You needed developers, infrastructure, technical co-founders, runway, a full team, just to build a working product that nobody wants. The barrier to entry was steep, and most people never made it past the idea stage.

AI changed that. Or at least, it’s changing it. Quietly, then all at once.

Today, you can have a working product in a weekend. With a database, user authentication, payment integration, and a decent UI. Most will argue that this is not scalable, but remember, you need at least one person to use your product first before you worry about scaling.

Lovable writes the code and claude helps you think through the logic. Which means a 32-year-old alcoholic can ship something that would have required a six-person engineering team three years ago.

And personally? I could start 10 companies per day right now if marketing and distribution were not a concern. Okay… I’m being dramatic, maybe 5. That’s still a lot, maybe 2?

I digress.

And Then You Finish Building…

You’ve shipped, the product works. You’re proud of it. You’re telling your nonchalant cousin about it, and funnily, this time, he seems chalant, then he asks, “So how do you get people to use it?”

And then it hits you.

“Oh. I never thought about that”, you naively replied.

And in a wave of sudden realisation, it dawned on you:

You need people to know this thing exists. Which means you need to reach people. Which means, God forbid, you need to market this thing.

Suddenly, the overwhelming list materialises in front of you like a curse.

• Social media presence… okay, which platforms? All of them?

• Content strategy… so you have to post consistently?

• Video… wait, you have to be on camera? Ha, please don’t tell me you have to dance on TikTok?

• Build in public… meaning tell strangers on the internet what you’re doing every day?

• SEO… you need to write articles that Google likes, but nobody will read?

• Paid ads… so you need to spend money to get users before you even know if they’ll pay you?

The product was the easy part. Nobody tells you that until you’ve already built the product.

Let’s Picture Some Dude Named ‘David’…

You know a David. Everyone knows a David.

The random dude with “Marketing Strategist | Growth Hacker | Brand Storyteller” in his LinkedIn bio. David talks a lot. He has strong opinions about ad funnels. And has a 2019 case study he references in every conversation.

David seems very confident.

You hire David. You give David $5,000 (your severance package), some creative freedom, and your genuine hope is that this will work.

David posts some things. David runs some ads. David sends you a report with a lot of numbers that somehow say nothing. David talks about “brand awareness” when you ask about customers. David has a theory about why the campaign underperformed that is very detailed and completely incomprehensible.

Zero customers. $5,000 gone. David is already pitching his next client and writing a case study about how he took your business from 0 to 100K in website traffic, leaving out the part where you got 0 sales.

The game is the game.

So What’s the Problem…

If building was hard before and is easier now, what happens next?

Everyone builds, obviously. Maybe not ‘everyone’, but still, the floodgates open. The number of products, apps, tools, platforms, SaaS startups, and AI wrappers launched per week will become genuinely overwhelming. It’s already getting there.

Which means the human you’re trying to reach is going to be absolutely suffocated with options.

Distribution gets harder when supply explodes. Because now you’re not just competing with the ten other companies in your category, you’re competing with every founder who woke up last Saturday and decided to build something.

So What Does This Actually Mean?

Unlike building a product, where you can genuinely get away with just Lovable and a weekend, marketing doesn’t have that equivalent yet. There is no “prompt your way to a successful GTM strategy” tool. No AI that can take your budget, shoot your videos, post your content, run your ads, read what’s working, adjust, and do it all over again, end to end, unsupervised. That doesn’t exist, yet. Or maybe I don’t know AI enough?

But my prediction is that the value of good marketers will go up in the near future.

Because the scarce resources will be people who can take what was built and make the world care about it, people who understand distribution, people who can create attention in a crowded room, people who know how to reach humans who are tired of being marketed to, and somehow still reach them.

Not that dude called David.

The world is about to be full of products and short on people who know how to move them.

If you’re already a good marketer, this is your moment. And if you’re thinking about what skill compounds in a world where everyone can build overnight… Just know, distribution is the new bottleneck.

This is the first monthly edition of the newsletter. And to be honest, I was looking forward to this one. I’ve been away for a big part of this new year, and it seems like for every hour I sleep, there’s something new happening in the world of AI, and I’ve not really had time to sit down and run through the news. So here I am. And by the way, this one is a long read, not sure I can particular tag it “in a nutshell”, but I’ll try 😃

What You Must Know

• Anthropic Told the Pentagon “No” and Got Blacklisted → This was the biggest story of the month, and maybe the biggest AI story of the year so far. The Pentagon demanded that Anthropic remove safety guardrails from Claude for military use, specifically for autonomous weapons and mass surveillance of US citizens. Anthropic refused. Trump got pissed and ordered all federal agencies to immediately stop using Anthropic. Trouble galore.

• OpenAI Signed the Same Pentagon Deal… with the Same Safety Rules → Hours after the Anthropic ban, Sam Altman announced OpenAI had secured a deal with the Pentagon for classified military systems. The confusing part was that the agreement reportedly includes the exact same guardrails Anthropic insisted on: no mass surveillance, human control over weapons. The internet had a field day with this. Dario got a blog post. Sam got the contract. People bashing OpenAI on Twitter. Interesting times 😎

• Claude Hit #2 on the App Store After Getting Banned → In possibly the greatest accidental marketing campaign of 2026, Claude surged to #2 on the App Store overnight after the supposed ban. Turns out, people now trust Claude more because of their honesty, and people now hate OpenAI more. At least, based on what Twitter is saying.

• OpenAI Closed the Biggest Private Funding Round in Tech History ($110 Billion) → I had to read this number multiple times. To put it in perspective, this single round is larger than the GDP of most countries. No wonder they needed that Pentagon deal 😄

• Claude Cowork Plugins Triggered a $285B Stock Crash (the “SaaSpocalypse”) → Back at the start of the month, Anthropic released 11 plugins for Claude Cowork, and software stocks lost $285 billion in a single day. The term “SaaSpocalypse” trended all month. The fear is that AI agents will replace entire SaaS products rather than just assist them. Companies like Salesforce and Intuit took double-digit hits. It’s early days, but Wall Street appears to be taking the threat seriously.

• Jack Dorsey Fired ~4,000 People and Said AI Did It → Block (the company behind Cash App) slashed roughly 40% of its staff, and Jack Dorsey didn’t sugarcoat the reason. He posted a memo on X explicitly saying AI has “fundamentally changed what it means to build and run a company.” Their stock went up 17%. If this doesn’t signal where things are heading, I don’t know what does. And my concern is, he just set the pace for other companies to follow.

• Karpathy Said “The Programming Era is Over → Andrej Karpathy (ex-Tesla AI, OpenAI co-founder) posted a thread describing how he built a full video analysis dashboard by giving AI agents instructions in plain English. It was done in 30 minutes with zero manual coding. His exact words: “You’re not typing code anymore, you’re spinning up AI agents, giving them tasks in English, and managing their work in parallel.” He called it “agentic engineering.” I know that whether I agree with the framing or not, the trend is undeniable.

• Anthropic Raised $30B at a $380B Valuation → Despite everything that happened later in the month, Anthropic closed a $30 billion fundraise at a $380 billion valuation, making it one of the most valuable private companies ever. It’s interesting that the same company that was blacklisted by its own government is also one of the most-funded AI companies on the planet.

• Perplexity Launched “Computer” → Perplexity dropped a new AI product called “Computer” that can browse the web, complete multi-step tasks, and interact with websites on your behalf. From my use of it, it sits pretty close to Claude Cowork and Openclaw.

What’s Good to Know

• Samsung Galaxy S26 Ships with Gemini Agents → Samsung launched the Galaxy S26 with (as they claimed) deeply embedded Google Gemini agents that can manage settings, schedule tasks, and control third-party apps. I’m not really motivated, considering how sloppy Siri or Apple Intelligence has been within iOS.

• Anthropic Unveiled Enterprise Agent Integrations → Claude Cowork now integrates with Slack, Intuit, DocuSign, and Gmail. The framing was that AI will augment enterprise tools, not replace them. Twitter peeps didn’t fully buy it, but it partly helped software stocks bounce back.

• Google Released Nano Banana 2 → Google dropped a lightweight image generation model designed for on-device use. The idea is real-time AI image creation on your phone without needing the cloud.

• A Study Says AI is Making Work Harder, Not Easier → Some Harvard Business Review study claimed that AI tools are actually making knowledge workers busier. Employees report spending more time managing and verifying AI outputs than on the tasks AI was meant to automate. I beg to differ. I agree that you’ll spend more time rambling around with AI if you don’t know how to use it, but that’s just a human skill issue, not an AI issue, in my opinion.

• Anthropic Acquired Vercept → Anthropic bought Vercept, a startup focused on AI-powered document understanding, signalling a deeper push into enterprise automation.

Kaizen Corner → Just my honest opinion

I’ve been writing this newsletter for 7 months now, and February 2026 was the first month when everything felt… different.

And I’m not saying this because the models got smarter (they did, but that’s normal now). But the consequences are beginning to look more real.

The repeated pattern I used to joke about (“new model, China did something, Europe regulates, AI Godfather warns us”) is gradually being replaced by something much harder to summarise.

