May 26, 2025 • 21 mins
Introducing Claude 4
AI Talent War in Silicon Valley Intensifies with Lucrative Offers and Creative Strategies
Jony Ive and Sam Altman Developing Compact AI Device for 2026 Launch
Canvas Just Became Google's Most Exciting AI Playground
New capabilities for building agents on the Anthropic API
Google's Gemini Diffusion: Fast and Accurate Language Model Using Diffusion Techniques
The AI That Codes Itself
#AI, #Claude4, #SiliconValley, #JonyIve, #Anthropic, #GoogleGemini, #SelfCodingAI
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
Welcome to Innovation Pulse, your quick no-nonsense update on the latest in AI.
(00:09):
First, we will cover the latest news.
Anthropicon Veil's advanced AI models, fierce competition for Silicon Valley talent,
Johnny Ive and Sam Altman's AI device, and Google's upgraded Canvas platform.
After this, we'll dive deep into Google DeepMind's revolutionary Alpha Evolve system.
(00:32):
Today, we're unveiling Claude Opus 4 and Claude Sonnet 4, the next generation of Claude models.
Claude Opus 4 is the best coding model globally, excelling in complex long-running tasks,
while Claude Sonnet 4 improves on its predecessor with advanced coding and reasoning capabilities.
Both models can now use tools like web search during extended thinking, enhancing their responses.
(00:59):
They support parallel tool execution and improved memory functions,
storing key facts to build knowledge over time. Claude Code, now widely available, integrates
into IDEs like VS Code and JetBrains, offering seamless coding collaboration.
Additionally, new API capabilities enhance developers' ability to build robust AI agents.
(01:24):
These models, available on the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI,
represent a significant leap in AI, supporting diverse applications from research to everyday tasks.
Pricing remains consistent with previous models.
For now, let's focus on the recruitment strategies in Silicon Valley.
(01:49):
The race for AI dominance is intensifying, focusing on superstar researchers.
Since the launch of ChatGPT in 2022, recruiting top AI talent has become fiercely competitive.
Companies like OpenAI and Google offer massive compensation packages,
sometimes over $10 million annually, to attract and retain these researchers.
(02:14):
The scarcity of elite AI talent, often called 10,000 times engineers, has led to innovative
hiring strategies, including using sports data analysis to find promising candidates.
Despite lucrative offers, some researchers, like Noam Brown, prioritize meaningful work
over financial gain. The recent departure of OpenAI's Mira Murati to start her own AI venture
(02:39):
has further heated the talent competition. As AI rapidly evolves, companies are seeking talent
from diverse fields, such as theoretical physics and quantum computing, to fuel advancements.
Details are emerging about Johnny Ive and Sam Altman's new AI device.
(02:59):
Analyst Ming-Chi Kuo suggests it will be slightly larger than humane's AI pin,
but as compact and elegant as an iPod shuffle. The device could be worn around the neck and
might lack a display, instead featuring built-in cameras and microphones for environmental detection.
It may connect to smartphones and PCs for computing and display functions.
(03:22):
This aligns with reports indicating the device will be aware of users' surroundings,
but isn't likely to be a pair of glasses. Recently, OpenAI announced purchasing Ive's AI
hardware company, IO, for $6.5 billion. This acquisition will influence the design of
OpenAI's software and hardware. Ive and Altman plan to launch their first devices in 2026.
(03:49):
Google has upgraded its Canvas platform with the Gemini 2.5 models,
transforming it into a powerful creation tool. Canvas now allows users to convert documents
and code into webpages, infographics, quizzes and podcasts in 45 languages.
(04:10):
The standout feature, Vibe Coding, lets users describe ideas to generate code for apps and games
without writing a single line. This makes app development accessible to anyone with an idea.
During a live demo, Tulsi Doshi from Google turned a sketch into a 3D web app in minutes
(04:31):
using Gemini. While there are some limitations, like workspace users not being able to share
content, the platform promises a new way to bring digital ideas to life. This innovation aims to
lower the barrier to creating software, allowing for rapid prototyping and interactive learning
experiences. Join us as we discuss the new API enhancements. Anthropoc has unveiled four new
(04:59):
API capabilities to enhance AI agent development, a code execution tool, an MCP connector,
a files API and extended prompt caching. These features, alongside Claude Opus 4 and Sonnet 4,
allow agents to execute complex code for data analysis, connect to external systems via MCP
(05:22):
servers and manage files efficiently. Developers can also maintain context for up to 60 minutes,
cutting costs significantly. For instance, a project management AI can integrate with
Asana to handle tasks, analyze progress and maintain context using these new tools.
