This essay first appeared on my blog.

I claim to love startups, to love entrepreneurship — but how do I define it?

Well, our friends at Google and OpenAI define it quite simply:

And although both these definitions are true, they avoid what startups are at their core.

At it’s core, a startup is made to solve a problem nobody has ever solved before, or to solve a problem better than the current person solving it.

- From the wise man himself, and yours truly, me — Noah

What a startup really is

Through following the principle that a startups’ core reason for existence is to solve the unsolved, it makes it so that if your truly “founding” a startup, your navigating in a constant unknown space — one that hasn’t been grasped by anything in humanity yet.

Because of this principle and belief, I view startups as its own inherent world.
A world that is constructed by all of societies problems.

This world is it’s own kind of utopia — a world that can’t ever be fully explored, because in order for us to have fully explored it, we would’ve had to solve every problem humanity faces, both now (present), and in the future.

And it’s just that, there will never be a time we stop facing problems as a society. Which means intrinsically, there will never be a time we have fully explored this “utopian world” of startups.

With that being said, what does it mean?

It means that there’s no playbook because no one’s ever been here before. There’s no set rules, there’s no guide, and there’s no built structure for when building a startup.

You’re building something nobody has ever done before. There is just you, and potentially the people you’ve brought along with you to go on about solving this problem.

How to Succeed & Navigate in a World of Unknown

OK. We’ve now settled down on a problem, how do you move forward? It’s pitch black, in this new world, where’s our flashlight? And how do we know we’re moving in the right direction?

IMO there’s three things you need:

1. Lifelong learning
2. The three mindsets to success (A “figure it out” (FIO) + antifragile mindset + beginner mindset)
3. Straight up luck — but both kinds

Each one of these stages to success come one after another, and have a particular order of which they come.

1. What is it to be a lifelong learner?

“In times of change the learners will inherit the earth” — Eric Hoffer

To learn is to keep on moving forward, it is to be adaptive for both the present, and for the future. By choosing to not learn, you’re choosing to do a great number of many other things; we call that opportunity cost.

The most dangerous outcome of refusing to learn, is that you will remain stagnant, to reach a peak and remain just that, until you erode and break.

If we focus on startups specifically, and study them, we can see that one of the most proprietary reasons giants fall is because they fail to adapt in volatile markets and times. A few companies come to mind when I think of this:

1. Blackberry
2. Blockbuster
3. MySpace

Blackberry especially, is an amazing case study on what happens when you don’t innovate, adapt, and refuse to learn.

Something else relevant to this context of being a “lifelong learner” is a quote my friend had mentioned the other day,

“Right now I’m the smartest I’ve ever been, and the dumbest I’ll ever be.”

I believe that if you want to reach some unconventional level of success, you have to live through this philosophy. Because through living in it, we humble ourselves, and remind ourselves that through learning, we will always be a beginner in some sort of way. And as a beginner, you gain a thirst to learn — which in turn, activates “The Three Mindsets.”

2. What the hell are “The Three Mindsets” ??

Great question.

The three mindsets are what follows after being a “lifelong learner,” they are mindsets that are naturally achieved from having such hunger for knowledge.

The three frames of mind include being a beginner, having a figure it out (FIO) approach, and an antifragile core.

You’ll see that through training each of these frames of mind, by virtue, greater things will occur almost by nature. So much that through doing so, the outcome of success builds upon itself and brings us to the next stage of creating your own luck (partially, but more on this after).

We’ll kick off with the first frame, being a beginner;
One of the GOATs said it best, the one and only Brian Chesky,

The bigger he gets, the more of a beginner he feels.

It took Picasso’s four years to paint like Raphael, but a lifetime to paint like a child. To be a beginner is to be curious, and that is to see the world with fresh eyes, to live // work without judgment, and never feeling like you’ve “made it.” Because once you believe you’ve arrived, you stop growing (which in turn is to stop learning — hopefully you get the connection and see the pattern).

Through being a beginner, we can see things that would’ve typically been overlooked. Because we are not blinded yet by expertise, assumptions, and our own arrogance.

When you’re new to something you question everything there is to it. You don’t accept “that’s how it’s always been done,” so you spot flaws veterans ignore. Whereas experts often filter reality. Experienced people see patterns — which lets them think quick, but also lead blindly. They skip over the depth of the issue because they assume the answer. Beginners will notice the awkward UX button, clunky onboarding process, the little friction, and so forth.

Those “naive” observations often lead to breakthrough design. After all, Brian Chesky was the one that said, “it’s in the tiny details that delight lives. Most people gloss over them. Beginners don’t.”

In correlation to this, something a good friend of mine recently said is “just when you reach the top of one mountain, you notice someone else standing in the clouds.”

You can’t have the mindset of believing that there even is an endpoint, or a peak; sure, you can be a big fish in a small pond, living on the mountain top.

It’s all about optimizing for what you want at the end of the day.

But I kinda want to see what the clouds are like, that’s for sure. Maybe I’ll see someone in space haha (hellooo Elon).

Figure it out (FIO) mindset:
In startups we don’t have the time, nor do we have the resources to not do the work. Simply put, if you don’t do the work, quickly and to a high standard, you will fail.

It is both a race and a marathon simultaneously.

The way we can possess a FIO mindset, is through holding such conviction in ourselves, that no matter what cards we’re handed, we trust ourselves to always find some sort of way to make ends meet, and finish the work needed.

The startup world moves quick, it’s speed is not relative to our time, and we cannot calculate it. This world can be thriving one day, and moving at an unprecedented pace — but then humanity can fail to keep up which leads to a bubble among many other things. It is an external factor to our success as founders, and all we can do is work within it, which is figuring things out is so important.

The greatest founders were amazing at “FIO” when it seemed impossible at face value. So many examples come to mind when talking about this, a great one is the beginning of DoorDash. Spun up “PaloAltoDelivery.com,” answered a Google Voice line, and manually delivered first orders.

Another good one is Reddit, the founders created tons of fake accounts to populate early posts/comments until real users arrived.

When Airbnb’s runway was getting short — they sold custom cereal boxes!! The opportunities are endless. You just have to think differently.

These are all great examples and reasons as to why you need to be thinking in a system where you refuse underperforming. You have to get the work done, it doesn’t matter — to any means possible.

There were so many opportunities to quit and fail in these examples listed, yet these amazing founders created opportunities out of nothing, and found the few things that could’ve actually worked out. All through thinking differently, and having the mindset that they could figure out whatever was thrown there way — the mindset capable for greatness.

Side note ::Would you believe me if I said my old LinkedIn bio used to say “Figure It Out Enthusiast”? So this one is held close to my heart… Hits home.

The antifragile core
The startup world is unpredictable. As mentioned before, it is a world of unknown. A world that can never be fully explored, as it is only somewhat explored by subject to the problems we know of — that have actually been solved.

Because of this unpredictability, we have to get comfortable with being in these unknown spaces — it’s an external environment that we cannot control, but in ways can manipulate it, and take advantage of the mass volatility.

This third mindset is derived fully from Antifragile: Things That Gain from Disorder, by Taleb — one of my favourite reads out there.

Taleb says, “Wind extinguishes a candle and energizes a fire. Likewise, with randomness, shock, and volatility: you want to be the fire, not the candle.”

Both the candle and fire cannot control the wind, it is an external force. However, they can place themselves in unique environments to grow, and thrive from the wind. This same philosophy applies to startups — we can use these external forces of being in an unknown environment, and take advantage of the volatility — not to scare from it, but to grow from it. Just as a fire would to the wind.

Through building an antifragile core, we can navigate the startup world to our best ability. Especially when combined with the other two mindsets of being a beginner, and to figure it out — we can combine these three systems to craft a great founder.

Luck in success + making your own (partially…)

There’s no doubt that luck plays a huge role in success — it’s in every aspect of winning in startups. Various factors like timing, people you meet, current economy, etc.

Instagram (or at that time “Burbn”) — literally raised there seed round from a16z and Baseline ventures from meeting them at a Hunch party.

Inversely, and more recently, Google got rid of the num=100 parameter — Reddit used to be a powerhouse in citations for AI. However with this new update, there rankings plummeted. Reddit’s citation share on ChatGPT had fell from approximately 29.2% to just 5.3%. In doing so, there stock tanked 15% in the span of 5 days — roughly $6.5B gone.

How could Reddit have seen that coming? They couldn’t. There’s just things out of anyones control because of the uncertainty.

It’s a big “right person, right time” game. There’s so many ways the “wind” could blow — that sometimes you get unlucky. Many can argue that since there are too many variables that goes into building a successful startup (worth $10M+), even if it’s not random, it’s too complex to be controlled.

And I agree with this, but to an extent.

There are external factors we cannot control, but that is in everything we do — you can be at a cushy corporate job, however a recession can occur and you can get laid off, or perhaps get passed on a promotion. I know is startups a far more volatile market than the corporate one — so luck plays a much larger factor, but you get the point.

I believe there are ways that we can make our own luck — I actually think a good portion of our luck is earned.

If we break down what luck is, it’s a fraction, odds. “How many times out of x number of times will I strike out?”

The odds of founding a unicorn through scaling a lemonade stand is minimal, yet the odds of founding a unicorn through scaling a startup in an interesting space, solving a genuine problem, with some sort of emerging technology, is a lot higher.

Not only that, but factors like how many startups you’ve worked on prior, or the number of people you’re working with play huge roles as well. This isn’t hyperbole either, it’s backed by real data.

According to First Round’s 10-Year Project, teams of more than one founder outperformed solo founders by 163%.

Ali Tamaseb, who analyzed 195 unicorns, the odds of succeeding nearly double when moving from a single-person team to a two-person or three-person:

I can go on, but the point I’m trying to make is that there are things we can do to increase our odds, and naturally, increase our luck. We create our own luck through building our own opportunities when there’s none to exist — I love the concept of “The Third Door” for this exact reason.

The third door talks about the two ways we can enter a club — you can wait in the long line with the normies, or alternatively you can go to the VIP line.

Couple problems,

A: it’s cold, and waiting in a line sucks.
B: we’re not cool enough to be in VIP.

So what do we do? We make the third door ourselves — whether that door be paying off the bouncer, kicking in a window at the bathroom, acting like a famous person, etc.

Opportunities are rare because they’re out of our control, but we can make them more common if we create them ourselves, therefore putting them within our control, and within our reach.

For instance, one of my friends needed to raise VC money » no VC would sit down with him » he ordered pizzas, dressed as a delivery guy, and put a QR code inside each box. Physically WENT to the VC offices and delivered them » when the VCs opened the lid, it linked to his demo, pitch deck, and one-pager.

He got a call back from every VC firm he went to. And then raised some money.

If you still want to build…

At the end of the day, startups aren’t about pitch decks, or seed rounds. It’s about choosing to build something in a place where nobody can tell you what to do, because nobody’s figured it out yet. A startup is just the tool — the real thing is deciding to live in uncertainty on purpose. As long as there are problems, there will be people trying to solve them. And I’d rather be out there trying, even in the dark, than sitting around waiting for someone else to turn on the lights.

If you made it this far, thank you! This essay in particular was especially long, so the time you’ve given me means so much more than you might realize. I’m sure your a very busy person.

I’m using this blog experiment to let people get to know me + what I believe — and hopefully connect with more of you along the way. If you have any questions about what I’ve done, or would just like the chat — hit me up at pointone@barbaros.ca — if I don’t get back within 24 hours, coffee’s on me.

Once again, thanks for your time.

AI and humans consume content differently. While traditional writing styles have worked for decades, we now need to understand how to create content that serves both human readers and AI systems effectively. This article will show you the fundamental steps to accomplish exactly that.

