Exploring AI so I don’t become obsolete | Based in 🇵🇹
Joined June 2025
- Tweets76
- Following74
- Followers85
- Likes77
Pinned Tweet
Prompt engineering has been replaced by loop engineering.
What is it? (Explained in 60 seconds)
For the past 2 years we have been prompting agents with individual tasks. That is starting to change.
So far, if you wanted an agent to build a dashboard for a client, you would give it a task, review the output, improve the prompt, and repeat the process until the work was done.
Looping changes that.
Instead of giving an agent individual tasks, you give it a goal and let it work through a recursive loop until that goal is met.
For example:
→ Research
→ Draft
→ Evaluate
→ Test
→ Improve
→ Repeat
The agent keeps cycling through the loop until it reaches the standard you defined.
Within loop engineering there are two main approaches:
1. Open Looping
You give the agent a goal and allow it significant freedom in how it achieves it.
This is powerful, but also expensive and harder to control.
2. Closed Looping
The human defines the architecture, constraints and evaluation criteria.
The agent is then responsible for executing, improving and iterating within those boundaries until the goal is reached.
The next evolution is orchestrated looping.
Instead of a single agent running a loop, one agent breaks the goal into smaller tasks and assigns them to specialist agents.
Each specialist runs its own loop and reports back.
In other words:
You move from one agent improving itself to an entire team of agents iterating together until the goal is achieved.
Books that survived wars, fires and centuries are now being destroyed to train AI. And its completely legal...
For years, we've talked about AI scraping the internet. Now it's buying physical books.
Court documents revealed that Anthropic purchased millions of books, cut off their spines, scanned every page at high speed, and then recycled the originals to build its training datasets.
The company even hired the former head of Google Books partnerships with the goal of acquiring "all the books in the world." A federal judge later ruled that this process qualified as fair use, provided the books had been legally purchased first.
The reason is surprisingly simple.
Books published before the recent AI boom are becoming increasingly valuable because they're far less likely to contain AI-generated text. As the web fills with synthetic content, high-quality human-created data is becoming one of the scarcest resources in AI.
What struck me most wasn't the legal ruling. It was the shift it represents.
For decades, books were things we collected to preserve knowledge. Now, in some cases, they're being treated as raw material for creating the next generation of AI models.
I don't think this is the end of the story.
If access to high-quality human knowledge becomes one of the biggest competitive advantages in AI, we're going to see an entirely new economy emerge around acquiring, digitising and creating data that models can't simply scrape from the internet.
🦔AI companies are bulk-buying rare books, scanning them through high-speed machines that cut the spines off, and shredding the originals. A service called ISBNdb facilitates orders of up to a million books and keeps buyers anonymous. Pre-2022 books are premium because they're free of AI-generated text. A federal judge ruled the practice is fair use because eliminating the original means only one copy exists at a time. Anthropic hired the former head of Google Books partnerships to obtain "all the books in the world."
My Take
This got to me. A bookseller told 404 Media that rare books with almost no surviving copies are being fed into this pipeline. Books that survived wars, fires, and centuries of handling are being shredded so an AI can learn to write a better marketing email.
ISBNdb's website literally says "'AI company destroys two million books' is not a headline that generates sympathy," and they still built an entire business around making it happen quietly. They offer NDAs as a feature. They coach clients to call it "digital preservation."
I've covered AI companies scraping the internet, torrenting libraries, and stealing music. This is worse because it's irreversible. You can re-upload a website. You can reprint a bestseller. You can't replace the last three copies of an 18th-century botanical text once someone shreds them for training data. And the judge said it's legal. So it's going to accelerate.
"We shred rare books and offer NDAs so nobody finds out" is a legitimate business model in 2026. What a timeline.
Hedgie🤗
If you think you're behind in AI, you're probably not. Fewer than 10% of companies have deployed AI agents at scale.
If you spend enough time on X, it can feel like every company has already automated half its business, has AI agents running everywhere, and is miles ahead.
The data tells a different story.
McKinsey found that 88% of companies are using AI somewhere in their business, but fewer than 10% have deployed AI agents at scale within any individual business function.
In other words, most companies are still experimenting.
It's easy to get caught up in the constant stream of new models, frameworks and tools, feeling like everyone else has figured it out while you're falling behind.
The reality is that almost everyone is still in the trenches, testing, breaking things and learning what actually creates value.
My advice would be simple. Don't optimise for trying every new AI tool that gets released, instead optimise for experimenting.