The code execution tool enables Claude to run Python code for financial modeling,
(05:47):
scientific computing and more. The MCP connector simplifies external tool integration,
while the files API streamlines document handling. Extended caching reduces expenses
and latency for long workflows. These tools provide a robust set for building advanced AI applications.
(06:11):
Google recently unveiled GeminiDiffusion, its first language model that uses diffusion techniques
instead of traditional auto-regressive methods. Unlike traditional models that generate text
sequentially, diffusion models refine noise step by step, allowing for rapid iteration and error
correction during generation. This enhances tasks like editing, particularly in math and code.
(06:37):
The key advantage is speed, as demonstrated when tasked with building a chat app, responding at 857
tokens per second. While independent benchmarks are not yet available, Google's landing page
claims it matches the performance of Gemini 2.0 flashlight at five times the speed.
(06:58):
Unlike image diffusion models, diffusion language models resemble BERT, using a transformer to
predict masked tokens in parallel. This method allows for generating sequences by gradually
reducing the masked tokens, refining outputs efficiently. And now, pivot our discussion
(07:18):
towards the main AI topic. Alright everybody, welcome back to Innovation Pulse. I'm Donna,
and today we're diving into something that honestly sounds like science fiction,
but is happening right now in Google's labs. We're talking about an AI system that doesn't just
(07:41):
solve problems, it literally invents new ways to solve them. That's right Donna, I'm Yakov Lasker,
and what we're looking at today, Google DeepMind's latest creation called Alpha Evolve.
And here's the kicker, this isn't just some research project sitting in a lab somewhere,
this AI has already been quietly running Google's data centers for over a year,
(08:01):
making them more efficient. And it just broke a mathematical record that had stood since 1969.
Wait, hold up. Since 1969, we're talking about something that mathematicians have been working
on for over half a century, and an AI just figured it out? Exactly, so let me paint the picture here.
You know how we've seen AI get really good at specific tasks, right? Like AlphaGo,
(08:25):
mastering the game of Go, or AlphaFold solving protein structures. Those were incredible,
but they were specialists, one trick ponies if you will. Right, like having a chess grandmaster who
can only play chess. Perfect analogy. But Alpha Evolve is different. Think of it as a mathematician,
computer scientist, and engineer, all rolled into one. Except it works at superhuman speed,
(08:49):
and never gets tired. The system combines Google's Gemini language models with what they call an
evolutionary approach. Okay, so break that down for me. What does evolutionary mean in this context?
Are we talking about digital natural selection? That's actually not a bad way to think about it.
Here's how it works. Alpha Evolve takes a problem and generates hundreds or thousands of potential
(09:11):
solutions. Think of these as different species of algorithms. Then it tests each one, sees which
ones perform best, and uses those winners to create the next generation of solutions. The weak ones die
off. The strong ones reproduce and mutate, creating even better solutions. So it's basically running
its own algorithmic survival of the fittest. That's wild. But I'm curious about the practical side.
(09:36):
You mentioned it's already running Google's data centers. Here's where it gets really impressive.
Alpha Evolve discovered a new scheduling algorithm for something called Borg.
That's Google's system for managing millions of servers worldwide. This one algorithm changed
freed up 0.7% of Google's total computing resources. That doesn't sound like much. Oh,
(09:57):
but it is at Google's scale. 0.7% represents hundreds of thousands of machines. We're talking
about millions of dollars in efficiency gains, running continuously 24 seven. It's like finding
free money in the couch cushions, except the couch is the size of a small country. Okay,
now I'm paying attention. What other problems has it tackled? This is where it gets really fun.
(10:19):
Remember that 56 year old mathematical record I mentioned? It involves something called
matrix multiplication. Basically a fundamental operation that underlies almost all of modern
computing, especially AI training. Matrix multiplication. That sounds like something
from high school math class that I tried to forget. Lay it on me. What exactly did it discover?