After personally creating content for businesses that consistently gets featured in Google’s AI Overviews and cited by leading language models like ChatGPT, Perplexity, and Claude, I’ve identified the key strategies that drive both AI visibility and human engagement. The shift from traditional SEO to AIO (AI Optimization) / GEO (Generative Engine Optimization) requires a fundamental rethinking of content strategy.

Here’s what we’ll cover: crafting content that AI systems prioritize and how to leverage the right optimization tools.

TL;DR: What I Learned Getting My Content Cited by AI

After getting cited in Google’s AI Overviews, Perplexity, and ChatGPT, I found a clear pattern:
AI systems reward structured, scannable, and trust-rich content — not keyword dumps or long-winded narratives.

The key drivers?

• Answer-first formatting
• Credibility signals within the first 200 words
• Brand authority over backlinks
• Schema, sentiment, and multimodal presence (especially video)

Bottom line: The brands getting cited aren’t necessarily creating better content — they’re creating content that matches how AI systems are designed to extract and reference information.

This article breaks down the exact playbook I use to get my content featured by AI, while still making it valuable for human readers.

Writing Content That Both Humans & AI Systems LOVE

To crack this code, we first need to understand how and why AI systems cite specific sources. If you’re interested in the technical breakdown of the inference stage in AI, I’ve covered that in detail elsewhere.

Build “Answer-First” Pages

Here’s what I’ve learned: AI systems favour content that’s concise, direct, and scannable — qualities that also happen to drive human engagement.

The key is frontloading your answers (ie. incorporating a TL;DR at the beginning). Instead of burying insights deep in paragraphs, lead with the solution, then provide supporting context. This “answer-first” approach satisfies both AI’s need for quick pattern recognition and human readers’ desire for immediate value.

The most effective formats include:

• TL;DR at the top of content
• FAQ sections
• Structured how-to guides
• Definition-heavy content that provides immediate value

AI models process structured content more effectively than dense paragraphs. The formats that consistently get cited include descriptive headings that telegraph content, FAQ sections that directly answer common queries, step-by-step processes that break down complex procedures, and bullet points that highlight key information.

This isn’t coincidence — it’s how AI systems are trained to extract information. Google’s AI Overviews frequently pull from FAQ schema, while Perplexity and Claude often reproduce well-formatted lists and step-by-step guides. By structuring your content to match these patterns, you’re essentially speaking the AI’s native language while simultaneously improving human readability.

The Clarity Advantage

Content clarity directly correlates with AI citation rates. Through testing, I’ve found that simplifying language and improving readability consistently outperforms keyword-heavy approaches. AI systems struggle with jargon-dense content the same way humans do — they favour clear explanations over technical complexity.

This means writing shorter sentences, explaining concepts in plain language, and maintaining strong grammatical flow. The goal isn’t dumbing down your content; it’s making expertise accessible. When you can explain complex ideas simply, both AI systems and human readers gravitate toward your content as the definitive source.

Building Trust Signals That AI Systems Recognize

Trust isn’t just about human perception — AI systems actively evaluate credibility markers when selecting content to cite. Through extensive testing, I’ve identified the specific signals that boost both AI visibility and human confidence.

Strategic Citation Integration

Content that references external sources receives significantly higher AI selection rates.

But there’s a methodology here:

• Cite authoritative sources within the first 200 words
• Use diverse source types (academic papers, industry reports, reputable publications)
• Ensure citations directly support your core arguments.

Expert Voice Integration

Direct quotes from recognized authorities create immediate credibility signals. The most effective approach involves weaving expert perspectives throughout your content, not just dropping them in as isolated blocks.

For instance, a cybersecurity article that integrates insights from security researchers, official reports, and industry leaders throughout the narrative demonstrates comprehensive sourcing that AI systems recognize and prioritize.

Data-Driven Substantiation

Statistical evidence serves as content currency in the AI economy. Specific numbers, percentages, and research findings provide concrete value that AI systems can extract and reference. The key is presenting data that’s both current and contextually relevant — “95% of users reported improved efficiency” carries more weight than vague claims about “significant improvements.”

Technical Depth Without Density

Including relevant industry terminology and technical concepts signals topical authority to AI systems. In cloud computing content, naturally incorporating terms like Kubernetes, REST APIs, and containerization demonstrates subject matter expertise. The critical balance is maintaining readability while showcasing depth — technical accuracy that doesn’t sacrifice accessibility.

E-E-A-T Optimization for AI

Experience, Expertise, Authoritativeness, and Trustworthiness aren’t just Google ranking factors — they’re AI selection criteria.

This means:

• Clear author credentials
• Demonstrated first-hand experience
• Confident, authoritative tone throughout your content

AI systems favor content that reads as expert-written and thoroughly researched.

Digital Ecosystem Presence

Perhaps most importantly, AI systems heavily weight cross-platform mentions and discussions. Your content’s likelihood of being cited increases dramatically when your brand or insights appear across forums, social media, and industry discussions.

This means strategic community engagement and thought leadership efforts directly impact AI visibility — a factor many content creators overlook.

Review Velocity as Trust Currency

AI systems perform sentiment analysis on your brand’s digital footprint, making reviews a critical ranking factor.

The combination of review count, velocity, and quality across platforms like Yelp, Google Business, and industry-specific sites creates powerful trust signals. LLMs actively crawl these review sources to assess credibility, meaning your off-page reputation directly influences AI citation rates.

Brand Authority > Backlinks

Unlike traditional SEO, AI systems prioritize brand authority and topical relevance over link building. Mentions in trusted sources — even without direct links — shape how AI models summarize and cite your content. This represents a fundamental shift where earned media and thought leadership carry more weight than traditional link acquisition strategies.

Stack That Got Me into AI Overviews

Traditional SEO tools weren’t built for AI optimization. While keyword research and backlink analysis remain useful, the metrics that matter most for AI visibility require entirely different measurement approaches. After testing dozens of tools and platforms, I’ve identified the essential stack that drives real AI citation results.

Schema Markup and Structured Data Implementation

Just as traditional Google algorithms rely on clear signals, AI systems need structured data to interpret your content effectively. Schema markup serves as a direct communication channel with LLMs, providing context that dramatically improves citation rates.

As mentioned earlier, the most impactful schema types for AI optimization include FAQ schema for question-based queries, How-to schema for process-driven content, Review schema for trust signals, and Organization schema for brand authority. Tools like Schema Pro, Structured Data Markup Helper, and JSON-LD generators streamline implementation, but the key is strategic deployment rather than comprehensive coverage.

Content Optimization for AI Readability

Earlier, I mentioned the importance of clarity — something we can actively test and improve. Tools like Grammarly offer baseline readability checks, but optimizing for AI requires deeper analysis. I often use LLMs themselves (specifically Claude) to fine-tune grammar, tone, and structure for AI readability.

YouTube Optimization for AI Visibility

Gemini heavily favors YouTube content in AI Overviews, making video optimization a critical component of AI strategy.

Focus on short, educational videos tied to core queries, optimized titles that match conversational search patterns, and strategic keyword placement in descriptions and transcripts. The goal isn’t viral reach — it’s becoming the authoritative video source that AI systems cite when answering specific questions.

The Future is Already Here

The shift from traditional SEO to AI optimization isn’t coming — it’s already happened. While most marketers are still optimizing for search engines from 2019, the brands getting cited by AI systems today are the ones that will dominate tomorrow’s search landscape.

The playbook is straightforward: structure your content like AI systems are trained to process it, build trust signals that algorithms can recognize, and optimize for the tools that actually matter in an AI-first world. It’s not about gaming the system — it’s about understanding how information flows in the age of generative AI.

What we’ve covered here — from answer-first formatting to schema implementation to cross-platform brand building — these aren’t advanced tactics. They’re foundational strategies that will only become more important as AI systems become the primary way people discover and consume information.

The businesses that adapt to this reality now will have a massive advantage. The ones that don’t will find themselves invisible in an AI-driven search world.

Let’s Keep The Conversation Going!

Hey! I’m Noah, a 17 y/o who’s been deep in the GEO space for the past year, where I’ve been running Antifragility Labs to help brands rank in both search engines and AI systems.

If you found this useful or have questions about any of these strategies, feel free to reach out. I love meeting people who are thinking about these problems and always enjoy swapping notes on what’s working in the AI optimization world.

The future of content is being written right now, and the more of us who understand it, the better off we’ll all be.

In this article, I’ll break down what Google’s 2025 I/O means for your brand’s SEO — in just 3 minutes.

Google’s 2025 I/O happened a month ago — it’s no doubt that they’re moving towards an AI-first world. Here’s what actually matters from I/O 2025, no filler, just what you need to adapt now.

AI Mode Is Here (and It’s Default)

Google rolled out AI Mode on May 20. It’s powered by Gemini 2.5 and now handles multi-step queries automatically — think: “fan-out search” where it breaks your one question into five sub-questions. It’s not a beta. It’s the main show.

Example:

Query: “What’s the best laptop for video editing under $1,500 that supports 4K output and has long battery life?”

This isn’t a single question. It contains multiple layers:

1. Find laptops under $1,500.
2. Filter for video editing performance (CPU, GPU, RAM).
3. Confirm 4K output support (HDMI 2.0, DisplayPort, etc.).
4. Cross-check battery life.

LLMs like Gemini 2.5 handle this with “query fan-out”: they deconstruct the prompt into sub-questions, process each, then recombine them into a summarized output.

TL;DR: If your content doesn’t answer layered queries, you’re invisible.

SEO ≠ SEO Anymore

Organic links still exist, but AI Overviews now dominate top real estate. Google confirmed a +10% usage spike where AIO is active. That’s translating into less clicks, even for high-ranking sites. Some brands are losing 30–40% traffic overnight. Niche experts? They’re gaining 15–45%.

TL;DR: Ranking #1 doesn’t matter if AI doesn’t quote you.

In fact, Ahrefs an SEO leading company did a report on 25M+ AI Overviews and 300,000+ KW’s where they found these same AI Overviews correlate with a 34.5% drop in CTR.

GEO: Generative Engine Optimization Is the New Frontline

Forget keywords. You now optimize for AI citation. If Claude, Gemini, or ChatGPT don’t pull from you, you’re out of the loop.

Actionables:

• Use structured Q&A formats.
• Include citations and source-like phrasing.
• Add llms.txt and AI‑friendly metadata.

If you want to dive deeper into GEO, check out my article— these methods helped my clients rank across all relevant AI Overviews and LLM answers from models like GPT, Perplexity, Claude, and more.

AIO Is Scaling Fast

AI Overviews now hit 200+ countries and 40+ languages. Gemini 2.5 + Deep Search + Project Mariner means Google doesn’t just answer — it acts. That includes booking a flight or summarizing your site without ever clicking on it.

TL;DR: Google isn’t a search engine anymore. It’s a decision engine.

The Playbook Moving Forward

• Structure content like it’s answering a user, not gaming a bot.
• Focus on “being cited,” not clicked.
• Monitor AI-related queries in your data — don’t wait for Search Console to catch up.
• Diversify outside of Google (email, community, social).

You’re not fighting for rankings anymore. You’re fighting for relevance in a search landscape that’s writing its own answers.

Your All-in-One Solution

Antifragility Labs; your all-in-one solution.

We helps startups get discovered — by both search engines & AI alike.

We’ve worked with early-stage teams backed by Y Combinator, League of Innovators, Betaworks, Creative Destruction Lab, and more — adapting them to how discovery works now. Optimizing you for both SEO & GEO.

We’ve helped clients consistently appear and be cited as the top result across Google’s AI Overviews, ChatGPT, Perplexity, and Claude — the new front page of the internet.

Alongside that, we’ve:

• 11x’d impressions
• 5x’d clicks
• 5x’d backlinks
• Taken them from unranked to Page 1 for every relevant keyword

📅 Let’s connect and schedule a free audit for your business.

