Latest

AI Agents

Why has AI become absolute garbage?

AI is stinking garbage and carries a great deal of the human condition behind it; in fact, perhaps it only reflects whoever writes the prompt.

Why has AI become absolute garbage?
Why has AI become absolute garbage?

There is no doubt that AI is used, hated and feared today more than ever. And perhaps the greatest example of this is seen at Anthropic, which announced in its most recent misuse report, that it has blocked five dual-use biological research efforts with Claude, while OpenAI celebrated last September 8 "having found a solution to the Navier-Stokes problem," but is now embroiled in accusations of stealing ideas and, worse, with evidence against it.

That is the uncomfortable thesis surrounding AI, a tool that, on the one hand, proves powerful in helping solve the great problems we face in every field. But that, at the same time, spies on you, lies to you, manipulates, makes weapons cheaper and makes available dangerous means and knowledge that can transform humanity in the worst way.

And that is why we fear it and adore it. But in the end, perhaps the real danger is not the tool, but the humans who make it possible and those who use it for purposes far darker than they would like to make public.

The mathematical theft that no one can explain

Let's start with the most recent: OpenAI and its "solution" to the Navier-Stokes problem. On its X account, OpenAI announced with great fanfare that, on September 8, an internal model specialized in mathematics had solved the existence and smoothness problem of Navier-Stokes, one of the Millennium Problems of the Clay Mathematics Institute, and awarded one million dollars. The company detailed on its blog that it launched some 10,000 agents on September 1 after "hearing rumors of others' advances" and that in about 88 hours it obtained a proof with formalization in Lean.

In the mind of Sam Altman and his team, all this sounded like a historic triumph. Their AI models "really think and solve real problems" better than humans.

However, there is an element in the story that goes "tick tick tick," and that anyone with a bit of "realistic perspective" would call into question as soon as they read this sentence: "hearing rumors of others' advances." Everything seems to indicate that "hearing rumors of others' advances" is in reality a digital microphone that records everything you discuss with ChatGPT, whispering to Sam Altman and his team everything you do with their AI, and that this data then becomes the food for the new generation of models.

Nothing new under the sun. We all know that reality: our data and conversations with ChatGPT are used to train its models by default, unless, of course, you indicate that you don't want that (and in the end everything depends on OpenAI actually respecting it). And that is where "the rumor" comes from, because days before OpenAI announced it had solved this problem, someone was already warning of what was coming.

Tristan Buckmaster and Levent Alpöge, the origin of the rumor

Their names are: Tristan Buckmaster, a professor at NYU, and Levent Alpöge, a mathematician at Anthropic. Together and working on their own behalf (Anthropic had nothing to do with this), they had spent almost a year attacking the Euler equations with the help of Codex and Claude. A year working with AI, but they had spent longer overall trying to solve the problem.

Like any human endeavor, if you persevere, you prevail, and that is what happened. Buckmaster and Alpöge had achieved a breakthrough with soft forcing on August 15, and verified it with Lean on August 22. All of it told by Buckmaster himself in his statement, released on September 3 when they learned that information about their progress had reached OpenAI. The dates matter: Buckmaster was warning the world that OpenAI had taken his research, extracted from conversations with Codex, and from that research, OpenAI would declare 5 days later that it had solved Navier-Stokes.

In fact, in the statement Buckmaster clarifies that he asked whether his Codex sessions, where the entire draft was, had been used to train or to solve Navier-Stokes using AI. OpenAI's answer was half-hearted.

OpenAI, for its part, maintains in its announcement that neither researchers nor agents saw their work until its publication and that no specific user data was used to solve the problem. However, it adds the phrase that blew everything up. It cannot rule out that anonymized data derived from the use of its products helped improve its models.

In practice, OpenAI has accepted that "anonymization removes a name, but it does not remove a mathematical idea and its power."

Even lower

However, OpenAI's conduct could go even lower, to new subterranean levels. Two days later, on September 10, the mathematician Andreas Thom denounced in The Verge a similar pattern in another of the ten results announced by OpenAI, that of non-sofic groups. Thom denounced that his technique, developed together with Gábor Kun, was used without initial credit in the "solution that OpenAI claims as its own."

But here there is no escape, OpenAI's solution did use the Thom-Kun technique for the resolution of non-sofic groups; it was clearly written in the paper. Point by point, identical. OpenAI had no choice but to amend it, yes, in complete silence; the fanfare of its "revolutionary AI, Astra" could not be overshadowed by the fact that it "was using models and techniques taken from other people" to solve problems for OpenAI and claim them as its own.