I don’t really have a neat way to explain it. I just know that when I sat down to write this edition, it felt deeper than usual. And I spent so much time going down the rabbit hole. Because I need to really understand this AI thing, I don’t know about you.

Nevertheless, stay curious and stay employable 😎

Here’s a gem I found that I think can be helpful → https://anthropic.skilljar.com/

Meme of the Month

That’s your month in AI.

If you learned something, tell a friend. And if you didn’t, blame yourself :)

Until next month,

Kay - your fellow human.

P.S. If this email lands in spam, that’s your inbox trying to stop you from staying plugged in. Fix it.

Quick Note: I’m not an expert, I’m just a curious normie navigating the world of AI (like you). I used to do this every Sunday morning before the kid wakes up. Now it’s once a month, but the ritual is the same: I go online, read about AI, and write down everything I found for anyone curious enough to read it. Why? Because I learn faster and understand deeper when I have to put it in writing. You’re welcome.

The project basically exploded overnight. People were buying Mac Minis just to run it. Tech Twitter was going crazy. And I couldn’t resist the urge to give it a spin.

Now… before I jump into something like this, I prefer to do a quick background check on everything related to the project. And below is what I found out. If you don’t like gossip, you can skip this next section :(

What I Found Out

OpenClaw is, more or less, an AI personal assistant with access to nearly anything you give it. It doesn’t just chat with you but can control your browser, read your emails, manage your calendar, execute shell commands, and you interact with it through WhatsApp, Telegram, Discord, wherever you already chat.

It was built by Peter Steinberger, a technical degen who makes shipping feel like drinking wine. Going through his profile, I can say he’s “that guy”. What really caught my attention was how much of a CLI and terminal lover he is. Everything he touches has this command-line vibe to it.

Anyway, late last year, he dropped the open-source project called “Clawdbot”.

The name “Clawdbot” came from that little claw monster you see when Claude Code is loading.

The project went viral. 60,000+ GitHub stars in like three months. You’d think it’s free marketing for Anthropic because everyone wanted to use Claude as the underlying model.

But Anthropic’s lawyers didn’t think so. So they sent a polite email: “Hey, your name is too similar to Claude. Trademark stuff. Please change it.”

Now, Steinberger thought he’d checked this. In a podcast three days before the rename, he said, “I looked it up, there’s no trademark for this”. But apparently “Clawd” sounded too close to “Claude,” and trademark law is trademark law.

So the man had to rebrand everything.

Clawdbot became “Moltbot” (because lobsters molt when they grow - honestly, pretty clever). The mascot “Clawd” became “Molty.” New website, new GitHub repos, new everything.

And of course, because this is the internet, during the renaming and account migration, some scammer managed to take over or squat related GitHub and X (Twitter) accounts tied to the project and started pushing fake “Clawd” meme coins. It was chaos for about 72 hours.

Eventually, after more drama, the project landed on “OpenClaw” - which is actually a much better name in my opinion.

My Experience With It

I always set out time around ungodly hours to play with new stuff like this.

At first, I wanted to install it locally on my computer, but then I thought... running an open-source AI bot on my computer is a bad idea. In fact, running it on any computer with actual data is a bad idea.

So I decided to spin up something isolated.

I fired up Render because Render is just too easy for these kinds of experiments. But I encountered a weird deployment error that seemed simple, but I was too lazy to look into it.

I gave up on the first day and moved it down my to-do list, setting the due date to February 7. We live to fight another day.

Two days later, I got an email from DigitalOcean saying it’s now available in their marketplace as a one-click install.

I thought: “This is interesting.”

So I opened my DigitalOcean account and followed this guide.

I don’t know if it’s because I’m an idiot, but it took me almost 3 hours to get this thing to work properly.

What OpenClaw Does

Someone called it “AI with actual hands”, so I think I’d just leave it at that.

1. It runs on a machine. It runs on your machine - Mac, Windows, Linux, doesn't matter. Or throw it on a VPS if you want it running 24/7 in the cloud. You can use Claude, GPT, Gemini, or even local models.

2. You talk to it through your normal chat apps. WhatsApp, Telegram, Discord, Slack, Signal, iMessage - pick your poison.

3. It has persistent memory. It builds a profile of you over time in markdown files that live in your workspace. So it remembers everything. Although I’m not quite sure how this will scale over a long period of time.

4. Extended system access. It can browse the web, fill out forms, scrape data, read and write files, run shell commands, and execute scripts.

5. Skills and plugins. The community has built 50+ integrations. Weather, GitHub, Notion, Slack, Google Calendar, Spotify, even controlling Philips Hue lights. You can install more from ClawHub, or build your own. The AI can even write its own skills.

6. Proactive notifications. It doesn’t just wait for you to ask questions. It can send you reminders, check in on you, run cron jobs, and handle background tasks.

The architecture is actually pretty clever. You’ve got the OpenClaw Gateway running 24/7, handling all the routing between your chat apps and the AI model. Everything goes through there.

Here’s What I Think

After spending time with this (I must confess, not a ton of time to be writing an opinion piece). Regardless, here are my unsolicited thoughts:

1. It’s technical. Only really adoptable by averagely technical people. Too many dependencies, too much setup, not plug-and-play at all. The DigitalOcean one-click install even required you to understand basic SSH and all this terminal drama. An average person will find this frustrating.

2. I don’t trust the security. And I say this as someone who was skeptical about connecting anything of value to it. If I have to set up a new email, a new WhatsApp, and a new everything just to use this safely… then we’re not quite there yet.

3. It can be very useful for certain people. If you’re a developer or someone technical who wants a truly personal AI assistant that you control completely, this is probably the closest thing to that vision right now. The WhatsApp integration alone is pretty magical when it works.

4. It offers a glimpse of what the future holds. Even if OpenClaw itself doesn’t become mainstream, it’s showing what’s possible. This is the “GitHub moment” for personal AI assistants - the open-source proof of concept that proves the category exists.

5. It chews API credits faster than anything. This is a real cost consideration. Because it’s always on and processes a lot with AI. So this kinda watered down my excitement, but it’s an AI assistant, it’s expected.

My Personal Conclusion

I think OpenClaw is ahead of its time, but it’s already paved the way for similar technology.

We can give it about 12 months. Either we’ll see more polished alternatives pop up (probably already happening), or one of the big AI companies - OpenAI, Anthropic, Google - will release something similar that people will naturally trust more and that will actually be usable for everyday people.

For now, if you’re technical and you love tinkering with this stuff, OpenClaw might be worth playing with. It’s genuinely impressive what one person built and what the community has added to it.

But for my day-to-day productivity? I’ll just stick to my simple dashboard.

P.S. - If you do try OpenClaw, run it on a completely isolated machine or VPS. Don’t be that person who gives an AI access to your main computer and then acts surprised when your crypto wallet is emptied :)

To me, the premise is simple enough: If you’re using ChatGPT for free, you’ll start seeing ads. If you’re paying for a subscription, no ads.

According to the tweet, the promise is that the ads won’t influence ChatGPT’s responses, that conversations will remain private from advertisers, and that everything will be clearly labeled.

Honestly… I’ve been expecting this…

OpenAI’s business model has always felt like a race to the bottom. They’re burning through cash faster than they can raise it, which is not sustainable. But I think it takes running a business to understand that things of value aren’t sustainable if they’re not generating enough revenue, unless you’re running a charity.

Then, my concern is… They’re only showing ads to free users. And free users are probably the worst group to show ads to. It’s like trying to monetize the people who’ve already told you they don’t want to pay for your product. But this is not uncommon; Spotify does the same.

Now, I know everyone’s losing their minds about ads “ruining” ChatGPT, but personally, I love seeing highly relevant ads. I mean ads that actually help. I’ve found some genuinely good stuff through Instagram ads, which I wouldn’t have discovered otherwise.

Plus, think about it: the conversations we have with ChatGPT are insanely high-intent. You're asking how to fix your car, plan a trip, learn to code, start a business. ChatGPT knows exactly what you need in real-time. Delivering targeted ads in that context is almost too easy.

Safe to say… If OpenAI can figure out how to show me helpful ads, not just random intrusive noise, I don’t have a problem.

What I’m really curious about… is whether other LLMs will follow the same approach.

If OpenAI can make ads work without destroying the user experience, we might see everyone else sprint toward the same direction. I know Google and Meta will love this idea, hehe, they are the gods of ads already. Anthropic will probably resist the longest, given their positioning around ethics and AI safety, but I’m not sure they’ll be able to run faster than the math eventually.

Nevertheless, we’ll see if OpenAI can pull it off without turning ChatGPT into the digital equivalent of Times Square.

My good guess is… They’ll figure it out, because they have to.

Last year, I found Missive (thanks to Linda), which does a really good job with the email management part. But it doesn’t have other things I need, like taking notes, or setting reminders, or any useful AI functionality. So I’m forced to use Evernote and combine it with Notion.