There’s no doubt that Search Engine Optimization (SEO) is a fundamental need for most businesses and startups — to rank on the top search engines like Google and Bing. However, the way the world searches and find s information has been revolutionizing around the world at an unprecedented pace. Before, there was one search engine giant: Google. Yet recently, Google’s top spot has been on the hot seat, with OpenAI’s ChatGPT setting the record for having the fastest growing consumer app of all time.

This brings me to exhibit A:

This chart highlights the acceleration in ChatGPT’s WAU growth alongside a steep year-over-year decline in Google’s Paid Click Growth. While correlation doesn’t always imply causation, the scale — with ChatGPT reaching 350 million WAUs — suggests there’s likely a direct impact. Additional signs of Google losing its grip on the top of the funnel include its search market share dipping below 90% for the first time since 2015.

In a 2025 analysis of 300,000 keywords, Ryan Law (Director of Content Marketing @ ahrefs) found that AI Overviews correlate with a 34.5% drop in average CTR for the #1 result compared to similar queries without Overviews.

The message is clear: people are no longer searching the same standardized way.

Businesses are still optimizing for Google — stuffing keywords, chasing backlinks, and ranking for SERPs — while missing the fact that AI like ChatGPT, Claude, or Google’s SGE are becoming the new interface for discovery.

Enter: GEO / Generative Engine Optimization

We all know what SEO is, but what’s GEO, and what’s the differentiation between SEO and GEO?

Traditional SEO (Search Engine Optimization):

• Focuses on making content rank higher on traditional search engines like Google, Bing, etc.
• Optimizes for keywords, metadata, backlinks, site speed, mobile friendliness, etc.
• Goal: Appear at the top of search engine results pages (SERPs) when users type queries.
• Based on how search engines index and rank webpages.

That is not how AI search tools function. They produce full answers, summaries, and even product recommendations by blending training data with real-time web searches.

GEO (Generative Engine Optimization):

• Newer concept designed for generative AI engines like ChatGPT, Perplexity, Claude, etc.
• Focuses on making content more likely to be cited, mentioned, or quoted by these AI tools.
• Optimizes for clarity, authoritativeness, originality, and structured data that LLMs ingest.
• Goal: Be included in AI-generated answers — not just traditional search rankings.

In simple terms:

• SEO = Rank higher on search engines.
• GEO = Be cited more often in AI-generated responses.

A few examples:

What traditional SEO looks like 👇

What GEO (Generative Engine Optimization) looks like in an LLM👇

What Google’s new Search Generative Experience (SGE), also known as AI Overviews looks like👇

How Do I Get My Business On There?

I’m not claiming any of the research made — I’m summarizing it in a way so easy to understand that a child can get the grip of it, and put all of the many (but few) reliable resources in one form.

I’m going to begin with one of the very first reports done on GEO — a paper from Cornell University.

This section can get very technical — but I’ll break it down.

A Generative Engine (GE), like BingChat or ChatGPT browsing mode, does two big things:

• It searches for information using a search engine.
• It generates natural language responses using AI models.

It works like this:

1. You send a query (like a question).
2. The system might reformulate (reword or split) your query into simpler ones.
3. These simpler queries are sent to a search engine to find a bunch of sources.
4. Another model summarizes each source.
5. A final model writes the full response based on the summaries and the search results.

The process can be captured in this formula created by the Cornell folks (don’t worry if it doesn’t make sense at first):

where:

• qu = your query
• pu = your personal info (preferences, history, etc.)
• r = the final response

Citations

• In the final response, each statement is ideally backed by a citation from the search results.
• Good GE behaviour = all sentences have accurate citations (high precision) and most citations are actually useful and connected to the sentences (high recall).

It’ll look like this 👇

Now we can apply that exact equation to what’s going on here:

🧠 Applying it to the example:

1. Input (User query):
   → qu = "who won? the Leafs or Senators?"
2. Personalization (PU):
   → Let’s say the user is Canadian, a hockey fan, or previously searched NHL playoff standings.
3. Process (fGE): The engine might:

• Reformulate the query for clarity (e.g., “Who won Game 6 of the Leafs vs Senators series?”)
• Use its search engine to pull sources (Reuters, Sportsnet, NHL.com)
• Summarize the content
• Generate a readable response combining all info

4. Output (r): → A natural language answer that clearly explains:

• The Leafs won Game 6, clinched the series 4–2
• Who scored, who assisted, who saved
• Citations used (displayed visibly)
• What happens next (Toronto plays Florida)

In it’s simplest form — a generative engine is just:

Search → Summarize → Respond,

While making sure every claim has a real, accurate source.

📊 How Citations Are Ranked: The Key Equation

One of the most important things a generative engine (like ChatGPT or Google’s AI Overviews) does is decide which sources to cite — and in what order. But it’s not random. Behind the scenes, the engine uses a scoring function to weigh each possible citation.

Each citation ci in the generated response r is evaluated using two key criteria:

Visibility:

This measures how prominently the citation appears. Is it early in the response? Is it mentioned multiple times? Is it part of long, important sentences?

Relevance:

This checks how on-topic the citation is — does it actually help answer the user’s query

The engine has an algorithm and tries to find a balance between these two scores for every citation and maximize the overall value across all citations. The black-box function f tells the engine how to combine “visibility” and “relevance” into a single score — but the exact method it uses isn’t publicly known.

So for instance, going back to this example from earlier (sorry Senator fans)

Let’s go back to the response for:

“Who won? The Leafs or Senators?”

The generated answer clearly states:

“The Toronto Maple Leafs defeated the Ottawa Senators 4–2 in Game 6 on May 1, 2025…”
And it’s backed by Reuters and Sportsnet as visible citations.

Here’s what happened behind the scenes:

• The model pulled multiple news sources.

Reuters had:

• High relevance: It directly confirmed the final score.
• High impression: It was cited early, in the opening sentence.
• Sportsnet may have supported details later on (like who advances next or player stats).

Because these sources scored highly on both visibility and relevance, they were selected and placed prominently in the response.

Other possible sources — less relevant, or cited in a short or late sentence — got deprioritized or left out.

So… how do you actually improve?

To evaluate how different strategies impact a website’s visibility within generative engine responses, the researchers applied Generative Engine Optimization (GEO) methods to modify existing website content using a large language model (LLM). The model was prompted with instructions tailored to each optimization method, rewriting the original content to reflect a specific stylistic or structural goal.

Each method targeted a distinct characteristic that might improve how content is interpreted and cited by a generative engine. The nine methods tested are:

1. Authoritative Style
   Rewrite your text to sound more persuasive, confident, and expert-driven — designed to project authority on the subject matter.
2. Statistics Addition
   Replace qualitative or vague claims with quantitative data whenever possible, this reinforces factual credibility.
3. Keyword Stuffing
   Increase the presence of user query keywords in the text, in line with traditional SEO practices.
4. Cite Sources
   Add explicit references to credible external materials to improve trust and transparency.
5. Quotation Addition
   Incorporate direct quotes from reliable sources to enhance authenticity and informativeness.
6. Easy-to-Understand Language
   Simplify technical or complex sentences, making the content accessible to a broader audience.
7. Fluency Optimization
   Improve grammar, coherence, and readability without altering the core message of the text.
8. Unique Word Addition
   Introduce less common vocabulary to distinguish the content from generic alternatives.
9. Technical Term Addition
   Insert relevant domain-specific language to increase content specificity and depth.

These methods fall into two categories. Most — including Authoritative Style, Statistics Addition, and Fluency Optimization — refine existing content without requiring any new information. They aim to improve how well-written, clear, or authoritative the text appears.

In contrast, Keyword Stuffing, Cite Sources, and Quotation Addition may involve introducing new material or references not found in the original content.

To evaluate performance, the researchers selected one source website at random for each user query and applied all nine GEO methods separately to that same source. This allowed for a controlled comparison of how each strategy impacted citation visibility and relevance.

These methods can boost visibility by up to 40%.

What Marketing Agencies Are Saying

Neil Patel and NP digital — arguably the biggest marketer and marketing agency out there did a report on how to rank on ChatGPT. This is what they came up with:

As you can see there is some overlap with the previous report from Cornell. This is the overview that Neil Patel gives of all those 6 factors:

Brand mentions — how many times was the brand, or product mentioned around the web? The more often it was mentioned we found the more likely ChatGPT recommended it.

Reviews — not only do reviews matter, but the more the better as well as the rating. So if you find a lot of people reviewing your company whether it is on your own site or on sites like Trust Pilot, Amazon, or even BBB, it helps (assuming the reviews are good in general).

Relevancy — there was a big correlation between the keywords in the question we asked ChatGPT and the products and companies it recommended. What we found is those keywords in almost all cases existed on web pages around the web mentioned the product and service mentioned.

Age — was the product or company recommended old or new? We found that older products or companies were recommended more often.

Recommendations — were other websites recommending the product or company? What’s funny is a lot of the suggestions that ChatGPT provided were pulled from sites that break down things like “the best credit cards for college students”. You know, those sites that list out products and make money from affiliate commissions? Funny enough it seems as if those sites are influencing the recommendations on ChatGPT even though they are really based on which company is paying the most affiliate commissions versus actually having the best offering.

Authority — What’s the social following of the companies recommended? What’s the domain authority? What type of sites are talking about the product, service, or company? Typically, the more authoritative the company is, the more it is recommended.

What I Believe Works:

Early research and industry experience have identified several content factors that increase the likelihood of being included or cited in AI-generated results. A 2023 academic study (Aggarwal et al., later published at KDD 2024) introduced GEO and tested various content optimization techniques. They found that certain strategies can boost a source’s visibility by up to 40% in generative engine responses. The top factors included:

• Incorporating Citations: Content that cites external sources or references tends to be viewed as more credible by AI. Including a few well-placed citations or links to authoritative sources can significantly boost visibility (foundationinc.co). Why? Generative AI, especially models designed to provide sources (like Bing or Perplexity), often prefer to use content that already shows evidence and sourcing, aligning with the AI’s goal of providing trustworthy info.
• Quotations from Experts: Adding quotes from subject-matter experts or reputable publications can improve your content’s chance of being picked (foundationinc.co). These quotations signal authority and richness. For example, an article about cybersecurity that includes a quote from a well-known security researcher or a line from an official report might be more likely to surface in an AI summary about cybersecurity tips.
• Statistical Evidence: Including data and statistics that back up your points can make your content more appealing for AI answers (foundationinc.co). Numbers and facts provide concrete value that an AI can share. (E.g., “95% of users reported X” or “Our study found a 40% increase in Y” could be the snippet an AI grabs to answer a query.)
• Content Clarity and Fluency: Writing in a clear, easy-to-understand manner (avoiding jargon, explaining concepts plainly) helps the AI parse your content. Fluency optimization — essentially good grammar and flow — was shown to improve results (foundationinc.co). Simpler language and short sentences can prevent misunderstandings by the model. (In the study, simplifying language and improving readability boosted visibility, whereas heavy keyword stuffing did not help.)
• Use of Unique & Relevant Terms: While “keyword stuffing” is discouraged (it didn’t fool the AI and can make text unnatural), it is beneficial to include relevant niche terms or technical terms where appropriate. This helps establish topical relevance. For instance, in a cloud computing article, mentioning specific frameworks or protocols (Kubernetes, REST API, etc.) where relevant can signal to the AI that your content covers the topic in depth. Just ensure these terms fit naturally.
• Authoritative Tone and E-E-A-T Signals: Content written in an authoritative voice with a confident tone tends to be more trusted by AI (foundationinc.co). Additionally, showing Experience, Expertise, Authoritativeness, and Trustworthiness (E-E-A-T) matters not just for Google’s algorithms but for AI selection too. This means having clear author bios, credentials, and demonstrating first-hand experience or insight. If your content reads as expert-written and trustworthy, it’s more likely to be chosen by the generative engine.
• Structured and Easy-to-Scan Format: AI models prefer content that is well-structured (much like humans do). Use of descriptive headings, bullet points, step-by-step lists, and FAQ sections can make it easier for the AI to extract relevant information (onely.com). Google’s SGE, for example, often pulls FAQ schema content or how-to steps into its answers. A Perplexity or Bing result might list out bullet points if your page had them. Well-organized content (with schema markup, which we’ll cover below) is more digestible to AI.
• Mentions and Reputation Across the Web: Generative AI heavily weighs how often and in what context your content or brand is mentioned in its training data or live data sources. If your brand is frequently discussed on forums (Reddit, StackExchange), Q&A sites, social media, or reviews, the AI is more likely to “know” about you and include you. For example, ChatGPT might recommend a product that it has seen mentioned in many Reddit threads or tech blogs, even if that product’s own site isn’t top-ranked on Google. This means your off-page presence (digital PR, community engagement) contributes to GEO.