Pick a real problem, build a solution, see what works, and repeat. Every workflow you test will teach you more than another hour spent reading about the latest model release.
Because right now, there isn't a playbook. We're all figuring it out in real time.
Scraping the internet to train AI models is old news.
The next trillion-dollar industry will be the task economy, let me explain...
@mercor , a startup that connects AI labs with domain experts who generate specialized training data, reportedly grew from a $1B to $2B annualized revenue run rate in just four months. The interesting part isn't how quickly they're growing, but why.
For years, AI companies competed by scraping more of the internet. Today, that data has largely become a commodity. The next generation of models is increasingly being improved by lawyers, doctors, engineers, consultants and other experts who create the training data the internet simply can't provide.
In other words, AI isn't just competing for compute anymore. It's competing for expertise.
I think this is where the next major shift in AI will happen. The companies with the biggest advantage won't necessarily be the ones with the largest models, but the ones that figure out how to collect niche, proprietary knowledge that has never been published online.
If the last race in AI was about scraping the internet, the next one will be about collecting the knowledge that only humans can create.
*24 months - Source: xitter.cf/BrendanFoody/status/20…
My AI agent just prevented a €2,000 Meta Ads mistake.
I've been building an AI agent that continuously monitors Meta Ads accounts instead of relying on someone to check dashboards a few times a day.
A few days ago it spotted an ad that had started bleeding budget. The CTR was dropping, the CPL was climbing, and the campaign was still spending.
Within minutes, the agent flagged the issue, explained what was happening, and suggested pausing the campaign before another €2,000 was wasted.
That made me realise something.
The biggest value of AI agents isn't that they replace experts. It's that they never stop paying attention.
While we were busy with other work, the agent was continuously monitoring performance, highlighting anomalies, and suggesting optimizations based on the data.
I think that's where AI agents will create the biggest impact over the next few years.
Not by making decisions for us. By making sure we never miss the ones that matter.
Alex Karp is going viral and just confirmed what I wrote about last week.
Companies don't want to rent intelligence anymore.
Last week I wrote that AI labs are becoming the gatekeepers of intelligence and that companies will increasingly want to own larger parts of their AI stack.
Now Palantir CEO Alex Karp is making the same point.
In yesterdays interview, he argued that enterprise customers are becoming increasingly uncomfortable relying entirely on frontier model providers.
Not because the models aren't good.Because they're realizing they're handing over the one thing that makes them unique.
Karp's argument boils down to three simple ideas:
→ If a model provider decides to compete in your market, they control the capabilities you're building on.
→ Your proprietary data is where your company's alpha lives. Handing it over weakens your competitive advantage.
→ If your entire business runs on someone else's models, you're no longer just renting infrastructure. You're renting intelligence.
I don't think companies will stop using frontier models.
But I do think they'll increasingly want to own the parts of the stack that matter most.
The companies with the biggest AI advantage won't necessarily have the smartest models, they'll have the most control.
Gojiberry's company page and founder account were banned.
LinkedIn's AI crackdown is just getting started.
Before that, tools like @useapolloio , @heyreach_io and @amplemarket had already run into issues with LinkedIn.
Whether you agree with it or not, I think the trend is becoming pretty clear.
Every major platform will fight AI automation. Not because they hate AI. Because they need to protect the user experience.
If you're building an AI product that depends on automating a third-party platform, I'd think twice.
Platform risk has always existed. AI is simply accelerating it.
The winners won't be the companies that push automation the furthest.
They'll be the ones that build within the rules or build products that don't depend on someone else's platform in the first place.
This is your reminder to spend 30 minutes today and set up your personal AI brain
The more context you give AI, the better it performs.
Yet most people's context is scattered across ChatGPT conversations, Slack messages, meeting notes, browser bookmarks and random documents.
AI can't connect dots it can't see. That's why I've been spending more time thinking about second brains.
The mistake many people make is thinking they need a complicated setup, but it’s simpler than you might think.
Most of my setup is just @obsdmd . A collection of markdown files organised and linked in a way that both I and my agents can understand.
As your knowledge grows, AI becomes dramatically more useful because it has access to years of context instead of whatever happens to fit into a prompt.
The best way to start?
Simply give @ChatGPTapp or @claudeai the goal to organise your files in Obsidian for you in a structure which makes sense to you.
From then on you can add every piece of new context into the vault and ask your AI to sort it where it belongs.
The real value is that every article, meeting, idea and insight becomes available to both you and your agents forever.
