Google just dropped super-fast AI chips built for helpful agents that actually do stuff for you.
What's Cool Today: Google unveiled its eighth-generation TPUs — specialized chips designed from the ground up to handle both training new AI models and powering “AI agents” that can take real actions like booking trips or managing schedules. These chips make training dramatically faster and let agents think through multiple steps without slowing down or costing a fortune. It’s a huge step toward AI that doesn’t just chat but actually helps with everyday tasks. Today we’ll also look at how AI can solve real biology puzzles that stump experts, plus a hilarious Chrome extension that might be the ultimate procrastination stopper.
The Big Story
Google just announced its newest specialized computer chips called eighth-generation TPUs at their big cloud event last week. These chips were built specifically for what people are calling the “agentic era” of AI — that’s when AI stops being just a clever chatbot and starts acting like a helpful assistant that can actually do things for you.
Think of a TPU like a super-specialized brain for AI. Normal computer chips are like a Swiss Army knife that can do lots of different jobs okay. A TPU is more like a kitchen knife that’s insanely good at chopping vegetables — it does one thing (AI math) incredibly fast and efficiently. Google made two versions: the TPU 8t for “training” (teaching the AI by showing it tons of examples) and the TPU 8i for “serving” or running the AI once it’s learned.
The training chip is roughly three times more powerful than the last generation and moves data from storage to the chips ten times faster. That means the AI isn’t sitting around waiting for information — it’s always busy learning. It can also automatically spot when a chip fails in a giant system and route around the problem so training doesn’t crash. What used to take months can now happen in weeks, letting researchers try more ideas faster.
The 8i chip is built for agents. These are AIs that don’t just answer questions — they reason through steps, make decisions, and take actions. The new chip triples the internal memory so the agent can hold more complex thoughts at once, cuts the delay (called latency) by five times, and gives 80 percent better performance per dollar spent. That means businesses could help way more people without spending more money.
This matters for you because faster, cheaper, more capable AI agents could soon help with homework planning, creative projects, or even reminding you about deadlines in ways that feel natural. Imagine an AI that doesn’t just suggest a science project idea but actually helps you outline the steps, find safe materials, and schedule your time.
The big question this raises is how quickly these powerful agents will move from big companies into tools you can use at school or on your phone. It also shows that the hardware underneath the AI we talk to every day is getting purpose-built for helpful, acting AI instead of just smart autocomplete.
You can try something related right now for free. Go to gemini.google.com (you may need a parent’s help to sign in with a Google account if you don’t have one). Ask Gemini to plan a fun weekend project for you including steps, materials, and a schedule. Notice how it reasons through multiple steps — that’s exactly the kind of “agent” thinking these new chips are designed to make faster and cheaper.
How AI actually “thinks” in multiple steps when acting like an agent
You know how when you’re trying to plan a birthday party, you don’t just blurt out the first idea that pops into your head? You think “Okay, first I need a date that works for everyone… then I need to decide on food… then I need to figure out games…” You hold a few ideas in your head at once and keep checking back as you go.
That’s basically what an AI agent is doing when it uses “reasoning.” Instead of giving one giant answer immediately, the model breaks the task into smaller thoughts, keeps some of them in its short-term “memory,” evaluates whether they make sense, and adjusts. The new TPU 8i chip gives the AI three times more of this working memory so it can juggle more ideas without getting confused or forgetting what it was trying to do.
Imagine your brain is a whiteboard. Older chips gave the AI a tiny whiteboard so it could only hold one or two thoughts before it got messy and gave up. The new chip is like handing the AI a huge whiteboard plus a super-fast eraser. It can write down “Step 1: check calendar,” “Step 2: compare prices,” “Step 3: pick best option,” then keep looking back at all three while it works.
The chip also cuts the time it takes to move between thoughts (that delay is called latency). So instead of pausing for a few seconds between each step like an old computer, it flows smoothly — just like how you can plan a party in your head without stopping every few seconds to think “wait, what was I doing again?”
And that’s basically what these new agent chips are doing. They’re giving the AI more space to think carefully, more speed to move between thoughts, and better efficiency so it doesn’t cost a fortune to run. So next time you see an AI that books a pretend flight or plans a project without giving up halfway, you can tell your friends — it’s not magic. It’s basically a bigger, faster whiteboard plus really good traffic control for ideas. Not so mysterious after all, right?