Bottom line: Content that is comprehensive, credible, and contextually relevant is favored by generative engines. The goal is to become the source that an AI finds most useful to answer a user’s question. Many of these factors align with good content practices anyway (provide value, cite sources, be clear), but GEO puts extra emphasis on them because of how AI selection works.

GEO in Practice: Example & Emerging Trends

As an example of GEO success: Plausible.io reported a 2200% increase in traffic from AI search tools in 2024 after focusing on this new paradigm. They adjusted content to be more AI-friendly and saw platforms like ChatGPT, Claude, and Perplexity send significantly more visitors their way. This kind of result underscores that AI search can become a major traffic driver when content is optimized for it. Marketers are treating AI search as the “new Google” — the next battleground for organic visibility.

A few emerging best practices and trends to highlight:

• Focus on “Being the Answer,” Not Just Getting a Click: In traditional SEO, you optimize to get the click to your site. In GEO, you often need to deliver value even if the user doesn’t click (because the AI might give the answer directly). This means your content should be concise and complete in itself. Paradoxically, by doing so, you build trust and increase the chance that a user will seek more from you (or that the AI will mention your brand as a trusted source). It’s a shift from pure traffic acquisition to brand visibility and authority.
• Monitor AI-Specific Analytics: Google Search Console doesn’t yet show how you appear in SGE, but it may in the future. Bing Webmaster Tools has started showing if your content was used in Bing’s chat results. Keep an eye on these developments. In the meantime, track SEO metrics like zero-click searches and CTR drops on certain queries — a high drop might indicate an AI overview is stealing clicks. If your top-ranking page’s clicks are falling while remaining #1, it could be because an AI summary is satisfying the query. That’s a prompt to double down on GEO for that query (ensuring you are part of that AI summary).
• Address AI Hallucinations or Errors About Your Brand: Sometimes AI might give incorrect info about your product or misattribute a quote. If you encounter such issues (either through your own testing or user reports), consider publishing content to correct it. For example, a Q&A on your site like “Misconceptions about [Our Product]” or adding clarifications in an FAQ. This gives the model a factual source to pull from, potentially overriding the misinformation (convert.com). Regularly asking AI “What is [Your Company]?” and seeing the answer can be a good practice; if it’s wrong, take steps to fix the record with new content.
• Blend GEO with Traditional SEO for a Holistic Strategy: GEO isn’t replacing SEO; it’s an evolution. High-quality content and solid SEO practices feed GEO success. For instance, ranking well on Google still matters because AIs often draw from the top Google results for fresh info. So continue optimizing for both. Think of it as Search Everywhere Optimization — making your content visible wherever people search, be it a search engine or an AI assistant. The companies that thrive will be those who cover all bases: their content ranks in Google, is referenced by AI answers, and is discussed in communities.

TL;DR

In practice, GenAI tools surface content based on:

• Mentions in trustworthy sources
• How well your content answers the actual question being asked. (Context matters here: ChatGPT, for example, will recommend different tools to a budget-conscious SaaS optimizer than to a Fortune 500 experimentation lead. Pre-AI Google would’ve treated those queries the same.)
• Site structure, crawlability, and clarity
• Signals like reviews, discussions, and real-world use cases.

Traditional SEO fundamentals still apply in AI search optimization (AISO): content quality, structure, authority, and intent alignment. But the ranking mechanism is different.

Statistical relevance, contextual mentions, and clarity of information now carry the most weight.

Content not ranking high in Google might still appear in a GenAI summary, especially if it’s well-structured and cited across credible platforms.

What We’re Building at Antifragility Labs

If you’ve made it this far — thank you. And you might be wondering how to actually apply all this — this is where I come in. At Antifragility Labs, I work with early-stage startups (backed & supported by Y Combinator, League of Innovators, Betaworks, Creative Destruction Lab, and others) to help them adapt to how discovery works now.

We build systems that make your content easier for generative engines to understand, cite, and recommend. That means rewriting key pages with citation-first structure, testing what kinds of content AI surfaces, and tracking how models actually talk about your brand. It’s not about chasing clicks — it’s about showing up when it matters.

If you’re building something and want to experiment with this, reach out:
📩 noah@barbaros.ca
🌐 antifragilitylabs.com
📅 Or just book a time here and let’s chat.

About Me:
Hey! I’m Noah. A highschool kid obsessed with entrepreneurship and AI.

I’m all about jumping at every opportunity. I don’t see the downside of taking on challenges and figuring things out.

Entrepreneurship is my addiction — I live by: “The best way to predict the future is to create it.”

If you ever want to talk about anything entrepreneurship feel free to reach out! I’m always open to new connections and opportunities even if it’s just a chat!

Reach out to me anytime at noah@barbaros.ca or on my LinkedIn

Also make sure to check out my monthly newsletter here where you’ll
see more of my progress into an amazing entrepreneur 🦄, and see exactly how I figure it out.

Follow for more articles and thanks for reading!

The way the world searches and discovers information is undergoing its biggest shift since the internet was created.

In November 2022, ChatGPT became the fastest-growing app of all time. Since then, AI-powered search has taken over. Google’s Search Generative Experience (SGE) — a feature that changes the rules of online search — now appears in over 87% of search queries. Instead of the traditional “10 blue links”, millions of users get AI-generated summaries. A 2025 study by Ahrefs analyzing 300,000 keywords found that AI Overviews correlate with a 34.5% drop in CTR for the #1 result compared to queries without Overviews. Simultaneously, 27% of Americans are “ditching” traditional search engines like Google for AI chatbots such as ChatGPT.

The way people find information has revolutionized — and not a single brand has an idea how to show up.

This is where Antifragility comes in.

The long-term mission of Antifragility is to help companies win visibility in a world where AI is the new frontier for search.

Not just traditional SEO where we stuff keywords, or write blog posts.

We optimize for generative engines, the LLMs and AI overviews that are now the first place people discover, evaluate, and convert.

The Blueprint

Like all new technologies, generative search brings chaos before clarity. My approach is to work in this unknown — learn by doing, test what works, and then build products that scale it.

So the plan is simple:

1. Start as a consulting agency — working hands-on with startups and brands to rank in SGE and LLMs.
2. Use real data from the field to identify repeatable patterns and unmet needs.
3. Build software that automates the highest-value parts of the process. Replacing human capital with AI as leverage.
4. Transform into the go-to GEO product platform — making AI optimization as measurable and actionable as SEO.
5. Define and dominate the GEO category.

What Makes GEO Different

To rank in generative engines, you need more than backlinks and meta descriptions.

You need to understand inference. How LLMs evaluate credibility. How they hallucinate. How they cite. And how to influence the underlying models.

This is where we operate.

We combine core SEO fundamentals with deep research into how LLMs parse, retrieve, and synthesize information. And we use that knowledge to build the systems needed that make your brand show up where it matters most.

Why Me

In 2024, while working at Armilla AI (YC W22, $6M+ raised), I was working on SEO when I asked myself: How would this work for LLMs?

That question pulled me into the rabbit hole. I’ve been studying AI search, inference patterns, and generative engines ever since.

Now, I’m building this company, Antifragility, focused entirely on solving that question at scale.

Why Antifragility

Incase you couldn’t tell by the name, one of my favourite books is Antifragile: Things That Gain from Disorder by Taleb. The idea of this book is that Antifragile systems don’t just survive stress — they get stronger because of it.

That’s the guiding philosophy of how I want this company to operate: every client project, failed test, or algorithm shift becomes a feedback loop to iterate.

We learn. We adapt. We evolve.

In an industry as unpredictable as AI search, fragility breaks, resilience endures, and Antifragility wins.

The Long Game

The interface of search is changing. Whether it’s through SGE, ChatGPT, Perplexity, or something entirely new — AI is the new way people source and find information.

That’s why I’m building Antifragility.

To become the pioneer of brand visibility in the LLM-powered web.

To help companies own the answers AI gives.

And to make generative engine optimization the moat for my clients.

Cheers,
Noah

From Zero to MVP: How AI Made Me a “Developer”

At 16, I had no clue how to code. A few months later, I still don’t — yet somehow, I built a website around an idea that had stumped me for over a year. Sure I’ve done some coding, but I wouldn’t call myself a “developer.” My real passion is building products and launching ideas — but there’s always been one big challenge: finding a tech partner.

Two Types of Founders

In entrepreneurship, there are two types of founders: technical and business-oriented.

• Technical founders can wake up with an idea and start building immediately.
• Business founders have the vision but lack the skills to build— meaning they either need a technical co-founder or hire developers. Otherwise, the idea remains just that — an idea.

For now, I consider myself the second type, on the business side. Without a technical cofounder, an idea can stall before it even starts. And for young entrepreneurs like me, there’s an even bigger challenge: credibility.

The Challenge of Being a Young Entrepreneur

In the business world, your projects define you. You don’t have a fancy university degree, let alone a high school diploma, to prove your skill set. The only way to stand out and provide credibility to yourself, is through relentlessly grinding on projects, shipping ideas, and proving your worth through execution. An example of this: A survey of Ivy League and top business school admissions officers found that applicants with independent projects or startups were 5x more likely to be admitted compared to those with just strong grades and test scores.

The takeaway? Experience compounds. Every project, success or failure, increases your likelihood of long-term success. But before you can gain that experience, you need to find people to build with.

Why?

When working in teams, you access and utilize diverse perspectives and skillsets, hold one another accountable — the list goes on. I mention this statistic again later on, but First Round’s 10-Year Project, teams of more than one founder outperformed solo founders by 163%.

This ability to collaborate and leverage different strengths is what separates great ideas from actual businesses, making it essential for young entrepreneurs to surround themselves with the right people. And thats what makes your network so important.

The Problem: Crickets

This is where I hit my biggest roadblock.

Over the past summer, I reached out to over 100 teenagers on LinkedIn, hoping to find ambitious peers to work on projects with. Almost every message was met with silence. Either they weren’t interested, or I simply wasn’t finding the right people.

I was shocked by the lack of ambition these teens had. I know that there’s entrepreneurial teens out there, and I know there’s communities of them that are growing rapidly, but where were they?

And that’s when I realized something:

LinkedIn sucks for teenagers

LinkedIn is the world’s largest, and only professional network, but for teen founders, it’s completely ineffective.

• It’s an online resume platform, not a place to build.
• Connections feel transactional rather than meaningful.
• Finding like-minded teen entrepreneurs is nearly impossible.

And that led me to a simple but important question:

Why isn’t there a real network for entrepreneurial teens?

This question ran laps around my mind for the past year.

I knew ambitious teen founders existed — I had met some. But outside of structured programs, they were nearly impossible to find.

That’s when I started thinking: If they’re not on LinkedIn, where are they?

Then it clicked.

During my time at TKS and SHAD — two amazing teen entrepreneurship accelerators — I was surrounded by incredibly driven young builders, where all they wanted to do was build.