In a world where AI is becoming a commodity, context will end up being the biggest advantage of all.
whats stopping you from building a second brain?
just install codex, connect it to all your tools
then run /goal mode with the loop below and have it cook for 12 hours building an obsidian vault
this loop will go through every detail from every connector you have
you will have the most perfect knowledge base for you and your agents
go experience personal AGI before everyone else
Max Stone retweeted
AI companies are becoming the gatekeepers of intelligence.
It's time to own your own AI stack.
Before Anthropic even launched Fable, they announced two changes:
→ Certain prompts may be retained for up to 30 days
→ Users may be downgraded to weaker models depending on how they use the frontier models
A few days later, Fable itself was taken down due to security concerns. Whether these decisions are right or wrong isn't really the point. The interesting part is what they reveal about the future of AI.
For the past decade, we've become comfortable building on infrastructure we don't own. Cloud providers, APIs and SaaS tools.
With AI, I think the equation is different. The capabilities of the model determine what your product, workflow or team can actually do. If those capabilities change, your business changes with them.
Imagine building critical workflows around a model, only for access, pricing or permitted outputs to change overnight.
You're no longer just renting infrastructure, you're renting intelligence.
My guess is that this is one of the reasons we will see more companies move towards owning larger parts of their AI stack.
So what does owning part of your AI stack actually mean?
→ Run critical workflows on models you control
→ Store your knowledge and data outside of a single provider
→ Build systems that can switch between models
→ Fine-tune models on your own processes and use cases
→ Treat AI as infrastructure, not just a tool
Fable may only have been online for a short period of time.
But it exposed a question every company building with AI will eventually have to answer:
How much of your intelligence stack do you actually control?
Just now: Anthropic is flying senior technical staff to Washington to repair its fight with the White House after export controls forced its top models, Mythos and Fable, offline.
The company is now trying to convince officials that the models can be safely controlled, turning this into a real-time test case for AI geopolitics.
Via Axios
Monday is getting more interesting by the minute.
AI companies are becoming the gatekeepers of intelligence.
It's time to own your own AI stack.
Before Anthropic even launched Fable, they announced two changes:
→ Certain prompts may be retained for up to 30 days
→ Users may be downgraded to weaker models depending on how they use the frontier models
A few days later, Fable itself was taken down due to security concerns. Whether these decisions are right or wrong isn't really the point. The interesting part is what they reveal about the future of AI.
For the past decade, we've become comfortable building on infrastructure we don't own. Cloud providers, APIs and SaaS tools.
With AI, I think the equation is different. The capabilities of the model determine what your product, workflow or team can actually do. If those capabilities change, your business changes with them.
Imagine building critical workflows around a model, only for access, pricing or permitted outputs to change overnight.
You're no longer just renting infrastructure, you're renting intelligence.
My guess is that this is one of the reasons we will see more companies move towards owning larger parts of their AI stack.
So what does owning part of your AI stack actually mean?
→ Run critical workflows on models you control
→ Store your knowledge and data outside of a single provider
→ Build systems that can switch between models
→ Fine-tune models on your own processes and use cases
→ Treat AI as infrastructure, not just a tool
Fable may only have been online for a short period of time.
But it exposed a question every company building with AI will eventually have to answer:
How much of your intelligence stack do you actually control?
Just now: Anthropic is flying senior technical staff to Washington to repair its fight with the White House after export controls forced its top models, Mythos and Fable, offline.
The company is now trying to convince officials that the models can be safely controlled, turning this into a real-time test case for AI geopolitics.
Via Axios
Monday is getting more interesting by the minute.
Max Stone retweeted
Prompt engineering has been replaced by loop engineering.
What is it? (Explained in 60 seconds)
For the past 2 years we have been prompting agents with individual tasks. That is starting to change.
So far, if you wanted an agent to build a dashboard for a client, you would give it a task, review the output, improve the prompt, and repeat the process until the work was done.
Looping changes that.
Instead of giving an agent individual tasks, you give it a goal and let it work through a recursive loop until that goal is met.
For example:
→ Research
→ Draft
→ Evaluate
→ Test
→ Improve
→ Repeat
The agent keeps cycling through the loop until it reaches the standard you defined.
Within loop engineering there are two main approaches:
1. Open Looping
You give the agent a goal and allow it significant freedom in how it achieves it.
This is powerful, but also expensive and harder to control.
2. Closed Looping
The human defines the architecture, constraints and evaluation criteria.
The agent is then responsible for executing, improving and iterating within those boundaries until the goal is reached.