(10:40):
So back in 1969, a German mathematician named Volker Strassen figured out a clever way to
multiply certain types of matrices faster than anyone thought possible. His method became the
gold standard. And for 56 years, nobody could beat it. Then alpha evolve comes along and finds a way
to do it with one fewer operation. One fewer operation doesn't sound revolutionary. But that's
(11:06):
the thing about mathematics. Sometimes the smallest improvements have huge implications.
When you're multiplying matrices billions of times to train AI models,
that one operation saved adds up. Alpha evolve actually used this discovery to speed up the
training of Google's own Gemini models by 23% for that specific operation, which translated to a
(11:30):
1% overall training time reduction. So it's literally making itself smarter, faster. Exactly.
It's optimizing the very systems that power itself. There's something almost poetic about that.
But here's what really blows my mind. When they tested alpha evolve against DeepMind's previous
specialized system called Alpha Tenser, which was built specifically for matrix multiplication,
(11:56):
alpha evolve actually outperformed it. A general purpose system beat a specialist at its own game.
That's like a decathlete beating a sprinter in the 100 meter dash. How is that even possible?
It comes down to the evolutionary approach. While Alpha Tenser was constrained by its specific
training, alpha evolve could explore solution spaces that humans might never have thought to look at.
(12:19):
It's not bound by our preconceptions about how problems should be solved.
This reminds me of those stories about AI finding unexpected strategies in games.
Like when Alpha Go made moves that professional players thought were mistakes,
until they realized they were genius. Perfect comparison. And speaking of unexpected discoveries,
Alpha evolve tackled something called the kissing number problem.
(12:42):
The what now? I know, I know. It sounds like middle school gossip, but it's actually a serious
mathematical challenge. Imagine you have a sphere and you want to surround it with as many other
identical spheres as possible, all touching the center sphere, but not overlapping each other.
Okay, I can visualize that like oranges stacked at a grocery store.
(13:02):
Exactly. In two dimensions, think circles on a flat surface. You can fit six circles around
one central circle. In three dimensions, it gets more complex and the number is 12.
But as you add more dimensions, the problem becomes incredibly difficult. In 11 dimensions,
the best known answer was 592 spheres could touch the central one.
(13:26):
And let me guess, Alpha evolve found 593?
Bingo. Now, that might seem like a tiny improvement, but in mathematics,
proving that even one more as possible can open up entirely new areas of research.
It's not just about the number, it's about the method used to find it.
So we've got data center optimization, matrix multiplication,
(13:47):
throughs and abstract mathematical discoveries. What else has this thing been up to?
Well, it's also redesigning computer chips. Alpha evolve proposed changes to Google's
tensor processing units. They're specialized AI chips. It found ways to eliminate unnecessary
operations in the chip's arithmetic circuits, making them more efficient without compromising
(14:09):
accuracy. Hold on. So this AI is now designing the hardware that will run future AIs?
You're starting to see the bigger picture. We're looking at a system that's optimizing
everything from the physical chips to the software algorithms to the data center management,
creating a feedback loop of continuous improvement.
That's simultaneously exciting and a little terrifying. How does this actually work under
(14:32):
the hood? Great question. Alpha evolve uses two different versions of Google's Gemini model working
together. Gemini flash handles the breadth, generating lots of ideas quickly. Gemini Pro
provides the depth, offering more sophisticated solutions when needed. But here's the crucial
part. It's not just generating random code and hoping for the best.
(14:57):
Right, because we know AI can hallucinate and make stuff up.
Exactly. That's where the evolutionary framework comes in. Every solution Alpha evolve proposes
gets automatically tested and scored. Can it solve the problem correctly? How fast does it run?
How elegant is the solution? Only the solutions that pass these objective tests survive to the
(15:20):
next generation. So it's like having a ruthless editor that immediately fact checks every idea
and throws out anything that doesn't work. Perfect analogy. And because the evaluation is
automated, Alpha evolve can test thousands of solutions in parallel, something human researchers
could never do at that scale or speed. Now I'm curious about the limitations. This sounds almost
(15:41):
too good to be true. You're right to be skeptical. Alpha evolve works best on problems that can be
clearly defined and automatically evaluated. It excels in domains like mathematics, computer
science and system optimization, where you can objectively measure success. But it's not going
to write poetry or solve social problems that require human judgment and nuance.