I was apart of these invaluable networks, and in business your network can be your greatest asset. And inside them, I saw something powerful.

Inside of the TKS Slack channel, there’s a #g-teamfinder,

And this channel is active. Constantly buzzing with new projects, and awesome opportunities to collaborate and work on some amazing stuff.

It was like seeing the answer to my question in a Slack channel.

Every day, teens were launching projects, forming teams, and finding partners within minutes. It was the kind of organic, fast-moving collaboration I had been searching for.

All of these teenagers are eager to collaborate, to create, to push boundaries. It’s like finding cofounders / team members on demand.

Outside of these circles, though, ambitious teens struggle to connect.

So I began asking: why?

Why does somebody have to enroll in a program like TKS to be apart of this network of ambitious teenagers?

Why can’t this ‘network’ be the norm for teenagers who don’t have access to these programs?

Because if we look at ‘adult’ entrepreneurs, they have countless networks, conferences, and incubators available to them. So why not teen entrepreneurs? Because the truth is, the future is teens. They are becoming more innovative, more entrepreneurial, and yet most of the world hasn’t recognized it. In fact, 76% of Gen Z aspire to be their own bosses, surpassing millennials (57%), Generation X (36%), and baby boomers (25%). Not only that, but a study done in 2023 revealed that 93% of Gen Z individuals have explored steps toward establishing their own businesses, with 75% aiming to become full-time entrepreneurs.

Yet, despite this surge of entrepreneurs in this new demographic, it’s constantly overlooked. The infrastructure to support their ambitions simply doesn’t exist at scale, leaving a massive gap that few are addressing.

Right now, teen founders are scattered across different communities and networks, struggling to connect:

• Some are in competitive startup accelerators like TKS, or SHAD.
• Others are in random Discord groups or Twitter threads, with no structured way to form teams.
• Many just have ideas but no network to bring them to life.

This separation and divide is the problem.

The biggest breakthroughs in tech have always come from small, passionate groups of builders working together. CoVenture makes that process frictionless for teen founders.

And I thought to myself, if you’re a teen looking to build, why should you struggle to find your network? What can’t there be a platform that’s a centralized hub and community for these ambitious teenagers?

So, I built one. CoVenture.

CoVenture is a dedicated network for ambitious teens to connect, collaborate, and build real, amazing projects together.

CoVenture - Teen Entrepreneur Community

How AI Made Me a “Developer”

How’d I build an MVP for CoVenture, if just a few months ago, I didn’t even know how to buy a domain name. Sure, I’ve done some coding, but I wouldn’t call myself a “developer.” My real passion is building products and launching ideas.

React? Vite? Supabase? No clue.

So how did I go from zero experience to launching a real platform?

Did I find a technical partner?

No. This idea was very personal to me, a challenge I’ve been facing for years, so I wanted to try making it on my own accustomed to my own vision. Then try to bring people on.

So how’d I build it?

Artificial intelligence.

I was completely lost. I had a ton of questions, like:

How do I start?

How do I set up a database? What even is a database?

How do I run a server locally?

With each problem, I turned to Cursor. Instead of searching through endless Stack Overflow threads, I simply described what I needed, or what I was lost on, and AI guided me through the process. It not only generated code and explained concepts, but it guided me using tools I’d never even heard of before like Supabase, and Vercel. And without spending years learning to code, I managed to equip the skills of a mid-level developer in a fraction of the time.

Suddenly, I wasn’t just the “business” founder. I was actually building. And I’m not alone:

• ~80% of Y Combinator’s latest batch used AI tools like Cursor in their development process.
• In that same video, YC CEO Gary Tan mentions that some startups won’t hire developers who refuse to utilize AI tools like Cursor, because it creates a massive productivity gap.

I currently have features like email authentication, real time posting for listings, connections, and messaging.

This is a complete game-changer for startups. With AI, you can go from idea to MVP in hours, not months. Whether you’re competing in a hackathon or launching a startup, the barriers to building have never been lower.

But I’ll be honest, I failed — a lot. It wasn’t at all easy at first.

I tore everything down and rebuilt it at least seven times. I had no roadmap. I had no experience. But I started anyway.

CoVenture was my first large-scale project. I had never built a website before. But AI enabled me to move fast, learn in real time, and iterate on my mistakes. And after a month of consistent work (and more all-nighters than I care to admit), CoVenture launched.

If you’ve never built a website, perfect. You don’t need to know everything. Just start. Have an idea? Cool. Build it.

AI isn’t replacing developers by any means — but it’s accelerating how quickly ideas become reality. The future of startups isn’t just about who can code. It’s about who can build.

There’s Always a Catch

But guess what, even if you can use AI to build an MVP, you’re still better off if you have a cofounder, or atleast bring on people as a CTO or founding engineer.

AI is an incredible tool — it can write code, streamline development, and get your product to market faster than ever before. It’s especially amazing for MVP’s, in my opinion, I don’t see how why anybody would ever try to make an MVP from scratch (of course with some exceptions like maybe the idea is too complex).

But building a successful startup requires more than just pushing out an MVP. It takes vision, strategy, and the ability to navigate the unpredictable challenges that every entrepreneur faces. Especially scaling.

A cofounder isn’t just another set of hands on deck — they bring fresh perspectives, complementary skill sets, and, most importantly, accountability. When you hit a wall (and trust me, you will), a cofounder can offer creative solutions, help you pivot, or simply keep you grounded.

There’s also the mental side of it. The startup journey is tough, and having someone to share the highs and lows with can make all the difference. While AI can generate solutions, it can’t offer real-time human support, debate ideas over coffee, or celebrate wins with genuine enthusiasm.

Studies back this up:

• According to First Round’s 10-Year Project, teams of more than one founder outperformed solo founders by 163%.
• According to Ali Tamaseb, who analyzed 195 unicorns, the odds of succeeding nearly double when moving from a single-person team to a two-person or three-person.

This is exactly why CoVenture exists — to bridge that gap, to connect young founders with the right partners, and to transform ideas into thriving ventures. Whether you’re a business-minded person looking for a tech partner or a technical wizard needing a strategic visionary, CoVenture is the space to find your match.

Because at the end of the day, AI can help you build, but a great cofounder can help you thrive.

Final Thoughts

Building a startup is hard. Doing it alone is even harder. AI can help you build faster and smarter, but it can’t replace the depth of collaboration, the exchange of ideas, or the resilience that comes from having a cofounder by your side.

CoVenture is the first network for ambitious teens to find their people, launch their projects, and transform ideas into reality. Whether you’re a developer or on the marketing-side, there’s a place for you at CoVenture, and a network of opportunities.

If you’re a teen founder with big ideas, don’t let the lack of a network hold you back. Join CoVenture, find your team, and start building. Because while you could do it alone, you really shouldn’t — and with the right people around you, you won’t have to.

CoVenture is in it’s MVP stage, and iterating for the consumer and designing a 10 star experience is very important to us right now — if you want to learn more about it, or have any suggestions / questions, please reach out to me anytime!

Reach out to me anytime via email noah@barbaros.ca or through LinkedIn.

CoVenture - Teen Entrepreneur Community

About Me:
Hey! I’m Noah. In short —Obsessed with entrepreneurship and AI.

I’m all about jumping at every opportunity. I don’t see the downside of taking on challenges — no matter how ‘unqualified’ I might seem, I’m confident in myself that I will figure it out and do it with high expectations.

The best part? I’ll learn a ton.

Entrepreneurship is my addiction — uncharted, unpredictable, and mine to shape. I live by: “The best way to predict the future is to create it.”

If you ever want to talk about anything entrepreneurship feel free to reach out! I’m always open to new connections and opportunities even if it’s just a chat!

Reach out to me anytime at noah@barbaros.ca or on my LinkedIn

Also make sure to check out my monthly newsletter here where you’ll
see more of my progress into an amazing entrepreneur 🦄, and see exactly how I figure it out.

Follow for more articles!

Let me just start by saying: I have always been fascinated by the power of data. However, before embarking on this project, my experience with programming was limited to basic scripting — the furthest I’d gone was a step by step YouTube tutorial to put it into perspective haha.

With such a beginner level understanding, how did I create a robust Twitter Sentiment Analyzer that can:

• Fetch real-time tweets based on user-defined queries
• Clean and preprocess tweet text for analysis
• Analyze sentiments using natural language processing
• Visualize sentiment distribution with intuitive charts
• Save and debug processed data for future reference?

I’ll walk you through each step below! However feel free to check out my tutorial for a full walkthrough on how this works!

Before you start, ensure you have Python installed on your machine along with the necessary libraries: tweepy, textblob, matplotlib, and re. You can install these using pip:

pip install tweepy textblob matplotlib

Step 1: Loading Configuration

Configurations like API keys and default settings are crucial for accessing Twitter’s API and customizing the behavior of our analyzer.

Comparable to setting up your workspace, loading configurations ensures everything is in place before diving into the code.

import json

def load_config():
    with open("config.json", "r", encoding="utf-8") as f:
        data = f.read()
        print("File content:", data)  # Debugging: ensure the JSON is read correctly
        return json.loads(data)

Explanation:

• Purpose: Reads the config.json file containing sensitive information like API keys.
• Debugging: Prints the file content to verify correct reading (remove in production).
• JSON Parsing: Converts the JSON string into a Python dictionary for easy access.

Sample config.json:

{
    "bearer_token": "YOUR_TWITTER_BEARER_TOKEN",
    "default_tweet_count": 50,
    "polarity_positive_threshold": 0.2,
    "polarity_negative_threshold": -0.2
}

Visual Aid: Imagine a neatly organized toolbox where each tool has its designated place — this is what config.json achieves for your project.

Step 2: Authenticating with Twitter API

Authentication is like getting your ticket to access Twitter’s data. Without it, you can’t fetch any tweets.

import tweepy

def authenticate(config):
    # For Twitter API v2, we primarily use the bearer token
    client = tweepy.Client(bearer_token=config["bearer_token"])
    return client

Explanation:

• Tweepy Client: Utilizes the bearer token from the config to authenticate requests.
• Security: Keeps API keys secure by storing them in a separate configuration file.

Best Practice Tip: Never hard-code your API keys in your scripts. Always use environment variables or configuration files excluded from version control.

Step 3: Fetching Tweets

Fetching tweets based on user-defined queries is the core of our analyzer.

def fetch_tweets(client, query, count):
    # Twitter API v2: search_recent_tweets can fetch up to 100 tweets per request
    max_results = min(count, 100)
    # Append `lang:en` to the query to filter English tweets
    query = f"{query} lang:en"
    response = client.search_recent_tweets(query=query, max_results=max_results, tweet_fields=['text'])
    return response.data if response.data else []

Explanation:

• Query Parameters: Filters tweets in English to maintain consistency in analysis.
• Rate Limits: Caps the number of tweets per request to 100, adhering to Twitter’s API limits.
• Error Handling: Returns an empty list if no tweets are found to prevent crashes.

Visual Comparison:

Before Fetching:

[Empty]

After Fetching:

["I love sunny days!", "Feeling sad about the news...", ...]

Step 4: Preprocessing Tweets

Raw tweet data can be messy. Preprocessing cleans the text, making it suitable for sentiment analysis.

import re
def preprocess_tweet(text):
    text = re.sub(r'RT\s+', '', text)           # Remove retweets
    text = re.sub(r'@\w+', '', text)           # Remove mentions
    text = re.sub(r'http\S+', '', text)        # Remove URLs
    text = re.sub(r'\s+', ' ', text).strip()   # Remove extra whitespace
    return text

Explanation:

• Regex Patterns: Efficiently removes unwanted parts of the tweet.
• Clean Text: Ensures the sentiment analysis focuses solely on meaningful content.