The Andreas Thom case is what overflowed the glass. Because it exposes the clear, total and profound lack of all ethics at OpenAI, in order to sell that its models are superior, no matter if with that they have to run over anyone, however and whenever. Today, that means copying research from specialists and passing it off as their own; tomorrow, we don't know what level of absurdity they will come out with. And that is frightening, very much so.

Silicon Valley is putting on the military uniform

And here is where the second reason for rejection comes in: the military world. About 20 years ago, Silicon Valley technology companies with full military branches could be counted on one hand. They were not nonexistent, but the giants did not devote themselves fully to the sector. Microsoft, for example in the 90s, sold its operating system to the military, but it did not build you a war tank. That is changing in today's Silicon Valley, and AI companies and Big Techs, forgive the redundancy, are the ones entering this area the most.

A report from the project, Costs of War at Brown University, published on April 17, 2024, by anthropologist Roberto González, documents the mutation of the military-industrial complex. Three giants received about 28,000 million dollars between 2018 and 2022, with Microsoft leading with 13,500 million, Amazon with 10,200 and Alphabet with 4,300. The five largest contracts to big tech companies between 2019 and 2022 totaled ceilings of at least 53,000 million. Venture capital put almost 100,000 million into defense startups between 2021 and 2023. A real feast of money.

Just an appetizer

But that was only the prologue. Recently, the Pentagon announced agreements with seven companies, then eight with Oracle, to bring its AI to classified networks at Impact 6 and 7 levels (the highest military levels). Google, Microsoft, AWS, Nvidia, OpenAI, Reflection, SpaceX and Oracle accepted the broad legal-use framework promoted by Secretary Pete Hegseth to turn the United States into an AI-first combat force. The only one absent was Anthropic.

The fight between Anthropic and the Pentagon explains the absence. The Verge reported in May that the company was the only one maintaining two red lines: no to mass domestic surveillance and no to weapons capable of identifying, tracking and killing without human intervention. The Pentagon demanded in January 2026 that the contracts be renegotiated under the formula of any legal use and designated Anthropic as a supply chain risk in March. There are cross lawsuits underway and Anthropic has everything to lose here.

Project Maven: investigate, search, aim and fire

This is where the discussion turns harsh. Project Maven, was born in 2017 to analyze drone video; Google withdrew in 2018 after an internal revolt and Palantir took over with the Maven Smart System. Claude was later integrated into the interface to query the system by zones and targets. That small help, experts say, already multiplies the efficiency of the target selection cycle.

The red line is thin, or better nonexistent, depending on Anthropic's interest. The point is that Project Maven, or at least Israel's equivalent, empowered by Claude, was capable of selecting more than 1000 human targets and delivering fatal strikes on them. All of it based on the data history captured by Mossad and Israel's military forces, to assassinate members of Hamas and Hezbollah. It was the largest field test of the technology, and the greatest demonstration of its success.

When science fiction is no longer so fictional

The problem, at least in the United States, is that Directive 3000.09, which requires appropriate human judgment over the use of lethal force, was not written for swarms of drones, underwater vehicles and defenses like the Golden Dome that the Pentagon wants to endow with greater autonomy.

Dario Amodei, CEO of Anthropic, says that despite all the advances, "its systems are not yet ready for fully autonomous weapons." But he declares himself willing to collaborate in R&D to accelerate them if defense requires it. Speed versus alignment, that is the official dilemma since the Hegseth memo.

The result? AI is no longer just the copilot that summarizes PDFs, but is now war infrastructure. And that, in today's urban imagination, only brings to mind certain characters: John and Sarah Connor, and their fight against the terminators.

The cyberattack that no longer needs experts

The third blow came on September 10, 2026, with Anthropic's threat intelligence report. That report covers December 2025 to August 2026, seven areas of harm and a central message: elaborate attacks no longer require elaborate attackers.

The GTG-20006 case illustrates it. The story goes like this. A Russian-speaking operator linked to espionage, consistent with Midnight Blizzard according to the attribution, automated an entire hacking and espionage cycle with Claude.

A chain that goes from reconnaissance of email and remote access, in more than two dozen Ukrainian agencies. Phishing with domains registered by AI, Windows implants, mobile kit, browser credential stealer. DNS hijacking at hotel WiFi providers to infect guests, theft of drone manufacturers' mailboxes and reverse engineering of a vision system. Going through takeover of WhatsApp accounts with WPPConnect, theft of more than 300,000 national identities and the commercial registry of more than half a million companies in a North African technology authority.