I’ve tried a lot of other tools, and I mean a lot, but none of them actually gives me everything I need precisely. And of course, why would they? They didn’t build the software for me.

This January, like I do every January, I was decluttering and prepping for the new year. Unsubscribing from emails, cleaning up my slack, deleting random notes I don’t remember writing. I was moving from Missive to Evernote to Notion.

Then I had a wild idea.

Why not just build exactly what I need? No more, no less.

Now, you’d ask, what do I even need?

Simple: A single dashboard that has tasks, resources/links I want to keep close, notes, reminders, a video gallery, and AI functionality like a daily briefing. Oh, and it needs to connect to my WhatsApp so I can capture things on the go.

I wrote out the rough draft of what I wanted.

Initially, I was going to do it on Windsurf. Then I thought of every pain I’d have to go through to get this thing working properly: frontend, backend, database, server, deployment. It’s going to be a lot of work, and although I was ready for the work, the wiser version of me thought... what if it could be easier?

I’ve used Lovable before, but only to build UI prototypes and play around with a fancy Netflix clone.

So I did some digging. The comments on Reddit weren’t even encouraging; people had mixed opinions. One guy even said, and I quote, “Only fools will think they can build anything useful with Lovable.” 😂

But I decided to go through Lovable’s documentation and saw that they have the entire ecosystem built in. The server, the AI, the database, everything.

“This should make everything easier”, I said to myself.

Then I concluded, “let’s give Lovable a shot”.

Now came the UI part. I didn’t want an ugly, scattered UI. I wanted it dark, simple, without any fancy elements. I love the UI of linear.app, so I figured I’d just copy their design.

Luckily, a good Samaritan published a Linear UI design kit on Figma. I duplicated that, plugged Figma MCP with Windsurf, and told Windsurf to give me the instructions I should give Lovable to achieve something close to that design. The lovable design theme functionality also made it smoother.

Then I got to work.

What I Built…

After 14 hours, I had a fully functional personal dashboard. Here’s a demo video:

I called it “Odogwu”, which is slang for “champion” or “big man”. Because why not?

The dashboard has:

Tasks → with drag-and-drop reordering, priorities, and the ability to tag tasks to different workspaces (which is something I was unable to do with Missive).

Notes → with rich text editing, tags, and reminders. If I mark a note with “remind me”, it’ll show up in my daily briefing and send me a WhatsApp message as a reminder.

Resources → for saving links and bookmarks I want to keep close.

Videos → a Netflix-style gallery for YouTube, Instagram, and TikTok videos I plan to watch. Once I mark them as watched, they disappear from the UI but stay in the database.

Calendar Integration → pulls from my Google Calendar, shows the next 7 days of meetings, detects clashes, and even auto-tags meetings based on attendee email domains.

Daily Briefing → AI-generated summary of the day, with tech news from the past day.

Just exactly what I need for now, and if anything else comes to my mind, I’ll add it, because the world is mine.

The WhatsApp Magic (and Frustration)

This was the feature I was most excited about, and also the most frustrating.

The idea is simple: I’m walking somewhere, and I remember something. Instead of opening a dashboard, I can just send a WhatsApp message like “I need to call my mum by 2 pm tomorrow”, and the AI adds it as a task and sends me a reminder at 2 pm the next day.

It worked.

And when I logged in to the dashboard, the tasks I'd sent on WhatsApp were already in the task list.

But here’s where the frustration came in.

To do anything with WhatsApp API, Meta requires this annoying Meta Business verification. I was hoping to avoid it, because this is just a simple personal project.

But my plan didn’t work. After 3 hours of suffering and trying, I eventually settled on using the Twilio sandbox number, which doesn’t require verification but does require me to reconnect it every 3 days by sending a message like “join kay-odogwu” to the Twilio WhatsApp number.

It’s annoying, but I’ll leave it that way for now.

What Surprised Me About Lovable

I didn’t need any deep prompt engineering to get Lovable to understand what I wanted. The prompts were very loose and lazy.

For example, when I wanted to add the videos module, this is exactly what I sent:

“Let’s add a videos module. It’ll take only either YouTube link, Instagram or TikTok link. Then we’ll have a page for it as videos I need to watch. It’ll be a fancy video gallery display, like Netflix. It’ll have a checkbox to mark it as completed. Once I mark it completed, remove it from the UI, but keep it in the database. If I send any YouTube, Instagram or TikTok link to the WhatsApp via Twilio, just know to add it as a video.”

It took a few corrections, but it worked just fine. I think this is what John meant when he said “there’s a new programming language, called ‘English’”.

Other Hiccups Along the Way

The WhatsApp situation wasn’t the only challenge. A few other things came up:

News scraping issues → I wanted the daily briefing to include tech news headlines. The initial implementation kept returning image URLs instead of actual article links.

AI response parsing → Sometimes the AI would wrap its JSON responses in markdown code blocks, which broke things.

Google Calendar authentication → This required some manual work with JWT creation and cryptographic signing. Not the most fun part.

Some points to reduce frustration:

• My rule of thumb is: once I've gone back and forth with it more than three times and it still hasn't fixed the issue, I’ll open the code and try to fix it myself. Which means having a basic understanding of programming makes it easier, but it's not required.

• Reference other sections/pages/files where something you’re trying to do has already been done properly e.g “On the videos page, the filter and button should be beside the search bar, similar to how it is on notes page.“

• If you have any UI reference or screenshot, attach it; it’s faster.

A Quick Disclaimer

This kind of project would require a much deeper look if I were pushing it out for public use, especially around optimisation, security, and scalability. But for a personal project? It’s solid. It does what I need it to do, and that’s all that matters.

The Takeaway

Interestingly, we’re in an era where you can build your own tools. Not just theoretically, but practically.

If there’s something you’ve been wishing existed, something that combines features from 5 different apps into one, you might be closer to having it than you think.

Sometimes, your best productivity tool is the one you build yourself.

Now, if you’ll excuse me, I need to go send “join kay-odogwu” to my errand boy on WhatsApp.

But this one’s different. No news roundup today. Just me, talking.

When I started this newsletter back in August, I had one selfish reason: I wanted to stay close to AI. Richard Feynman said something about developing an obligation to talk about something if you want to learn it. So I thought, why not write about it?

Every Sunday morning, before my kid wakes up, I sit down and dig through the AI chaos of the week. It’s become my version of what my dad used to do with newspapers, except my newspaper is the internet.

And it worked. I’ve learned more by writing this newsletter than by passively reading ten others. And every now and then, one of you subscribers replies with something like “this helped me” or “I finally understand this now, and I go: “Oh, interesting…”

But I’ve noticed something about the AI news part of this newsletter.

The weekly news has become… predictable.

It’s either:

• A new model just dropped

• China did something

• Europe is regulating something

• Or AI Godfather issued another warning

Rinse and repeat every week.

Don’t get me wrong, important things are still happening. But the weekly cadence is starting to feel like I’m reading the same story with different titles attached.

So here’s what I’ll change in 2026:

My weekly AI news digest will now happen monthly.

This actually gives me time to find things worth writing about. Not just “another model, another benchmark”.

However, I’ll still drop random AI gems whenever I stumble on or think of something interesting.

Thank you for reading. Thank you for subscribing. Thank you for the emails telling me this is useful.

See you in 2026.

And Happy New Year from your fellow human, Kay

One, “this is interesting”.

Two, AI should absolutely not be trusted with the “Run Society” button, yet.

The piece, from Fundamental Research Labs (FRL), was about researchers creating a tiny simulated village populated entirely by AI agents.

It’s like The Sims. But everyone is powered by large language models.

According to what I read, these agents had memories, goals, jobs, and opinions. They formed relationships and planned events.

Which I think is cute. And terrifying. I’d say mostly terrifying, but in a cute way, like if my toddler was holding a knife and saying “I love my daddy” at the same time.

So what actually happened in this AI village?

The researchers dropped these AI agents into a simulated town and let them… exist.

Each agent had its own personality, background, and knowledge. Some were shopkeepers. Some were artists. Some were just there...

Over time, they remembered past interactions, training and adjusted behaviour. And they collaborated without being told.

Basically, they did what humans do.

At one point, the agents collectively decided to organise a party. No one prompted them. They just remembered that humans do this thing called social bonding and went, “Yeah, let’s do that.”

However, the agents sometimes fell into endless loops of polite agreement or chased unattainable goals, typical AI drama.

Also, at times, the AIs were frustratingly independent; they’d ignore requests and say, “I want to do my own thing” to pursue their own agendas. Humans do this, too, though.

The most interesting part…

The emergent behaviour was the most interesting part for me.

The agents weren’t explicitly programmed to create culture. But the culture happened.

Which means when we build systems complex enough, we don’t fully control what they become. We only influence them.

My human takeaway

I think watching AI agents recreate society is interesting, and we’re not even in the AGI era yet. Over a long period, they might be able to do a better job of it than humans would because the problem with human systems isn’t intelligence. It’s incentives, fear, ego, and vibes. Bad vibes, mostly.

If artificial agents can cooperate without these bad vibes, maybe, over time, a society created by 1000 AI agents will beat a society created by 1000 humans?

