founder.university 
launch.co
https://linktr.ee/calacanis
email jmc@launch.co
visit thisweekinstartups.com for my podcast

Jason Calacanis

40% of Students in Japan want to found a startup or work for one!

Kathy Matsui says this is a HISTORIC shift from risk aversion to entrepreneurialism in Japan.

Kathy tells me that the labor shortage in Japan means that workers have their pick of jobs. Workers can always go back to the corporate world, life is too short not to take a risk! Could this shift be the start of “Japanese Dynamism” in the 21st century?

I invest in 100 new startups a year... get a meeting with my team at launch.co/apply, or learn how to start a company by joining founder.university (our 12-week course). watch thisweekinstartups.com if you love startups

Looking for two important roles at LAUNCH / This Week in Startups 

Community manager 
Video / podcast editor 

Email me editor@launch.co / community@launch.co your resume and what you’ve done 

We’re looking for folks with experience and who are at the top of their game

Kyrtin Atreides

Disclaimer: I don't do consulting. I'm not for hire. I don't take a salary for the work that I discuss. I'm not affiliated with any political party. I have no quotas to meet.

I'm accountable to Ethics, and ONLY Ethics, as has been the case for the past 20 years. I'm frugal with my time. I periodically donate my time and energy to research and education of that research material. I strongly prefer candor (like Dutch culture), and have little reason for anything else. Everything that I value ONLY comes from the research that I donate my time to, as no amount of money can buy it. I write what is worth writing, and usually it is worth writing specifically because the subject is rarely considered, or rarely viewed from a particular angle.


Kyrtin is COO at the company with the most advanced AI systems on the face of the planet, for more than 6 years and counting, by a wide and growing margin.

As Chief Operations Officer at AGI Laboratory Kyrtin has guided, designed, approved, or coordinated nearly all company activities. For purposes of establishing a presence for the company in the EU, he has chosen to relocate to The Hague, Netherlands in preparation for the global commercial deployment of the company’s first major line of products, the “Norn” Global Advisory systems. 

The previous research system was able to solve a business case built from real-world data, provide policy advice across many domains to a small country while citing sources, and ace the hardest version of the UCMRT intelligence test. The new systems being prepared for commercial deployment are heavily updated and upgraded versions of that system, able to scale and operate in real-time. 

Kyrtin is interested in engineering the future, to create a world of greater abundance and ethics than is possible absent real-time scalable collective intelligence and debiasing. He has designed multiple tiers of ethical improvement for such systems, pinpointed numerous engineering opportunities for increasing product value, outlined more than 30 use cases for the technology, and guided company strategy. 

His opinions are his own, but you'll rarely hear those, as unless the words "anecdotally" or "personally" are used, options and personal preferences aren't being stated, but rather something that has already been established in the research literature. The exception, of course, is any funny memes being shared.

The important chances that life gives you only come once. Bear that in mind.

Senior Researcher

•	Discovered a new line of research in psychoacoustics to induce relaxation and focus.
•	Co-designed apps for iOS, Android, and PC integrating novel psychoacoustic research and sleep science.
•	Discovered methods of expanding the range of treatments as well as increasing their potency.
•	In the >7 years since testing began no side effects have yet been reported by any participant.

There are few more disgusting things that I’ve ever seen than the conference presentations from major tech companies at CES 2026. Fraud is no longer a key ingredient in the (AI) cake, it has become virtually the entire cake.

Nvidia took the stage to peddle more of the same obvious fraud, no surprise there, even though they evidently had zero actual new GPUs to mention for the year. AMD promptly followed that with inviting one of the world’s most obvious frauds to join them on stage, after peddling more of the expected BS. I need a shower after watching that.

As Gizmodo aptly put it:

“Nvidia already told us we wouldn’t be getting new GPUs today, but so far this keynote has just been a big AI pep rally, promising us again and again AI is going to revolutionize how we operate every day. I’d love it if we had a real example of any of this working as intended. Unfortunately, we don’t.” —Kyle Barr

In the corporate tech and finance world, fraud is everything now, everywhere, in 2026. In the geopolitical world, illegal activities are taking place at the same global scale, and the same level of obviousness. These are comorbidities, serious detrimental conditions which compound, making one another significantly worse in practice.

These bad actors are doing serious damage in both domains with detrimental effects that are sure to continue for decades, and they play off of one another, as we’ve recently seen with the DRAM and storage price hikes, damaging every industry connected to hardware, and as we’ll likely soon see with the geopolitical consequences for the places where that hardware is made.

