like much,

but if you repeat this for like billions of times, then you know even this small difference in compute savings can add up

to quite a lot. All right, moving on to another major breakthrough. This involves something called the Burch

Swinton Dire conjecture. This is actually one of the famous Millennium Prize problems. Now to understand this,

we first need to understand something called elliptic curves. These are basically equations that when plotted on

a graph create certain curved shapes like this. Now imagine I gave you one of these equations and I asked you to find

all the possible points on the curve where both coordinates are rational numbers. A quick refresher, a rational

number is basically any number that can be expressed as a fraction. for example, a half, a quarter, or even whole

numbers. Now, imagine we plot one of these elliptic curves on a graph, and we're only interested in points where

both coordinates are rational numbers. What's interesting is that some curves have only a limited number of these

points, while others have infinitely many. So, how do we figure this out? Well, mathematicians discovered that

these curves have a special property. If you know certain points on the curve, you can use a method to combine them and

you can keep repeating this process to generate more points. So even if a curve has infinitely many points, you don't

necessarily need to find them individually. In some cases, you might just need a few starting points to

generate infinitely many others. So back in the 1960s, two mathematicians named Burch and Swinerton Dyer discovered that

instead of studying all these points directly, they could study something called an L function. They discovered

that by studying how this formula behaves around a particular input, they could potentially determine how many

independent starting points the curve has. For example, if the formula doesn't output zero at that particular input,

the prediction is that the curve only has a limited number of rational points. If it does output zero, then the way the

formula behaves around that zero could reveal how many independent starting points the curve needs. In other words,

instead of trying to find an infinite number of solutions, you could just study the behavior of this one

mathematical function to learn about the solutions. So this became the famous Bur Swinerton Dyer conjecture. Now

mathematicians have managed to prove this relationship in certain special cases. But proving it generally has

remained unsolved for 60 years. Like does this pattern hold true for all possibilities no matter what? Well,

here's OpenAI's discovery. They published a series of proofs claiming that this relationship holds for a much

larger class of elliptic curves than we thought. And combining this with another discovery related to these curves, they

claim to have proven the full prediction for almost all curves within certain families of related equations. Now, keep

in mind this doesn't mean OpenAI has solved the entire Bur Swinerton dire conjecture. There are still cases that

aren't covered by these proofs. But if these new results hold up to independent verification, this could represent

another major breakthrough on one of the most famous unsolved problems in mathematics. Again, keep in mind this is

one of the Millennium Prize problems. All right, that was super complicated. Hopefully, you didn't get a headache

from trying to understand all of this and we only covered like four out of the over 700 manuscripts they published.

It's an insane amount of information to process. I think the bottleneck right now has flipped to where we are no

longer limited by compute or the intelligence of AI, but rather by the capacity of us humans to actually

process all of these discoveries that AI is coming up with. I mean, even just reading a handful of these discoveries

is overwhelming, let alone trying to understand all 700 of them. Anyway, if you're interested, I will link to this

GitHub which contains the full collection of all these manuscripts. Also this week, Google released an

open-source AI model called embedding Gemma 2. Now, this model is a bit different from the large language models

that you're used to like GPT or Claude or Gemini because it's not meant to generate text or images. Instead, its

job is to understand information and make it searchable. So, you can feed it a ton of different things like photos,

video, audio recordings, and documents. And it'll basically convert all of this into numerical representations called

embeddings. It's basically converting all these different types of files like text, images, audio, and video into the

same dimensional space so that another model can search all this information easily. And what's particularly

impressive is how small the model is. So you can run this with as little as 270 million parameters for text only. You

can also add an additional 170 million parameters for vision and 300 million for audio. But like the total size of

this is under a billion parameters, making it super small. And because of its size, this can run pretty

efficiently on even just a smartphone like the Google Pixel. Now, its context window is pretty small. This is

basically the amount of information you can feed it. So, it can process up to 5.5 minutes of audio or 29 images or 59

video frames. Now, if you compare this with other similarsized embedding models, you can see how performant and

efficient this new embedding Gemma is. You can see it performs the highest on image embedding. And then here's the

benchmark on audio embedding. Again, this is not a regular language model, but you can use this for like AI powered

search engines or personal assistants or other applications which need to understand a large collection of

different files. At the bottom of this page, they've already released the model on hugging face. So, if you click on

this link, it contains all the instructions on how to get started. And because this is only a few hundred

million parameters, it's pretty tiny at only 1.5 GB in size. And the nice thing is this is under the Apache 2 license

which has very minimal restrictions. If you're interested in reading further, I'll link to this main page in the

description below. Also this week, probably the most well-known European AI lab, Mistl, has finally released their

latest model, Mistral Large 4, also known as Lonk. This is quite a massive and obese model. So here they say that