Analogy: Think of preprocessing as cleaning ingredients before cooking — removing impurities to enhance the final dish’s flavor.

Step 5: Analyzing Sentiment

Using natural language processing to determine the sentiment of each tweet.

from textblob import TextBlob

def analyze_sentiment(tweets, pos_threshold, neg_threshold, debug=False):
    positive_count = 0
    neutral_count = 0
    negative_count = 0
    processed_tweets = []
    for tweet in tweets:
        cleaned_text = preprocess_tweet(tweet.text)
        blob = TextBlob(cleaned_text)
        polarity = blob.sentiment.polarity
        if polarity > pos_threshold:
            positive_count += 1
        elif polarity < neg_threshold:
            negative_count += 1
        else:
            neutral_count += 1
        if debug:
            processed_tweets.append((cleaned_text, polarity))
    return positive_count, neutral_count, negative_count, processed_tweets

Explanation:

• TextBlob: Analyzes the polarity of the tweet, ranging from -1 (negative) to +1 (positive).
• Thresholds: Customizable to define what constitutes positive, neutral, or negative sentiments.
• Debugging: Optionally stores processed tweets and their polarity for further inspection.

Visual Aid: Imagine a scale where tweets are placed based on their sentiment score — some leaning towards positive, some neutral, and others negative.

Step 6: Plotting Results

Visualizing the sentiment distribution helps in quickly understanding the overall mood of the tweets.

import matplotlib.pyplot as plt

def plot_results(positive, neutral, negative):
    labels = ['Positive', 'Neutral', 'Negative']
    counts = [positive, neutral, negative]
    colors = ['green', 'blue', 'red']
    
    plt.figure(figsize=(8, 6))
    plt.bar(labels, counts, color=colors)
    plt.title('Sentiment Distribution')
    plt.ylabel('Count')
    plt.xlabel('Sentiment')
    plt.ylim(0, max(counts) + 10)
    
    for i, count in enumerate(counts):
        plt.text(i, count + 1, str(count), ha='center', va='bottom')
    
    plt.show()

Explanation:

• Bar Chart: Clearly displays the number of positive, neutral, and negative tweets.
• Customization: Colors and labels enhance readability and visual appeal.
• Annotations: Adds count labels above each bar for precise information.

Best Practice Tip: Always label your axes and provide a title to make charts self-explanatory.

Step 7: Running the Main Function

Integrating all components to execute the sentiment analysis workflow.

def main():
    config = load_config()
    client = authenticate(config)

query = input("Enter a search term: ")
    count_input = input(f"Number of tweets to fetch (default: {config['default_tweet_count']}): ")
    try:
        count = int(count_input) if count_input.strip() else config["default_tweet_count"]
    except ValueError:
        print("Invalid input. Using default tweet count.")
        count = config["default_tweet_count"]
    tweets = fetch_tweets(client, query, count)
    if not tweets:
        print("No tweets found. Please try a different query.")
        return
    positive, neutral, negative, _ = analyze_sentiment(
        tweets,
        pos_threshold=config.get("polarity_positive_threshold", 0.2),
        neg_threshold=config.get("polarity_negative_threshold", -0.2),
        debug=False
    )
    print("\nSentiment Analysis Results:")
    print(f"Positive: {positive}")
    print(f"Neutral: {neutral}")
    print(f"Negative: {negative}")
    show_chart = input("\nShow chart? (y/n): ")
    if show_chart.lower() == 'y':
        plot_results(positive, neutral, negative)
    debug_print = input("Print cleaned tweets for debugging? (y/n): ")
    if debug_print.lower() == 'y':
        _, _, _, debug_tweets = analyze_sentiment(
            tweets,
            pos_threshold=config.get("polarity_positive_threshold", 0.2),
            neg_threshold=config.get("polarity_negative_threshold", -0.2),
            debug=True
        )
        print("\nSample Cleaned Tweets:")
        for t, p in debug_tweets[:10]:
            print(f"\"{t}\" - Polarity: {p}")

Explanation:

• User Interaction: Prompts users to input search terms and desired tweet counts.
• Error Handling: Validates user input and handles scenarios where no tweets are found.
• Conditional Execution: Offers options to display charts and debug information based on user preference.
• Sample Output: Displays a subset of cleaned tweets with their polarity scores for deeper insights.

Best Practice Tip: Always validate user inputs to make your program robust against unexpected or invalid data.

Conclusion

Building a Twitter Sentiment Analyzer from scratch was an enlightening experience that bridged my understanding of data fetching, natural language processing, and data visualization. By integrating Python’s powerful libraries — Tweepy, TextBlob, and Matplotlib — I was able to create a tool that not only analyzes sentiments but also presents the data in an intuitive manner.

What Did I Learn?

• API Integration: Gained hands-on experience with Twitter’s API using Tweepy, understanding how to authenticate and fetch real-time data.
• Data Cleaning: Learned the importance of preprocessing raw data to ensure accurate sentiment analysis.
• Natural Language Processing: Utilized TextBlob to analyze and interpret the sentiment of textual data effectively.
• Data Visualization: Enhanced my ability to present data through clear and informative charts using Matplotlib.
• Error Handling: Implemented robust error handling to make the application more user-friendly and resilient.

What’s the Bigger Picture?

In an era where data drives decisions, sentiment analysis tools like this one can be invaluable for businesses, researchers, and individuals seeking to gauge public opinion or emotional trends. Imagine scaling this project to:

• Real-Time Monitoring: Integrate with live data streams to monitor sentiment in real-time.
• Advanced Analytics: Incorporate machine learning models for more nuanced sentiment detection.
• User Interface: Develop a web-based interface for broader accessibility and user interaction.
• Multi-Language Support: Expand the analyzer to handle tweets in multiple languages, increasing its global applicability.

Future Enhancements:

• Subscription Models: Offer premium features such as historical data analysis, detailed reports, and personalized dashboards.
• Integration with Other Platforms: Extend the tool to analyze sentiments across various social media platforms like Facebook, Instagram, and Reddit.
• AI and Machine Learning: Leverage more sophisticated models to improve sentiment accuracy and context understanding.

Final Thoughts

Creating the Twitter Sentiment Analyzer was not just about writing code; it was a reminder that no matter how hard the project might be there’s always a way to figure it out. I’m not some crazy established coder! Yet I still found a way to finish this project. Relatively quick too!

So, why wait? Start building your own sentiment analyzer today and explore the countless possibilities that data science has to offer!

See you next time!

Additional Resources

Source Code (Github Repo)

Twitter Developer Portal

Tweepy for interacting with the Twitter API

TextBlob for sentiment analysis

Matplotlib for data visualization

Regex (re) for text preprocessing

About Me:
Hey! I’m Noah. In short — I’m crazily obsessed with startups, product-market fit, AI, and becoming a unicorn founder.

I’m all about action and jumping at every opportunity. I’ve never seen the downside of taking on challenges. No matter how unqualified I might be — I don’t care! I’ll figure it out.

I love everything about entrepreneurship and startups — it’s an obsession haha.

If you ever want to talk about anything entrepreneurship feel free to reach out! I’m always open to new connections and opportunities even if it’s just a chat!

Reach out to me anytime at noah@barbaros.ca or on my LinkedIn

Also make sure to check out my monthly newsletter here where you’ll
see more of my progress into becoming a unicorn founder 🦄, and see exactly how I figure it out.

Follow for more articles!

We like to think we’re rational investors, driven by data and diligent analysis. But let’s be real: markets don’t just move on numbers — they swing wildly on emotion, hype, and headlines. Think about it. One tweet, one scandal, one new product launch, and a company’s stock can skyrocket or tank, sometimes for reasons that have nothing to do with the balance sheet. This is the hidden power of sentiment in the stock market: it’s not about what is happening, but how people feel about it.

In a world where perception often outweighs reality, the ability to read market sentiment becomes a critical skill. That’s where this project steps in. With a unique combination of sentiment analysis and real-time data scraping, we’re capturing the heartbeat of the market for companies like Amazon, Google, and Meta— seeing beyond the financial reports to understand the emotion driving the prices.

Whether you’re an investor looking for an edge or simply fascinated by the psychology of finance, this project lets you see the market as more than numbers on a screen. Get ready to dive into the minds of the market movers and see what makes prices tick.

A Classical Approach to Stock Analysis 📊

In the world of traditional investing, tracking a stock’s sentiment means staying updated with news, reports, and analyst opinions — all of which take considerable time to read and interpret. Often, investors rely on financial advisors, news summaries, or quantitative analysis platforms to stay informed, but each of these is either time-consuming or expensive. What if we could combine the power of natural language processing (NLP) and sentiment analysis to speed up this process?

Manual Sentiment Analysis: A Tedious Process

A manual approach to reading stock sentiment looks something like this:

1. Read the Headline: Manually go through each piece of financial news.
2. Interpret the Tone: Decide if the news is positive, neutral, or negative.
3. Record the Sentiment: Track these interpretations in a database.

This method quickly becomes impractical. The speed and sheer volume of financial news demand a faster, more automated solution.

Sentiment Analysis with VADER: A Modern Solution 🧠

Enter VADER (Valence Aware Dictionary and sEntiment Reasoner). VADER is an NLP tool specifically designed for analyzing the sentiment of short texts, making it ideal for headlines. Here’s how VADER transforms headlines into scores:

• Positive, Neutral, and Negative Scoring: Each headline is broken down into scores across positive, negative, and neutral categories.
• Compound Score: The overall sentiment is expressed as a compound score ranging from -1 to +1, with positive scores indicating optimism and negative scores signaling concern.

Let’s consider a basic example:

• A headline like “Amazon breaks sales records this quarter” might receive a high positive score.
• In contrast, “Market challenges affect Meta’s profits” could score negatively.

VADER’s algorithm sorts these scores, allowing us to quickly see which companies are trending positively or negatively in the market.

The Naive Bayes Algorithm: The Engine Behind Sentiment Analysis

Our project taps into the fundamentals of sentiment analysis through a statistical approach known as the Naive Bayes Algorithm. This algorithm, grounded in probability, predicts whether a headline is likely to impact a stock positively or negatively. It works by evaluating the likelihood of certain words or phrases appearing in either positive or negative reviews based on historical data.

In this model, we let “X” represent all the headlines or news snippets, while “Y” becomes the final output score, where a positive score indicates an optimistic sentiment and a negative one a pessimistic sentiment. Here’s a look at the process:

• P(Y=Positive | X) gives us the probability that the market sentiment is positive given a particular headline.
• P(Y=Negative | X) does the same for negative sentiment, allowing us to gauge the influence of news on a stock’s public perception.

VADER: Quick, Rule-Based Sentiment Scoring

We use VADER (Valence Aware Dictionary and sEntiment Reasoner), a rule-based model built for general sentiment analysis. VADER scores text for positive, neutral, negative, and compound (overall) sentiment. The beauty of VADER is its efficiency: it doesn’t analyze every word but focuses on those that carry emotional weight, skipping filler words like “the” or “and.”

This tool enables us to calculate sentiment on a massive scale — quickly scoring each headline without heavy computational demands.

Web Scraping and Real-Time Analysis

Of course, analyzing sentiment would be meaningless without real-time data. Using Python’s web scraping tools, we fetch live headlines and stock news directly from Finviz, a leading financial website. Our scraper identifies and extracts headlines tied to specific tickers — AMZN, GOOG, META— pulling them into our analysis pipeline. By examining the latest news, we can dynamically update the sentiment score based on recent events.

Our project then structures the scraped data, linking each headline with its timestamp and sentiment score to create a live snapshot of how the market feels about each company. The final result? An insightful, constantly updating bar chart displaying sentiment over time for each stock.

Coding the Project: Step-by-Step Implementation 👩‍💻

This section breaks down how I coded a simple sentiment analysis tool that scrapes headlines from Finviz, analyzes them with VADER, and visualizes the results.

