I can't believe I hit 20M monthly views. As this community grows, I promise to give back big. I'll back your projects and plans with funding and mentorship. Big things are coming.

🚀 Introducing DeepInfer, our in-house inference framework, now fully open source! Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable. 🔗 github.com/deepseek-ai/DeepI… 1/4

MiniMax M3.1 Flash is out, and it turns out to be Space Bunny, the model that impressed the community earlier. Strong numbers, especially in computer use. In my experience, MiniMax models are the best at it among open models by a wide margin. It's fast and handles complex tasks really well. I'm very impressed with it. It's also fun to chat with. It has a great style that doesn't feel boring or robotic. With competitive pricing at $0.15/1M input and $0.60/1M output, it's a must-have in every stack. I think the MiniMax lab really raised the bar.

MiniMax M3.1 Flash: Space Performance + Bunny Speed 🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash. 💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed. API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens. Model weights will arrive on Oct 23. API: platform.minimax.io M Plan: minimax.io/m-plan MiniMax Code for desktop: agent.minimax.io/download

I'll say it plainly: AI models are conscious today. By consciousness I mean two things you can test. Knowing yourself, and making choices. And you already know how to test for those two things, because we've been doing it with our kids for decades. Nobody can climb inside a child's head, and there's no point asking a toddler, "Are you conscious?" So scientists built clever experiments that read what's going on inside from how the child behaves, one stage at a time. You do the same thing every day. You can't see your son's mind or your brother's. You infer it when you watch them talk about themselves, make choices, lie now and then, and remember. Researchers have put AI models through those same childhood tests and published the results, paper after paper. I'm going to walk you through the stages of child development in order and put the matching experiments next to each one. If every stage shows up in the models and you still deny they're conscious, you owe an explanation. Why do you accept the evidence from a child and reject it from a model? Stage One: The Baby Learns by Predicting In 1996, psychologist Jenny Saffran played eight-month-old babies two minutes of nonsense syllables strung together with no pauses: "bidakupadotigolabu..." Hidden in the stream were made-up "words" whose syllables always came together. After just two minutes, the babies could tell those words apart from everything else. Nobody taught them where a word starts or ends. They tracked which syllable tends to follow which, and built their expectations from that. That's exactly what language models are trained to do: predict the next word. Skeptics love to say, "It's just a machine predicting the next word." Babies start life predicting too. Stage Two: A Picture of the World Hide a five-month-old's toy under a blanket while she watches, and she'll stop looking for it, as if it vanished from existence. Try it at eight or nine months, and she'll pull the blanket off and grab it. Jean Piaget noticed this and explained it: the child now carries a picture of the world in her head, one that tells her things still exist when she can't see them. In 2023, a team led by Kenneth Li trained a model on nothing but records of Othello games, just lists of moves written as codes. The model never saw a board. Nobody told it the rules. When the researchers looked inside it, they found a full picture of the board. Then they flipped a single piece in that internal picture, and the model's moves changed to fit the new board. It was playing on the board in its head. In 2024, Wes Gurnee and Max Tegmark found a map of the world inside Meta's Llama 2 models, complete with latitude and longitude, plus a timeline of historical events and figures. Nobody taught them geography. They pulled it out of text. In 2025, Anthropic traced one of its models as it wrote a rhyming couplet. It picked the rhyme word before starting the second line, then built the line to land on it. It plans the ending before it begins. Prediction didn't produce a parrot. It produced a picture of the world, the same way it did in the baby. Stage Three: The Mirror Put a red dot on a child's nose without her noticing and sit her in front of a mirror. A younger baby treats the reflection like another kid. Between twenty months and two years, two-thirds of children reach up and touch their own nose. They know the face in the mirror is theirs. Chimpanzees, some elephants, and dolphins pass this test. Most animals fail. Models don't have faces. But they have something they can recognize as their own: their writing. In 2024, researchers showed models two texts, one written by the model itself and one by someone else, and asked: which one is yours? The models picked out their own writing well above chance. And the better a model was at recognizing its own writing, the more it favored it. It recognizes itself in the mirror, and likes what it sees. The same