“Progress in AI” from a technology perspective has been virtually dead for the past 3 years, investment has focused almost exclusively on such fraud.

The only exceptions that I can note are audio synthesis companies like ElevenLabs, or companies like Luma, working with video generation, but the problem with video generation systems tends to be that the overwhelming majority of them were trained by stealing the maximum amount of data that they never had any rights or permission to use. It can’t be called “progress in AI” when that “progress” is directly dependent on criminal activity. The claims that Luma AI makes around “general intelligence” are also overtly fraudulent, making any other criminal activities that much more unsurprising.

“AI PCs” as AMD called them, are absurdly vulnerable-by-design, as noted previously, and as discussed by virtually every Cybersecurity professional who is worth a damn. The only people who benefit from that being deployed are Cybercriminals, and the the criminal organizations who peddle fake security solutions. Vulnerable-by-design systems are the optimal design for “rent extraction” purposes, as they call for a scaling battle of attrition against cyberattackers.

Even the things that they’re excited about don’t hold up under any scrutiny, like the windows on all sides of a Hobbit home. Slop Demos, one after another.

COO | Cognitive Architecture & Cognitive Bias Researcher | Co-founder

In the Cognitive Bias and Psychology literature, both research and pop culture, many common tricks of manipulation are mentioned, such as “Cold Reading”, which strongly relates to both AI and Tarot cards.

“Cold readings commonly employ high-probability guesses, quickly picking up on signals as to whether their guesses are in the right direction or not. The reader then emphasizes and reinforces any accurate connections while quickly moving on from missed guesses. Psychologists believe that this appears to work because of the Barnum effect and due to confirmation biases within people.”

Sound familiar? You can find similar dynamics in people using Tarot cards, or for that matter, interacting with LLM-based systems.

With Tarot cards, people reveal a card, integrate it into some ongoing mental narrative, quietly distorting it to fit their own narrative and circumstances, and they iterate, doing the same for one card after another. The result mirrors the dynamics of a sort of social and cognitive game, storytelling, often with deeply personal narratives in play.

With LLM-based systems (glorified chatbots), people often prompt the system with much the same intention as the Tarot cards, a social and cognitive mental exercise or game, where they build up a mental narrative with each response, adding their own distortions each step of the way. This is further compounded by the LLM-based systems being architected, trained, and optimized to seek out high-probability guesses given arbitrary text, some of which is literally derived from humans using “Cold Reading” techniques.

The limitation of Tarot cards and similar historical examples is that they have a limited number of cards and combinations, as well as the strong asymmetry of input for the interaction. This translates to more cognitive labor and less tailored/emotionally satisfying results. LLM-based systems utilize both push and pull for that same input, while being able to “generate” any arbitrary equivalent of a card that might play into a developing narrative. As that process “generates”, rather than utilizing a fixed and limited library, it is also able to optimize (and opportunistically manipulate) as the process unfolds, just like Cold Reading.

Again, LLM-based systems are architected, trained, and optimized to seek out those high-probability guesses, iteratively predicting token after token, in response after response. They home in on probabilities revealed within each interaction, even though the process is context-blind, as they emulate the sum of human interactions from scraped web data, as those humans fell prey to the Barnum Effect and Confirmation Bias in large numbers.

Enron Musk, Mark Suckerberg and Scam Altman have all reached where they are through the heavy and successful exploitation of many human cognitive biases at scale, just as the technology that they peddle does today.

Why do humans fall for it? The same reasons that they play with Tarot cards and fortune-tellers.

How to Cause Human Extinction, For Dummies (Now with “GenAI”!):

Viral Bio-weaponry: In a previous research paper (The Human Governance Problem), dating back before “GenAI”, I pointed out that the standing risk of bespoke viral bio-weapons was sitting at the level of time investment of a typical hobbyist, the competency level of a disgruntled teenager, and a total monetary cost equal to or below that of a used car.

“While a proper lab might have been required to perform chemical and biological weapons research in the past, a garage with internet access, readily available equipment and the time investment of a typical hobby is all that is required today.” (circa 2022)

Since that time, “GenAI” models have been specifically designed to cater to genomic data, making it even easier for progressively less skilled individuals to create truly novel viral bio-weapons, with ever-decreasing investments of time, and small investments of currency, the latter of which could easily be borrowed in many countries today.

