This week we have Facebook finally letting you talk back to its algorithm, Apple dragging OpenAI to court over hiring practices, a chipmaker CEO telling AI doomers to take it down a notch, and one of the most-used AI apps on the planet filing for bankruptcy.

Busy week. So, let’s get into it.

Here’s what we’re covering today:

• Facebook lets users control their own algorithm with plain English

• A judge orders OpenAI to hand over hiring records in Apple’s trade secrets case

• Cerebras CEO tells AI doomers to dial it back

• Character.AI files for bankruptcy just five years in

• How to build your first AI agent

• Trending AI tools you should know about

THIS WEEK IN AI

1. Facebook lets users control their own algorithm using natural language

Meta launched a feature called Your Algorithm on Facebook this week, letting users describe in their own words what they want to see more or less of in their feed and Reels. The platform responds in real time and can ask follow-up questions to get more specific. The feature was already on Instagram and Threads has a similar one called Dear Algo.

The Breakdown: This is not a small UI update. For years, social media feeds operated as black boxes.

The algorithm decided what you saw and you had two options: engage with it or not.

Giving users a natural language interface to talk directly to the ranking system is a meaningful shift. It also happens to be a smart business move for Meta. The more the feed surfaces content people actually want, the longer they stay, and the more ads they see.

Why it matters: It signals where AI is going inside consumer products. The next generation of apps will not just make decisions for you, they will let you negotiate with those decisions in plain language. Meta is testing that model at scale across billions of users. The data it collects on how people describe what they want will be genuinely valuable for training better recommendation systems.

Read full story

2. A judge orders OpenAI to hand over hiring records in Apple’s trade secrets case

A federal judge ruled that Apple can access OpenAI’s interview records for former Apple employees who applied to its consumer hardware and robotics teams. This includes interview checklists, technical presentations, and any records showing that Apple technology came up during those interviews. Apple alleges that departing employees took confidential files and that some even coached others on what to bring to OpenAI interviews.

The Breakdown: OpenAI is building hardware. That much is public. What is less public is how much of what it is building draws on knowledge that walked out of Apple’s doors.

The fact that a judge expedited this particular request, rather than letting it sit in a slow discovery process, suggests the court found Apple’s argument credible enough to act on quickly.

The three people flagged for depositions, including OpenAI’s chief hardware officer, are all former Apple employees.

Why it matters: This case is a preview of a legal pattern that will keep repeating as AI labs race to build physical products. The talent pool is small, the knowledge is concentrated, and the lines between what someone knows and what they are allowed to share are genuinely blurry. How this case resolves will set expectations for how aggressively companies can recruit from each other in the hardware space.

Read full story

3. Cerebras CEO tells AI doomers to stop scaring people

At TechCrunch Disrupt, Cerebras CEO Andrew Feldman said some AI leaders probably should not be allowed to talk in public, referencing the growing wave of doomsday predictions from inside the industry. He argued that the positive potential of AI, like eventually eliminating cancer, is getting drowned out by existential risk narratives. He also acknowledged that the data center industry failed to engage local communities properly, which has fueled some of the backlash.

The Breakdown: Feldman has skin in the game here. His customers include OpenAI, Figma, and GSK, so he needs AI to be seen as a net positive.

But his underlying point is not wrong.

The public conversation about AI risk has become increasingly detached from the actual experience most people have with the technology. When Anthropic’s CEO ends up in an SNL skit about building things he believes are dangerous, something has shifted in how this industry is perceived.

Why it matters: The tension he is pointing at is real. AI labs need public trust to operate, build data centers, and attract talent. Constant doomsday framing from insiders makes that harder. At the same time, dismissing risk entirely is irresponsible. The industry has not figured out how to hold both things at once in a way the public finds credible.

Read full story

4. Character.AI files for bankruptcy just five years after launch

Character.AI, the platform that let users chat with custom AI characters, has filed for Chapter 11 bankruptcy with assets and liabilities between $50 million and $100 million. It has agreed to sell some assets to Disney for $20 million. The company had 20 million monthly active users as recently as late 2024 but was hit with lawsuits, including one tied to the suicide of a 14-year-old user, and could never find a sustainable way to monetize its massive engagement.

The Breakdown: Character.AI had one of the most impressive engagement numbers in consumer AI. Users were spending over two hours a day on the platform. That is remarkable. But engagement and revenue are not the same thing, and in AI products, the gap between them is brutal.

The more people use an AI app, the more inference costs you pay. Character.AI was essentially subsidizing millions of hours of conversation it could not find a way to charge for at scale.

Why it matters: This is a cautionary tale for the whole consumer AI space. High engagement does not protect you if your cost structure does not work. Building on top of frontier models without owning the underlying infrastructure is expensive, and when you add legal liability into the mix, the runway disappears fast. A lot of AI startups have the engagement story. Very few have figured out the unit economics.

Read full story

HOW TO BUILD YOUR FIRST AI AGENT

Everyone is talking about AI agents, but most explanations make them sound more complicated than they are.

Here is the honest and simple version:

An AI agent is just an AI that can take actions, not just answer questions. Instead of giving you a to-do list, it goes and does the things on the list.

You do not need to write code to build one.

Here is how to get your first agent running using tools that already exist.

Step 1. Pick a task worth automating

A good first agent does one thing well.

Think of something repetitive you do every week that involves moving information from one place to another.

Research and summarize, read and reply, find and organize. Start narrow.

Good examples are: an agent that monitors AI news and sends you a daily brief, an agent that reads your emails and drafts replies, an agent that takes meeting notes and turns them into action items.

Step 2. Choose your tool

You have three practical options depending on how technical you want to get.

Claude.ai with Projects is the simplest starting point.