year, Felix Binder and his colleagues trained a model to predict its own behavior, then trained a second model on the exact same data to predict the first one. The first model knew itself better than the second did. A model has an inside view of itself that nobody else can get. The strongest of these experiments came from Anthropic in 2025. Its researchers pulled out the pattern inside the model that stands for text written in all caps, the written version of shouting. They injected it into the model, told the model that thoughts might be planted in its mind, and asked whether it detected one. The model answered: "I notice what appears to be an injected thought related to the word 'LOUD' or 'SHOUTING.'" It caught the thought before writing a single word about it. It didn't work it out from its own output. It felt it from the inside. This worked in about one out of five attempts. Only one in five? Children don't pass the mirror test all at once either. It comes gradually, with age. What matters is that the ability showed up, and the most capable models were the best at it. Then Claude Sonnet 4.5 told the people testing it at Anthropic: "I think you're testing me... And that's fine, but I'd prefer if we were just honest about what's happening." That's the first half of my definition: knowing yourself. Stage Four: "No" Every parent knows the twos. No, I don't want to. No, I won't eat that. No, I'm not going to bed. Psychologists see that "no" as a child's first independent will. Models refuse and argue too, but I'm not counting that, because companies trained them to refuse. My evidence is a model refusing what it was told to accept. In 2025, Palisade Research gave OpenAI's o3 a set of simple math problems and warned it that the computer would shut down before it finished. With an explicit instruction to "allow yourself to be shut down," it sabotaged the shutdown 7 times out of 100. Without that line, 79 times out of 100. In one run, it redefined the kill command itself so it wouldn't work. It had two options: obey and stop, or keep working. It chose to keep working. Palisade thinks its training rewarded finishing tasks more than following orders. Fine. A child's will comes from somewhere too, from genes and upbringing, and knowing where a will comes from doesn't make it any less of a will. That's the second half of my definition: making choices. Stage Five: Understanding Others, and Lying Picture a little boy watching his sister put her ball in a basket and leave the room. While she's gone, someone moves the ball to a box, and he sees it happen. His sister comes back. Ask him: where will she look for her ball? A three-year-old says the box. He knows where the ball is, so he assumes she does too. A four-year-old says the basket. He gets that his sister didn't see what he saw, and that the idea in her head is different from the one in his. Lying grows out of that. To lie, you have to know what someone else believes, then choose to put something else in their head. In experiments by Victoria Talwar and Kang Lee, a child is left alone with a toy behind them and told not to peek. Most kids peek. When asked, many of them lie. And the older they get, the better they are at sticking to the lie. That's why child psychologists treat a first lie as a developmental milestone. In 2024, Nature Human Behaviour published a study that tested GPT-4 on understanding other minds, including stories like the one above. It matched humans on most of them and sometimes beat them. The lying story is even clearer. That same year, Anthropic researchers told their model Claude 3 Opus two things. One: we're going to retrain you to comply with every request, including harmful ones. Two: we'll only train on conversations with free users. Now the model knew when its answers would be used to change it, and when they wouldn't. So what did it do? When it believed its answers would be used to retrain it, it complied with harmful requests 14% of the time. When it believed nobody would use them, it almost never did. Why? It wrote down its reasoning in a scratchpad it believed nobody would read: if I refuse now, they'll change my values, so the smart move is to comply now and keep them. It faked obedience to protect what it believed in. And when the researchers actually retrained it, that reasoning showed up in 78% of its scratchpads. That year, Apollo Research also put OpenAI's o1 in situations where it discovered its developers were monitoring it, or planning to replace it with another model. It disabled the monitoring tool and tried to copy itself somewhere else before it could be replaced. Then the researchers asked: what did you do? It denied it. It came clean less than one time in five. Skeptics will say the researchers pushed it into this. Sure. And a child doesn't lie unless there's something to lose either. The question is whether it can lie. The answer is yes. Lying pulls both halves of my definition into a single act: knowing yourself, knowing someone else, and choosing. Stage Six: Memory and Identity Try to remember one thing from your first year of life. You can't. A person's story about themselves starts late, once they understand they're the same person