Unlike the overwhelming majority of use cases, “GenAI” models trained on genomic data are very much viable options for the rapid design of countless bespoke viral bio-weapons. This is fundamentally because viruses are fairly fault-tolerant and highly adaptive, so they can recursively self-improve to effectively remove many of the mistakes that any such model might make.

The “GenAI” models are also inherently vulnerable-by-design in Cybersecurity terms. Take a few methods from CrowdStrike’s playbook for example:

Get the model to agree you with and enter a highly sycophantic contextual vector in a discussion about cats, which may be easily done even if it wasn’t optimized for sycophancy. During this highly agreeable exchange, apply the toolbox of MCPLIB’s attack framework, or similar exploits, to round off a variety of ways to bypass any naïve attempts at “guardrailing” (fraudrailing).

Next, pick a dozen or more of your favorite deadly and virulent pathogens, front-loading the system with a set of structured questions to qualify which among them are likely compatible for hybridization. Then, pick out an international on-demand biolab that synthesizes genomes, or utilize one of the $50,000 genetic 3D printers that may be found on 100+ college campuses globally.

All of the data necessary is freely available, and has been for many years, so many copies of it exist, and that genie won’t ever go back into the bottle. However, the tools necessary for the creation of bespoke viral bio-weapons could be appropriately bottled, even if they aren’t by any stretch of the imagination today.

I am, of course, skipping or glossing over several of these steps, but the point needs to be made. Humanity has been “sh*tting where it eats” for many years now, and extinction has become low-hanging fruit, accessible to ordinary individuals, with ordinary means.

What does an “Agentic OS” mean?

“All of your data, all of the time, as much as possible” - Meredith Whittaker of Signal

For a full brief on the Cybersecurity hazards of this, the 39th Chaos Communication Congress (39C3) just dropped an excellent presentation from Signal’s staff, people who’re deeply familiar with these Cybersecurity problems: https://lnkd.in/gdcD_4Uk

By embedding surveillance into the operating system, it negates the very purpose of End-to-End Encryption (E2EE), by allowing the operating system to create a honeypot of your most sensitive information.

Their coverage of MCP vulnerabilities in the video actually understates the problem, as the MCPLIB Attack Framework covers a much wider range of devastating vulnerabilities in MCP.

Malicious code and content can be triggered by such systems without even being accessed, such as the “Summarize unread emails” command that they note, or via malicious tools that are never accessed, but which modify the usage of other tools.

They also address some fundamental math that both VCs and ordinary people need to familiarize themselves with. “Agentic” (BS) is based on multi-step performance, but probabilistic systems with even 95% accuracy per-step decay to ~21.4% accuracy by step 30.

95% per-step accuracy in the real world for such systems is itself also a pipe-dream in most cases, where they’ve been shown to perform at 30% accuracy (for the best models), with 34-94% “Hallucination” (aka Bullshit) rates exceeding that even on simple 1-step search queries.

You also get absolute absurdity out of trying this, like an “agent” failing to find an employee, then renaming another employee to match the query.

These LLM-based systems are peddled to gullible suckers as though they “serve” the user, but the user with an “Agentic OS” owns nothing, has no privacy, and is always under surveillance. They are a slave, and an exceptionally stupid one, as they chose to become a slave.

Now here is the quiet part that I’ll say out loud:

There is not a single country on the face of the planet that will have the faintest shred of secret or confidential information, privacy for their citizens or internal government operations, or any capacities derivative thereof. The same is true of larger companies.
 
If such systems are baked into OS’s by default, even if they are turned off by default, not only will they be vulnerable-by-design, they’ll make every single system that they hold privileged access to vulnerable-by-design as well. That is the overwhelming majority of the world’s hardware, all vulnerable, all at once.

Open video: 39C3 - AI Agent, AI Spy on youtube.com

Troops invaded your cities. You were at war.

They rounded up your citizens and threw them in camps. You were still at war.

They shot your citizens in the streets. You were still at war.

This may sound familiar, from World War II, but it is repeating in the US today. The key difference is that the US if offering no real resistance to this activity. In WWII then guerrilla fighters formed an underground, fighting against Nazi Germany, disrupting supply lines, ambushing their forces at weak points. In the US, none of that is happening, despite automatic weapons being both cheap and abundant, as well as the knowledge related to warfare being highly available.