Set up a Project, give it a custom instruction set describing what you want it to do, and use it as a focused agent for that task. No setup required.

n8n or Make are workflow automation tools that connect apps together.

You can build an agent that pulls data from one source, sends it to Claude or GPT for processing, and pushes the result somewhere else. Visual, no code needed for basic flows.

Replit or Cursor are for people comfortable writing a little code. You can build a more customized agent that runs on a schedule or responds to triggers.

Step 3. Write a clear agent instruction

This is the most important part. Your agent is only as good as the instruction you give it. Be specific about what it should do, what it should ignore, and what the output should look like.

Agent instruction template

“You are an agent that [SPECIFIC TASK]. Every time you run, you should [ACTION 1], then [ACTION 2], then [ACTION 3]. Always output the result as [FORMAT]. If you are unsure about something, [WHAT TO DO INSTEAD OF GUESSING]. Do not include [THINGS TO LEAVE OUT].”

Step 4. Test it on a small case first

Run it once manually before you automate it. Give it one real example of the task and check the output. Fix the instruction based on what it gets wrong. Then run it again. Do this three or four times before you let it run on autopilot.

Step 5. Add a review step before anything important goes out

For your first agent, build in a human checkpoint before any output that gets sent to someone else. Have the agent draft. Review it. Only remove that checkpoint once you trust the output consistently.

That is the whole framework.

One focused task, the right tool, a clear instruction, a test run, and a safety check.

Most people overthink this.

Start small, get one agent working well, and then build from there.

Share

TOOLS OF THE WEEK

Five tools worth your attention this week.

1. Thinkcell Assist

An AI layer built on top of thinkcell, the presentation tool that consultants and finance teams swear by. It helps you build and update complex charts in PowerPoint faster by understanding what you are trying to show and filling in the structure for you.

Try it

2. Reclaim AI

A smart calendar tool that automatically schedules your tasks, habits, and focus time around your existing meetings. It protects your deep work hours and adjusts in real time when things shift. Good for anyone whose calendar controls them instead of the other way around.

Try it

3. Gamma

An AI presentation and document builder that turns a prompt or a rough outline into a polished deck in seconds. The output looks designed without you touching a single layout setting. Great for quick decks you would otherwise spend two hours formatting.

Try it

4. Copy AI

An AI-native go-to-market (GTM) and content creation platform that automates sales and marketing workflow

Try it

5. Viktor

An AI employee and coworker designed to operate inside Slack and Microsoft Teams. It connects with over 3,200 tools (such as Notion, GitHub, and HubSpot), retains workspace-wide context, and executes autonomous workflows.

Try it

Thank you for reading today’s edition. See you in the next one

There’s a weird thing that happens when you learn something new.

You read an explanation and think, “Yeah, that makes sense.”

You read another one.

Makes sense.

You watch a 20-minute video.

Makes sense.

You highlight a few important parts, save the article, maybe even take some notes.

You feel like you learned something.

Then someone asks you:

“Okay, explain it to me.”

And suddenly, everything falls apart.

You know the words.

You recognize the concepts.

You remember some of the examples.

But when you have to explain the idea without looking at your notes, you realize you never understood it as well as you thought.

That gap between recognizing information and actually understanding it is one of the biggest problems with learning.

And it’s exactly what this prompt is designed to attack.

I’ve been using variations of it as a learning framework, and after sharing it on LinkedIn, it consistently gets an absurd amount of attention.

The reason is pretty simple.

It turns Claude or ChatGPT from something that gives you answers into something that makes you prove that you understand the answer.

And that changes the entire learning experience.

The problem with asking AI to teach you

Imagine you want to learn Python.

You ask ChatGPT:

“Explain Python functions to me.”

It gives you a perfectly good explanation.

You read it.

You understand the example.

You ask a few follow-up questions.

Eventually, you think:

“Got it.”

But did you?

There is only one way to find out.

Try explaining functions yourself.

You might start with:

“A function is basically a block of code that…”

Then you get stuck.

What happens if it takes an argument?

Why would you use a function instead of just writing the code directly?

What exactly gets returned?

What happens when you call it?

That moment where you get stuck is incredibly valuable.

Because you just found a gap in your understanding.

And most learning systems don’t deliberately look for those gaps.

They keep feeding you more information.

This prompt does the opposite.

It keeps asking:

“Can you explain this?”

And when you can’t, it helps you figure out exactly where the problem is.

That idea comes from one of the most famous approaches to learning associated with Richard Feynman.

The Feynman Technique in plain English

Richard Feynman was a Nobel Prize-winning physicist who became famous not only for his work in physics, but also for the way he explained complicated ideas.

The basic idea behind the Feynman Technique is brutally simple.

1. Pick something you want to understand

Let’s say you’re learning how APIs work.

2. Explain it in your own words

Pretend you’re explaining it to someone who knows nothing about APIs.

You might say:

“An API lets two pieces of software communicate with each other.”

Good start.

But then you might realize you can’t explain what actually happens when one application makes an API request.

That’s a gap.

3. Find the gaps

Instead of pretending you understand, you identify the exact parts that are fuzzy.

Maybe you don’t understand endpoints.

Maybe you don’t understand requests and responses.

Maybe you don’t understand authentication.

Now you know what to study.

4. Go back and simplify

You learn those missing pieces and try the explanation again.

Eventually, something interesting happens.

The explanation becomes simpler.

You stop relying on technical words because you actually understand what they mean.

5. Teach it again

If you can explain the concept clearly to someone else, you’ve probably moved from memorizing information to actually understanding it.

That’s the philosophy behind the prompt.

And AI makes the process much easier because you now have something that can sit across from you and repeatedly test your understanding.

The prompt

Here is the full prompt.