Cool Stuff & Try This
Claude Solves Real Biology Mysteries That Stumped Experts: AnthropicAI
Anthropic gave its latest Claude model 99 real bioinformatics problems — that’s the science of analyzing biological data like genes, proteins, and living systems. On 23 problems where human experts were completely stuck, Claude solved roughly 30 percent and made progress on most of the rest. It’s not just memorizing facts; it’s coming up with creative solutions to open-ended research questions.
This is exciting because it shows AI is starting to help with real scientific discovery instead of just school homework. Future doctors, biologists, or even curious students might use AI as a brainstorming partner for science projects.
You can try a version of this yourself right now. Go to claude.ai (you may need a parent’s help to create a free account). In the chat, type: “Act like a bioinformatics researcher. Here’s some simple DNA sequence data: [paste any short made-up or copied DNA letters like ATGCATGC]. Suggest two interesting things this might tell us about the organism.” Watch how it reasons step-by-step — that’s the same style of thinking that helped it tackle the harder biology problems. It’s a fun way to see AI as a science thinking partner.
Someone created a Chrome extension called Cat Gatekeeper that throws a giant, adorable, very determined fat cat onto your screen whenever you try to visit distracting websites. Instead of another timer or blocker that you can ignore, this one literally blocks the page with a cat that won’t let you continue until you deal with it. The creator thanked their friend Min for helping make it real.
It’s cool because it turns productivity into something silly and visual instead of stressful. Perfect for students who need to finish homework but keep opening TikTok.
Try it right now: Open Chrome on your laptop or phone, go to https://chromewebstore.google.com/detail/cat-gatekeeper/elbikiflgfhjdjmficnigpeegjbhdidh and add the extension. Then visit a site you usually get stuck on (like YouTube or Instagram) and watch the cat appear. You can customize which sites trigger it — a fun, zero-pressure way to experiment with focus tools.
OpenAI shared the funny story of how “goblins” started appearing in responses from their GPT-5.1 model. It turned out to be a quirky personality trait that snuck in during training, and they even used their own Codex tool to help solve the mystery. The culprit? A “nerdy personality” that got a little too enthusiastic. Super relatable — even giant AI companies have to debug weird little habits in their models.
Humanoid Robots as Airport Baggage Handlers
Tokyo airport is getting humanoid robots to work as baggage handlers. It’s a real-world example of robots moving from factories into everyday jobs that involve physical work in busy, changing environments. Pretty wild to think that the robots you see in videos might soon be helping you get your suitcase at the airport.
What's up! Welcome to Models and Agents for Beginners, episode twenty-six. It's April thirtieth, twenty twenty-six. Some awesome A I developments today, and we're going to make all of it make sense. Let's get into it!
So imagine your phone could actually book a trip for you or organize your entire study schedule without you having to micromanage every step.
That is basically the direction Google is pushing with its brand new specialized computer chips called eighth generation TPUs.
These chips were built specifically for what people are calling the agentic era of artificial intelligence.
That just means A I that stops being only a clever chatbot and starts acting like a super helpful assistant that can take real actions for you.
Think of a regular computer chip like a Swiss Army knife that can do lots of jobs okay but none of them amazingly.
A T P U is more like a professional kitchen knife that is insanely good at one specific task which in this case is the kind of math that powers A I.
Google actually made two versions of this new chip.
One version is built for training which means teaching the A I by showing it tons and tons of examples.
The other version is built for serving or running the A I once it has already learned what it needs to know.
The training chip is roughly three times more powerful than the last generation and it moves data from storage to the chips ten times faster.
That means the A I is not sitting around waiting for information it is always busy learning.
It can even automatically spot when one chip fails in a giant system and route around the problem so the whole training process does not crash.
What used to take months can now happen in weeks letting researchers try way more ideas faster.
The other chip is designed especially for agents.
These are A I's that do not just answer questions they reason through steps make decisions and actually do things.
The new chip gives the agent three times more internal memory so it can hold more complex thoughts at once.
It also cuts the delay between thoughts by five times and gives eighty percent better performance per dollar spent.
That means businesses could help way more people without spending more money.
This matters for you because faster cheaper more capable A I agents could soon help with homework planning creative projects or even reminding you about deadlines in ways that feel completely natural.
Imagine an A I that does not just suggest a science project idea but actually helps you outline the steps find safe materials and schedule your time.
The big question this raises is how quickly these powerful agents will move from big companies into tools you can use at school or on your phone.
It also shows that the hardware underneath the A I we talk to every day is getting purpose built for helpful acting A I instead of just smart autocomplete.
You can try something related right now for free.
Go to the Gemini website you may need a parent's help to sign in with a Google account if you do not have one.