Or maybe I’m overthinking it, and this is just digital ants doing ant things with better marketing.

Regardless, we’ll be alright.

Summary for extra-lazy people: 1,000 AI agents built their own Minecraft civilisation, Claude launched as a Chrome extension, you can now pin chats in ChatGPT, Google lets you verify AI videos in Gemini, and developers can submit apps to ChatGPT.

What You Must Know

• You Can Now Pin Chats in ChatGPT → ChatGPT now lets you pin chats to the top of your list. It works on iOS, Android, Web, and Atlas.

• 1,000 AI Agents Left to Build Their Own Village → So this company ran an experiment where they dropped 1,000 AI agents into Minecraft and just let them do their thing for days. They built farms, markets, governments, and even religions. Some became leaders, others priests, and a few got corrupt. It was weirdly human, complete with cliques and gossip. The whole point was to see if AI agents could actually work together, but turns out they were too independent. Most users just wanted them to follow orders.

• Claude is now on Chrome as an Extension → Claude now has a Chrome extension. It can see your screen, click buttons, fill forms, basically use the browser as you would. In theory, I find it similar to ChatGPT Atlas, just less capable because it’s just a Chrome extension.

• You Can Now Verify Google AI-Generated Videos in Gemini → Google added a feature to Gemini where you upload a video and ask if it’s AI-generated. It scans for SynthID watermarks and tells you which parts are AI. The issue is that it only works for videos generated with Google’s AI. I tested it with a Sora video, and it couldn’t detect anything because Sora doesn’t use Google’s watermarks. I don’t know if it already exists, but a tool that tells whether any content is AI-generated (with great accuracy) will have a good use case in this age.

• Developers Can Now Submit Apps to ChatGPT → OpenAI opened app submissions for ChatGPT. There’s now an app directory at chatgpt.com/apps where developers can submit apps using the Apps SDK.

What’s Good to Know

• OpenAI Adds Teen Safety Rules to Model Spec → OpenAI updated ChatGPT’s rules for users under 18. Teen safety now comes first, even over being helpful. My first thought was, how would they know who is what age?

• EU Drafts Rules to Label AI-Generated Content → The EU published a draft to mark AI-generated content with watermarks and labels. It’s part of the AI Act, which requires AI providers to mark synthetic stuff and deployers to label deepfakes. I believe some providers already do this e.g Google.

• OpenAI Ships GPT-5.2-Codex for Long-Context Coding → OpenAI released GPT-5.2-Codex for long coding sessions without losing context. It’s optimized for refactors, migrations, and big feature builds. I’ve never been a fan of CODEX cos my first experience with it was annoying, but I’ll keep this in mind.

• Anthropic Makes Agent Skills an Open Standard → Anthropic released Agent Skills as an open standard, as they did with MCP. ICYDK, Skills are folders containing instructions that tell the AI how to perform specific tasks. Microsoft already adopted it in VS Code and GitHub. OpenAI quietly copied the same structure in ChatGPT and Codex CLI. Anthropic just gave it away for the culture.

Money Moves This Week

💰 Neurable raised $35 million to advance its AI-powered brain-computer interface technology.

💰 CuePilot AI raised $1.8 million to enhance its AI capabilities and expand globally.

💰 Mythic Inc. raised $125 million to challenge NVIDIA.

💰 NobodyWho raised €2 million to power on-device AI models.

💰 FINNY raised $17 million to help great advisors help more people.

Kaizen Corner → How to Self-Host n8n if You’re a Lazy Person

Using n8n extensively on its cloud option is kinda expensive compared to hosting it yourself, and that’s aside from data privacy.

Now, there are a lot of ways to self-host n8n, but I always tend to go with the laziest way, which takes less than 5 minutes if you don’t have shitty internet.

The approach is: Using render.com and the n8n Docker image. Just follow this guide: https://render.com/docs/deploy-n8n

Render handles the deployment for you, and you’re up and running without dealing with server configs or terminal commands. It’s the easiest path if you just want to get n8n working without the usual DevOps headache.

Meme of the Week

That’s your week in AI.

If you learned something, tell a friend. And if you didn’t, blame yourself :)

Until next Sunday,

Kay - your fellow human.

P.S. If this email lands in spam, that’s your inbox trying to stop you from staying plugged in. Fix it.

Quick Note: I’m not an expert, I’m just a curious normie navigating the world of AI (like you). Every Sunday morning, I wake up super early (before the kid wakes up), go online, and read about AI, like my dad used to read newspapers back in the day. Then I write down everything new I discovered and turn it into this newsletter for anyone else who’s curious enough to read it. Why? Because I learn faster and understand deeper when I have to explain it in writing. I’m welcome.

Summary for extra lazy people: OpenAI launches GPT-5.2 after declaring “Code Red”, Disney strikes $1B Sora deal with OpenAI, Trump signs executive order to preempt state AI laws, and Google launches deepest research agent yet.

What You Must Know

• OpenAI launches GPT-5.2 → After declaring ‘code red’ as I mentioned last Sunday, OpenAI threw GPT-5.2 at us as its new flagship model family. From what I read on TechCrunch, the focus is better reasoning, longer context handling, and improved tool use. It’s obviously positioned as a response to Google’s recent Gemini upgrades. Tit for Tat 😎

• Google releases Disco, an AI browser-like workspace → Disco is a Google Labs experiment that tries to understand what you’re doing online (researching, planning, learning) and then builds AI tools on top of your open tabs. I had to rewatch the demo like 5 times to fully get the concept, and it’s still waitlist-only 😒

• Cursor adds a visual editor that writes real code from UI edits → Cursor now lets developers visually tweak UI elements (drag, resize, adjust spacing) and converts those changes directly into real source code. I’ve seen mixed reactions online; some love it, others think we should be past drag-and-drop workflows by now. I lean slightly toward the second camp.

• Disney signs a $1B Sora content and investment deal with OpenAI → Disney agreed to license hundreds of its characters to OpenAI’s Sora video model and also committed roughly $1 billion as part of the partnership. So now it’s no longer “media companies vs AI”, it’s “media companies embedding AI”.

• Trump signs executive order to preempt state AI laws → The executive order pushes for a single national AI framework instead of state-level AI laws. Although I read in another article that some policy groups say this could weaken consumer protections. I consider having state-level AI laws a step backward, but what do I know about politics?

What’s Good to Know

• Google launches its deepest AI research agent yet → Gemini Deep Research is now more powerful and available to developers over the API. Google doesn’t seem to be stopping anytime soon.

• McDonald’s pulls AI-generated Christmas ad after backlash → McDonald’s Netherlands removed an AI-generated holiday ad after criticism over its quality and tone. I personally thought the video was dope, but my definition of ‘dope’ seems to be questionable.

• A Reddit user trains TimeCapsuleLLM using only 1800s data → The experiment uses 19th-century London texts to reduce modern societal bias. It’s more of a research curiosity than a product, but still an interesting and creative approach to LLM training.

• Claude Code launches on Android → Anthropic’s coding-focused Claude experience is now available on Android, extending access beyond desktop and web.

Money Moves This Week

💰 AIR raised $6.1 million to modernise real-time credit intelligence with AI.

💰 Harness raised $240 million to automate AI’s ‘after-code’ gap.

💰 Gracia AI raised $1.7 million for AI-powered ultra-photorealistic volumetric videos.

💰 NeoSapien raised $2 million to scale India’s first AI-native wearable.

💰 Medra raised $52 million to build physical AI scientists.

💰 Matta raised $14 million to develop “sentient factory” technology.

🧠 Kaizen Corner → Argument: AI Doesn’t Create New Ideas, It Just Repackages Old Ones?

I saw this argument floating around on Reddit recently. The claim was that: Today’s AI doesn’t really create new ideas, it just makes old ones look new.

The person's logic was that models like GPT predict the next word based on massive amounts of past text, so every output is shaped by what already exists. From that view, AI isn’t inventing, it’s remixing history with very good packaging.

But the counter-argument I saw in the comments was just as interesting. According to the argument, humans don’t create ideas from nothing either. We remix experiences, culture, and things we’ve seen before. That’s why a toddler cannot invent. You need knowledge, experience, and perspectives to be able to invent new ideas.

So my deduction was: No idea is born out of nothing, even if it’s a new one; everyone and everything needs some past data to make a new one. Safe to say, we’re not that different from the machines we’re building.

Meme of the Week

That’s your week in AI.

If you learned something new, tell a friend. And if you didn’t, blame yourself :)

Until next Sunday,

Kay - your fellow human.

P.S. If this email lands in spam, that’s your inbox trying to stop you from staying plugged in. Fix it.

Quick Note: I’m not an expert, I’m just a curious normie navigating the world of AI (like you). Every Sunday morning, I wake up super early (before the kid wakes up), go online, and read about AI, like my dad used to read newspapers back in the day. Then I write down everything new I discovered and turn it into this newsletter for anyone else who’s curious enough to read it.

Why? Because I learn faster and understand deeper when I have to put it in writing. I’m welcome.