(16:06):
Makes sense. It needs clear rules of the game to play effectively.
Exactly. And there's another important point. While Alpha evolve is more general than previous
systems, it still requires human experts to set it up properly, interpret its results,
and decide which solutions to implement. It's augmenting human intelligence, not replacing it.
(16:27):
Speaking of humans, what does this mean for researchers and engineers? Should they be
worried about job security? Here's the thing. The researchers at DeepMind emphasize that this
is about collaboration, not replacement. Matei Balog, one of the lead researchers,
mentioned that traditionally, you couldn't build a scientific tool and immediately see real world
(16:50):
impact at this scale. Alpha evolve is freeing up human experts to focus on bigger picture problems
while handling the tedious optimization work. So it's more like having a super powered research
assistant? Right. And Google is already planning to share this capability. They're developing a
user interface and launching an early access program for academic researchers. The goal is to
(17:15):
democratize access to this kind of algorithmic discovery. That could be huge for universities
and smaller research institutions that don't have Google's resources. Absolutely. Imagine a
graduate student being able to tackle optimization problems that would have taken months of manual
work or a researcher in material science using alpha evolve to discover new algorithms for
(17:36):
molecular simulation. You mentioned material science. Where else might we see this technology
applied? The researchers mentioned drug discovery as a major target. Think about it. Drug development
involves massive optimization problems. Which molecular structures might work? How do you
optimize for effectiveness while minimizing side effects? These are exactly the kinds of
(17:59):
problems alpha evolve could help solve. And probably much faster than traditional methods.
Exactly. We could be looking at accelerated timelines for everything from new medications
to more efficient solar panels to better battery designs. Any field that relies on complex algorithms
could benefit. This feels like one of those moments where we're witnessing a fundamental
(18:21):
shift in how science gets done. I think you're right. We're seeing AI move from being a tool
that executes human design solutions to one that discovers entirely new solutions. And here's the
kicker. As language models continue to improve, alpha evolves capabilities will grow alongside them.
It's like we're watching the early stages of AI assisted scientific discovery becoming mainstream.
(18:45):
And the timing is crucial. We're facing massive challenges. Climate change, energy efficiency,
computational demands of AI itself that require novel solutions. Having a system that can explore
solution spaces humans might never consider could be exactly what we need. So what's the bottom line
(19:06):
here? What should our listeners take away from this? The big takeaway is that we're entering an
era where AI isn't just getting better at existing tasks. It's beginning to discover new ways to
approach fundamental problems. Alpha evolve represents a shift from narrow AI to more general
problem solving systems that can augment human creativity and expertise. And the practical
(19:29):
impact is already being felt. Google's data centers are running more efficiently, their chips are
being redesigned, and mathematical problems that stumped researchers for decades are being solved.
Right. This isn't some far future concept. It's happening now. And the implications are enormous.
Whether you're a researcher, an engineer, or just someone interested in how technology shapes our
(19:52):
world, Alpha evolve shows us what's possible when we combine human insight with AI's ability to explore
vast solution spaces. For anyone working on optimization problems or algorithmic challenges,
this might be worth keeping an eye on, especially as Google rolls out broader access.
Absolutely. And even if you're not directly involved in these fields, understanding how AI is
(20:15):
evolving, pun intended, helps us all prepare for a world where human AI collaboration becomes the
norm rather than the exception. Well, this has been fascinating, Yaakov. Thanks for walking us
through what might be one of the most significant AI developments we've seen this year. Always a
pleasure, Donna. The pace of innovation in this space continues to amaze me. And to our listeners,
(20:38):
this is exactly the kind of breakthrough that reminds us why we do this show, to catch these
pivotal moments as they happen and help make sense of what they mean for all of us. Thanks for
tuning into Innovation Pulse. And until next time, keep your eyes on the future. Because apparently,
the future is busy writing its own code. I love it. See you next time, everyone.
(21:04):
We've explored Anthropics innovative AI models, enhancing coding and reasoning skills,
and Google's advancements with Gemini and DeepMind's Alpha Evolve, pushing the boundaries of digital
content creation and problem solving. Don't forget to like, subscribe, and share this episode with
your friends and colleagues so they can also stay updated on the latest news and gain powerful
(21:28):
insights. Stay tuned for more updates.