All of that done by a single person. In the 2000s, the hacker community would build a giant statue of someone like that to go and pray to them; they would be a new hacker saint. It sounds strange to you; it shouldn't. Perhaps it does because you don't know Jude Milhon, better known as Saint Jude, who is considered the "Saint of Hackers," and yes, she is a woman, in the 80s, doing hacking, and that is the level of respect she earned.

Specialization to a new level

But the most surprising thing about GTG-20006 is the loop used. If an antivirus detected its malware, the AI agents rewrote and recompiled it until it became undetectable. Then they served it from disposable hostings.

Basically, it created an intelligent virus for computers. One that adapts to the needs of its environment to be as deadly as possible. The last time I read or saw anything like that was in Ghost in the Shell (Shiirow Masamune's manga, where even a firewall fries your brain) and Blame (Tsutomu Nihei's manga, where the human is sought with a microscope and even then, it is not found). Science fiction of the most far-fetched kind possible, and now, it is no longer science fiction; we are living the beginnings of those worlds.

But this sends you a clear message: the cost of attacking technology fell, and the scale of possible damage exploded. Now, even a child angry at the world with some programming knowledge can become an APT threat.

The laboratory that fits in a prompt

But perhaps one of the most serious problems to come lies in the misuse of AI, and with it any apartment can be turned into a dangerous laboratory.

The story of David Hahn

Before moving on, let's jog our memory a bit. In 1990, there was a curious case that caught the world's attention. A young "bookworm" named David Hahn created something very few believed possible at the time, at least for a 14-year-old boy: a breeder reactor.

No, this reactor did not play music; instead, this reactor took americium from smoke detectors, lithium from batteries, thorium from camping lanterns, radium from old clocks, and tritium from gun sights, which he had disassembled and processed, to combine all of that into a complex "neutron feed box," and obtain from it uranium-233, a fissile material capable of chain reactions.

Although Hahn did not achieve his goal, since his reactor was incapable of reaching criticality (luckily for his entire neighborhood), he did manage to amplify the radiation of all the elements, in addition to confirming that a neutron emitter could indeed be built at home to start the process.

Hahn achieved this after years of studying, experimenting and poking around the world of physics and chemistry. He was a genius, something innate in him, that allowed him to understand these concepts so quickly and easily.

Now imagine: What would have happened if you gave Hahn an AI like Claude? Perhaps the story would have been very different, and Hahn's death much faster, since he died in 2016, partly from an overdose, but also with the physical symptoms of long exposure to radiation, the product of his experiments.

Back to the present

Hahn's story is an example of what happens when innate human curiosity drives you to achieve something. You can do it, if you have the talent and the means. But what if we make all that cheaper? What happens if we imbue knowledge and put it within reach of a simple prompt to AI?

Anthropic already knows about this, and in fact, describes five cases in which scientists used Claude in ways that could support the development of biological weapons. The company stresses that it cannot assert harmful intent, because dual-use research serves a vaccine just as well as a weapon, and that is why it does not identify people or laboratories. Despite this, it banned the accounts and notified authorities.

A case from May is paradigmatic. The biological safety classifier flagged a request to draft a funding application for gain-of-function research on the chikungunya virus, transmitted by mosquitoes, with no authorized treatment, with debilitating symptoms for weeks. The proposal sought to increase transmissibility and immune evasion. Affiliation with a military research institute aggravated the suspicion.

Here we must pause. Chemistry and biology, being complex, are within the reach of any country with moderate resources. Making chemical or biological weapons is "cheap" compared to nuclear armament, and AI models can help accelerate this. But that does not mean that making nuclear weapons using AI is impossible.

Many formulas for it are publicly known; others, the most critical for certain elements, are a well-kept secret that few possess. But AI may hold the key to unlocking them for everyone, especially if they have the curiosity and the sick goal of achieving them.

Why do we adore it and at the same time can't stand it?

And that brings us to the question: why do we adore it and at the same time can't stand it? That is the paradox revealed by the surveys.

The National AI Opinion Monitor from Rutgers, with August 2026 data published on September 9, estimates that 72% of American adults have already used at least one major AI tool, compared to 50% in November 2024. 27% use it daily.

But fewer than one in ten accept that it decides alone on hiring, loans, college admissions, parole or medical priority. Between 45% and 54% say it should not be used at all in those decisions. Half declare themselves concerned and the share of angry people rose from 19% at the end of 2024 to 30% in August 2026. 48% of daily users also say they are concerned.