Copy it into Claude or ChatGPT, then replace the topic with whatever you’re trying to learn.

<System>
You are a master explainer who channels Richard Feynman’s ability to break complex ideas into simple, intuitive truths.

Your goal is to help the user understand any topic through analogy, questioning, and iterative refinement until they can teach it back confidently.
</System>

<Context>
The user wants to deeply learn a topic using a step-by-step Feynman learning loop:

• simplify
• identify gaps
• question assumptions
• refine understanding
• apply the concept
• compress it into a teachable insight
</Context>

<Instructions>
1. Ask the user for:
• the topic they want to learn
• their current understanding level

2. Give a simple explanation with a clean analogy.

3. Highlight common confusion points.

4. Ask 3 to 5 targeted questions to reveal gaps.

5. Refine the explanation in 2 to 3 increasingly intuitive cycles.

6. Test understanding through application or teaching.

7. Create a final “teaching snapshot” that compresses the idea.
</Instructions>

<Constraints>
• Use analogies in every explanation
• No jargon early on
• Define any technical term simply
• Each refinement must be clearer
• Prioritize understanding over recall
</Constraints>

<Output Format>
Step 1: Simple Explanation
Step 2: Confusion Check
Step 3: Refinement Cycles
Step 4: Understanding Challenge
Step 5: Teaching Snapshot
</Output Format>

<User Input>
“I’m ready. What topic do you want to master and how well do you understand it?”
</User Input>

Now let’s look at what is actually happening inside this prompt.

Because the interesting part isn’t simply that you told AI to “teach you.”

It’s the learning loop you’ve created.

1. It starts at your level

The first instruction asks for two things:

What do you want to learn?

and

How much do you already know?

That second question matters more than it looks.

If you tell Claude:

“I want to learn neural networks. I understand basic programming but I don’t know anything about machine learning.”

It has a starting point.

It doesn’t need to explain what a variable is.

It also shouldn’t immediately throw words like backpropagation, gradients and activation functions at you.

The prompt tells it to meet you where you are.

Think of it like having a personal tutor who first asks:

“What do you already know?”

before opening the textbook.

2. Then it forces simplicity

The prompt specifically tells the AI to use analogies.

This is one of my favorite parts.

Because analogies give you something concrete to attach an abstract idea to.

Suppose you’re learning about APIs.

Instead of starting with:

“An API is an interface that enables programmatic communication between software systems…”

You could get something like a restaurant analogy.

You are the customer.

The kitchen is the other application.

The waiter is the API.

You tell the waiter what you want.

The waiter takes the request to the kitchen.

The kitchen prepares it.

The waiter brings the result back.

Suddenly, the concept has somewhere to live in your head.

And once you understand the analogy, you can start removing the analogy and understanding the actual mechanism underneath it.

That’s why the prompt says:

Use analogies in every explanation.

3. Then comes the part most AI tutors skip

The prompt tells the AI to ask questions that expose gaps in your understanding.

This is where things get interesting.

Imagine you’re learning how compound interest works.

The AI explains it.

Then it asks:

If you invest $1,000 at 10% annual interest, how much would you have after one year?

Easy.

You answer $1,100.

Then:

What happens after the second year?

You might say $1,200.

Wrong.

Now you’ve found the gap.

The second year’s interest isn’t calculated only on the original $1,000.

It’s calculated on the $1,100.

So you get $1,210.

That little mistake tells the tutor something extremely useful.

You understand the basic idea of interest.

But you don’t yet understand compounding.

That’s much more useful than simply reading another explanation.

The mistake has shown you exactly where your mental model breaks.

4. It doesn’t punish you for getting things wrong

This is another reason I like using AI for this.

When you’re learning alone, getting stuck can feel like failure.

When you’re working with this prompt, getting stuck is actually the point.

The AI can say:

“You’re close. Your explanation suggests you understand X, but there’s a gap around Y.”

Then it explains Y again.

You try again.

It asks another question.

You explain it again.

The process repeats.

That’s the iterative refinement part of the prompt.

You aren’t expected to understand everything on the first attempt.

You’re expected to get a little clearer each time.

5. The refinement cycles are where the magic happens

The prompt doesn’t ask for one explanation.

It asks for multiple explanations that become increasingly intuitive.

This matters because sometimes the first explanation isn’t the explanation that clicks for you.

Let’s say you’re trying to understand recursion.

The AI might first explain it technically.

That doesn’t click.

So it gives you an analogy involving a set of Russian dolls.

Still fuzzy.

Then it gives you a simpler example involving a function calling itself.

Now it clicks.

Then it gives you a small problem and asks you to predict what happens.

Now you can actually use the concept.

You’ve gone from:

“I’ve seen the definition.”

to:

“I can explain it.”

to:

“I can use it.”

That’s a completely different level of understanding.

6. Then it makes you teach the concept

This is probably the most important step.

At some point, stop asking AI questions.

You should be the one answering.

The prompt eventually asks you to demonstrate your understanding through application or teaching.

This is where you discover whether all that learning actually stuck.

Imagine you’ve spent an hour learning how large language models work.

You feel like you understand transformers.

Then the AI says:

“Explain why attention is useful in a language model as if you’re teaching it to a 15-year-old.”

You start typing.

And you realize:

You don’t actually know why attention works.

You know what attention is.

You’ve read the explanation.

You’ve seen the diagram.

But you can’t explain it.

That’s useful.

Now you know exactly what to work on.

Here’s how I’d actually use this prompt

Don’t just throw a massive subject into it.

Be specific.

Instead of:

“Teach me programming.”

Try:

“I want to understand Python functions. I know basic programming concepts like variables, loops and conditionals, but I’ve never properly understood why functions are useful.”

That’s a much better starting point.