Welcome to Innovation Pulse, your quick no-nonsense update on the latest in AI.
(00:09):
First, we will cover the latest news.
Anthropicon Veil's advanced AI models, fierce competition for Silicon Valley talent,
Johnny Ive and Sam Altman's AI device, and Google's upgraded Canvas platform.
After this, we'll dive deep into Google DeepMind's revolutionary Alpha Evolve system.
(00:32):
Today, we're unveiling Claude Opus 4 and Claude Sonnet 4, the next generation of Claude models.
Claude Opus 4 is the best coding model globally, excelling in complex long-running tasks,
while Claude Sonnet 4 improves on its predecessor with advanced coding and reasoning capabilities.
Both models can now use tools like web search during extended thinking, enhancing their responses.
(00:59):
They support parallel tool execution and improved memory functions,
storing key facts to build knowledge over time. Claude Code, now widely available, integrates
into IDEs like VS Code and JetBrains, offering seamless coding collaboration.
Additionally, new API capabilities enhance developers' ability to build robust AI agents.
(01:24):
These models, available on the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI,
represent a significant leap in AI, supporting diverse applications from research to everyday tasks.
Pricing remains consistent with previous models.
For now, let's focus on the recruitment strategies in Silicon Valley.
(01:49):
The race for AI dominance is intensifying, focusing on superstar researchers.
Since the launch of ChatGPT in 2022, recruiting top AI talent has become fiercely competitive.
Companies like OpenAI and Google offer massive compensation packages,
sometimes over $10 million annually, to attract and retain these researchers.
(02:14):
The scarcity of elite AI talent, often called 10,000 times engineers, has led to innovative
hiring strategies, including using sports data analysis to find promising candidates.
Despite lucrative offers, some researchers, like Noam Brown, prioritize meaningful work
over financial gain. The recent departure of OpenAI's Mira Murati to start her own AI venture
(02:39):
has further heated the talent competition. As AI rapidly evolves, companies are seeking talent
from diverse fields, such as theoretical physics and quantum computing, to fuel advancements.
Details are emerging about Johnny Ive and Sam Altman's new AI device.
(02:59):
Analyst Ming-Chi Kuo suggests it will be slightly larger than humane's AI pin,
but as compact and elegant as an iPod shuffle. The device could be worn around the neck and
might lack a display, instead featuring built-in cameras and microphones for environmental detection.
It may connect to smartphones and PCs for computing and display functions.
(03:22):
This aligns with reports indicating the device will be aware of users' surroundings,
but isn't likely to be a pair of glasses. Recently, OpenAI announced purchasing Ive's AI
hardware company, IO, for $6.5 billion. This acquisition will influence the design of
OpenAI's software and hardware. Ive and Altman plan to launch their first devices in 2026.
(03:49):
Google has upgraded its Canvas platform with the Gemini 2.5 models,
transforming it into a powerful creation tool. Canvas now allows users to convert documents
and code into webpages, infographics, quizzes and podcasts in 45 languages.
(04:10):
The standout feature, Vibe Coding, lets users describe ideas to generate code for apps and games
without writing a single line. This makes app development accessible to anyone with an idea.
During a live demo, Tulsi Doshi from Google turned a sketch into a 3D web app in minutes
(04:31):
using Gemini. While there are some limitations, like workspace users not being able to share
content, the platform promises a new way to bring digital ideas to life. This innovation aims to
lower the barrier to creating software, allowing for rapid prototyping and interactive learning
experiences. Join us as we discuss the new API enhancements. Anthropoc has unveiled four new
(04:59):
API capabilities to enhance AI agent development, a code execution tool, an MCP connector,
a files API and extended prompt caching. These features, alongside Claude Opus 4 and Sonnet 4,
allow agents to execute complex code for data analysis, connect to external systems via MCP
(05:22):
servers and manage files efficiently. Developers can also maintain context for up to 60 minutes,
cutting costs significantly. For instance, a project management AI can integrate with
Asana to handle tasks, analyze progress and maintain context using these new tools.
The code execution tool enables Claude to run Python code for financial modeling,
(05:47):
scientific computing and more. The MCP connector simplifies external tool integration,
while the files API streamlines document handling. Extended caching reduces expenses
and latency for long workflows. These tools provide a robust set for building advanced AI applications.