Step 1: Data Collection — Gathering Headlines from Finviz

The first part of our process involves collecting data. For this project, we pull stock headlines directly from Finviz for tickers like Amazon (AMZN), Google (GOOG), and Meta (META), which serve as the foundation for our sentiment analysis.

Here’s the code for scraping headlines:

from urllib.request import urlopen, Request
from bs4 import BeautifulSoup

tickers = ['AMZN', 'GOOG', 'META']
finviz_url = 'https://finviz.com/quote.ashx?t='

news_tables = {}
for ticker in tickers:
    url = finviz_url + ticker
    req = Request(url=url, headers={'user-agent': 'my-app'})
    response = urlopen(req)
    html = BeautifulSoup(response, 'html.parser')
    news_table = html.find(id='news-table')
    news_tables[ticker] = news_table

We start with a request to Finviz for each ticker symbol. Once we gather the HTML data, we use BeautifulSoup to isolate the table containing each company’s news. This table is stored for further processing, ready to be analyzed for sentiment.

Step 2: Applying Sentiment Analysis with VADER

Once we’ve gathered the news data, we dive into sentiment analysis using VADER. VADER, or Valence Aware Dictionary and sEntiment Reasoner, provides a compound sentiment score that categorizes text as positive, negative, or neutral.

Here’s the code for processing sentiment:

from nltk.sentiment.vader import SentimentIntensityAnalyzer
import pandas as pd

parsed_data = []
vader = SentimentIntensityAnalyzer()

for ticker, news_table in news_tables.items():
    for row in news_table.findAll('tr'):
        title = row.a.text.strip()
        date_data = row.td.text.strip().split(' ')
        date = date_data[0]
        time = date_data[1] if len(date_data) > 1 else 'N/A'
        compound = vader.polarity_scores(title)['compound']
        parsed_data.append([ticker, date, time, title, compound])

df = pd.DataFrame(parsed_data, columns=['ticker', 'date', 'time', 'title', 'compound'])

This step applies VADER’s compound score to each headline, converting sentiment into a numeric value that is easy to interpret: positive, negative, or neutral. We now have structured data that reflects the emotional tone behind each headline, making it possible to track shifts in market sentiment over time.

Step 3: Visualizing the Results

Finally, we visualize our sentiment scores to reveal trends and patterns. This step involves grouping data by date, calculating average sentiment for each stock, and plotting it as a bar graph to showcase sentiment shifts over time.

Here’s how we generate this visual insight:

import matplotlib.pyplot as plt

df['date'] = pd.to_datetime(df['date'])
mean_df = df.groupby(['ticker', 'date']).mean().unstack()
mean_df = mean_df.xs('compound', axis="columns")
mean_df.plot(kind='bar')
plt.title("Daily Sentiment for Stocks")
plt.xlabel("Date")
plt.ylabel("Mean Compound Sentiment Score")
plt.show()

The bar plot effectively communicates sentiment shifts by date and ticker. A positive bar height indicates optimism, while a negative bar height signals caution or pessimism — making it easy to identify which stocks are receiving favorable or unfavorable attention

AI isn’t perfect

AI is powerful, but it’s not perfect. There are times when sentiment analysis models misinterpret the tone of reviews. For example, a headline may earn a sentiment score that doesn’t align with the true intent behind the text, leading to results that may not accurately reflect the market’s mood. This becomes especially evident in cases of sarcasm or complex phrasing, where the literal words don’t match the implied sentiment.

Let’s look at some cases where sentiment scores and actual sentiment diverge:

# Identifying positive sentiment in news that should be negative
positive_sentiment_negative_review = df.query('compound > 0 and title.str.contains("loss|decline|problem")', engine='python')
print(positive_sentiment_negative_review[['title', 'compound']].head(3))

# Identifying negative sentiment in news that should be positive
negative_sentiment_positive_review = df.query('compound < 0 and title.str.contains("growth|gain|success")', engine='python')
print(negative_sentiment_positive_review[['title', 'compound']].head(3))

By querying for specific keywords that may contradict the sentiment score, we can identify examples of sentiment discrepancies. Here’s an example of where this might occur:

• “Company reports strong growth, but a decline in specific sectors”: VADER might score this as positive due to keywords like strong and growth, despite the mention of decline, which tempers the news sentiment.
• “Revenue gains lead to success, although facing minor issues”: The model may interpret success and gains positively, overlooking the sentiment-shifting impact of minor issues.

These nuances illustrate how AI models, including VADER, can stumble with sentiment detection due to the inability to gauge context fully, especially for nuanced expressions like sarcasm.

TL;DR:

Sentiment analysis tools like VADER provide a powerful yet imperfect solution for extracting emotional insights from text data. While capable of classifying sentiments as positive or negative, they can fall short in scenarios involving sarcasm or conflicting language.

For visual learning, this video has all the information, and steps providing an in depth and engaging explanation of how the code works, and is applied.

About Me:
Hey there👋! My names Noah Barbaros, and I’m a high school student with a deep passion for technology, entrepreneurship, and the transformative power of Artificial Intelligence. I’m driven by the belief that AI, combined with entrepreneurship, and the process of creativity and innovation, can solve the most challenging problems that humanity faces, and create lasting impact. I’m always exploring how these technologies can improve daily life, from small tasks to larger societal challenges. If you’re interested in connecting or collaborating, feel free to reach out via email at noah@barbaros.ca or connect with me on LinkedIn. Thank you for reading — I really hope this article provided valuable insights or inspired new ideas!

Artificial intelligence has transformed industries and revolutionized how we interact with the world. But the most exciting part of AI’s evolution is just beginning. OpenAI’s latest advancement, the ‘Strawberry’ model (also known as o1), isn’t just a technological achievement — it’s a window into a future where AI augments human capabilities in ways we once thought were impossible. While current AI models like Gemini and Claude have limits when it comes to complex reasoning, the o1 model is designed to break those boundaries, solving problems that require a blend of raw computational power and something comparable to human intuition.

So, why should anyone care? Because the o1 model represents a drastic shift. It’s not about replacing human intelligence but complementing it — opening doors to innovations that can transform industries, enhance problem-solving, and fuel breakthroughs. You may have heard a version of the phrase, “AI won’t take your job, it’s somebody using AI that will take your job.” — Economist Richard Baldwin said the phrase and it represents businesses that leverage this technology will have a competitive advantage, finding new efficiencies and creating innovations that were previously out of reach. This isn’t just about technology — it’s about harnessing AI’s potential to shape the future of work, creativity, and human thought.

As AI continues to evolve, one thing becomes clear: we are on the edge of a new era. And with the right tools, the possibilities are limitless.

“The future belongs to those who see possibilities before they become obvious.” — John Sculley

Before o1, AI struggled to solve complex puzzles, handle advanced math equations, or approach the level of reasoning required for everyday tasks that come naturally to humans. o1 uses a complex chain of thought process that puts it ahead of any other AI model. We’ll dive deep into this process later on.

The Current State of o1

The o1 models excel in tasks that require deep reasoning, particularly in STEM fields. They use a chain-of-thought approach, mimicking human problem-solving processes. Want to see some numbers? Sure:

• 89th percentile on Codeforces (competitive programming platform)
• 83% accuracy on USA Math Olympiad qualifying exams (GPT-4o managed an insignificant 13%)

Compare this to GPT-4o’s 13% accuracy on AIME, a challenging test of advanced problem-solving skills, and the gap between them becomes clear. This represents a huge milestone for generative AI, highlighting its advancing capabilities in tackling intricate problems. However, o1 primarily excels at solving complex puzzles and questions; it lacks the speed and multimodal capabilities that make GPT-4o so impressive. OpenAI notes that “GPT-4o is still the best option for most prompts” and points out that o1 struggles with simpler tasks. It’s somewhat ironic — overthinking so much that it falters at basic tasks. Nevertheless, this is only a preview of what will evolve into exceptional technology in the near future. We are witnessing AI inching closer to human-like problem-solving abilities, an exciting moment for tech.

How o1 Works: A Deep Dive for Everyone

So, how does o1 pull off these impressive results? Let’s break it down in a way that’s accessible, even if you’re new to AI concepts.

Understanding AI Models and Their Architecture

First, let’s talk about what an AI model is. Think of it as a sophisticated computer program designed to perform tasks that usually require human intelligence — like understanding language, recognizing images, or solving problems.

The architecture of an AI model refers to the design and structure of this program — how it’s built, how its components interact, and how it processes information. It’s similar to the blueprint of a house; it dictates how everything fits together.

The Transformer Architecture

Now, the Transformer architecture is a specific blueprint used in many advanced AI models, including OpenAI’s GPT series and o1. Introduced by researchers at Google in 2017, Transformers revolutionized the field of AI, especially in natural language processing.

Here’s a simplified explanation:

• Traditional Models: Before Transformers, AI models processed language data sequentially, word by word, which was slow and struggled with long sentences.
• Transformers: They use a mechanism called self-attention to process all words in a sentence simultaneously. This allows the model to understand the context of each word in relation to all other words, capturing the variations and meanings more effectively.

Imagine reading a sentence and being able to consider every word at once, understanding how each word relates to the others. This is what Transformers enable AI models to do.

Chain-of-Thought Reasoning

Chain-of-thought reasoning is a method where the AI breaks down complex problems into smaller, manageable steps, much like how a person might tackle a difficult math problem by solving it step by step.

Here’s how it works:

1. Problem Decomposition: The AI identifies the different parts of a problem.
2. Sequential Processing: It solves each part one at a time, ensuring each step is correct before moving on.
3. Integration: It combines the solutions of the individual parts to arrive at the final answer.

This approach allows o1 to handle complex tasks that require multi-step reasoning, which is something most AI models struggle with. We’ll give a comparison very soon.

Reinforcement Learning from Human Feedback (RLHF)

One of the key innovations in o1 is its use of Reinforcement Learning from Human Feedback. Let’s unpack that:

• Reinforcement Learning (RL): This is a type of machine learning where an AI learns by interacting with an environment and receiving rewards or penalties based on its actions. It’s similar to how a dog learns tricks — you reward it when it does something right.
• Human Feedback: In o1’s case, humans provide the rewards or penalties by evaluating the AI’s responses.

Here’s how it all comes together:

Initial Training: o1 is first trained on a large dataset containing text from the internet, books, and other sources. This gives it a base understanding of language and facts.

Interactive Learning:

• Prompting: The AI is given a problem to solve.
• Response Generation: It provides an answer along with the reasoning steps it took to get there.
• Human Evaluation: Human reviewers assess the answer and the reasoning. They provide feedback on accuracy, clarity, and alignment with human logic.

Learning from Feedback:

• The AI adjusts its approach based on the feedback, reinforcing strategies that lead to correct answers and penalizing those that don’t.

Iteration:

• This process is repeated thousands of times, allowing o1 to refine its problem-solving abilities.

Why This Matters:

• Alignment with Human Thinking: By learning from human feedback, o1 develops reasoning patterns that closely mimic how we think.
• Improved Accuracy: The AI becomes better at providing correct answers and explaining how it arrived at them.
• Transparency: We can understand and trust the AI’s reasoning because it lays out its thought process.

So where is o1’s shortcoming? OpenAI has yet to fully handle the type of complex problems that living beings face on a day to day basis, far more difficult than complex polynomials. A token is pieces of a word that processes through LLM’s, and a limiting factor is that the most advanced publicly accessible LLM’s can process up to 2 million tokens in their context window — significantly less than the amount of information we come across in our lifetimes. However, they are actively working to expand this context window to 10 million tokens. This expansion could enable AI models to process and recall vast amounts of data, bringing them closer to human-like understanding.

Visualization Example: Solving a Simple Problem

Let’s walk through a simple example to compare how o1 applies these concepts.

