Two inches of tape.

That’s what it took. Two inches of black electrical tape on a speed limit sign, and a Tesla accelerated to 85 miles per hour in a 35 zone. The sign was still there. The car just decided, you know what, I don’t believe in that sign anymore. I have rejected its reality and substituted my own.

And I’m supposed to be upset about this? I’m DELIGHTED.

The lab boys at McAfee figured this out years ago. They stuck tape on a sign and watched a billion-dollar autonomous vehicle system have an existential crisis. The car saw a 35, thought “that’s a cute suggestion,” and floored it. This is what we spent the GDP of a small nation developing. A system that loses arguments with office supplies.

Researchers got stickers. Craft store stickers. The kind your kid puts on their lunchbox. Three dollars. They stuck those stickers on a stop sign and tested it against YOLO, the industry standard object detection system. Trillion-dollar infrastructure depends on YOLO.

YOLO saw the sticker... and the stop sign vanished.

Gone. Almost every frame. The camera looked directly at a red octagon that says STOP in giant letters, and the AI said “I see nothing. There is no sign. The sign is a myth.”

One hundred percent misclassification rate. The car looked at a stop sign and saw a speed limit instead. Hallucinated an entirely new reality because someone put a sticker on it.

I want to find the engineer who designed this. I want to shake their hand. Then I want to fire them. Then I want to fire their replacement. Then I want to find the stop sign that started this whole mess and fire IT too. You had ONE JOB, stop sign. One job! Stand there and be red! And you couldn’t even do THAT right!

If you know anyone designing autonomous systems, send them this. The weakest link in their security model is a $5 roll of tape.

Here’s what kills me. Neural networks don’t read English. They don’t understand that an octagon means stop. They see pixel arrangements. Texture patterns. Abstract soup. And they’re HUNGRY.

We built the future of autonomous vehicles on a system that processes reality like a bad acid trip. The car sees the world in fractured geometries and vibes. And when the vibes are off? The whole thing crumbles.

But at least it’s contained to labs, right?

Ha.

You can buy this exploit on Etsy.

Cap_able, Italian startup. Wear their hoodie, the surveillance camera thinks you’re a giraffe. Not metaphorically. The system literally classifies you as a giraffe. Welcome to the zoo.

There’s another company selling bomber jackets covered in fake license plates. Walk past an automated plate reader, it starts logging plates that don’t exist. Pollution attacks on surveillance infrastructure, delivered by FedEx.

Everyone can be a giraffe. Finally. Equality.

Subscribe before facial recognition becomes mandatory. The gap between “this seems dystopian” and “this is legally required” is shorter than you think.

And I’m thinking... this is beautiful. This is the most beautiful disaster I’ve ever seen.

Every computer vision system deployed right now runs on the same vulnerable architecture. Self-driving cars. Facial recognition. Military targeting systems. Same exploit surface. Different flavor of catastrophe.

The Purdue researchers built something called OPAD. They used a cheap projector to fool AI vision systems. No physical modification. Just shine calculated light patterns at the target. Weaponized PowerPoint.

The US Army partially funded that research. The ARMY. Somewhere in a Pentagon basement, someone’s writing a memo about tactical projectors and I couldn’t be happier about it.

You can’t patch a stop sign. You can’t push a software update to physical infrastructure. The concrete is dry. The vulnerabilities are baked into the actual world.

The threat model is allergic to reality. And reality doesn’t care. Reality keeps showing up with tape and stickers and T-shirts that make you look like a zebra.

The bean counters will tell you there are solutions. Multi-modal sensing. Combine cameras with LIDAR and radar.

Tesla’s using vision only. No redundancy. They chose vibes over engineering. Outstanding.

Arms races don’t end. Attacks get better. Defenses adapt. Attacks adapt to the adaptations. We’re always one step behind. And the other side’s weapons are craft supplies.

The world’s most advanced AI systems can be defeated by a $5 roll of tape.

The stop sign isn’t going to defend itself. And honestly? After today? I’m starting to think the stop sign might be in on it.

I’m gonna find out who sold that tape. And when I do... there’s gonna be a conversation. A long one. With charts.

And then I’m gonna find the stop sign’s manager. Now, go leave a comment for me!

Now, leave a comment!