(06:11):
Google recently unveiled GeminiDiffusion, its first language model that uses diffusion techniques
instead of traditional auto-regressive methods. Unlike traditional models that generate text
sequentially, diffusion models refine noise step by step, allowing for rapid iteration and error
correction during generation. This enhances tasks like editing, particularly in math and code.
(06:37):
The key advantage is speed, as demonstrated when tasked with building a chat app, responding at 857
tokens per second. While independent benchmarks are not yet available, Google's landing page
claims it matches the performance of Gemini 2.0 flashlight at five times the speed.
(06:58):
Unlike image diffusion models, diffusion language models resemble BERT, using a transformer to
predict masked tokens in parallel. This method allows for generating sequences by gradually
reducing the masked tokens, refining outputs efficiently. And now, pivot our discussion
(07:18):
towards the main AI topic. Alright everybody, welcome back to Innovation Pulse. I'm Donna,
and today we're diving into something that honestly sounds like science fiction,
but is happening right now in Google's labs. We're talking about an AI system that doesn't just
(07:41):
solve problems, it literally invents new ways to solve them. That's right Donna, I'm Yakov Lasker,
and what we're looking at today, Google DeepMind's latest creation called Alpha Evolve.
And here's the kicker, this isn't just some research project sitting in a lab somewhere,
this AI has already been quietly running Google's data centers for over a year,
(08:01):
making them more efficient. And it just broke a mathematical record that had stood since 1969.
Wait, hold up. Since 1969, we're talking about something that mathematicians have been working
on for over half a century, and an AI just figured it out? Exactly, so let me paint the picture here.
You know how we've seen AI get really good at specific tasks, right? Like AlphaGo,
(08:25):
mastering the game of Go, or AlphaFold solving protein structures. Those were incredible,
but they were specialists, one trick ponies if you will. Right, like having a chess grandmaster who
can only play chess. Perfect analogy. But Alpha Evolve is different. Think of it as a mathematician,
computer scientist, and engineer, all rolled into one. Except it works at superhuman speed,
(08:49):
and never gets tired. The system combines Google's Gemini language models with what they call an
evolutionary approach. Okay, so break that down for me. What does evolutionary mean in this context?
Are we talking about digital natural selection? That's actually not a bad way to think about it.
Here's how it works. Alpha Evolve takes a problem and generates hundreds or thousands of potential
(09:11):
solutions. Think of these as different species of algorithms. Then it tests each one, sees which
ones perform best, and uses those winners to create the next generation of solutions. The weak ones die
off. The strong ones reproduce and mutate, creating even better solutions. So it's basically running
its own algorithmic survival of the fittest. That's wild. But I'm curious about the practical side.
(09:36):
You mentioned it's already running Google's data centers. Here's where it gets really impressive.
Alpha Evolve discovered a new scheduling algorithm for something called Borg.
That's Google's system for managing millions of servers worldwide. This one algorithm changed
freed up 0.7% of Google's total computing resources. That doesn't sound like much. Oh,
(09:57):
but it is at Google's scale. 0.7% represents hundreds of thousands of machines. We're talking
about millions of dollars in efficiency gains, running continuously 24 seven. It's like finding
free money in the couch cushions, except the couch is the size of a small country. Okay,
now I'm paying attention. What other problems has it tackled? This is where it gets really fun.
(10:19):
Remember that 56 year old mathematical record I mentioned? It involves something called
matrix multiplication. Basically a fundamental operation that underlies almost all of modern
computing, especially AI training. Matrix multiplication. That sounds like something
from high school math class that I tried to forget. Lay it on me. What exactly did it discover?
(10:40):
So back in 1969, a German mathematician named Volker Strassen figured out a clever way to
multiply certain types of matrices faster than anyone thought possible. His method became the
gold standard. And for 56 years, nobody could beat it. Then alpha evolve comes along and finds a way
to do it with one fewer operation. One fewer operation doesn't sound revolutionary. But that's
(11:06):
the thing about mathematics. Sometimes the smallest improvements have huge implications.
When you're multiplying matrices billions of times to train AI models,
that one operation saved adds up. Alpha evolve actually used this discovery to speed up the
training of Google's own Gemini models by 23% for that specific operation, which translated to a
(11:30):
1% overall training time reduction. So it's literally making itself smarter, faster. Exactly.