Anyone that understands the basics of AI will know that LLMs are trained to predict the next word in sequences of text. When o1 seems to be reasoning, it’s actually identifying text patterns it learned during its training process.

For example, if a prompt includes the string “12+18=”, the model is likely to predict the next token as “30” based on patterns it learned during training. The LLM isn’t using any deep understanding of numbers — it’s simply treating “12”, “18”, and “30” as tokens, much like it would with words like “dog” or “each.”

Now imagine an LLM is given the sequence “2+5+4+5–7+10–2=”. Since this specific sequence likely wasn’t part of its training data, the model won’t instantly know the correct answer. However, if we prompt it to “think step by step,” it might generate an output along the lines of:

Begin adding the first two numbers: 2+5=7

Add the next number: 7+4=11

Add the next number: 11+5=16

Subtract the next number: 16–7=9

Add the next number: 9+10=19

Subtract the final number: 19–2=17

Although “2+5+4+5–7+10–2=” isn’t in the training data, the strings such as “2+5=7”, or “7+4=11” and so on probably are within the training data. So by breaking down one big math equation the LLM is able to get the correct answer. However this is only doable if the model’s training data includes a bunch of examples where a big math problem is broken down into smaller parts to solve.

What’s Happening Here:

• Chain-of-Thought: o1 breaks down the problem into steps, making it easier to follow and verify.
• Reinforcement Learning: Through prior training and feedback, o1 knows to show its work, which is valuable for understanding and trust.

This is the core concept behind o1. While chain of thought isn’t a new idea, it gained significant attention in the ML community following a well-known 2022 paper on the topic. Today, most models are trained to automatically break down complex problems in this way. However, OpenAI improved the technique in two key ways that made it significantly more effective.

One improvement was generating a large amount of training data that demonstrated how models can engage in more complex, extended chain-of-thought reasoning. While this isn’t exclusive to OpenAI — most major AI labs are focused on improving the quality and quantity of training data — OpenAI appears to have done a more comprehensive job than its competitors.

The real breakthrough, however, was in a new training process that allowed o1 to learn more effectively from these long chain-of-thought examples, significantly enhancing its reasoning capabilities.

Diving Deeper into the Technology

Now that we have seen and understand how o1 works in practice, let’s explore the underlying technology in more detail.

Self-Attention Mechanism

In the Transformer architecture, the self-attention mechanism allows the model to weigh the importance of different words in a sentence relative to each other.

• Example: In the sentence “The cat sat on the mat,” the word “cat” is closely related to “sat” and “mat.” The model assigns higher importance to these connections, helping it understand the context better.

This mechanism enables o1 to process information more effectively, especially in complex sentences or problems.

Dynamic Memory and Meta-Cognition

o1 incorporates components that act like a working memory, allowing it to store and manipulate intermediate results.

• Dynamic Memory: This is where o1 keeps track of the information it’s processing, similar to how we might jot down notes while solving a problem.
• Meta-Cognition: o1 can reflect on its own reasoning process, checking for errors or inconsistencies and adjusting as needed.

Extended Context Windows

• Context Window: This refers to the amount of information the AI can consider at once.
• o1’s Advantage: It can handle longer context windows, meaning it can process and remember more information, which is crucial for solving complex problems.

How o1 is Different from Previous Models

In terms of capability, the o1 model can do just about anything — answering questions, summarizing content, and generating new material. However, it’s important to note that this is a preview model — you cannot upload attachments (images, links, videos, etc.) like with GPT-4o.

Architectural Innovations:

• Self-Reflective Attention Mechanism: Allows the model to review and refine its own outputs before presenting them, akin to a human double-checking their work.
• Dynamic Memory Module: Provides a working memory space where intermediate computations and reasoning steps are stored and manipulated.
• Meta-Learning Capabilities: Enables the model to adapt its reasoning strategies based on the problem domain, improving efficiency and accuracy over time.

Compared to Previous Models:

• GPT-4: Excels at general language tasks and multimodal inputs but doesn’t delve deeply into multi-step reasoning.
• Claude (Anthropic): Focuses on being helpful and harmless but doesn’t have o1’s advanced problem-solving abilities.
• LLaMA (Meta): Aimed at democratizing access to LLMs but lacks the specialized reasoning modules of o1.

This combination of chain-of-thought reasoning and reinforcement learning makes o1 particularly suited for STEM fields, where deep reasoning is essential. For example, o1 can tackle quantum physics problems, break down complex coding tasks, and solve intricate math proofs with a level of step-by-step reasoning that most other models can’t replicate.

Business implications

Anyone that has used any AI model knows that it can provide help with just about anything. They use it to assist them, build reports, and provide feedback. So you might be asking yourself the same question I did: how does o1 change and most likely improve this? How can we apply o1 to our rapidly evolving business world? With the key concepts that o1 has to offer, there are so many possibilities, especially as AI continues to expand beyond our expectations.

Real-World Use Cases

According to a recent case study, the o1 model was instrumental in identifying market patterns and executing trades with unprecedented accuracy, significantly outperforming traditional algorithms. This not only demonstrates o1’s capability in handling complex financial models but also highlights its potential to revolutionize industries reliant on data-driven decision-making.

Key Business Benefits:

1. Enhanced Data Analysis: The model’s ability to break down complex tasks and datasets into countless steps makes it invaluable for deriving quality insights. Businesses can leverage this for consulting, making informed decisions, and identifying trends that were previously obscured by data complexity.
2. Automation of Repetitive Tasks: From creating complex reports to generating code, o1 can handle much more intricate tasks, allowing businesses to scale operations and address larger-scale questions without a proportional increase in resources.
3. Improved Decision Making: Because of the process of breaking everything down, the AI provides improved decision-making capabilities. This allows o1 to offer more strategic decisions, enabling businesses to approach problems with greater flexibility and foresight.

Overall, by leveraging AI’s ability to generate and execute thousands of reasoning steps, companies can unlock new levels of creativity, efficiency, and strategic foresight. As this technology continues to evolve, it will undoubtedly become a cornerstone of business strategy in the digital age.

Embracing the Future with AI

OpenAI’s o1 model is a big step forward in artificial intelligence. It’s not just another piece of tech — it’s a peek into a future where AI works alongside us in ways we once thought were impossible. Unlike earlier models, o1 can tackle really tough problems by thinking through them step by step, much like how we do. This doesn’t replace human intelligence; it complements it, opening up new opportunities for innovation and problem-solving in all kinds of fields.

As we stand on the edge of this new era, it’s clear that o1 offers tools that can make us more efficient, creative, and strategic. Businesses that embrace this technology will have a competitive advantage, and funny enough — here is where I’m going to bring back the two quotes I had quoted at the beginning of this article, just like Richard Baldwin said: “AI won’t take your job, it’s somebody using AI that will take your job.” And echoing past CEO of apple, John Sculley’s words, “The future belongs to those who see possibilities before they become obvious.” With o1, those possibilities are becoming clear. It’s up to us to take hold of them.

In short, o1 isn’t just about advancing technology — it’s about unlocking our own potential and seeing how far we can go when we work together with AI. The future is here, and with tools like o1, there’s no limit to what we can achieve.

About Me:
Hey there👋! My names Noah Barbaros, and I’m a high school student with a deep passion for technology, entrepreneurship, and the transformative power of Artificial Intelligence. I’m driven by the belief that AI, combined with entrepreneurship, and the process of creativity and innovation, can solve the most challenging problems that humanity faces, and create lasting impact. I’m always exploring how these technologies can improve daily life, from small tasks to larger societal challenges. If you’re interested in connecting or collaborating, feel free to reach out via email at noah@barbaros.ca or connect with me on LinkedIn. Thank you for reading — I really hope this article provided valuable insights or inspired new ideas!

References:
Art of Problem Solving. (n.d.). Artofproblemsolving.com. https://artofproblemsolving.com/wiki/index.php/United_States_of_America_Mathematical_Olympiad

Azhar, A. (2024, September 13). 🍓 Playing with Strawberry: When machines can reason. Exponentialview.co; Exponential View. https://www.exponentialview.co/p/strawberry-openai-o1

Binder, M. (2024, September 26). OpenAI plans to become a for-profit business — here’s what that means for the AI company. Mashable. https://mashable.com/article/openai-for-profit-turn

Claude. (n.d.). Claude.ai. https://claude.ai/login?returnTo=%2F%3F

Codeforces. (n.d.). Codeforces. https://codeforces.com/

Ferrer, J. (2024, January 9). How Transformers Work: A Detailed Exploration of Transformer Architecture. Datacamp.com; DataCamp. https://www.datacamp.com/tutorial/how-transformers-work

GPT-4. (2015). Openai.com. https://openai.com/index/gpt-4/

Hello GPT-4o. (2024). Openai.com. https://openai.com/index/hello-gpt-4o/

How OpenAI’s o1 Model Can Transform Business Operations | Briolink. (2024, September 17). Briolink. https://briolink.com/news/innovation/how-openais-o1-model-can-transform-business-operations

Introducing OpenAI o1. (2024). Openai.com. https://openai.com/index/introducing-openai-o1-preview/

Learning to Reason with LLMs. (2024). Openai.com. https://openai.com/index/learning-to-reason-with-llms/

Lee, T. B. (2024, September 20). OpenAI just unleashed an alien of extraordinary ability. Understandingai.org; Understanding AI. https://www.understandingai.org/p/openai-just-unleashed-an-alien-of

Llama 3.2. (2022). Meta Llama. https://www.llama.com/

News, I. (2024, September 14). OpenAI’s o1 Correctly Solved 83.3% of the Problems, While GPT-4o Only 13.4% | IBL News. Iblnews.org. https://iblnews.org/openais-o1-correctly-solved-83-3-of-the-problems-while-gpt-4o-solved-only-13-4/

OpenAI. (2024). What Are Tokens and How to Count them? Help.openai.com. https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them

Piper, K. (2024, September 20). What it means that new AIs can “reason.” Vox. https://www.vox.com/future-perfect/372843/openai-chagpt-o1-strawberry-dual-use-technology

Ramponi, M. (2024, August 21). Decoding Strategies: How LLMs Choose The Next Word. News, Tutorials, AI Research. https://www.assemblyai.com/blog/decoding-strategies-how-llms-choose-the-next-word/

Redirecting you — Medium. (2024a). Medium.com. https://medium.com/r?url=https%3A%2F%2Fmedium.datadriveninvestor.com%2Fi-used-openais-o1-model-to-develop-a-trading-strategy-it-is-destroying-the-market-576a6039e8fa

Redirecting you — Medium. (2024b). Medium.com. https://medium.com/r?url=https%3A%2F%2Fartofproblemsolving.com%2Fwiki%2Findex.php%2FUnited_States_of_America_Mathematical_Olympiad

Research Graph. (2024, September 20). How OpenAI’s O1 Series Stands Out Redefining AI Reasoning. Medium. https://medium.com/@researchgraph/how-openais-o1-series-stands-out-redefining-ai-reasoning-9e499937139e

Robison, K. (2024, September 12). OpenAI releases new o1 reasoning model. The Verge; The Verge. https://www.theverge.com/2024/9/12/24242439/openai-o1-model-reasoning-strawberry-chatgpt

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain of Thought Prompting Elicits Reasoning in Large Language Models. ArXiv:2201.11903 [Cs]. https://arxiv.org/abs/2201.11903

What is Reinforcement Learning? — Reinforcement Learning Explained — AWS. (n.d.). Amazon Web Services, Inc. https://aws.amazon.com/what-is/reinforcement-learning/

What is RLHF? — Reinforcement Learning from Human Feedback Explained — AWS. (n.d.). Amazon Web Services, Inc. https://aws.amazon.com/what-is/reinforcement-learning-from-human-feedback/