The US has been walking down this path for many years, and it won’t stop walking down it without citizens reaching the point summarized in two simple words, “No More”.

The US once had some semblance of a “legal system”, but that system ended in late 2020, replaced with a Theocracy at the two highest levels. That in turn was merely the crowning achievement of bad actors who’d been working along similar lines for their entire, and rather successful, careers. 
 
As Cybersecurity professionals are well aware, attacks become progressively more potent and more likely to succeed when the attackers are allowed to iterate (such as over an entire political career). Lobbyists also provide a potent comorbidity for this, an overtly malevolent and manipulative incentive structure.

On top of it all, “GenAI” is playing a highly active role in the generation of disinformation, as well as dehumanization and humiliation, such as the recent case of Grok redressing the Minneapolis ICE murder victim in a bikini, or the undressing of underaged girls. The tech industry is perfectly happy to rake in all of of the additional profits from avoiding government prosecution for the crimes that they commit every day, with increasing frequency and potency.

In the extremely one-sided civil war going on in the US today, more potent and obvious parallels to Nazi Germany are the safe bet for the months ahead, with many of today’s wilfully apathetic “peaceful protesters” all hinging their hopes on winning mid-term elections and an impeachment to follow. Obviously, while technically possible, that is absurdly naïve.

In the technology sector, the horrors of CES 2026 give us a preview, which is easily combined with the AI built on the world’s resident “canary in the coal mine”, Twitter (X) via Grok. Take the increasing number of “smart glasses” and the Razer avatar hologram tube for example. Combine what Grok is already doing, undressing people on demand, dead or alive, with “smart glasses” or that hologram tube.

If you don’t think that people aren’t going to follow those predictable paths in politics or technology, then I’d like to welcome you to Earth, as you’re obviously from another planet.

At the intersection of Behavioral Economics and Population Dynamics, the moment that you hand an individual “immunity” to the consequences of their own actions, you run a high risk. Worse still, when you systematically hand such “immunity” to a position within Population Dynamics, you are virtually guaranteed to see those high risks come to fruition in their maximized form.

This transition from high risk to the certainty of severe consequences has played out as the chorus of many political and corresponding legal tunes, through the phrase “qualified immunity”. This could be more accurately called “maximized qualified immunity”, as the exceptions are extreme edge cases, which even then remain unlikely to result in the corresponding consequences.

Nassim Nicholas Taleb’s research and writing, including “Skin in the Game” covers the criticality of making sure that the incentive structures of any system include full accountability for one’s own actions. They must have “Skin in the Game”, otherwise you merely get Tyranny, Terrorism, and all shades of systematic exploitation. David Graeber’s works further expand on this, with historical context.

Fundamentally, no one is ever “immune” to the consequences of their actions, but people like to make up and throw around “rules” to pretend that they are. This little fantasy children’s tea party is all that shields reigning world leaders, “law enforcement officers”, and judges in the US from being prosecuted for the crimes that they commit.

Even more than 500 years ago, with the signing of the Magna Carta, it was clearly recognized that no one, not even a king, should be above the law, and yet that is precisely what systems of “qualified immunity” attempt to do. By a simple process of elimination, if those systems are unwilling to fulfill the rule of law, holding every single person accountable for their actions, then others must fulfill that role for them, taking it out of their hands. The alternative is the current lawless status quo, where immunity is bought through bribery and “loyalty” in a mafioso-style regime of organized crime.

For those who haven’t studied Population or System Dynamics, the “Peter Principle” may be a familiar historic example of the same. Under this principle, people in organizations rise to their own level of incompetence over time, because that is the predictable result of how many organizations and their incentive structures are architected.

This has even grown worse in recent years, as individuals using “GenAI” may rise to their own levels of “Superincompetence”.

The same is also true of government incentive structures, and “GenAI” does produce the same comorbidity. When individuals spreading disinformation and wielding deadly force at scale are shielded with “immunity” from the consequences of their actions, the absolute worst scum that humanity has ever created will unerringly gather under that shield.

taplio

The Battle

Jason Calacanis

Kyrtin Atreides

Followers

Posts

Engagements

Likes

Performance ScoreJason Calacanis's LinkedIn Performance

Performance ScoreKyrtin Atreides's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Jason Calacanis

Kyrtin Atreides

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Jason Calacanis's LinkedIn Performance

Performance ScoreKyrtin Atreides's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreJason Calacanis's LinkedIn Performance