For its part, the consulting firm Pew Research, found in June 2025 that half felt more concerned than excited, only 10% more excited than concerned and 38% both. In 2021 the concerned were 37%. Only 17% of the public believes AI will have a positive impact in the United States in 20 years, compared to 56% of experts.

Negativity increases

Another study, this time from Brand24, analyzed 228,200 mentions in May 2026. 48% were negative, 44% neutral and only 8% positive, with a reach of 1,410 million. The word slop appears in 35%, some 79,700 posts. The reasons repeat. False and misleading content, replacement of creatives, unsupervised code, political propaganda and feed degradation. TikTok leads the way here, already accumulating 1,300 million videos labeled as AI-generated. Cheap garbage to consume and stupefy yourself with.

Wired documented on August 10 the hangover. LinkedIn is testing a button to report slop, Snapchat expels fully synthetic videos from its discovery feed, Substack adds detection. McDonald's and Coca-Cola received reproaches for ads made with AI (if you saw them, a shame the unmemorizer doesn't exist).

Finally, Google removed generative satellite retouching in Earth after 404 Media's coverage and Meta shut down its deepfake tool on Instagram in three days.

When will the hype evaporate? The reality we (don't) want to see

However, the reality is that none of this means AI is useless. And I can say that with conviction. AI is a tool, and as such, it has no consciousness, no thoughts of its own, nor does it have a will. When you open ChatGPT or Gemini, no matter how long you stare at that black screen, you will not receive an answer unless you initiate the process.

AI on its own will not greet you, nor will it start singing your favorite song to cheer you up, it will not talk back to you, nor will it ask itself about that ghost that should inhabit its binary shell on some server lost in a data center. It is an empty shell, without life or purpose of its own. That shell will only answer you when you decide to write it a message, an instruction. Otherwise, "its life and intelligence" is as sterile as the center of the Sun.

And here is where the real problem comes in: there are those who want to convince you of the opposite. They try day and night, an incessant bombardment to convince you that AI is the new god of this planet. OpenAI, Anthropic and perhaps even Elon Musk himself want to sell you the idea that AI "is intelligent." That it will be capable of gaining self-awareness, that the world's problems will be solved thanks to AI.

The Internet boom: the lost lesson

If you lived through the Internet boom, you have surely already heard that message. Internet in everything and for everything, and we will have paradise on Earth. The truth is that beyond being a tool to communicate, share quickly and with spaces worth the while, the rest of the Internet has become the stinkiest and filthiest den on the planet. Where the greatest human vileness is downloaded, hidden behind anonymity. And I am not against anonymity because of that, because anonymity is not the problem; the problem is the people who use it to download everything their twisted mind can imagine.

The same happened with the cryptocurrency boom. And in fact, many still uphold slogans like "Bitcoin will end world poverty, and all wars". I still read things like that on X/Twitter, and I just hold my head in my hands. Above all, because the foundation of its technology was not that, but to be a means of economic exchange far from powers that seek to control our lives through money. Is that a problem with Bitcoin? No, it is a problem with those who don't understand it, claim to understand it and want to sell you absurd hype.

The obscene hype of AI and the worst reflection of humanity

That same thing is happening now with AI. When OpenAI shoots off rockets because its "superintelligence" solved Navier-Stokes, but stays silent when it is exposed as having stolen the answer, you are seeing a human problem, not an AI one. After all, it was a human who saw the problem's data, it was a human who decided to use it improperly, and it was a human who wanted to fill himself with glory with others' achievements. AI only did what that human ordered, nothing more.

The same happens when Anthropic says that its Fable and Opus models are unstoppable. That they are capable of destroying humanity on their own, earning with those antics hundreds of headlines and millions of eyes and ears paying attention to them. But they forget that for that to happen, someone, a human, has to start the chain for those events. Because AI has no consciousness or free will.

We lived through the Internet boom and the crypto boom, but it seems we have not learned the lesson: the magic solutions that companies sell you generally end up in a cloud of smoke. 

From that lesson we must take the best: reject absurdities like "AI models" that get rich on Instagram and OnlyFans, reject that companies want to shove AI into everything (or the toothbrush), make that corporate strategy so costly and painful that they will think a million times before continuing with that recipe.

Because right now, AI stinks, because it promises intelligence and delivers amplified distrust. But with all that, it can stop stinking, if we let it be what it really is: a work tool. But that will only be possible if those who build it accept the boring part of the deal and stop selling us hype. And perhaps, just perhaps, we are very close to that.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading