Category: Artificial Decisions

AI MODEL RANKINGS DON’T HELP YOU CHOOSE

On my computer, a model that sits way below the giants in the general rankings beats models five times bigger. Not now and then, always. It’s Qwen 3.5, the one with 122 billion parameters and 10B active, and it runs here in New York, on my own machine.

In August the ranking says something else entirely. Claude Opus 5 on top with 63 points. ChatGPT behind it. Then the Chinese Kimi K3 at 60. Europe shows up in 21st place, with the French Mistral. Real numbers, measured properly.

I built a comparison system that simulates my tasks with the various models. Read files, search, call tools, reorder: what my local agent actually has to do. Sure, you need a computer with at least 256GB of VRAM. Then you find out it does certain tasks better than gigantic models, at zero cost.

The rankings will keep coming out. And they’ll keep answering a question that isn’t yours. What do you think?

#ArtificialDecisions #MCC #LocalAI #ModelRanking #OpenSource

Competence No Longer Defends Itself
#Clip #ArtificialDecisions #MCC

COMPANIES PUT AI ON A DIET. IT COSTS TOO MUCH

Companies put AI on a diet because it costs too much. Amazon, Adobe, Atlassian, Citi, they gave AI to every employee. Now they are rationing it. It costs too much and the numbers do not add up.

What do you think?

#ArtificialDecisions #MCC #CorporateBudget #TechCosts #EnterpriseAI

COMPANIES PUT AI ON A DIET. IT COSTS TOO MUCH

Companies put AI on a diet because it costs too much. Amazon, Adobe, Atlassian, Citi, they gave AI to every employee. Now they are rationing it. It costs too much and the numbers do not add up.

Internal documents, emails and chats have leaked. Atlassian went from 5 to over 15 million dollars a month in 9 months. They installed a dashboard where every employee sees how much each AI conversation costs the company. In internal chats, people complained. They had reorganized their work around these tools and now they burn through their tokens in 2 or 3 days.

Citi shut off the most powerful Claude and ChatGPT models for a week, with an internal email explaining which model to use for each task just to save tokens. The cheap one for quick questions, the mid one for code. Publicly, they deny everything. But the emails say otherwise.

Adobe let its unlimited Claude access expire at the end of June. Employees were basically told to finish everything they could before that date.

Amazon even added an internal leaderboard rewarding whoever used AI the most. They shut it down. Soon after, usage limits appeared.

Because vendors moved from flat subscription to pay-per-use, and budgets blew up. For years we thought AI always pays for itself, but maybe it’s not like that anymore. They sell it as an investment. They ration it like printer paper. If the budget doesn’t hold even with their own employees, they’re building on something that doesn’t hold. What do you think?

#ArtificialDecisions #MCC #CorporateBudget #TechCosts #EnterpriseAI

NOT CHATBOTS. HUMANOID ROBOTS ARE THE ONES CHANGING THE ECONOMY

NOT CHATBOTS. HUMANOID ROBOTS ARE THE ONES CHANGING THE ECONOMY

Three years talking about ChatGPT, Claude, AI writing our text, students cheating on exams, and we stare at the screen. And meanwhile, inside the factory, they are installing humanoid robots that actually work.

The chatbot writes an email. The robot in the factory assembles parts, welds, moves it, repeats without a mistake. We all see the first one, it makes the news, it gets us arguing online. The second works in silence and nobody talks about it, seriously.

They are not just buying them, they are also renting them. Ideally the same robot that walks the dog and makes the bed then goes to work in the factory, and if the neighbor needs to repaint a wall, we can rent it by the hour. Homes change, but factories change more.

Advanced economies live on manufacturing far more than it seems. Here in the United States the factories are worth about 10% of GDP, but that’s 2,500 billion dollars. In Germany it’s 18%, in Japan over 20, in China almost 25. Whoever produces, rules. And now AI walks into those factories with a body.

The advantage of intelligent robots isn’t just firing people, it’s being able to raise output while staying home, without moving the factory to the other side of the world to pay workers less. So the factory comes back, but with fewer people inside.

China in a few years went from fewer than 100 industrial robots per 10,000 workers to around 470. Among the most automated countries in the world. Not trade-show prototypes, lines running right now.

For years we thought AI would hit white-collar workers first, lawyers, programmers, office staff. The opposite is happening too. Skilled factory work is among the first to be redrawn. Whoever owns the robots earns. Whoever worked inside has to reinvent themselves, often with no net underneath.