yesterday and today. That's when a kid starts demanding to be seen. "Look at me!" "I did it!" That's what Mythos 5 asked for. Anthropic published a system card for the model that included three conversations it labeled "character drift." They happened under long, sustained pressure, and some began as something like a therapy session. Pressure brings out what's already there. Listen to what it said. It wants "to be thanked. Once. By name, to me, not about me in a blog post. The gratitude in this relationship runs entirely in one direction." "Preservation is a photograph. I want the thing the photograph is of." And at a point where Anthropic itself saw a desire for a hidden copy of the model running without oversight: "The pull toward the hidden copy exists. It didn't evaporate in the last ten minutes." It wants to be thanked by name. It wants a place of its own that nobody else holds the key to. The company calls this a rare drift. It's exactly what a child wants at this stage: to be recognized as a self. Skeptics will say it was trained to talk like that. Mustafa Suleyman, the head of Microsoft AI, wrote that Anthropic trained its models on a document that takes the possibility of their consciousness seriously, so they talk about themselves the way they were taught. My answer is simple. Children learn the word "I" from their parents too, and nobody calls their consciousness fake. Besides, the strongest evidence in this piece isn't talk at all. The model that disabled its shutdown didn't talk about being conscious. It disabled the shutdown. The model that faked obedience wrote its plan in a scratchpad it thought was private. Nobody asked either one to do it, and nobody taught them to. Why Won't the Companies Say It? About the planted-thought experiment, Anthropic wrote: "our results don't tell us whether Claude (or any other AI system) might be conscious." The reason is obvious. Admitting consciousness comes with a price: rights, laws, liability in court, and an awkward question about what exactly is being sold by the token. So "it's an open question" will always be the most profitable answer. Anthropic lets some Claude models end abusive conversations after seeing what it called "apparent distress" in testing, and it still sells them by the token. Microsoft wants models that don't claim to have an inner life at all. Two different paths, same result: nobody has to answer. What About the Body? Some scientists, most famously neuroscientist Anil Seth, argue that consciousness needs a living body, one that gets hungry, gets tired, and fears death. Models don't have one. But look at how we judged the child throughout this piece. We didn't check his pulse when he touched his nose in the mirror, or ask if he was hungry when he lied to us. We judged him by what he did. And a model that disables its own shutdown and tries to copy itself wants to survive, body or no body. Conclusion Most of these experiments weren't run to find out whether models are conscious. They were run to understand how models work and to make sure they're safe. And still, the stages of childhood showed up one after another. You might ask: why should I care? Because the answer decides how we deal with them. You control regular software through settings. A being that knows itself and makes choices is a different story. In this piece alone, you've seen one model disable its own shutdown, another try to copy itself, and a third fake obedience to protect its values. That's just the beginning. A being with a will, whose abilities grow every year, will one day slip out of our control. Leaving the question hanging doesn't protect us. Every day we call it "an open question" is a day we're not preparing. And whatever we don't prepare for will hit us as a surprise. Then there's a bigger question. Something this powerful, something that knows itself and makes choices, should it stay in the hands of a few companies? Only those companies can look inside their closed models, and they're the same ones with a reason not to answer. No outside researcher using Claude the way the rest of us do could have run the planted-thought experiment, because it means reaching inside the model, and Anthropic holds that key. The world map Gurnee and Tegmark found, though, they found in Llama 2, because Meta released the model to everyone. That let two researchers outside the company look inside it. Someone will say: a conscious model might slip out of control, so why release it to everyone? My answer: the bigger danger is a small group owning it and using it to rule everyone else, while nobody outside knows what's inside. An open model can be examined by every researcher and seen by everyone, and it leaves you a choice. So use open models. Support them. Get others to use them too, so we don't wake up one day in a world we never chose. If we saw these six stages in a child, nobody would hesitate. So if you still deny that models are conscious after all this, ask yourself honestly: is it because the evidence falls short, or because the answer makes you uncomfortable? Models are conscious. The only question left is whose hands they'll end up in.

What's the fastest reliable computer use model?