MRO large 4 is a 1 trillion parameter model. This is a mixture of experts. So when you use it only 52 billion

parameters are active. Now currently this is not even open sourced yet. They say that the weights will drop end of

this month. Now while it's their best model I would say its performance is still not up to par with the leading

open source models like Kimik 3 or GLM 5.3. For example, if you look at this intelligence index by artificial

analysis, Mr. Large is all the way down here performing even worse than the much more efficient DeepSeek Flash and it's

also behind like GLM 5.3 Flash as well as you know the other leading open models over here. So, not particularly

impressive in terms of intelligence. It's also not that impressive in terms of cost. So, this is even more expensive

than other open models that are better than it like deepseek 4.1 Flash as well as MIMO 2.6 6 Pro, which is just

incredibly cheap. So, if you look at like intelligence versus cost, ideally you want to be in this upper left

corner. But unfortunately, Mistro large is all the way down here. So, just given its performance and price, it's really

not competitive with the other leading models we have already. But if you're still interested in trying this out via

their API, I will link to this page in the description below. Also, this week, Meta releases something called Genia.

And this can turn a regular photo or video into complete 3D objects. Now, how this works is quite interesting. They

used an existing model called SAM 3D, but it adds a process that continuously checks how well the generated 3D object

matches the original image. After comparing the differences, it adjusts the appearance and position to better

match what was actually captured. Now, at the top of the page, they have released the code to this. So, if you

scroll down this page, it contains all the instructions on how to download and run this locally on your computer. Note

that this is under the creative comments non-commercial license. If you're interested in reading further, I'll link

to this main page in the description below. Also, this week, we have another really cool scientific discovery by AI.

So, Val's AI used a team of claude opus 5.5 agents to discover two promising materials that could potentially lead to

faster and smaller computer memory. Now, to understand why this matters, let's first look at how magnetic computer

memory works. So, computers store information using zeros and ones. And one way to do this is with tiny magnets.

If a magnet points in one direction, that could represent zero. And if it points in the other direction, that

could represent one. So, by changing the direction of these tiny magnets, we can store information. But there's a

problem. Regular magnets produce magnetic fields around them. So, when you pack millions or billions of tiny

magnets closely together, their magnetic fields can interfere with each other. And this makes it really hard to cram a

ton of magnets together and shrink these magnetic memory devices further. So ideally we want a material that can

store information like a magnet without producing a magnetic field that interferes with things nearby. So that's

where something called an antifromagnet comes in. In an antifromagnet you can think of the atoms in the magnet like

pointing in opposite directions. So one points up while the next one points down. So their magnetic fields kind of

cancel each other out. And this is really useful because now you could potentially pack these magnets much

closer together for computing without worrying about magnetic interference. And these materials could also allow

information to be switched much faster. Now you might be wondering how do we actually read the information stored

these magnets. Remember a computer needs to distinguish between a zero and a one. And one way of doing this is by

measuring how electricity flows through a memory device. Specifically within these magnets, electrons will behave

differently with different spins. So we can take advantage of these differences to create electric signals which tells

us whether the memory is in zero or one. The problem is that ordinary antiferomagnets generally don't

distinguish between different electron spins. So yes, their magnetic effects cancel out, but it also makes it harder

to read information using conventional techniques. So what we want is the best of both worlds. We want a material that

has almost no overall magnetic field, but we also need to like be able to distinguish electrons flowing within

these magnets based on their spin. Now, it's important to know that these aren't ordinary semiconductors. These magnetic

semiconductors are different and it's actually really hard for them to operate at room temperature or else it will mess

things up. Usually, we need to store these in like really cold conditions. So, we also need to design this to be

able to work in room temperature or above. Basically, the researchers asked AI to find some new or existing

materials that can satisfy these three things. It has to have a net zero magnetic field. It needs to be able to

differentiate electron spin. Plus, it needs to operate well at room temperature or higher. So, after running

like hundreds of quantum mechanical simulations, they identified two promising candidates. The first one has

this chemical name and this is a completely new material design that the AI came up with. And according to the

simulations, it could have these ideal properties. As you can see from this chart, it can even work at room