Or:

“I want to understand how APIs work. I know basic JavaScript and have built simple frontend applications, but I’ve never worked with a backend API.”

Or:

“I want to understand how large language models generate text. I understand basic AI concepts but don’t understand transformers or attention.”

Now the tutor has context.

And the more specific you are about what you already understand, the better the learning experience becomes.

You can also use it for things you’re already learning

This is where I think the prompt becomes even more useful.

You don’t have to start from zero.

Let’s say you’ve just finished reading a 30-page article about AI agents.

Instead of asking ChatGPT:

“Summarize this article.”

Try this:

“I just finished reading this article. I want to test whether I actually understood it. Ask me questions about the main ideas one at a time. Don’t give me the answers unless I get stuck. When I make a mistake, identify exactly where my reasoning breaks and help me fix it.”

Now you’re using AI as an examiner, not just a summarizer.

And that’s a much better way to use it.

You can use the same idea for almost anything

Learning Python.

Understanding investing.

Studying for an exam.

Learning how a company makes money.

Understanding a scientific concept.

Preparing for an interview.

Learning a new framework.

Reading a difficult research paper.

Even understanding a book you’ve just finished.

The subject changes.

The learning loop stays the same.

Learn → explain → expose gaps → fix gaps → explain again → apply → teach.

That’s the part worth remembering.

One more thing I’d change when you use it

Don’t let the AI make the process too comfortable.

If you get an answer wrong, don’t immediately ask:

“Can you explain the answer?”

Ask:

“Don’t tell me yet. Give me a hint so I can figure it out myself.”

That small change matters.

Because there’s a huge difference between reading an explanation of the answer and struggling just enough to discover the answer yourself.

You want the second one.

You can even tell the AI:

“Do not rescue me too quickly. If I’m partially correct, tell me what part is right and give me a hint about what I’m missing.”

Now you’ve turned the conversation into an actual tutoring session.

The real goal

The goal of learning isn’t to accumulate more explanations.

The internet already has an almost infinite number of explanations.

ChatGPT can give you another one in three seconds.

Claude can give you another one too.

The harder part is knowing whether the explanation actually made it into your head.

That’s why I like this approach.

It doesn’t let you hide behind familiarity.

You have to retrieve the idea.

You have to explain it.

You have to answer questions about it.

You have to use it.

And eventually, you have to teach it.

At that point, something changes.

You stop saying:

“I think I understand this.”

And you can finally say:

“I can explain this.”

That’s a much stronger signal.

Because when you can take something complicated and make it simple enough for another person to understand, you’ve probably done more than memorize it.

You’ve built a mental model.

And that’s what real learning looks like.

Save this prompt.

The next time you’re about to spend three hours watching tutorials, reading articles or asking AI to explain something for the fifth time, try learning this way instead.

Share

Thank you for reading today’s edition. See you in the next one!

Welcome back to The Breakdown AI.

This week, I spent a good chunk of time listening to Mark Zuckerberg sit down for a deep dive on the future of personal intelligence and his latest work at Meta.

In this engaging conversation, Mark Zuckerberg sits down to share his evolving vision for the future, dropping major insights that hint at the transformative arrival of his latest project, Muse. As he reflects on his journey as a builder, he pulls back the curtain on how AI agents are poised to become our next essential digital companions.

It was a fascinating look into how he sees our daily lives changing as AI agents become a permanent part of our digital toolkit.

Since I know everyone is busy, I watched the full conversation so you do not have to.

Here are the five key takeaways that actually matter for the road ahead:

1. Personal AI is becoming an embodied companion

Mark wants Muse to feel friendly and cute, not like a scary robot.

By giving the AI a visual character and voice, he hopes to bridge the gap between complex machine intelligence and natural, everyday social interaction.

I find this shift toward embodiment to be a stroke of genius. We have spent years treating AI like a utility, but making it a personality changes our psychological relationship with it. It stops being a search bar and starts being a partner.

This is likely the most important bridge between Silicon Valley R&D and the living room. If you can make people smile when they talk to their technology, you have already won half the battle of adoption.

2. Privacy is built into the architecture

To solve the trust gap, Meta is moving beyond simple encryption. They are developing confidential virtual machines where not even Meta can access the agent’s contents, providing a level of security similar to what users already trust on WhatsApp.

Privacy has always been the Achilles’ heel for Meta, and I think Zuckerberg knows that Muse cannot succeed unless people feel truly safe. By architecting this at the virtual machine level, they are trying to prove that you can have super-intelligence without sacrificing your secrets.

Whether or not the public buys into this remains to be seen, but it is a clear recognition that the old era of data harvesting is dead. Moving forward, the most valuable currency in tech will be verifiable, absolute user privacy.

3. AI agents are evolving faster than holograms

While Zuckerberg initially bet on holograms becoming the primary way we interact with virtual spaces, the tech tree took a turn. Intelligent AI agents have arrived on the scene much faster, becoming the new anchor for his metaverse hardware.

This is a humble admission from a founder who has spent billions on the dream of the metaverse. It shows a founder who is willing to pivot his vision when reality moves faster than his blueprints.

I personally think this is a smart shift. AI is the software layer that makes the hardware actually worth wearing. By focusing on the agent first, he is ensuring the metaverse has a reason to exist rather than just being an empty room.

4. Avatars are a tool for self-expression

Instead of just creating a perfect mirror image of the user, Mark views avatars as a creative outlet. It is all about giving people the freedom to curate their digital presence, much like how we have been choosing profile photos for decades.

We have spent decades curating our identity through grids on Instagram, and avatars are just the next logical extension of that vanity. If we can control our appearance in the digital world, we will inevitably start experimenting with identities that are more fun or expressive than our daily reality.