The next evolution is orchestrated looping.
Instead of a single agent running a loop, one agent breaks the goal into smaller tasks and assigns them to specialist agents.
Each specialist runs its own loop and reports back.
In other words:
You move from one agent improving itself to an entire team of agents iterating together until the goal is achieved.
I was worried I was falling behind in AI. So I built an AI professor.
Here's how to do the same:
1. Context
Peter first interviewed me to understand:
→My goals
→Technical level
→Available time
→What I'm currently working on
Based on that, he created a personalised AI curriculum.
2. Daily Learning
Every day I get:
→1 hour of technical learning
• 1 hour of automation-focused learning
→All notes are automatically structured and pushed into Notion, creating my own AI textbook over time.
3. Adaptive Learning
This is my favourite part.
Peter continuously adapts the curriculum based on:
→My understanding
→My progress
→Available time
He also monitors AI news and X posts, evaluates what is relevant to my goals, and updates the curriculum automatically.
The result is I spend less time deciding what to learn and more time actually learning.
Setting this up might sound complicated.
But you can build a simplified version in 10 minutes.
I've dropped the prompt in the comments.
42°C+ temperatures this summer. I asked @claudeai what happens next, the answer wasn't reassuring.
For those unfamiliar:
El Niño occurs when a massive area of the Pacific Ocean warms by several degrees.
That may not sound like much. But the Pacific stores more heat than the entire atmosphere. And when that heat gets released, the effects can be global.
This year is predicted to be the biggest El Niño we have ever seen.
Here are some of the impacts Claude predicted and experts are watching closely:
→ More frequent and intense heatwaves across Europe, with temperatures potentially exceeding 40°C in multiple regions.
→ Major atmospheric river events, dumping huge amounts of rain across parts of North and South America and increasing flood risk.
→ Record-breaking heatwaves and drought across Argentina, Chile and Brazil. As the world's largest agricultural exporter, disruptions in Brazil can quickly push food prices higher globally.
→ Increased risk of monsoon failures in India, affecting up to 115 million farmers and putting pressure on food supplies for more than 1.5 billion people.
→ Increased drought and wildfire risk across Australia, Indonesia and parts of the Amazon.
The image below compares ocean temperature anomalies during the historic 1877 event and what is developing today.
Whether this becomes a record-breaking El Niño remains to be seen.
It may also be a reminder that what feels exceptional today could become increasingly normal.
In 2 years there will be more agents using @SlackHQ than people.
Most software is not ready for this.
That might sound ridiculous
But I actually think Slack is right.
We are moving toward a world with two kinds of software:
1. Software designed for humans
2. Software designed for agents
And the second category is going to win very quickly.
Because agents do not care about:
-beautiful UI
-onboarding flows
-dashboards
-animations
They care about:
-APIs
-permissions
-context
-integrations
-structured data
-execution reliability
The interesting shift is that software is slowly becoming infrastructure for other software.
Humans become supervisors.
Agents become operators.
That’s also why MCP, integrations and “agent compatibility” suddenly matter so much.
The companies that understand this early will become the default layer agents interact with.
The ones that don’t may slowly become invisible.
Slack understands this.
Most companies still don’t.
Just had coffee with a top AI security researcher, what he said shocked me
Expect that 6-9 months from now open source models will be at the levels of Anthropic’s mythos.
What that means: Online privacy will be a thing of the past. Any data you have stored online will be accessible and exploitable.
There is no way to fight this, either you go off the grid now or you accept it.
Max Stone retweeted
The week isn’t even over and we’ve already had:
-Google DeepMind publish a report showing agents can be manipulated by websites displaying information humans can’t even see.
-Anthropic share their new model Mythos, which can hack major infrastructure with an 80% first-try success rate.
-Most Fortune 500 companies partnering with Anthropic to mitigate these attack vectors.
Times are incredibly exciting, yet it feels like we’re one wrong move away from something breaking in a catastrophic way.
It is insane to think that Anthropic essentially holds a zero day to any company in the world.
They are more powerful than any government in the world.
The week isn’t even over and we’ve already had:
-Google DeepMind publish a report showing agents can be manipulated by websites displaying information humans can’t even see.
-Anthropic share their new model Mythos, which can hack major infrastructure with an 80% first-try success rate.
-Most Fortune 500 companies partnering with Anthropic to mitigate these attack vectors.
Times are incredibly exciting, yet it feels like we’re one wrong move away from something breaking in a catastrophic way.