It's optimizing the very systems that power itself. There's something almost poetic about that.
But here's what really blows my mind. When they tested alpha evolve against DeepMind's previous
specialized system called Alpha Tenser, which was built specifically for matrix multiplication,
(11:56):
alpha evolve actually outperformed it. A general purpose system beat a specialist at its own game.
That's like a decathlete beating a sprinter in the 100 meter dash. How is that even possible?
It comes down to the evolutionary approach. While Alpha Tenser was constrained by its specific
training, alpha evolve could explore solution spaces that humans might never have thought to look at.
(12:19):
It's not bound by our preconceptions about how problems should be solved.
This reminds me of those stories about AI finding unexpected strategies in games.
Like when Alpha Go made moves that professional players thought were mistakes,
until they realized they were genius. Perfect comparison. And speaking of unexpected discoveries,
Alpha evolve tackled something called the kissing number problem.
(12:42):
The what now? I know, I know. It sounds like middle school gossip, but it's actually a serious
mathematical challenge. Imagine you have a sphere and you want to surround it with as many other
identical spheres as possible, all touching the center sphere, but not overlapping each other.
Okay, I can visualize that like oranges stacked at a grocery store.
(13:02):
Exactly. In two dimensions, think circles on a flat surface. You can fit six circles around
one central circle. In three dimensions, it gets more complex and the number is 12.
But as you add more dimensions, the problem becomes incredibly difficult. In 11 dimensions,
the best known answer was 592 spheres could touch the central one.
(13:26):
And let me guess, Alpha evolve found 593?
Bingo. Now, that might seem like a tiny improvement, but in mathematics,
proving that even one more as possible can open up entirely new areas of research.
It's not just about the number, it's about the method used to find it.
So we've got data center optimization, matrix multiplication,
(13:47):
throughs and abstract mathematical discoveries. What else has this thing been up to?
Well, it's also redesigning computer chips. Alpha evolve proposed changes to Google's
tensor processing units. They're specialized AI chips. It found ways to eliminate unnecessary
operations in the chip's arithmetic circuits, making them more efficient without compromising
(14:09):
accuracy. Hold on. So this AI is now designing the hardware that will run future AIs?
You're starting to see the bigger picture. We're looking at a system that's optimizing
everything from the physical chips to the software algorithms to the data center management,
creating a feedback loop of continuous improvement.
That's simultaneously exciting and a little terrifying. How does this actually work under
(14:32):
the hood? Great question. Alpha evolve uses two different versions of Google's Gemini model working
together. Gemini flash handles the breadth, generating lots of ideas quickly. Gemini Pro
provides the depth, offering more sophisticated solutions when needed. But here's the crucial
part. It's not just generating random code and hoping for the best.
(14:57):
Right, because we know AI can hallucinate and make stuff up.
Exactly. That's where the evolutionary framework comes in. Every solution Alpha evolve proposes
gets automatically tested and scored. Can it solve the problem correctly? How fast does it run?
How elegant is the solution? Only the solutions that pass these objective tests survive to the
(15:20):
next generation. So it's like having a ruthless editor that immediately fact checks every idea
and throws out anything that doesn't work. Perfect analogy. And because the evaluation is
automated, Alpha evolve can test thousands of solutions in parallel, something human researchers
could never do at that scale or speed. Now I'm curious about the limitations. This sounds almost
(15:41):
too good to be true. You're right to be skeptical. Alpha evolve works best on problems that can be
clearly defined and automatically evaluated. It excels in domains like mathematics, computer
science and system optimization, where you can objectively measure success. But it's not going
to write poetry or solve social problems that require human judgment and nuance.
(16:06):
Makes sense. It needs clear rules of the game to play effectively.
Exactly. And there's another important point. While Alpha evolve is more general than previous
systems, it still requires human experts to set it up properly, interpret its results,
and decide which solutions to implement. It's augmenting human intelligence, not replacing it.