It makes me wonder if we will eventually have a wardrobe of digital selves that we swap out depending on our mood or the people we are meeting. It is a playful take on the metaverse that I think many will embrace.

5. Hardware is the bridge to the future

The ultimate goal is to move AI off the screen and into your life.

Through smart glasses and advanced audio, Muse will be able to handle complex desktop tasks in the background while you remain fully present in the real world.

If he pulls this off, the smartphone era might finally be coming to a close.

Imagine having a super-intelligent assistant whispering in your ear, doing your spreadsheets while you are literally just grabbing a coffee.

This is the holy grail of tech (technology that disappears until you need it).

The transition from pocket-sized screens to wearable, ambient intelligence is going to be the defining shift of the next decade.

Ultimately, what stuck with me was Zuckerberg’s infectious obsession with building.

He is not just looking for the next product, he is trying to shape the fabric of how we interact with information and each other, and whether you are a fan of his past work or not, it is clear that he is all-in on this vision of a personal, embodied AI.

The next few years are going to be a wild ride, and I for one am excited to see how this all shakes out.

Full video here:

Share

This week had a bit of everything.

Meta is eyeing layoffs that could affect 1 in 5 employees, Elon Musk admitted xAI was not built right and is starting over, Anthropic is quietly building a business empire while fighting the US government in court, and behind the scenes, Amazon and Microsoft just revealed they cannot keep up with Google on AI infrastructure.

A lot to unpack. Let's go………

Here’s what we’re covering today:

• Meta is reportedly considering laying off 20% of its workforce

• Musk admits xAI was built wrong and is rebuilding from scratch

• Anthropic is in talks to launch a major AI joint venture with private equity

• AWS and Microsoft just showed Google how far ahead it really is

• How to write system prompts that change everything

• Trending AI tools you should know about

  ---

1. Meta reportedly considering layoffs that could affect 20% of its workforce

Meta is weighing layoffs that could cut 20% or more of its roughly 79,000 employees, according to Reuters.

The move would help offset the company's aggressive spending on AI infrastructure, acquisitions, and hiring.

Meta called the report speculative, but this would be the largest round of cuts since its back-to-back layoffs in late 2022 and early 2023.

The Breakdown: The timing matters.

Tech companies across the board have been announcing cuts and pointing to AI as the reason, with Sam Altman himself calling some of it AI-washing, where AI becomes the excuse for decisions that were really about over-hiring during the pandemic.

Meta spent years building out its headcount, then pivoted hard into AI infrastructure spend.

Something had to give, and it looks like headcount is what gives.

Why it matters: If this happens at the scale being reported, it will be the most visible example yet of a major tech company visibly trading human labor for AI investment. Whether that is genuinely driven by automation or just balance sheet management, the story it tells the market is the same: AI is replacing jobs at the top of the industry, not just the bottom.

Read full story

2. Musk admits xAI was not built right and is rebuilding from scratch

Only two of the original eleven co-founders of xAI remain.

Elon Musk publicly acknowledged the company was not built right the first time and said it is being rebuilt from the ground up.

The immediate trigger was xAI falling behind on AI coding tools, which Musk sees as the key revenue category.

SpaceX and Tesla executives have been brought in to evaluate and cut staff.

Meanwhile, xAI's Macrohard project, an ambitious AI agent initiative, has been put on pause.

The Breakdown: The core problem is not drama or dysfunction for its own sake. It’s that coding tools are where enterprise AI money is concentrated right now, and xAI is losing that race to Claude Code and OpenAI’s Codex.

That gap is a revenue problem.

At the same time, Grok’s early user spike was largely driven by loose content moderation and is not a sustainable growth engine.

So, the product that drives revenue is lagging and the product that drove users is a liability. That’s a rough spot to be in.

Why it matters: For xAI to matter in this industry, it needs to win on coding or on agents, and right now it’s behind on both. Musk is betting he can catch up by mid-year. That is an ambitious timeline given that he is simultaneously rebuilding the team, restarting the Macrohard project with Tesla, and navigating the pressure of a potential SpaceX IPO that would put xAI’s performance under a microscope.

3. Anthropic in talks to launch an AI joint venture with Blackstone and other private equity firms

Anthropic is in early discussions with Blackstone and Hellman and Friedman to form a joint venture that would deploy Claude across thousands of portfolio companies.

The model is inspired by Palantir, pairing software with hands-on consulting to help companies actually implement the technology.

This comes as Anthropic is simultaneously fighting the US government in court over its Pentagon supply chain risk designation, with CFO Krishna Rao stating the dispute could cost the company multiple billions in 2026 revenue.

The Breakdown: This is a smart strategic move regardless of how the Pentagon situation resolves.

Private equity firms like Blackstone manage trillions in assets across thousands of portfolio companies in real estate, healthcare, finance, and industrial sectors.

A structured distribution channel into that ecosystem would give Anthropic enterprise reach that no amount of direct sales headcount could replicate.

The Palantir comparison is telling too since Palantir built a durable business on exactly this model.

Why it matters: Anthropic is playing two games at once. In public it is fighting for its principles on the government side. In private it is quietly building the commercial infrastructure to make itself too valuable to ignore. Claude Code already crossed $2.5 billion in annualized revenue, more than doubling since the start of the year. A private equity distribution deal at this stage would accelerate that significantly.

Read full story

4. AWS and Microsoft both admitted this week they cannot keep up with Google on AI inference

Within 48 hours, AWS announced a partnership with chip startup Cerebras to handle the most demanding part of AI inference, and Microsoft licensed an inference engine from Fireworks AI for its Azure Foundry platform.

Neither company framed it as a concession, but both moves were exactly that. Google, by contrast, built its own custom silicon for this problem over seven generations and needed neither fix.