Summary for extra-lazy people: OpenAI declares “Code Red” as Google threatens to take over, Mistral unveils open-source Mistral 3 models, Anthropic plans to use AI to interview its own users, and HSBC partners with Mistral.

What You Must Know

• Anthropic Said Their Engineers Now Use Claude for 60% of Tasks → They claimed that internally, employees self-report using Claude in 60% of their work and achieving a 50% productivity boost. In my last article, I asked a question relating to this. You should check it out.

• OpenAI Declares “Code Red” for ChatGPT → OpenAI CEO Sam Altman sent an internal memo declaring “Code Red”, i.e, halting non-essential projects to focus on improving ChatGPT. Why? Google’s Gemini 3 has been hot on the charts, beating ChatGPT on several benchmark tests. I remember years ago when Google declared its own “Code Red” upon ChatGPT's launch. It’s a battle for the fittest 😎

• Mistral Unveils New AI Models → Mistral announced Mistral 3, their most capable model yet, with 41B active and 675B total parameters. I’ve actually never gotten to retry Mistral since their initial launch, and the only thing I remember is that I liked their brand colour 😃. I guess it’s worth checking out now.

• Anthropic Will Use AI to Interview Its Own Users → In a week-long pilot, an AI system will interview users about how they use AI and what they worry about, with each session lasting 10-15 minutes. I find it interesting… using AI to understand how people feel about AI 😎

• HSBC Signs Multi-Year Deal with Mistral AI → HSBC (my favourite bank, actually 🙂) announced a multi-year partnership with Mistral to integrate AI tools across the bank’s operations. I’m guessing they chose Mistral due to the European data-seriousness culture and Mistral’s open-source advantage.

What’s Good to Know

• Runway Unveils Gen-4.5 Text-to-Video Model → They claim it has “unprecedented” accuracy for physics, motion, and visual detail. To put that in non-alien words, it can produce more realistic and consistent clips from prompts.

• Google Releases Deep Think for Gemini → This allows the model to spend more time reasoning through complex problems before responding. Similar to OpenAI’s o1 reasoning models. According to Reddit people, it performs well on difficult tasks that require multi-step thinking.

• OpenAI Partners with Cursor’s Agent to Improve Codex Models → Cursor has become one of the most popular AI coding assistants, and I believe this partnership means OpenAI will get better data on how developers actually use AI for code.

• NotebookLM Expands Character Limit for Chat Customization → From 500 to 10,000 characters. This means you can give much more detailed instructions on how you want the AI to behave when analyzing your documents.

Money Moves This Week

💰 Lumoo raised €6.5 million in angel funding to develop its AI retail and e-commerce platform.

💰 Imper.ai raised $28 million in equity funding to advance its cybersecurity solutions.

💰 Lumia Security Inc. raised $18 million in seed funding to build its AI-powered cybersecurity platform.

💰 Hyperpilot raised £2 million in pre-seed funding to develop its AI automation technology.

💰 Guide Labs raised $9 million in seed funding to build its AI platform.

Kaizen Corner → Unpopular Opinion: Notion’s AI is Underwhelming

I’ve been using Notion every day for years, and when they added AI features, I was excited.

But actually, it’s just... meh.

When it summarises, it’s basic. If you need it to generate a database, it hallucinates. And above all, it’s just TOO EXPENSIVE.

Don’t get me wrong, it works. The problem isn’t that it’s bad. It’s just okay in a world where “okay” doesn’t cut it anymore.

Maybe I’m being too harsh. Maybe the integration is the value, having AI right there in your notes without switching tabs. But honestly? I can achieve the same with ChatGPT Atlas and other AI browsers. Regardless, AI tools should genuinely be better at their jobs, not just conveniently located.

I hope it gets better; it doesn’t necessarily affect my Notion usage. In fact, I’m writing this in Notion :)

Meme of the Week

That’s your week in AI.

If you learned something new, tell a friend. And if you didn’t, blame yourself :)

Until next Sunday,

Kay - your fellow human.

P.S. If this email lands in spam, that’s your inbox trying to stop you from staying plugged in. Fix it.

Quick Note: I’m not an expert, I’m just a curious normie navigating the world of AI (like you). Every Sunday morning, I wake up super early (before the kid wakes up), go online, and read about AI, like my dad used to read newspapers back in the day. Then I write down everything new I discovered and turn it into this newsletter for anyone else who’s curious enough to read it. Why? Because I learn faster and understand deeper when I have to put it in writing. I’m welcome.

On the bright side, it gives people like me a solid excuse to rewrite the whole thing into a compact, relatable summary, which is what we have here. You’re welcome.

So… Anthropic recently dropped a study, titled “Estimating AI productivity gains from Claude conversations”, they analysed 100,000 actual conversations with Claude to answer a simple question: how much time is AI really saving people?.

What they found…

• Without AI, the average task in their dataset would take around 90 minutes.

• With Claude, people got the same work done 80% faster.

• Meaning, many tasks that took 90 minutes could be done in roughly 18 minutes with AI (on average).

• When they broke this down by job type, they found some variations:

  • Some tasks (e.g. preparing school curriculum), which might take humans hours, could be done by AI in minutes.

  • Other, simpler tasks (food prep, basic maintenance, etc.) showed smaller gains or didn’t make a big difference.

• After all their fancy modelling and economic cross-checking, they concluded that if AI like Claude were widely adopted, it could raise annual productivity by 1.8%. For context, that’s double the usual productivity growth rate in the U.S. since 2019.

Important caveats…

The authors (Alex Tamkin and Peter McCrory) were honest enough to say, “Relax, this is not gospel truth”.

Here are some of the limitations they pointed out:

• Their method assumes that all time saved by using AI gets turned into productive work. But in reality, sometimes the time saved equals “let me think about my life for 44 minutes”, which we can’t exactly allocate to productivity.

• Claude’s time estimates are based on AI’s opinion of how long humans spend on tasks. And AI isn't exactly known for being street-smart.

• The analysis doesn’t cover all types of work. It only uses tasks actually done with Claude (so by people choosing to use the AI), which may bias toward tasks that are already good fits for AI, i.e, tasks AI suck at would have taken longer.

• And of course… no research is 100% accurate.

My deductions…

• Yes, AI can significantly speed up work. We already knew this, but it’s nice to see numbers attached instead of vibes.

• But faster ≠ better. If you save 60 minutes only to spend 45 minutes correcting what the AI wrote, the “gain” is just an anger management session.

My big question…

If AI really makes the average person more productive…

I believe the natural consequence of that is that companies will need to hire fewer people.

Which means… As individuals, we celebrate the speed boost → “My work is faster now! AI is saving my life!”

But on the other side of that same coin… → “Ah, fewer people needed. AI is taking our jobs.”

Same technology. Different beneficiaries. One cup, two sips.

I know productivity and employment don’t necessarily move in a straight line.

But the question is: Would this be a fair deal?

Summary for extra lazy people: Anthropic claims Claude boosts productivity by 80% and now has a better context window, Amazon commits $50B for government AI infrastructure, OpenAI integrates voice directly into ChatGPT, Sora gets rate-limited, and both OpenAI and Perplexity launch AI shopping assistants.

Quick Note: I’m not an expert, I’m just a curious normie navigating the world of AI (like you). Every Sunday morning, I wake up super early (before the kid wakes up), go online, and read about AI, like my dad used to read newspapers back in the day. Then I write down everything new I discovered and turn it into this newsletter for anyone else who’s curious enough to read it. Why? Because I learn faster and understand deeper when I have to explain it in writing. I’m welcome.

What You Must Know

• Anthropic Estimates Productivity Gains from Claude → Anthropic released a research analyzing time savings from Claude conversations, they estimated an average 80% speedup on tasks. I’m always skeptical when companies measure their own product’s impact; it’s a “judge and jury” situation 😀, but if even half of this is true, it’s a big deal.

• OpenAI and Perplexity Roll Out AI Shopping Assistants → I saw on Techcrunch that both OpenAI and Perplexity launched AI shopping features inside their chatbots to help users research and compare purchases. I believe “shopping” is an everyday use case, and I previously wrote an opinion piece about why I think AI Agent for everyday use is “not there” yet, but I’m hopeful to see how using AI for shopping will evolve over time.

• OpenAI Integrates ChatGPT Voice → OpenAI has integrated ChatGPT Voice directly into the main chat interface, so we no longer need to use a separate mode. That means we can now talk, view responses, and access earlier messages in real time.

• Claude Can Now Have Longer Context Windows → Everyoneeeee hates Claude’s context window limit. But now, Anthropic has updated Claude to auto-compact older messages and prevent context limit interruptions mid-chat. This means you can have longer conversations without hitting the dreaded “context window full” wall.

• OpenAI Suffers Data Breach Through Third-Party Mixpanel → Within the week, I received several emails about the Mixpanel (an analytics tool) breach. But OpenAI’s was the most worrisome. In their clarification, they mentioned that only basic profile information was exposed. Not a major breach, but still embarrassing for a company that handles so much sensitive data. Now, word on the street is that OpenAI has dumped Mixpanel.