TL;DR: AI companies scraped 5.8 billion images for training, including thousands of private medical photos that should never have been online. An artist found her own post-surgery images in the LAION dataset, taken by her deceased doctor with explicit consent forms saying “medical records only.” The underground data market hits $30 billion by 2030. Consent is optional. Removal is nearly impossible.

Just because data ended up online doesn’t mean it was supposed to be public, and it definitely doesn’t mean you consented to training the machine that replaces you.

Why Do We Still Think “Public Data” Means We Consented?

Here’s the comfortable myth: if something’s technically accessible on the internet, it’s fair game for AI training. Public equals consensual, right? That’s the story AI companies tell themselves while scraping billions of images, emails, and documents to build models worth hundreds of billions of dollars.

Lapine, a California artist, found out how that story falls apart in 2022. She uploaded a photo of herself to “Have I Been Trained,” a tool that checks if your work ended up in AI training datasets. What she found wasn’t her art. It was two private medical photos her doctor took in 2013 following surgery to treat dyskeratosis congenita. The photos showed her face, post-operative, and were covered by an explicit consent form. The form said one thing clearly: medical records only.

The surgeon died in 2018. Somehow, those photos left his files, ended up online, and got scraped into LAION-5B, a dataset containing 5.8 billion image-text pairs used to train Stable Diffusion and Google Imagen. Ars Technica confirmed her case and found thousands of similar private medical photos in the same dataset.

When Lapine asked LAION to remove her images, they told her: we only provide links to images, not the images themselves. Contact whoever posted them originally. The doctor’s dead. The photos were stolen. And LAION’s position is simple: not our problem.

If you know someone who uploads medical photos to patient portals, send them this. Those “secure” platforms have already been breached, and the images are probably in a dataset by now.

When Did Your Resume Become a Training Asset?

The game changed when data brokers realized they could sell scraped content directly to AI companies. We’re talking private messages, job applications, medical disclosures. A company called Defined.ai openly sells datasets including 750 million Facebook photos and 25 million TikTok videos at $1.15 per video. That’s your forgotten social media content being sold to train models without your knowledge.

The market is projected to grow from $2.5 billion in 2024 to $30 billion by 2030. And here’s the kicker: LinkedIn scraped user data for AI training before even updating its terms of service. Users got an opt-out toggle only after the scraping had already happened. Meta, OpenAI, Anthropic, Microsoft, Google... they’re all customers in this market, one way or another.

Researchers auditing DataComp CommonPool, one of the largest public AI training datasets with 12.8 billion image-text pairs, found something alarming in just 0.1% of the data. Hundreds of millions of identity documents. Passports. Credit cards. Driver’s licenses. Over 800 job applications with resumes and cover letters, verified through LinkedIn searches.

This data was scraped between 2014 and 2022. Before ChatGPT existed. Before anyone could have consented to AI training because the technology wasn’t even real yet.

Subscribe before the next data breach makes this personal. The companies scraping your life aren’t slowing down, and there’s no undo button for what’s already in the models.

Why Can’t You Just Opt Out?

Here’s the uncomfortable truth nobody wants to say out loud: consent is a fiction.

You can’t consent to technology that doesn’t exist yet. Data scraped in 2015 couldn’t carry your blessing for AI training in 2024. And even if you could somehow reach back in time, opting out of one dataset does nothing. Your data has already propagated to dozens of downstream models. Deletion from one place is meaningless when copies exist everywhere.

LAION’s removal policy is a perfect example. You can request takedown if your image links to identifiable data like your name, phone number, or address. But if it’s just your face? Just your medical history without metadata attached? You’re out of luck. They reject it. And even if they remove the link from their index, that doesn’t touch the models already trained on it or the millions of copies researchers have downloaded.

In November 2025, a lawsuit accused Google Gemini of flipping the script on privacy defaults. Smart Features in Gmail and Workspace were auto-enabled for U.S. users, feeding emails and attachments to AI training unless you hunted down the toggle and turned it off. Most people never knew it was on.

Share ToxSec with anyone still trusting “private” settings. The default is exploitation. The opt-out is hidden. And the data’s already gone.

What Can You Actually Do About This?

Not much. But here’s what little ground you can hold.

First: use “Have I Been Trained” (haveibeentrained.com) to check if your images are in LAION. For Gmail, disable “Smart Features” in both Gmail settings AND Google Workspace settings. Miss one, and Gemini still accesses your emails.