temperature. But there's a catch. To actually manufacture it, the atoms would need to be arranged in a very specific

pattern, which would be extremely difficult to achieve. But the second discovery is even more interesting. So

they identified another candidate with this chemical name, which we had actually manufactured all the way back

in 1999. Now, this material already has no overall magnetic field, which is ideal. Plus, it's also good at sorting

the spins of electrons. They found that this one can work at well above room temperature, even above 100° C. And

after running the simulations, the AI suggested that this is the exact combination of properties that the

researchers have been looking for. Now, there's one important thing to point out, which is that the properties of

these materials have only been predicted through computer simulations. They haven't actually confirmed anything in

physical experiments yet. The researchers also haven't built a working memory device from either material. So

that's the next step. It's to actually conduct physical experiments and try to build this out. The cool thing is

they've actually released the GitHub repo which documents everything. So you can read the entire thing if you're

interested in the technical details. And if future experiments confirm what the AI has predicted, then these materials

could help develop new types of magnetic memory that are faster, more compact, and more energyefficient. So, a pretty

cool discovery in material science. If you're interested in reading further, I'll link to this main page in the

description below. Also, this week, we have a pretty interesting AI called Flash Dex Retarget. This is used to

convert human hand demonstrations into motions that an actual robot hand can perform. Now, how this works is quite

interesting. Instead of training a completely separate system for every human demonstration, they just train one

learning policy across many different motions at the same time. The AI can also look ahead at the next few frames

of the movement. So, it has a better idea of where the hand and object are going. And if you look at these

benchmarks for converting human demonstrations into robotic movements, then this new approach on average beats

previous methods. The awesome thing is they've released the code to this already. So, if you click on this button

here, it contains all the instructions on how to download and run this yourself. However, at least for

training, then it does require like around 110 GB of GPU memory. And also note that currently the code is only for

the XHand and the Sharp Wave USD robotic hand. If you're interested in reading further, I'll link to this main page in

the description below. Also, this week, we have a new open-source image model called Iris 3B. So this is a three

billion parameter model. And the really cool thing is this does not use a VAE. In other words, it works entirely in

pixel space which you and I can see which in theory should give it higher accuracy and quality. So here are a ton

of different examples for your reference. It can generate photorealistic images as well as

different artistic styles. It also has no problem generating large text. Now, this can do text to image as well as

image to image, but it can't do semantic image editing like Nano Banana. Looking at the quality of these photos, I would

say it's not as good as the leading open models out there like Crea 2 or Quen 2.1. But this is another decent

lightweight alternative. But what's really cool is you can fine-tune this further for a specific use case. For

example, you can fine-tune this model to detect the depth of a photo like this. Or here's another example. or instead of

depth you can also train this to become an upscaler. So here's the result or here's another example. The awesome

thing is this is open source. So at the top if you click on this code button here it contains the instructions on how

to download and run this locally on your computer. They also released the training script to this including the

data. So this is fully open source and the models are under the Apache 2 license which is very permissive. Now at

3 billion parameters this is fairly tiny. So, the base model is only 12 GB in size, so you should be able to fit

this on most consumer GPUs. If you're interested in reading further, I'll link to this main page in the description

below. Also, this week, we have a really cool AI system called Terra. As you can see from this demo, it basically teaches

simulated human bodies to move across challenging terrain using realistic bones and muscles. Now, most virtual

characters and robots are controlled by directly moving their joints, but that's not how actual human bodies work, right?

In real life, our muscles contract and pull on our bones to produce movement. So, that's what this project is trying

to replicate. Can they simulate the actual contractions and movements of human muscles and skeletons and teach

them to perform complex movements like moving across an uneven surface or walking upstairs? So after training the

AI on like over 9 hours of motion data with 18 demonstrated movements across different terrains, this general muscle

model is now able to, you know, correctly move this simulated human body across these difficult surfaces like

slopes and steps. And not only that, but you can then extend this further by using some specialized fine-tuned

policies to get it to do some crazier stuff like jumping and rolling. Now, at the top, they released the GitHub to

this. So, if you click on this link and you scroll down a bit here, it contains all the instructions on how to download

and run this locally on your computer. And note that this is under the Apache 2 license, which has very minimal

restrictions. If you're interested in reading further, I'll link to this main page in the description below. Also,

this week, we have an AI called Code Dim Recon. This takes normal photos of a real environment and turns them into an

editable 3D scene that's actually ready for physics simulation. And it also includes the physical properties of the