What’s Good to Know

• Alphabet Nears $4T Valuation on AI Optimism → Alphabet’s (the parent holding company of Google) market value is on track to hit around $4 trillion due to the market pricing Google as one of the main long-term winners of the AI race.

• Jony Ive and Sam Altman Say They Finally Have an AI Hardware Prototype → I read on The Verge that Jony Ive and Sam Altman now have a working prototype for OpenAI’s first hardware device, which Ive mentioned will potentially arrive in “less than” two years. The details are still vague, and I'm not sure if it’ll be a playful AI gadget or a phone-shaped device. Given Ive’s design legacy, people are expecting something interesting.

• Sora Gets Rate-Limited from High Demand → OpenAI throttled generation limits on Sora, mentioning huge demand over the holiday period. In simple terms, people are hammering these models so hard that the company had to hit the brakes, which says a lot about current AI usage levels.

• Character.AI Launches “Stories” Mode for Teens After Banning Open Chat → After announcing it would cut off open-ended chats for teens, Character.AI rolled out a new “Stories” feature pitched as a more structured, safer way for under-18s to interact with its AI. Teens get guided, story-like experiences instead of free-form conversations. I’d say it’s better, in my opinion.

• Amazon Plans Up to $50B for US Government AI Supercomputing → Amazon said it could invest up to $50 billion building high-security AI and high-performance computing data centers for U.S. government agencies.

Money Moves This Week

💰 Onton raised $7.5M to expand its AI shopping platform.

💰 Model ML raised $75M to automate junior bankers’ work with AI.

💰 Uxia raised €1 million in pre-seed funding to develop its AI software and SaaS platform.

💰 AskCipher raised an undisclosed amount in strategic funding to advance its AI and SaaS solutions.

💰 Ranketta raised €1 million in pre-seed funding to build its AI and SaaS platform.

💰 Bonus Find → 49 US AI startups have already raised $100M+ in 2025

Kaizen Corner → Lenny’s Product Pass Might Be Worth Checking Out

I found something on Lenny’s Newsletter that caught my attention: Product Pass.

It’s basically a bundle subscription that gives you access to a bunch of product management and tech tools in one package.

For an annual subscription to Lenny’s Newsletter (around $200-350/year), you can get a free year of 15+ products, including things like Devin, Superhuman, Linear, Perplexity, Lovable, and more. The total value is supposedly over $10,000.

I didn’t need it because I already have subscriptions to most of them, but if you use a lot of these tools or want to try them out, this could be a solid deal.

Note: Some products might run out of codes since they’re limited, and you need an annual subscription (not monthly) to access the bundle.

Check it out here if you’re interested.

Meme of the Week

That’s your week in AI.

If you learned something, tell a friend. And if you didn’t, blame yourself :)

Until next Sunday,

Kay - your fellow human.

P.S. If this email lands in spam, that’s your inbox trying to stop you from staying plugged in. Fix it.

It later dawned on me that I was just fooling myself.

My sister already knew how to use GPT better than I thought, and the employees all used it lowkey anyway. The earlier I tuned in, the better.

So… do I think AI makes us forget how to think?

The short answer is yes, but the long answer is no, which is weird because ‘no’ is shorter than ‘yes’. You see what I did there? 😉

Here’s what I think…

Let me start with the obvious: AI is definitely changing has definitely changed how we think, just as every tool humans have ever invented has changed how we think.

Calculators didn’t make us forget how to do arithmetics, but they did free us from long division so we could focus on understanding mathematical concepts.

Google Maps didn’t make us forget how to navigate, but it did free us from memorising routes or looking up at the sky to figure out the direction. (Actually google maps made me forget how to navigate but that’s not the topic for today).

Grammarly didn’t make us forget how to spell, but it did shift our focus from perfect spelling to clearer communication and better ideas.

So… No. AI might not make us “forget” how we think, but it can make us “shift” how we think.

How to use AI with your brain…

For the shift to be effective, I believe we need to understand the difference between collaborative assignment and dependency.

Example 1

Let’s say you have a debate…

If you write the argument and use AI to help structure it with extra contexts, you might actually learn something about argumentation and the best approach to take arguments from → That’s collaborative assignment.

But if you let AI generate the entire argument while your brain is on flight mode, you learn nothing. → That’s dependency.

Example 2

Let’s say you’re debugging code...

If you identify where the bug might be, form a hypothesis about what’s wrong, and use AI to help validate your thinking or suggest solutions you can evaluate, you’re getting better at debugging → That’s collaborative assignment.

But if you just dump your broken code and say “fix this” without trying to understand what was broken or how the fix works, you’ll likely write that same bug again next week. → That’s dependency.

In summary, using the collaborative assignment approach makes you change what you think about, rather than not thinking at all.

But also thinking of it from another angle…

Maybe we’re romanticising the kind of thinking we’re supposedly losing. A lot of what we consider “thinking” is actually just pattern matching, retrieving information from what we know, or applying known solutions. AI is extraordinary at this, and to be fair, most of us weren’t that great at “thinking” to begin with; we’re all idiots (haha).

The question is, if AI can think better than me, why would I need to think? In this case, perhaps we’re free to focus on other important things, like touching grass, feeding cats, and spending time with the fam.

To wrap up this epistle…

In my opinion, it’s neither black nor white; what we need is more intentionality about how we use AI.

But… If you’re one of those people who currently cannot function effectively if put in a room without AI and asked to “deliver”. Maybe it’s not a good position to be in. Or maybe it actually doesn’t matter.

And that, my friend, is why I said the answer is “yes” and “no”.

Summary for extra lazy people: Google launches Gemini 3, NVIDIA beats earnings by a landslide, Yann LeCun leaves Meta after 12 years to start his own AI company, and Stack Overflow becomes an AI data provider.

What You Must Know

• Google Launches Gemini 3 → Google dropped Gemini 3, which comes with better reasoning, multimodal understanding, and a new coding platform called Antigravity. I’ve seen a lot of hype around it online, and I can see that it’s been topping the LMArena leaderboard. Sadly, I’ve not had enough time to breathe, so I haven’t played with it yet. I’ve always expected Google to take the lead in this AI race, and I'm glad to see them living up to that expectation.

• NVIDIA Beats Earnings, AI Fears Ease → NVIDIA reported $57B in revenue (up 62%) and forecast $65B for next quarter. About two newsletters ago, I mentioned that if the earnings don’t go as expected, it might hurt the stock market. Thank goodness it did go well. This has kind of relaxed market fears about the AI bubble.

• Yann LeCun Leaves Meta, Challenges LLMs → One of AI's pioneers, Yann LeCun, announced his departure from Meta after 12 years to launch his own startup focused on “Advanced Machine Intelligence.” His big argument is that LLMs are a “dead end” for true intelligence, and the field should move beyond them. This comes after Meta sidelined him by bringing in younger “GEN Z sort of” executives and shows how even the most prominent AI companies are struggling with direction.

• xAI Announces Grok 4.1 Update → xAI updated Grok to version 4.1, which they claimed has faster processing, fewer hallucinations, and better agent capabilities. It ranks close to Gemini 3.0 on LMArena. I gave it a quick spin with a coding task, the result was decent, not decent enough to make me switch from claude sonnet 4.5 though 😉

What’s Good to Know

• Stack Overflow Pivots to AI Products → Stack Overflow announced new products to turn its Q&A web forum into structured AI data for enterprise use. In simpler words, they are now selling access to their knowledge base to help companies train AI with real human expertise. Smart move, considering their traffic has been declining in the AI era. As the elders say, “If you can’t beat them, you join them”.

• Google Announces Nano Banana Pro → Google launched Nano Banana Pro, built on Gemini 3. It can generate images up to 4K resolution and includes text rendering in multiple languages, among other things. I even saw some funny, almost-realistic images someone generated with it of big tech CEOs chilling together in a room 🤣. If not for the absurdity of the idea, you couldn’t possibly tell the picture is AI-generated.

  

• Jeff Bezos Creates A.I. Start-Up → Called Project Prometheus, the company is focusing on artificial intelligence for the engineering and manufacturing of computers, automobiles, and spacecraft. Bezos is returning as co-CEO, just like Elon; Bezos is steadily grinding.

• EU Proposes Loosening GDPR and AI Act → On November 19, the EU moved to relax GDPR and AI Act rules. The changes aim to boost innovation while maintaining safeguards. In all fairness, I believe it’s pretty difficult for tech companies to compete when privacy rules are too strict, but privacy rules also need to be strict. Blue pill, red pill.

Money Moves This Week

• Stratis raised an undisclosed amount in seed funding to advance its AI data analysis platform.

• Genspark raised $275 million in Series C funding at a $1.25 billion valuation to scale its AI workspace platform.

• Cavela raised $6.6 million in seed funding to develop AI for global product sourcing.

• Flexion raised $50 million in Series A funding to expand its robotics and automation technology.

• Redrob raised $10 million in Series A funding to intensify development of its AI platform.

• AlphaXiv raised $7M to bridge the AI research-to-practice divide.