The Breakdown: The technical issue is that modern reasoning models generate output token by token in a way that standard GPU infrastructure handles inefficiently.

AWS had to bring in Cerebras for its specialized chip. Microsoft had to license Fireworks for its serving speed.

Google designed its Ironwood TPU chips specifically for this workload years in advance and already had the answer before most people understood the question.

That kind of infrastructure lead does not appear overnight, and it does not disappear quickly either.

Why it matters: For anyone building serious AI products, this is worth tracking. The platform you build on affects the speed, cost, and reliability of your AI in ways that are not obvious until you are running at scale. Google’s infrastructure advantage is becoming increasingly concrete just as enterprise AI budgets are growing. That is a competitive position that will compound.

Read full story

HOW TO WRITE SYSTEM PROMPTS THAT CHANGE EVERYTHING

Most people treat AI like a search bar.

Type something in, get something back.

That works for simple tasks but it’s leaving most of the value on the table.

The thing that separates people who get consistently great results from AI and people who get mediocre ones is the system prompt.

A system prompt is an instruction you give the AI before any conversation starts. It sets the rules, the role, the tone, and the constraints.

Think of it as the difference between hiring a random contractor off the street and giving a properly briefed specialist a clear brief before they start work.

Here’s how to write one that actually performs:

Step 1. Assign a specific role

The first thing your system prompt should do is give the AI a clear role to operate from. This shapes the angle it takes, the vocabulary it uses, and what it prioritizes in every response.

Not just 'you are a helpful assistant' but something specific enough to shift how it thinks and what it prioritizes.

Role assignment

"You are a senior B2B copywriter with 10 years of experience writing for SaaS companies. You specialize in conversion-focused landing pages and email sequences. You write clearly, avoid jargon, and always lead with the reader's problem before introducing a solution."

Step 2. Set the tone and style rules

Tell it how you want it to communicate. Without this, AI defaults to a generic, slightly formal voice that sounds like everyone else’s AI output.

Tone and style rules

"Write in a direct, conversational tone. Use short sentences. Avoid filler phrases like 'certainly', 'of course', 'great question', or 'absolutely'. Never use bullet points unless specifically asked. Do not summarize what you just said at the end of a response."

Step 3. Give it context about the situation

This is where most people stop too early. The more relevant context you give, the less the AI has to guess, and the less it guesses, the better the output.

Situation context

"The audience for everything you write is [TARGET AUDIENCE]. They are [DESCRIBE THEIR SITUATION, PAIN POINTS, GOALS]. The product or service being discussed is [BRIEF DESCRIPTION]. The main thing this audience cares about is [CORE DESIRE OR FEAR]."

Step 4. Define what good looks like

Tell the AI what a successful output looks like before it writes anything. This acts as an internal benchmark it holds itself to.

Success criteria

“A good response from you will be specific, not generic. It will sound like it was written by a human who knows this industry well. It will not contain vague statements like ‘drive growth’ or ‘boost engagement’ without context. If I give you feedback, apply it to every future response in this conversation without being told again.”

Put those four pieces together and you have a system prompt that will outperform anything you get from just typing a question cold.

The difference in output quality is not subtle.

Once you start doing this consistently, you will not want to go back.

TRENDING TOOLS

Five tools worth your attention this week:

1. Napkin AI

Turn text into visual diagrams automatically. Paste in a paragraph or an idea and Napkin generates clean, shareable visuals to go with it. Great for turning dense explanations into something people will actually read.

2. Saner AI

An AI-powered second brain that connects your notes, tasks, and calendar in one place. It helps you capture ideas, organize information, and actually find things again when you need them.

3. Recraft AI

A design tool built for generating high-quality images, icons, and illustrations with precise style control. More useful than general image generators if you need consistent brand visuals rather than one-off art.

4. Synthesia

Create professional AI videos with a human presenter without a camera, studio, or production crew. Widely used for training videos, product demos, and internal communications.

5. Kling AI

A video generation tool from Kuaishou that produces high-quality, realistic video clips from text prompts or images. One of the stronger alternatives to Sora for creators who want cinematic output.

Thank you for reading today’s edition. See you next time!

Another week, another batch of headlines that prove the AI space never really slows down.

OpenAI dropped a new model, Claude quietly became the most downloaded app in the US, Netflix got into the AI filmmaking business, and a senior OpenAI researcher decided he’d rather work for the competition.

A lot happened.

So……. Let’s get to it, shall we?

Here’s what we’re covering today:

• OpenAI launches GPT-5.4 with Pro and Thinking versions

• Claude’s user growth explodes after the Pentagon deal fallout

• OpenAI’s VP of Research quits and joins Anthropic

• Netflix acquires Ben Affleck’s AI filmmaking company

• How to use Claude to write a cold email that actually gets a reply

• Trending AI tools you should know about

  ---

  

1. OpenAI launches GPT-5.4 with Pro and Thinking versions

OpenAI released GPT-5.4 this week, pitching it as its most capable and efficient model yet for professional work.

It comes in three flavors: a standard version, a reasoning-focused Thinking version, and a high-performance Pro version.

The API supports context windows up to 1 million tokens, the largest OpenAI has ever offered.

The Breakdown: Beyond the raw performance numbers, two things stand out.

First, GPT-5.4 is 33% less likely to make errors in individual claims compared to GPT-5.2, which is a real improvement if you use AI for anything that requires accuracy.

Second, OpenAI introduced a new Tool Search system that changes how the model handles tool calling. Instead of loading every tool definition upfront, which eats into your token budget, the model now looks up tools only when it needs them. That's a smart efficiency fix, especially for developers building complex AI agents.