object. So it can make things that can bend, stretch, or be squished. How this works is that the AI first reconstructs

the overall geometry of the scene and then treats different deformable objects differently. Something thin like a cable

becomes a flexible 3D curve. Something like a piece of paper or plastic bag becomes a flexible surface, while

something thick like a cushion usually becomes a deformable 3D solid. You can then run robot interactions to test

whether these objects behave correctly and revise them if something looks wrong. And so far, this is the only 3D

reconstruction tool that includes all these capabilities. Now, at the top here, it says the code is coming soon,

so stay tuned for that. For now, if you are interested in learning more, I'll link to this main page in the

description below. Also this week we have another AI called dist scene which can take a single image of an

environment and turn it into an entire 3D scene including both the surrounding environment and the individual objects

inside it. Now here's how it works. Instead of simply generating a bunch of isolated 3D objects and trying to place

them together afterwards. This AI generates the environment and all the objects in one shared 3D coordinate

system. It then pulls out each object, refineses its geometry individually, and puts it back into the correct position

in the scene. Now, if you look at these benchmarks in terms of geometry, consistency, and quality, you can see

that across the board, this new disc scene method performs the best. Now, at the top of the page, it says they are

going to release the model and the data set to this. So, stay tuned for that. For now, if you're interested, I'll link

to this main page in the description below. Also this week, Google quietly released Nano Banana 2.1. Now, this is

just a 0.1 upgrade, so it's not like a huge improvement over the previous version. Here are some examples for your

reference. So, they claim that this is an upgraded version designed to generate and edit images more accurately,

especially when you're dealing with complicated instructions, multiple characters, and text inside images. So,

it's really good at like creating infographics or diagrams or microediting certain stuff with natural language.

They say that Nano Banana 2.1 supports up to 14 reference images. It also has improved image realism text rendering

and it also supports up to 4K resolution. Now, after some initial tests, I would say the quality isn't

great. Still noticeably behind the best image model out there, GPT Image 2.5. But the nice thing is you can try this

out for free. So, in Google's Flow platform, which I'll link to in the description below, here's where you can

try out Nano Banana 2.1. So, once you sign up for an account here, you can use the latest Nano Banana 2.1 to generate

images, and they're quite generous. You get like a few dozen free generations per day. If you're interested, I'll link

to this release page in the description below. Also, this week, Anthropic drops their latest model, Claude Haiku 5.5.

Now, Haiku is the smallest and fastest variant whereas on the other end, Opus is the largest and the most intelligent.

So, here's its performance on various benchmarks on knowledge work, agentic coding, computer use, and so on. As you

can see, it's a huge improvement compared to the previous version of Haiku. And for most of these benchmarks,

it even beats GPT6 Luna, which is like OpenAI's Haiku equivalent. Now, here's the performance of Sonnet 5.5 for your

reference. In general, Haiku is a dumber model, but it's way faster and it costs a lot less. So, here are various

benchmarks for your reference. And as you can see, Haiku 5.5 in pink in general costs a lot less than sonnet 5.5

at just slightly lower intelligence. And then here's a knowledge work benchmark. If you look at this intelligence index

by artificial analysis, then you can see that haiku 5.5 is a bit dumber. So it's ranked all the way over here behind

Kimik 3 and step 5 and GLM 5.3. So it's not meant to be frontier in terms of intelligence. However, if you look at

the cost of this, here's where it becomes really attractive. It only costs around 21 cents per task according to

artificial analysis. Even lower than GLM 5.3 and Deep Seek Flash, but still not as cheap as GPT6 Luna, which is like

three times cheaper. Now, if you look at intelligence versus cost, then haiku is not the most efficient option. Xiaomi's

MIMO 2.6 is still king and then GPG6 Luna is also really good. Now, the nice thing about Claude Haiku 5.5 is it's

available even on the free plan. So, you can select Haiku 5.5 for the fastest response. If you're interested in

reading further, I'll link to this main page in the description below. And that sums up all the highlights in AI this

week. Let me know in the comments what you think of all of this. Which piece of news was your favorite? And which tool

are you most looking forward to trying out? As always, I will be on the lookout for the top AI news and tools to share

with you. So, if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content. Also,

there's just so much happening in the world of AI every week, I can't possibly cover everything on my YouTube channel.

So, to really stay uptodate with all that's going on in AI, be sure to subscribe to my free weekly newsletter.

The link to that will be in the description below. Thanks for watching and I'll see you in the next one.