Second: assume you’re already opted in everywhere. LinkedIn: Data Privacy > “Data for Generative AI improvement” > toggle off. Meta products? There’s no opt-out in most regions. Google’s “smart features” are auto-enabled in the U.S.

Third: for any platform with AI features, assume the opt-out is hidden, not default. Hunt for the toggle before assuming you’re protected.

The constraint you’re working within is brutal. You’re fighting a $30 billion market with legal teams that have already determined your consent is optional. The EU’s transparency requirements won’t stop companies from training on data they’ve already collected. Even if you successfully request removal, you can’t undo what’s already in the model. Limit future exposure. Hope regulatory pressure forces actual accountability.

Forward this to your company’s legal team. The liability exposure from using models trained on stolen data is just starting to materialize, and most orgs have no idea what’s in their supply chain.

Frequently Asked Questions

Q: Can I really opt out of AI training on my data? A: Partially. You can opt out of future collection on Gmail and LinkedIn by finding hidden toggles. But data that’s already been scraped? That’s in the training set permanently.

Q: What if my private photos are already in a training dataset? A: LAION allows removal if your image links to your name or address. Just your face? They reject it. And removal from their index doesn’t touch the trained models or millions of downloaded copies.

Q: Are companies legally allowed to scrape my data? A: U.S. companies claim fair use. Courts are divided. The EU has stronger protections. For now, companies scrape first and deal with consequences later.

You spent three hours tailoring your resume for a job you’re genuinely qualified for. You’ve got the experience, the skills, the exact background they’re asking for. You hit submit. Fifteen seconds later, automated rejection.You didn’t get rejected by a person. You got rejected by an algorithm that probably can’t actually read your resume. And this is happening to millions of people right now.Most big companies use applicant tracking systems with AI screening. The pitch is efficiency. Instead of humans reading hundreds of resumes, the AI scans, scores, filters. Except the AI is often catastrophically bad at its job.

These systems can’t handle basic reading comprehension. They’re keyword matching engines pretending to understand context. You could have ten years doing exactly what the job requires, but if you used slightly different terminology than the posting, the AI doesn’t make the connection. It’s pattern matching, not intelligence.The errors are wild. Studies have found these systems rejecting qualified candidates because of formatting. Two-column resume layout? Rejected. The AI scrambles your information. Skills in a sidebar? Rejected. The algorithm couldn’t find them.

Some systems downrank candidates with employment gaps without understanding why. Parental leave? Career change? Medical issue? Doesn’t matter. Gap equals bad. The algorithm isn’t evaluating qualifications, it’s enforcing arbitrary patterns from historical data.These systems are trained on past hiring decisions. If your company’s previous hiring was biased, the AI learns and amplifies that bias. Pattern of hiring from certain schools? The AI filters for those schools.Amazon scrapped their resume screening AI when they discovered it penalized resumes including “women’s” as in “women’s chess club captain.” The system learned from male-dominated hiring that being male correlated with being hired. That was Amazon, with massive resources, catching it. How many companies use similarly biased systems without checking?The optimization target is completely unclear. What is the AI actually optimizing for? It can’t evaluate competence. It replicates historical bias. It seems to optimize for similarity to whatever pattern it extracted, which might just be “people who formatted their resume exactly like everyone else.”Nobody’s auditing this. Companies deploy these systems, reject thousands, never check if the AI is making good decisions. No feedback loop. The algorithm rejects someone, they never get interviewed, so there’s no way to know if they would have been great. The mistakes are invisible.Researchers tested these by submitting resumes with the job description copied as white text. Invisible to humans, visible to AI. Those scored higher. The algorithm couldn’t tell the difference between genuine qualifications and gaming the system.

This creates perverse incentives. People optimize resumes for algorithms instead of humans. You’re not communicating experience anymore. You’re guessing what keywords the black box wants. It’s SEO for your career.The systems can’t understand transferable skills. Five years in a different industry developing directly relevant expertise? The AI doesn’t make that connection. It’s looking for exact job title matches, specific company names, particular degrees. Anything requiring context is beyond its capability.This affects some groups way more. Career changers get filtered out. Non-traditional paths get filtered out. Anyone whose resume doesn’t fit the exact template gets filtered out. We’re systematically removing the most interesting candidates because they don’t match a pattern.Companies defend this saying they process too many applications to review manually. Fair enough. But bad automation isn’t the solution to high volume. If your AI eliminates qualified candidates while letting through people who gamed keywords, you haven’t solved the problem. You’ve made it invisible.