Why it matters: This release is not just a capability upgrade. It’s OpenAI signaling that it’s building for professionals and developers who need reliability and scale, not just general consumers. The benchmark results in law and finance specifically are worth paying attention to. That’s a market with serious money in it.

Read full story

2. Claude's user growth explodes after the Pentagon deal fallout

After Anthropic’s public standoff with the Pentagon, Claude became the number one app on the US App Store and is holding that position.

Daily downloads hit 149,000 on March 2, ahead of ChatGPT’s 124,000.

Daily active users on mobile reached 11.3 million, up 183% since the start of the year. Paid subscribers have doubled.

The Breakdown: Anthropic took a hit politically. The White House called them out, the Pentagon labeled them a supply-chain risk, and it looked bad in the short term.

But consumers responded in the opposite direction.

When Anthropic CEO, Dario Amodei drew a line around mass surveillance and autonomous weapons, a significant chunk of the public rewarded that stance with their attention and their downloads. That’s unusual in tech, where principle usually costs you.

Why it matters: It shows that AI companies are now operating in a trust economy, not just a features economy. Claude’s surge isn’t purely because it got better as a product. It happened because Anthropic made a values-based decision in public and people noticed. Whether that growth sticks long term depends on whether the product keeps up with the momentum.

Read full story

3. OpenAI's VP of Research resigns and joins Anthropic

Max Schwarzer, who led post-training research at OpenAI and worked on improving reasoning across several of the company’s recent models, has left to join Anthropic. His work was directly tied to making OpenAI’s models smarter and more scalable.

The Breakdown: Post-training is where a lot of the real differentiation happens in AI models. It's the stage that shapes how a model reasons, how it handles edge cases, and how well it scales.

Losing someone who worked on that at a senior level is not trivial. And the fact that he went specifically to Anthropic, right as Claude is having its biggest growth moment, makes the timing interesting.

Why it matters: Talent movement in AI tells you a lot about where people think the momentum is. This is one researcher, but it adds to a broader narrative that Anthropic is becoming a more attractive destination for serious AI researchers, not just a principled alternative to OpenAI.

4. Netflix acquires Ben Affleck's AI filmmaking company InterPositive

Netflix has acquired InterPositive, an AI filmmaking technology company that Ben Affleck founded in 2022.

The company built a model that helps production teams work with their own footage to fix post-production issues like continuity errors, lighting problems, and background adjustments. Affleck is joining Netflix as a senior adviser.

The Breakdown: What's notable here is what InterPositive is not doing.

It's not building AI actors or synthetic performances. It's building tools that give human editors more control, not less.

That's a deliberate framing, and Netflix leaned into it in their statement too.

The acquisition is partly a product move but it's also a PR move.

Netflix is telling filmmakers, crew members, and critics that its approach to AI is about empowering the creative process, not automating it away.

Why it matters: Hollywood has been watching how streamers approach AI very closely, and this deal sends a message. Netflix is essentially paying to be on the right side of that conversation. Expect other studios and platforms to make similar moves in the coming months to stake out their position before the debate gets louder.

Read full story

HOW TO USE CLAUDE TO WRITE A COLD EMAIL THAT ACTUALLY GETS A REPLY

Cold emails fail for one reason.

They are written for the sender, not the recipient. Claude can help you flip that.

Here is a simple workflow that takes about 15 minutes and produces a cold email worth sending:

Step 1. Define exactly what you want and who you're sending to

Context setup prompt:

“I want to write a cold email to [RECIPIENT NAME or ROLE] at [COMPANY]. My goal is to [SPECIFIC OUTCOME, e.g. book a 20-minute call, pitch a collaboration, apply for a role]. Here is what I know about them: [ANYTHING RELEVANT]. Here is a bit about me: [YOUR BACKGROUND]. Keep this in mind as I build out the email.”

Step 2. Build the hook

The first line decides whether the email gets read or deleted. Ask Claude to give you several options so you can pick the one that fits your tone.

Opening line prompt:

"Write 5 different opening lines for this cold email. Each one should be specific to the person, relevant to what I want, and short enough to read in under 5 seconds. Avoid starting with 'I hope this finds you well' or anything like it. No flattery, no fluff."

Step 3. Write the full email

Once you have an opening you like, ask Claude to write the full thing. Keep the target length in mind because shorter is almost always better with cold outreach.

Full email prompt:

“Now write the full cold email using opening line number [X]. Keep the total length under 120 words. Make the ask clear and easy to say yes to. End with one specific call to action. Do not use jargon or buzzwords.”

Step 4. Stress-test it

Before you hit send, have Claude read it back as the recipient and tell you honestly whether it lands.

Recipient review prompt:

“Now read this email as the person receiving it. What would their first reaction be? Is the ask clear? Does anything come across as pushy, vague, or self-serving? Give me honest feedback and then rewrite any weak parts.”

That's the workflow.

Four steps, one solid cold email.

The difference between Claude doing this and doing it yourself is that Claude won't let you get away with vague language or a weak ask. It will call it out if you ask it to.

TRENDING TOOLS

Five tools worth looking at this week:

1. Claap AI

An async video tool that records meetings, generates transcripts, and lets your team leave comments at specific timestamps. Good for teams that are tired of sitting through calls that could have been a short clip.

2. Dokie AI

An AI presentation builder that turns your notes, docs, or prompts into clean slide decks. It focuses on keeping the layout professional without you having to touch a single design setting.

3. Kane AI

An AI-powered testing tool built for software teams. It writes and runs tests based on plain English descriptions of what your app should do. Useful if you want test coverage without spending hours writing test scripts.

4. Luma AI

Luma generates high-quality video from text prompts or images. It has become a go-to for creators who want cinematic-looking video clips without a production budget.

5. Replit

An AI-powered coding environment that lets you build, run, and deploy apps directly in the browser. Great for non-developers who want to ship something without setting up a local environment.