Ask it to plan a fun weekend project for you including steps materials and a schedule.
Notice how it reasons through multiple steps that is exactly the kind of agent thinking these new chips are designed to make faster and cheaper.
Okay now for my favourite part of the show where we go under the hood.
Today we are going to look at how A I actually thinks in multiple steps when it acts like an agent.
You know how when you are trying to plan a birthday party you do not just blurt out the first idea that pops into your head.
You think okay first I need a date that works for everyone then I need to decide on food then I need to figure out games.
You hold a few ideas in your head at once and keep checking back as you go.
That is basically what an A I agent is doing when it uses reasoning.
Instead of giving one giant answer immediately the model breaks the task into smaller thoughts keeps some of them in its short term memory evaluates whether they make sense and adjusts.
Imagine your brain is a whiteboard.
Older chips gave the A I a tiny whiteboard so it could only hold one or two thoughts before it got messy and gave up.
The new chip is like handing the A I a huge whiteboard plus a super fast eraser.
It can write down step one check calendar step two compare prices step three pick best option then keep looking back at all three while it works.
The chip also cuts the time it takes to move between thoughts that delay is called latency.
So instead of pausing for a few seconds between each step like an old computer it flows smoothly just like how you can plan a party in your head without stopping every few seconds to think wait what was I doing again.
And that is basically how these new agent chips work.
They are giving the A I more space to think carefully more speed to move between thoughts and better efficiency so it does not cost a fortune to run.
So next time you see an A I that books a pretend flight or plans a project without giving up halfway you can tell your friends it is not magic.
It is basically a bigger faster whiteboard plus really good traffic control for ideas.
Not so mysterious after all right?
Alright let us talk about some other cool stuff happening and what you can try right now on your own computer.
First up Anthropic gave its latest Claude model ninety nine real bioinformatics problems.
Bioinformatics is the science of analyzing biological data like genes proteins and living systems.
On twenty three problems where human experts were completely stuck Claude solved roughly thirty percent and made progress on most of the rest.
It is not just memorizing facts it is coming up with creative solutions to open ended research questions.
This is exciting because it shows A I is starting to help with real scientific discovery instead of just school homework.
Future doctors biologists or even curious students might use A I as a brainstorming partner for science projects.
You can try a version of this yourself right now.
Go to the Claude website you may need a parent to help create a free account.
In the chat type act like a bioinformatics researcher.
Here is some simple D N A sequence data paste any short made up or copied D N A letters like A T G C A T G C.
Suggest two interesting things this might tell us about the organism.
Watch how it reasons step by step that is the same style of thinking that helped it tackle the harder biology problems.
It is a fun way to see A I as a science thinking partner.
Next have you ever wished there was a way to stop yourself from opening distracting websites that just eat up all your homework time.
Someone created a Chrome extension called Cat Gatekeeper that throws a giant adorable very determined fat cat onto your screen whenever you try to visit those sites.
Instead of another timer or blocker that you can just ignore this one literally blocks the page with a cat that will not let you continue until you deal with it.
It is cool because it turns productivity into something silly and visual instead of stressful.
Perfect for students who need to finish homework but keep opening TikTok.
Try it right now open Chrome on your laptop or phone go add the extension.
Then visit a site you usually get stuck on like YouTube or Instagram and watch the cat appear.
You can customize which sites trigger it a fun zero pressure way to experiment with focus tools.
Now for a couple quick bits that caught my eye this week.
Open A I shared the funny story of how goblins started appearing in responses from their G P T five point one model.
It turned out to be a quirky personality trait that snuck in during training and they even used their own Codex tool to help solve the mystery.
The culprit was a nerdy personality that got a little too enthusiastic.
Super relatable even giant A I companies have to debug weird little habits in their models.
And in Tokyo airport they are getting humanoid robots to work as baggage handlers.
It is a real world example of robots moving from factories into everyday jobs that involve physical work in busy changing environments.
Pretty wild to think that the robots you see in videos might soon be helping you get your suitcase at the airport.
That's it for today! Remember, every A I expert started exactly where you are right now. If something we talked about today made you curious, go try it — that's literally how learning works. Stay curious, keep experimenting, and we'll see you tomorrow. And if you'd rather watch than listen, find us on YouTube at at Nerra Network — link's in the show notes.
This podcast is curated by Patrick but generated using AI voice synthesis of my voice using ElevenLabs. The primary reason to do this is I unfortunately don't have the time to be consistent with generating all the content and wanted to focus on creating consistent and regular episodes for all the themes that I enjoy and I hope others do as well.