Kaizen Corner - What is AGI?

If you’ve spent some time online, you’ve probably heard of “AGI” by now. So what does it actually mean?

AGI stands for Artificial General Intelligence.

That’s a fancy phrase, but here’s the normal-person version: AGI means a computer brain that can learn and think about anything the way a human can.

AI today is really talented specialists (like ChatGPT writing or Midjourney making images). It’s like hiring someone who’s incredible at chess but can’t cook an egg. AGI would be the person who can play chess, cook, fix your car, write poetry, and learn brain surgery over the weekend.

And why is everyone obsessed?

Because AGI is the holy grail, it’s the point where AI stops being a tool and becomes... well, something that thinks like us.

Some people believe we’re 2-5 years away. Others think we are decades away. A few believe we’ll never get there with current approaches.

But who knows? We never know.

Meme of the Week

That’s your week in AI.

If you learned something, tell a friend. And if you didn’t, blame yourself :)

Until next Sunday,

Kay - your fellow human.

P.S. If this email lands in spam, that’s your inbox trying to stop you from staying plugged in. Fix it.

Summary for extra lazy people: OpenAI releases GPT-5.1 with “Instant” and “Thinking” modes, Claude helps stop the first AI-powered cyberattack, Microsoft launches an AI superfactory, and Cursor raised $2.3B.

What You Must Know

• OpenAI Released GPT-5.1 → GPT-5.1 brings two new modes, Instant for fast replies and Thinking for deeper reasoning. I haven’t tested it yet, I think I’m beginning to develop model fatigue, there’s so much “now even smarter!” I can take in a week, but I’ll probably try it once my curiosity beats my exhaustion.

• Chinese Hackers Used Claude Code for an AI Cyberattack → Anthropic caught the first known AI-orchestrated cyber-espionage campaign, allegedly run by Chinese hackers using Claude Code. It’s a chilling preview of what AI can do when misused. On the bright side, Anthropic caught it fast, so at least someone’s paying attention. I’ll read more on it and write a summarised piece later this week.

• Cursor Raises $2.3B to Build Its Own AI Model → The coding assistant startup Cursor raised $2.3B at a $29.3B valuation to build its own foundation model for developers. How better can they make an already great software? The future looks interesting.

• Anthropic Shares Its Bias-Busting Toolkit → Anthropic has open-sourced a framework to measure the political bias of models. This will be useful for people who argue with ChatGPT about politics 😅

• Baidu Unveils ERNIE 5.0 and AI Agents → China’s Baidu dropped ERNIE 5.0, a multimodal model that can process text, voice, and video, plus a suite of AI agents built for offices, classrooms, and factories.

What’s Good to Know

• ChatGPT Group Chats Now In Asia-Pacific → OpenAI is testing group chats in Japan, New Zealand, South Korea, and Taiwan. You can now rope your friends or coworkers into one conversation with ChatGPT. Not sure how that’ll work, but it sounds interesting.

• Wall Street Questions the AI Hardware Gold Rush → As tech lords race toward $1 trillion in AI infrastructure spending, investors are asking: how long before all these shiny GPUs become junk? Could the AI hardware boom be the next crypto mining rig story? I doubt.

• OpenAI and Microsoft Partner with U.S. Attorneys General on AI Child Safety → The two companies are teaming up with U.S. state AGs to tackle AI-related child safety issues, like deepfakes and grooming risks.

• NotebookLM Now Has ‘Deep Research’ Mode → Google’s NotebookLM can now handle more file types and dig deeper into sources, basically giving users a mini research assistant.

• ”Writing Might Be AI’s Worst Use Case” - Sitepoint → SitePoint argues that using AI for writing is counter-productive: you spend more time prompting and rewriting than drafting, and lose the ability to express your own thoughts clearly. I kinda agree, in my opinion, the best use case for AI in writing is proofreading.

Money Moves This Week

• Sim.ai raised $7 million in Series A funding to expand its AI agent platform.

• Dialog raised €3.74 million in seed funding to build AI shopping agents for ecoms.

• Netic raised $23 million in Series B funding to scale its AI-driven infrastructure.

• Obello raised $8.5 million in seed funding to enhance its AI creative software.

• Lative raised $7.5 million in seed funding to develop smarter AI sales planning.

Kaizen Corner - What the Heck Is “AI Inference”?

I wrote about this last week, but here’s the short, brain-friendly version.

“Inference” is basically the moment when an AI uses what it has already learned to give you an answer.

Think of it like this: training is when the AI studies, and inference is when it finally writes the exam.

So when you type a prompt into ChatGPT, it’s not “learning” anything new. It’s just looking at what it already knows, doing some abracadadra math, and spitting out its best guess.

If you want a longer breakdown with an analogy, read it here.

Meme of the Week

That’s your week in AI.

If you learned something, tell a friend. And if you didn’t, blame yourself :)

Until next Sunday,

Kay - your fellow human.

P.S. If this email lands in spam, that’s your inbox trying to stop you from staying plugged in. Fix it.

You’ve probably seen the word “inference” floating around AI articles and news.
It’s one of those mysterious words AI folks love to drop in every conversation.

And in case you’re wondering what it actually means, which you probably aren’t, but I’ll explain it anyway.

So, What’s “Inference”?

In human terms, inference just means “the part where the AI actually uses what it learned to generate an answer.”

You can think of it this way:

• Training is like teaching the AI; that’s where it studies, learns from data, and gathers all the wisdom it’ll later pretend to have.

• Inference is like using what it learned to provide an answer; when you ask a question and it responds with “Actually…” (in that confident AI tone).

You can stop reading here if you have short attention span.

Practical Example: A Robot Chef

Let’s say you own a fast-food restaurant with a robot chef (yeah, you know we’re in the future now).

1. You spend months teaching it all your secret recipes. That’s training.

2. Every time someone orders a burger, it still needs to make it fresh, flipping patties, toasting buns, everything. That’s inference.

The robot already knows how to make the burger, but each order still needs mixing, cooking, ingredients, and a few minutes of work to cook it. We call it the “doing phase“.

So, if anyone tries to make “AI inference” sound complicated, just nod wisely and say, “Oh, you mean the part where the robot does stuff based on its training?”

You’ve been enlightened. Now go forth and spread the knowledge.

But today, I could no longer deny the curiosity. So, I did what every normal person does on a lazy Sunday: I started digging.

Apparently, AI models can understand your needs better when you use XML-style tags. Why? Because they create clear boundaries. When you use normal texts, headings, and bullet points, the AI still has to play FBI to guess where one idea ends and another begins. XML tags eliminate that uncertainty.

Practical Example

Example → Regular English Style:

I’m building a mobile app for restaurants. Must work offline, sync when online and must collect tip from everyone by force. Like Toast POS but simpler.

Example → XML Style:

<context>
I’m building a mobile app for restaurants. We have 50 locations
and spotty internet in kitchens.
</context>

<requirements>
- Must work offline with local data storage
- Sync orders when connection restores
- Must collect tips whether the person likes it or not
</requirements>

<examples>
Toast POS offline mode 
</examples>

That shift makes the AI go, “Ahhh, I see what you did there”, instead of “Wait, what are we doing again?”

Advanced Techniques

1. Nested tags for hierarchy:

<project>
  <background>
    <business_context>We’re a B2B SaaS...</business_context>
    <technical_context>Current stack is React/Node...</technical_context>
  </background>

  <task>
    Evaluate whether to migrate to Next.js
  </task>
</project>

This helps when your prompt starts sounding like confusion.

2. Sequential processing:

<step_1>
First, analyze this data: [data]
</step_1>

<step_2>
Then, using only the findings from step 1, create a summary
</step_2>

That forces the AI to follow instructions in order, rather than jumping to step 47 because it got excited.

When It Can Be Useful

The effectiveness of XML tagging seems higher when you have complex prompts (multiple sections, many examples, nested logic, or machine‐readable structure. It’s also helpful when making requests over the API.

Essentially, when there’s potential for confusion, XML can be a safety net.

This blog article speaks more about it.

My Actual Conclusion

After all the digging and experimentation, I’ve come to conclude that the benefit stems from the structure rather than specifically XML tags. Which means if you can structure your regular English prompts properly, it works just as fine.

Sure, XML makes things cleaner in edge cases, but the newer LLM models are so good, they’ll figure it out anyway. Unless you’re doing something extremely structured, like parsing multiple files or separating logic layers, XML might just be overkill.

So yeah… For now, I’ll stick to my usual English prompting like a normal human.

Maybe, one day, in the future, I’ll find a real use for XML tags.

Summary for extra lazy people: Apple to pay Google $1B a year to upgrade Siri, ChatGPT quietly kills job ads, AI is interviewing people now, and Amazon goes after Perplexity’s shopping bots.

What You Must Know

• Viral: Redditor Gets Interviewed by an AI → A Reddit post went viral after someone realized their entire “job interview” was with an AI. The debate now is: if companies use bots to hire, can candidates use bots to answer? 🤔

• Apple Pays Google $1B a Year to Smarten Up Siri → Apple will pay Google $1 billion annually to integrate Gemini AI into Siri. The new version should make Siri sharper.