(16:27):
Speaking of humans, what does this mean for researchers and engineers? Should they be
worried about job security? Here's the thing. The researchers at DeepMind emphasize that this
is about collaboration, not replacement. Matei Balog, one of the lead researchers,
mentioned that traditionally, you couldn't build a scientific tool and immediately see real world
(16:50):
impact at this scale. Alpha evolve is freeing up human experts to focus on bigger picture problems
while handling the tedious optimization work. So it's more like having a super powered research
assistant? Right. And Google is already planning to share this capability. They're developing a
user interface and launching an early access program for academic researchers. The goal is to
(17:15):
democratize access to this kind of algorithmic discovery. That could be huge for universities
and smaller research institutions that don't have Google's resources. Absolutely. Imagine a
graduate student being able to tackle optimization problems that would have taken months of manual
work or a researcher in material science using alpha evolve to discover new algorithms for
(17:36):
molecular simulation. You mentioned material science. Where else might we see this technology
applied? The researchers mentioned drug discovery as a major target. Think about it. Drug development
involves massive optimization problems. Which molecular structures might work? How do you
optimize for effectiveness while minimizing side effects? These are exactly the kinds of
(17:59):
problems alpha evolve could help solve. And probably much faster than traditional methods.
Exactly. We could be looking at accelerated timelines for everything from new medications
to more efficient solar panels to better battery designs. Any field that relies on complex algorithms
could benefit. This feels like one of those moments where we're witnessing a fundamental
(18:21):
shift in how science gets done. I think you're right. We're seeing AI move from being a tool
that executes human design solutions to one that discovers entirely new solutions. And here's the
kicker. As language models continue to improve, alpha evolves capabilities will grow alongside them.
It's like we're watching the early stages of AI assisted scientific discovery becoming mainstream.
(18:45):
And the timing is crucial. We're facing massive challenges. Climate change, energy efficiency,
computational demands of AI itself that require novel solutions. Having a system that can explore
solution spaces humans might never consider could be exactly what we need. So what's the bottom line
(19:06):
here? What should our listeners take away from this? The big takeaway is that we're entering an
era where AI isn't just getting better at existing tasks. It's beginning to discover new ways to
approach fundamental problems. Alpha evolve represents a shift from narrow AI to more general
problem solving systems that can augment human creativity and expertise. And the practical
(19:29):
impact is already being felt. Google's data centers are running more efficiently, their chips are
being redesigned, and mathematical problems that stumped researchers for decades are being solved.
Right. This isn't some far future concept. It's happening now. And the implications are enormous.
Whether you're a researcher, an engineer, or just someone interested in how technology shapes our
(19:52):
world, Alpha evolve shows us what's possible when we combine human insight with AI's ability to explore
vast solution spaces. For anyone working on optimization problems or algorithmic challenges,
this might be worth keeping an eye on, especially as Google rolls out broader access.
Absolutely. And even if you're not directly involved in these fields, understanding how AI is
(20:15):
evolving, pun intended, helps us all prepare for a world where human AI collaboration becomes the
norm rather than the exception. Well, this has been fascinating, Yaakov. Thanks for walking us
through what might be one of the most significant AI developments we've seen this year. Always a
pleasure, Donna. The pace of innovation in this space continues to amaze me. And to our listeners,
(20:38):
this is exactly the kind of breakthrough that reminds us why we do this show, to catch these
pivotal moments as they happen and help make sense of what they mean for all of us. Thanks for
tuning into Innovation Pulse. And until next time, keep your eyes on the future. Because apparently,
the future is busy writing its own code. I love it. See you next time, everyone.
(21:04):
We've explored Anthropics innovative AI models, enhancing coding and reasoning skills,
and Google's advancements with Gemini and DeepMind's Alpha Evolve, pushing the boundaries of digital
content creation and problem solving. Don't forget to like, subscribe, and share this episode with
your friends and colleagues so they can also stay updated on the latest news and gain powerful
(21:28):
insights. Stay tuned for more updates.
Popular Podcasts
Boysober
Have you ever wondered what life might be like if you stopped worrying about being wanted, and focused on understanding what you actually want? That was the question Hope Woodard asked herself after a string of situationships inspired her to take a break from sex and dating. She went "boysober," a personal concept that sparked a global movement among women looking to prioritize themselves over men. Now, Hope is looking to expand the ways we explore our relationship to relationships. Taking a bold, unfiltered look into modern love, romance, and self-discovery, Boysober will dive into messy stories about dating, sex, love, friendship, and breaking generational patterns—all with humor, vulnerability, and a fresh perspective.
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com