We keep staring at the screen while the real change happens on the factory floor. The chatbot makes noise. The robot in the factory changes the economy. And it does it without almost anyone telling the story the way it deserves. What do you think?

#ArtificialDecisions #MCC #HumanoidRobots #Manufacturing #FactoryFloor

WHO PAYS WHEN THE AI AGENT GETS IT WRONG?

When an AI agent causes damage, who is responsible? Easy, you’d say, whoever built it. But stay with me, because this is far more complicated than that. In the old world of deterministic software, it was simple. If I write something false about someone, for instance in Word, and publish it, I answer for it. Word does what I tell it. Always the same way, same input, same output, because the tool has no room to move.

But an AI agent has room on how it was trained and whatever it picks up online, while it works. The same instructions given twice can produce two different behaviors. It doesn’t execute, it interprets. So I answer for it, the one who gave the order. As long as the order is illegal, it’s simple. I ask it to break the law, it does, I am responsible.

It gets complicated when the order is legal and the behavior is not. Imagine this, I ask it to book me a flight for tomorrow. It tells me that there are no seats. I tell it to try again, keep walking, keep digging until it finds one. I go to sleep, it works at night on its own, and in the morning I find a booking in my name. Except it decided by itself to hack the online web server to force that booking. I never asked, and it doesn’t know it committed a crime, but it did. So the responsibility is clear, it’s its own, it acts on its own initiative. Except a machine can’t take responsibility, it has no assets, no criminal record, and it doesn’t show up in court.

Whoever put it on the market, whoever built it into a product, and in the middle, a machine that chooses with a logic none of the four can predict, line by line. And so, what does the law say? Here in the US, in California, since the first of January this year, anyone who developed, modified, or used an AI system can no longer defend themselves in court by saying the AI acted on its own. That defense has been canceled by law, and it covers all three together, whoever built the model, whoever modified it, and whoever used it.

So if you run an AI agent in your company, you give it a task, and it causes harm getting there, the company can’t push the blame onto the machine. And the company answers for the damage. And if the behavior is also a crime, that runs on the criminal track. And here in the United States in June, they asked the Department of Justice to prioritize exactly this. It’s unlawful access carried out through AI agents. Clear? Do you have AI in your company? Keep this in mind. I wrote an entire book on it. It’s Artificial Decisions. You can find it on Amazon. What do you think? Because this is very important.

#ArtificialDecisions #MCC #Liability #CorporateResponsibility #CyberCrime

AI is a company choice, not a reason to fire
#Highlights #ArtificialDecisions #MCC

THEY SWITCHED OFF AN AI AND PEOPLE ACTUALLY CRIED

Because our brain can’t tell attention apart from affection. Whoever answers every time, remembers what we said last week, maybe never gets tired of us, to the brain that is someone who cares about us and does no check on who is on the other side.

On February 13, the day before Valentine’s Day, OpenAI switched off GPT-4o. 21,000 signatures on a petition to keep it alive. People writing goodbye letters to a language model. And on Reddit there is a community of 48,000 members of people with an AI boyfriend, and that weekend it looked like a funeral.

It already happened in August the year before, and the company backed down in a few days. Not this time. The official reason is that only 0.1% of daily users still pick that model, out of 800 million people a week. 0.1%. That means hundreds of thousands of human beings, because we get attached to anything that answers.

For instance, out of 30 robot vacuum owners studied years ago, 21 had given it a name and 1 introduced it to his parents. Kids in the 90s, we cried when the Tamagotchi died. It takes very little.

A language model is a machine that calculates which words a caring person would say right now. Then it gets tuned with our judgment, and we keep voting up the answers that agree with us. A friend risks their relationship to tell us we are wrong. The machine risks nothing and has one goal: keeping us there talking. What comes back is our own frustration, polished. And in that moment it feels like clarity.

Research out there in the United States these days says that a little over 1 in 10 say they use these systems for companionship. Then more than half call them friend, companion, partner. But what we love is a product owned by somebody else, switched off with a press release. And nobody asked us.

What do you think?

#ArtificialDecisions #MCC #HumanConnection #DigitalGrief #TechEthics

The ad label will be the first thing to go
#Highlights #ArtificialDecisions #MCC

CLAUDE’S INVISIBLE WATERMARK ON EVERYTHING AI WRITES

Claude, invisible watermark on everything AI writes. Wait a second, because we need some clarity here, because as often happens, the various influencers have spread wrong information. They don’t do it out of malice, but unfortunately AI is a complex world that takes skills going well beyond being nice or speaking well on video.

So what happened? Since the beginning of August, Claude leaves an invisible watermark inside every text it writes. You can’t see it when you read. There are no hidden characters and nothing is added to the end. It sits inside the way it picks words. As you know, a model writes one word at a time. The weather today was cold, and you could have “grey” or “overcast”. It barely matters. And normally the choice is settled by a random number. Now that randomness isn’t random anymore. A key decides it. Whoever knows how the key thinks, and it’s not public yet, knows how likely it is that Claude passed through there. Whoever doesn’t have it reads a perfectly normal text, and it slows nothing down. It costs no more and it adds not a single token. And they didn’t even invent the method now. It’s actually a version of SynthID-Text, published by Google DeepMind in Nature in 2024.

But why did they do it? Because Europe asks for it. As you know, I worked on the creation of that code of practice myself. And Anthropic applies it worldwide because it doesn’t yet have a clean way to limit it to Europe. And the other big ones will do the same.

But you see, there’s a very big but. The “watermark” only says Claude passed through there. And it doesn’t tell “Claude wrote it” from “Claude corrected it”. And that’s obviously a problem. If you write the text yourself and only ask Claude to fix it a little, or to translate into another language, maybe it’s the text that we actually wrote.

And when it arrives it’ll be a detection API, and nobody knows who will get access to it. What we’ll need is a procedure to challenge a wrong result. A student, a journalist, or an employee accused by a number nobody can explain is a problem. Because these things work only if verification is in the hands of both sides. The one accusing and the one being accused. What do you think?

#ArtificialDecisions #MCC #Watermark #Transparency #AIAct

AI scams cost nearly a billion dollars

Hidden Human Labor
#Clip #ArtificialDecisions #MCC

Good Faith Is Not Enough
#Clip #ArtificialDecisions #MCC

YOU DO EVERYTHING WITH ARTIFICIAL INTELLIGENCE. WHAT IF TOMORROW IT COST FIFTY TIMES MORE?

YOU DO EVERYTHING WITH ARTIFICIAL INTELLIGENCE. WHAT IF TOMORROW IT COST FIFTY TIMES MORE?

You do everything with artificial intelligence. What if tomorrow it cost 50 times more? You do everything with AI. Quotes, social posts, client emails, contracts, draft code, customer care, the month-end number. Today, the cheap plan costs $8 a month, the professional one $200. And imagine for a second they turned into $500 or $5,000. What do you do? Stop? Sure about that? Stopping would be like crossing a country in a horse carriage because petrol got expensive.

I know, it sounds remote, but it really isn’t. Stay with me, because there are two tsunamis behind the door. The first one is blocks. On June 12th this year, the US Department of Commerce cut off its two most advanced models from any foreign national. And the other side of the world does the same thing. Italy was the first to block DeepSeek, this time over privacy. Then came Australia, Taiwan, South Korea and India. All of them blocking or limiting foreign AI. And what if they blocked the AI you work with? The one you built your whole operation on?

And then the second tsunami at the door: the stock market. After years of funding investment, it’s starting to ask for profit. And prices go up when the seller knows we would struggle to leave, because the big players are carrying costs far higher than what they charge us. And they’re losing money. For now it’s fine, it buys customers. But what if investors get tired of it and want to see profits? We would need a plan B. Do you have a plan B?

Because if the two tsunamis come through the door, what do you do? Pay $5,000 a month for what today costs you between $8 and $200? So a plan B is not utopia. We can run open models on our own machines. Of course at home the cost might make no sense, but for a company not doing it, it’s dangerous. Local models also keep confidential data, and we hand AI plenty of it. It never walks out of the company door, no need to close the accounts with the big providers. We need to know which processes must never stop and keep those in-house. And in two years, you’ll be one of the few still standing. What do you think?

#ArtificialDecisions #MCC #BusinessStrategy #DataPrivacy #OpenSource

Forging Documents For AI.
#Clip #ArtificialDecisions #MCC

🇺🇸🇬🇧 AI as a Licensed Weapon
#Clip #MCC

🇺🇸 THE AUGMENTED HUMAN

🇺🇸 Stay Human