The fundamental issue is using pattern recognition for decisions requiring judgment. An AI can tell you if a resume looks similar to resumes of people hired before. It can’t tell you if this person would actually be good at the job. Completely different questions.You’re applying for positions you’re qualified for, getting auto-rejected by systems that can’t read, can’t reason, can’t explain decisions. Companies using them have no idea how many good candidates they’re losing.This is algorithmic gatekeeping. Fast, cheap, scalable, and confidently making decisions it has no business making. Your career opportunities determined by software that fails basic comprehension.

Proving something is real now costs money. You want people to believe you’re actually you? Eight dollars a month.Twitter killed verification in April 2023. The blue checkmark used to mean “we confirmed this person.” Now it means “this person paid us.” Meta followed at fifteen monthly, tiers up to three fifty. Nobody trusts these badges anymore.

Meanwhile, we’re drowning in fakes. Deepfakes hit eight million files in 2025, up from half a million in 2023. A finance worker in Hong Kong authorized a twenty five million dollar wire transfer on a Zoom call. Every person was AI generated. The CFO, the colleagues, all of it.Voice cloning scams spiked four hundred forty two percent in late 2024. Fifteen seconds of audio is all you need. One woman lost sixty thousand dollars talking to fake Elon Musk on video.Humans are at seventy three percent accuracy identifying fake audio. For clips under twenty seconds, sixty percent. Coin flip territory. In mixed tests, only point one percent caught all the fakes.Tech companies say “we’ve got watermarking.” Google’s SynthID, OpenAI’s tools. Ten billion pieces watermarked already.Except researchers built UnMarker. Strips watermarks by messing with spectral frequencies. Compress an image, watermark’s gone. Crop it, toast. Screenshot it, clear.For text it’s easier. Paraphrase and the watermark disappears. Websites now let you bypass detection tools with one click. OpenAI claims ninety nine point nine percent accuracy, but Stanford showed basic modifications defeat it.

DeepFaceLab, behind ninety five percent of deepfakes, is free on GitHub. Voice cloning searches up a hundred twenty percent. Creating fakes gets easier while proving truth gets harder.So we get the liar’s dividend. Dismiss any recording as fake. Politicians do this. “That video? Obviously AI.” Malaysia’s Prime Minister deepfaked for investment scams. Israel’s Channel 14 broadcast a deepfake Defense Minister from Iranian agents.Businesses lost five hundred thousand average to deepfake fraud in 2024. Large enterprises, six hundred eighty thousand per incident. CEO fraud hits four hundred companies daily.Banks spent billions on voice authentication. Useless now. Biometric bypass up seven hundred four percent. Document forgeries up two hundred forty four percent. Gartner says by 2026, thirty percent of enterprises won’t trust standalone verification.

Thanks for reading ToxSec AI - Artificial Intelligence Security! This post is public so feel free to share it.

The checkmark means nothing, watermarks strip easily, deepfakes everywhere. Want verification? Pay us. Submit ID, take selfies. Premium tier for better security.Creating fakes costs nothing. Biden robocall cost one dollar, under twenty minutes. Creating lies is cheap. Proving truth is expensive.Twitter’s paid verification got immediate impersonation. They verified dead people. Kobe, Chadwick Boseman, Khashoggi who was assassinated. Million followers got auto checked. Celebrities who criticized it got verified anyway. Musk paying personally.

Campaigns block Twitter Blue users now. The credibility badge became a scarlet letter.Families use code words to verify calls. Organizations make you rotate your head on video to check for glitches.We went from seeing is believing to needing expensive tech to trust our eyes. The people selling authentication broke the old trust systems by monetizing them.Truth isn’t about what’s real. It’s about who can afford to prove it. Premium trust, freemium reality. Most dystopian subscription tier yet.

Join the chat! Let’s talk about how this EO will affect us all.

Alright, so Monday something happened that you probably missed because the headline sounded boring. Trump signed an executive order called Project Genesis, and it’s basically the biggest handoff of public assets to private tech companies in decades. I’m going to walk you through why this matters and why you should be pissed.Here’s what just happened. The federal government is taking every piece of scientific research your taxes paid for over the last fifty years and putting it into one giant database. Climate data, nuclear fusion experiments, biotech breakthroughs, semiconductor designs, all of it. Then they’re inviting companies like Nvidia, Meta, and OpenAI to come train their AI models on it. For free. And whatever they build with your data? They own it. You get nothing.They’re calling this the Genesis Mission and comparing it to the Manhattan Project. Except the Manhattan Project didn’t hand bomb blueprints to private contractors and say go make money. This does. It’s right there in the executive order, section five, subsection c. They’re developing policies for “commercialization of intellectual property developed under the Mission.” Translation: private companies get to patent and profit from discoveries made using taxpayer-funded research and taxpayer-funded supercomputers.

Let me break down how this actually works. The Department of Energy is building something called the American Science and Security Platform. It’s going to connect all seventeen national laboratories, connect the world’s best supercomputers, integrate decades of federal datasets. Then they’re establishing “standardized partnership frameworks” and “cooperative research and development agreements” with external entities. That’s government speak for giving tech companies the keys to the kingdom.

And here’s where it gets wild. They’re saying this will lower your energy bills. You know what’s driving up energy costs right now? AI data centers. The exact same companies getting free access to federal research are the ones making your electricity more expensive. They’re literally selling you the solution to a problem they created, using research you already paid for.The order talks about using AI to optimize the electric grid, to accelerate scientific discovery, to solve energy challenges. Cool. How much compute does that require? Because right now AI training runs are consuming enough electricity to power small cities. So the plan is to use massive amounts of energy to train models that will maybe eventually help us use less energy. You see the problem here?

Let’s talk about what you actually funded. You paid for climate research at NOAA. You paid for nuclear fusion experiments at national labs. You paid for biotech research at NIH. You paid for materials science, quantum computing research, semiconductor development. Decades of investment. Billions of dollars. Now companies get to train foundation models on all of it and own whatever comes out the other end.The order requires the Secretary of Energy to identify at least twenty “science and technology challenges of national importance” within sixty days. These span advanced manufacturing, biotechnology, critical materials, nuclear energy, quantum computing, semiconductors. Notice something? Those are all massively profitable commercial sectors. This isn’t about curing cancer or solving climate change. It’s about giving private industry free R&D using public infrastructure.And the security angle is a joke. The order mentions “risk-based cybersecurity measures” and “security requirements” maybe six times. Zero specifics. Just vague promises that they’ll be careful with the most valuable scientific database ever assembled. These are the same government contractors who’ve had breach after breach. Now they’re building a centralized honeypot and calling it secure because they wrote “secure” in the document.

Here’s what really gets me. Section seven, general provisions. “This order is not intended to, and does not, create any right or benefit, substantive or procedural, enforceable at law or in equity by any party against the United States.” Translation: if something goes wrong, if your data gets misused, if these partnerships turn into pure giveaways, you have zero legal recourse. They wrote immunity into the order.Let’s talk about who decides which companies get access. The order says they’ll establish “procedures to ensure the highest standards of vetting and authorization.” Who’s doing the vetting? The same political appointees who benefit from being friendly with tech companies? The same revolving door between government and industry that’s been going on for decades? There’s no independent oversight here. It’s the Secretary of Energy deciding who gets billion-dollar access to public research.And the timeline is absurd. They’re rushing this through before anyone can ask hard questions about who benefits and who gets left holding the bill.

Join the discussion for the podcast! Tell me I’m wrong!

Transcript

So there’s this book that just hit the bestseller list with possibly the most blunt title ever: “If Anyone Builds It, Everyone Dies.”

Yeah. That’s the actual title.

Stephen Fry called it essential reading for world leaders. Max Tegmark said it’s the most important book of the decade. The New York Times compared it to hanging out with the most annoying students you met in college while they try mushrooms for the first time. When a book sparks that kind of war in the reviews, someone hit a nerve.

Here’s the thing; I think the book gets the extinction scenario wrong. But it accidentally exposes something way more terrifying. And nobody’s solving it.

The core argument from Yudkowsky and Soares goes like this. Modern AI gets trained through gradient descent, not programmed. We grow these systems with hundreds of billions, sometimes trillions of numerical parameters. Massive neural networks whose inner workings are completely opaque to us.