• ChatGPT Shrinks Job Ads by 32% → A Stanford study found ChatGPT’s rise led to a 32% drop in job postings, hitting young workers hardest. So yeah, AI is taking some jobs, just not mine (yet).

• Google’s AI Can Now Make Slides from a Note → Google’s Gemini Canvas now turns your rough notes into full presentations. I’d call it “PowerPoint for people with better things to do” 😅

• Amazon Sues Perplexity Over AI Shopping Agents → Amazon’s lawyers want Perplexity’s AI shopping agents to stop making purchases on their own.

What’s Good to Know

• Parallel Search Launches High-Accuracy Web API for AI → Parallel launched a web search API designed specifically for AI models, claiming “the highest accuracy on the web.”

• Maye Musk Uses AI to Relive Her Runway Days → Elon Musk’s mom, Maye, used Grok AI to recreate her first grandmother modeling job at the age of 42, and the result looks surprisingly real.

• Breaking Bad’s Creator Roasts AI Laziness → Vince Gilligan joked that people using AI for every little thing are “forgetting how to think”. Fair point, I must say.

• Google Updates Vertex AI Agent Builder → Google improved Vertex AI Agent Builder, making it easier for developers to move from prototype bots to production-ready agents.

• Chinese Lab Moonshot AI Unveils Kimi K2 Model → China’s Moonshot AI released an open-source reasoning model that “they claim“ matches (and sometimes beats) GPT-5 and Claude Sonnet 4.5 on key benchmarks.

Money Moves This Week

• Giga raised $61 million in Series A funding to scale its AI technology platform.

• Fastbreak AI raised $40 million in Series A funding to bring AI innovation to sports technology.

• IndustrialMind raised $1.2 million in pre-seed funding to develop AI for manufacturing optimization.

• Subtle Computing raised $6 million in seed funding to advance its general AI infrastructure.

• Inception Labs raised $50 million in seed funding to build diffusion-based AI models.

Kaizen Corner → The Best AI Users Are Professional Question-Askers

When people say “ChatGPT gave me nonsense”, it usually (at times) means they threw a lazy question at it and hoped for magic. That’s like walking into a restaurant and shouting, “Food, please!”

You’ll get something, but probably some dish from Kenya.

The best AI users act like detectives. They:

• Ask specific questions (“Explain like I’m new to marketing” is better than “Explain marketing”)

• Follow up when answers are vague (“Give me examples from fintech”)

• Clarify goals (“I’m trying to write a headline, not an essay”)

You don’t need to command AI; you need to converse with it. Treat it like a curious intern: it knows a lot, but it needs direction.

Meme of the Week

That’s your week in AI.

If you learned something, tell a friend. And if you didn’t, blame yourself :)

Until next Sunday,

Kay - your fellow human.

P.S. If this email lands in spam, that’s your inbox trying to stop you from staying plugged in. Fix it.

I was curious about how their login flow worked, a Web3 wallet that uses social login instead of seed phrases. I could’ve kept reading, but I’ve learned that the fastest way to understand anything tech-related is to actually build it.

So I fired up Windsurf, teamed up with Claude Sonnet 4.5, and over a 15-hour marathon, we built a working Web3 crypto wallet from scratch (video below).

I’m talking Web3Auth integration, across three blockchains, USDT token support, receiving/sending transactions, dedicated wallet pages, transaction history, and even a horizontal-scrolling dashboard, because, why not make it pretty? 😉

The project ended up with over 100 files, 17 docs, 16 configs, 47 frontend, 10 backend, and about 15 others. Not bad for a single day’s build with an AI co-pilot.

And honestly? It was a blast. Here’s the recap.

The Setup

The last time I used Windsurf was months ago, to build a Bybit trading bot with Claude 3.5 Sonnet. It worked, but it felt like working with a smart friend who needed everything explained three times. It didn’t feel like that then, but now it does.

Back then, I’d prepare meticulously, like it was an exam, gather all links, libraries, and documentation, and dump them into Windsurf as “rich context,” as I’d call it. This left nothing to guess, which reduced the chance of hallucination.

But this time, I just raw-dogged it. No prep, no context files, just vibes. Maybe it was laziness, or maybe I just felt it was a playground project. What could possibly go wrong?

Where Claude Sonnet 4.5 Shines

1. Needs Less Context

Usually, AI coding models need some sound context engineering to understand your task. Not so much for 4.5.

I gave lazy prompts like, “Make the USDT wallet page similar to the others”, and it just did it.

I’ve always believed AI’s biggest challenge is context, so it’s good to see that it’s beginning to need less of it to do a good job.

2. Understands Web3’s Security

I tested it by proposing moving transaction signing to the backend “for better control and logging”.

It caught the ‘bad’ in this idea, explained why storing or handling private keys server-side was a massive security violation, and then walked me through the correct client-side approach.

That’s the kind of awareness that saves you from becoming a “how we lost $2M in user funds” Medium post.

3. Less Back-and-Forth

Normally, working with AI feels like playing table tennis, endless back and forth.

This time, it was smoother (not perfect).

At first, I thought it was because I was better at using it, but later, I realised I was actually doing a nonsense job; the model was just better.

4. Better Taste in Design

Most AI-generated UIs look like a purple alien landing page (I’ve written about that).

I expected the same here, but 4.5 gave me a sleek dark theme. When I asked why, it said “dark suggests luxury and privacy, ideal for Web3 users.”

Honestly? It’s not wrong.

5. One-Click Deploy

This is a new feature on Windsurf (not Claude’s); I saw it for the first time.

Deploying straight to Netlify was refreshingly easy. There was no GitHub setup, no Render deployment (which I’d normally use), just click, build, and done.

Hopefully, they add more deployment options soon.

6. Documentation Was Clean

The whole point of this 15-hour marathon was to learn, not just code. So documentation mattered even more than the build itself.

I told Claude to document everything, every step, every decision, so I could revisit it later like a personal dev diary. And it did great: clear, organised, and surprisingly thoughtful.

Of course, it’s not perfect…

No AI story is complete without a few “what the hell” moments.

1. Too eager sometimes. At one point, it imported a non-existent TransactionHistory component because it assumed I’d want a generic one. Nice guess, wrong move.

2. Memory issues. As the project grew, it forgot we’d already installed a confetti library… and tried to add another one. Two confettis, one app. What an idiot.

3. Goes on a foolish search for bugs. It sometimes chased the wrong issue. I’d tell it, “The problem isn’t the internal server error, check further down”, and it’d reply, “Ah, you’re absolutely right”. Classic AI humility.

The Bottom Line

I learned a lot.

Is Claude 4.5 perfect? Nope.

Will it replace developers? Definitely not anytime soon.

But will it make a good developer insanely more productive? Absolutely. If I could use it productively, I wonder what crazy developers can do with that thing.

Video of the wallet’s build is below.

And of course, I won’t be publishing it live; I’m only curious, not foolish. 😄

Claude Sonnet 4.5 is impressive, but not that impressive. It’s not “ship a Web3 wallet to production and sleep peacefully” impressive.

Read: Can you actually build a production-ready app with prompt-to-code tools…. and should you?

And since the best way to truly understand something is to re-explain it in plain English, here’s me doing just that, breaking it down so you don’t need to rack your brain to get it.

1. The Gist

The paper is about “introspection”, a fancy word for looking inside one's own head. Humans do it all the time (e.g., “Why did I just open the fridge again?”).

Anthropic asked: Can an AI model do that, too? Can it notice what’s happening inside its own circuits and describe it?

2. Why Bother?

They believe that if AI models can notice themselves, we can make them safer and more transparent, like a self-driving car that says, “Uh oh, my sensors are glitching”.

They mentioned that it’s not about consciousness. It’s about diagnostics. Although I see a blurry line between those two, but let’s not digress.

3. What They Did

Their researchers poked Claude’s brain by secretly tweaking its inner state, kind of like slipping in a hidden thought (“betrayal,” for example).

Then they asked: “Hey Claude, feel anything different?”

Sometimes, Claude replied:

“I sense intrusive thoughts about betrayal.”

When that matched the injected concept, that’s introspection.

But… it didn’t happen every time.

Sometimes Claude totally missed it or guessed wrong.

4. What It Means (and Doesn’t)

• Claude can sometimes detect and describe internal changes.

• Claude is not self-aware. Half the time, it’s like a car yelling “Engine fault!” when you just turned on the radio.

5. Why It’s Cool Anyway

If it gets refined, we could have AIs that ‘actually’ explain themselves.

Note: Some reasoning models can already show their reasoning process, but that’s externally generated explanation, not internal self-detection.

Imagine asking your model, “Why’d you say that?”, and it can actually show its reasoning from an internal perspective.

That’s huge for safety, debugging, and trust.

Bottom line:

Claude isn’t self-aware. But it’s starting to show signs of noticing what’s going on inside, like a toddler realizing it has hands, and that’s both fascinating and a little weird.