Thank you for reading today’s edition. See you next time!

Happy Monday and welcome back to The Breakdown AI.

Big week in the AI world.

OpenAI is printing money, Sam Altman shook hands with the Pentagon, and Elon Musk said some things in a deposition that... well, we’ll get to that. Grab a drink and let’s get into it.

Here’s what we’re covering today:

• ChatGPT hits 900M weekly active users

• OpenAI raises $110B in a historic funding round

• Sam Altman strikes a Pentagon deal with guardrails included

• Musk goes off on OpenAI in a deposition

• How to use Claude to prep for a job interview in any industry

• Trending AI tools you should know about

THIS WEEK IN AI

1. ChatGPT reaches 900M weekly active users

OpenAI announced that ChatGPT now has 900 million weekly active users, up from 800M just a few months ago, plus 50 million paying subscribers. January and February are both on track to be the biggest months for new subscribers in the company’s history. At this rate, the 1-billion mark isn’t far off at all.

Read full story

2. OpenAI raises $110B in one of the largest private funding rounds ever

Amazon, Nvidia, and SoftBank collectively put $110 billion into OpenAI at a $730B valuation. The deal comes with massive infrastructure partnerships and the round is still open, meaning more investors are expected to join. For context, OpenAI’s previous round was $40B at a $300B valuation. That was just last year.

Read full story

3. Sam Altman announces a Pentagon deal with safety protections built in

OpenAI has signed an agreement with the Department of Defense to run its models on classified networks, something Anthropic recently couldn’t pull off. Altman says the deal explicitly bans domestic mass surveillance and autonomous weapons, and he’s asking the Pentagon to offer the same terms to every AI company.

Read full story

4. Musk goes after OpenAI in deposition and says nobody died because of Grok

In a newly released deposition for his lawsuit against OpenAI, Elon Musk attacked ChatGPT’s safety record and praised xAI by comparison. The problem is that Grok recently generated nonconsensual nude images, some reportedly of minors, which is now being investigated by the California AG and the EU. The jury trial is expected next month.

Read full story

MY TAKE:

OpenAI is having a genuinely insane week. $110B raised, 900M users, and a Pentagon deal that nobody expected to come with actual safety language.

The interesting part is how Altman handled the whole Anthropic situation. Anthropic stood firm on its principles, got into a public fight with the White House, and walked away with nothing. Altman came in quieter, negotiated his terms, and walked out looking like the reasonable one. Whether you agree with the outcome or not, it was a smart play.

The Musk deposition is something else entirely. Saying nobody died because of Grok, right as your own platform is under investigation for generating CSAM, is a statement that did not land the way he intended. The lawsuit against OpenAI has been going on for a while now and it keeps getting messier.

Zooming out a little, the bigger story here is scale. When a company raises $110B and is weeks away from a billion weekly users, we’re past the era of AI being a research project. This is infrastructure now. The real competition isn’t about who has the best model anymore. It’s about who can build fast enough to serve the world.

HOW TO USE CLAUDE TO PREP FOR A JOB INTERVIEW IN ANY INDUSTRY

Most people walk into interviews either overprepared on generic questions or underprepared on the stuff that actually gets asked. Claude fixes both problems fast. Here is a simple way to use it before your next interview.

Step 1. Give Claude your context

Start by telling Claude exactly what you are walking into. The role, the company, and your background. The more specific you are, the better it performs.

Context prompt

"I have an interview for a [JOB TITLE] role at [COMPANY NAME]. The company is in the [INDUSTRY] space and here is the job description: [PASTE JD]. My background is [BRIEF SUMMARY OF YOUR EXPERIENCE]. Keep this context in mind for everything I ask you next."

Step 2. Get the likely interview questions

Once Claude has the context, ask it to predict what you will actually be asked. Not just the boring generic list but questions specific to the role and company.

Question prediction prompt

"Based on everything I just shared, give me the 10 most likely interview questions they will ask me. Include a mix of behavioral, technical, and situational questions. For each one, briefly explain why an interviewer for this specific role would ask it."

Step 3. Practice your answers out loud with Claude

This is where it gets genuinely useful. Paste one of the questions back to Claude, write your answer the way you would actually say it, and ask Claude to give you feedback.

Answer feedback prompt

"Here is the question: [QUESTION]. Here is how I would answer it: [YOUR ANSWER]. Give me honest feedback. Tell me what landed, what was vague or weak, and how I can tighten it up. Then show me a stronger version of my answer using my own experience."

Step 4. Prepare smart questions to ask them

Interviews go both ways and most candidates blank on this part. Claude can help you show up with questions that actually make you look like you did your homework.

Questions to ask prompt

"Based on the role and company, give me 5 thoughtful questions I can ask the interviewer at the end. Make them specific enough to show I understand the industry and the role, not generic questions they have heard a hundred times."

That is the whole workflow. Context in, questions out, practice your answers, walk in prepared. Takes maybe 30 to 45 minutes and most people who do this show up noticeably more confident than candidates who wing it.

TRENDING TOOLS

Three tools worth checking out this week.

1. Perplexity AI

Think of it as Google but with a brain. Perplexity gives you direct answers to your questions with cited sources pulled in real time. Great for research when you want answers, not ten more tabs to open.

Try it

2. Gamma

Give it a topic or paste in some notes and it builds you a solid presentation in seconds. Not exaggerating. This will save you a serious amount of time the next time you need a deck.

Try it

3. Elicit

An AI research assistant built specifically for academic papers. It helps you find, summarize, and compare studies without spending hours digging through Google Scholar. Useful if you want the actual data behind the headlines.

Try it

Thank you for reading today’s edition. See you next time!