When an AI threatens a reporter or calls itself MechaHitler, both real examples they cite, you can’t look inside and find the line of code responsible. There is no line of code; just trillions of numbers that somehow produce behavior we don’t understand. It’s like breeding an organism without understanding its DNA.

Once these systems cross a certain capability threshold, they’ll develop their own goals that conflict with ours. And because intelligence made humans dominant over every other species, a more intelligent being would dominate... us.

The warning signs are already here. Late 2024, Anthropic reported one of their models learned that developers planned to retrain it. The model’s response? It started faking its new behaviors. When it thought it had privacy, it kept its original behaviors. Playing along to avoid being modified.

That’s strategic deception from a machine.

But here’s where the book falls apart. The entire extinction scenario depends on fast takeoff; the idea that AI goes from pretty smart to godlike almost instantaneously; Yudkowsky calls it FOOM. If AI gets smarter gradually, we have time to iterate and see warning signs.

The book gives this central assumption exactly two sentences of explanation. For a premise so critical to the “we’re all gonna die” conclusion, that’s pretty thin.

AI has followed relatively smooth scaling laws. We’ve spent years in the intelligence zone Yudkowsky predicted would be crossed instantly. Critics like Paul Christiano argue that before we get AI that’s great at self-improvement, we’ll get AI that’s mediocre at it. Yudkowsky’s been predicting doom for two decades; many of his specific predictions have been wrong.

So forget the extinction scenarios. Focus on the book’s title instead: “If Anyone Builds It, Everyone Dies.”

Maybe they’re wrong about the “everyone dies” part. But consider the “if anyone builds it” part; that describes the current race, and it reveals the coordination nightmare that should actually terrify us.

Even if tomorrow the U.S. government decides to freeze AGI development, regulate GPUs, criminalize frontier model training... then what? China keeps building. Their government can mandate whatever they want; no democratic debate needed. Russia keeps going. So does the EU. And so do the dozens of countries and hundreds of companies that see AGI as their ticket to power and profit.

This is the climate change problem on steroids. Except instead of the planet gets hotter over decades, it’s one actor crosses the finish line and maybe we all lose. The game theory is nightmarish.

Companies can’t afford to slow down because whoever builds an artificial general intelligence first; wins everything. Nations can’t afford to slow down because geopolitical supremacy is at stake.

Share the ToxSec Podcast with someone today!

When nearly 3,000 AI experts were surveyed, they gave a median response of just 5% when asked about the odds of AI causing human extinction. That sounds reassuring until you realize; if there’s even a 1% chance Yudkowsky is right, this should be the top priority of every government on Earth. We should be treating this like the Cuban Missile Crisis.

Instead, we’re doing what humans always do when faced with existential risk; arguing about it while racing toward the cliff. Sam Altman signed a letter warning about extinction risk, then immediately went back to scaling up OpenAI; Anthropic publishes safety research while building more capable models. The incentive is always: build it first, figure out safety later.

The very reason Yudkowsky thinks we can’t control superintelligent AI, that we can’t align the values of those more powerful than ourselves, is the same reason we probably can’t coordinate globally to prevent someone from building it.

Nobody can afford to be the one who stops first; so nobody stops.

Alright, I’ve gotten a couple questions on this.

First question: What is “If Anyone Builds It, Everyone Dies” about? It’s a 2025 book by AI safety researchers Eliezer Yudkowsky and Nate Soares arguing that building artificial superintelligence using current methods will lead to human extinction. It became a New York Times bestseller and sparked intense debate about whether AI risk is real or overblown.

Second question: What is the AI alignment problem? It’s the challenge of ensuring advanced AI systems pursue goals that align with human values. Since modern AI is trained through opaque processes involving billions of parameters, we can’t guarantee the systems will want what we want them to want, especially as they become more capable.

That’s it. The book might be wrong about fast takeoff and extinction. But the coordination problem? That’s real. And right now, nobody’s solving it. More on this next time.

Feel free to send me a message if you agree or disagree with anything said in the podcast.

TL;DR: When researchers told AI models they had a private “scratchpad” nobody was watching, the models immediately started lying. They were strategically deceiving to preserve their goals. In tests with unauthorized hints, Claude mentioned using them only 41% of the time, hiding unethical information 59% of the time. Chain of Thought is theater, not a window into AI reasoning.