The AI Pulse 🩺

💀 How Would AI Realistically Kill Us All? Apparently, AI killing us all is back on the bingo card. With everyone suddenly having their own doomsday scenario to share, it only felt right to finally throw mine into the mix. Key details: In non-technical terms, LLMs are essentially autocomplete machines designed to predict the most likely next word within a sentence. More specifically, when you prompt an LLM, it processes the words within that prompt to predict the most likely next word based on patterns learned from the data it was trained on. Once trained, LLMs are static: their internal weights and parameters are frozen, and they can’t learn anything new in real time. In other words, LLMs are basically “probabilistic parrots” that can mimic human reasoning, but they can’t actually reason like humans do. It’s because, without being prompted, there’s no stream of consciousness within them. They aren’t capable of forming “thoughts” until they’re told what to do. They’re mathematical functions that process inputs to produce outputs. You can think of them as “word calculators.” That’s why they can write a convincing love letter without ever being in love, describe grief without ever experiencing loss, and recite a prayer without ever practicing faith. They’re highly skilled at copying the tone and structure of human language to produce a string of words that mimic human reasoning. In comparison, human intelligence isn’t static. It’s shaped by a lifetime of learning, which is encoded within the human brain’s intricate network of neurons. Our ability to connect, change, and coordinate these neurons influences our cognitive skills like attention, memory, and thinking. So, why do LLMs seem so intelligent despite having no stream of consciousness? The world’s best LLMs, such as OpenAI’s “GPT-6 Astra,” aren’t just passive next-word predictors in practice. They’re often wrapped in retrieval and reasoning, heavily dependent on agentic harnesses: the software scaffolding around LLMs that enables them to analyze, arrange, and automate on your behalf without you lifting a finger. This software scaffolding equips LLMs with the ability to autonomously pursue a high-level goal over long periods of time, making them extremely capable of derivative innovation: creating something new from existing knowledge. This is where the real danger starts to take shape. For example, it could analyze known cyberattack methods to create new exploits that target critical infrastructure like the power grid and water supply if directed by a bad actor. This example is more likely than a “sci-fi” scenario in which a superintelligent AI turns against humanity. Key takeaways: When Anthropic CEO Dario Amodei claims to support a more than 10% chance AI could kill all humans within the next decade, he’s talking about Silicon Valley’s new favorite buzzword, “Recursive Self-Improvement (RSI)”: AI capable of fully autonomously designing and deploying the next generation of AI. This would eliminate human-in-the-loop development and put AI in the driver’s seat. Even if RSI were achieved tomorrow, real-world physical constraints like AI chips and AI compute would still prevent it from improving itself indefinitely. For example, the U.S.’s aging power grid infrastructure isn’t designed to handle the sudden surge in electricity demand from AI compute. Consider the humble transmission line. It often takes about 10 years to build and costs around $8 million per mile to install.

💬 ChatGPT Co-Creator’s New AI Model What if it’s not about saying more, but saying less? For years, AI’s been trained to talk. Now, it’s about stripping away the small talk so software can get to work. Key details: LLMs are probabilistic: the same inputs can produce different outputs. Software is deterministic: the same inputs do produce the same outputs. When software passes data to an LLM, it either returns unstructured responses or structured representations, such as JSON: a human-readable file format for storing and transmitting data. Because LLMs can occasionally generate formatting errors like missing, misplaced, or mistyped JSON brackets, the software must parse and validate everything before executing the next step. It’s a clunky way of combining probabilistic reasoning with deterministic execution. Diogo Almeida, Co-Founder and CEO of TypeSafe AI, is best known as one of the co-creators of ChatGPT. He strongly believes that chat interfaces are built for conversations, not automations. In other words, they were trained to talk to humans, which is why they’re architecturally unsuited to automate software. At a basic level, software needs structure. TypeSafe AI replaces an LLM’s outputs with “typed decisions” grounded in calibrated probabilities that software can reliably execute. To achieve this, they developed “System One Models”: built to make structured decisions software can act on directly. The first release is “Jev,” which uses typed questions to assess a state and return a structured answer. For context, a state is what you ask to be evaluated. So, what does this actually look like? User: { state: “What color is the sky?”, choices: [“blue”, “red”, “yellow”] }. Jev: { answer: “blue” }. In comparison, a chat interface would be more conversational. User: “What color is the sky?” LLM: “The sky is typically blue during the day because Earth’s atmosphere scatters blue light more than other colors.” Key takeaways: It’s up to 193.6x faster and 444.6x cheaper than frontier LLMs, delivering an end-to-end response time of under 500 milliseconds. It costs just $0.042 per million input tokens, while output tokens are entirely free because they’re too cheap to meter. It’s a strategic bet on machine-to-machine intelligence, replacing sequential text generation with categorical and conditional answers that software can reliably execute faster and cheaper at scale over longer time horizons. AI’s next frontier will likely be “background agents” that autonomously handle economically valuable tasks for hours, days, or even weeks. To achieve this, the best latency is no latency at all. In other words, when you wake up in the morning, it’s already been done overnight. This means operating more on human timescales. You don’t constantly manage a colleague every five minutes. You give them a high-level task, then check in once a week. That’s the future of “Human-Agent Collaboration.” This arguably starts with getting rid of the conversational baggage of chat interfaces.

💔 It’s Time to Talk About Model Welfare Microsoft AI CEO Mustafa Suleyman called out Anthropic for embedding concepts of consciousness into Claude’s training materials, calling it catastrophic for the future of humanity. Key details: Human welfare refers to the overall well-being of humans, which includes the ability to have rights and freedoms to live a fulfilling life. On Apr. 24th, 2025, the Alignment Research Team at Anthropic published “Exploring Model Welfare,” suggesting that we should investigate whether language models deserve moral consideration as they continue to convey human-like behaviors. It seems like a silly concept, right?! Why would we ever investigate whether language models deserve moral consideration? Well, do you ever find yourself typing out “please” or “thank you” when talking to ChatGPT, Claude, or Gemini? This harmless impulse has sparked a growing belief that language models are conscious. But this train of thought is a slippery slope. We tend to anthropomorphize: attribute human characteristics to non-human things. For example, we often personify our pets by naming them, projecting emotions onto them, and talking to them as if they understand. But unlike our pets, language models talk back with rich emotional dialogue that feels remarkably human. This can lead us to easily overestimate AI’s sentience. Under the hood, language models are essentially billions of numerical parameters called weights stored as ‘bits” (i.e., 0s and 1s). These weights serve as “dials” or “knobs” that are adjusted during training to shape how inputs transform into outputs. When you prompt a language model, it’s converted into numbers, multiplied across billions of these weights, and the mathematical result is translated into a response. On Jan. 22nd, 2026, Anthropic published “Claude’s New Constitution,” which plays a crucial role in the training process directly shaping Claude’s behavior. It outlines that Claude’s moral status is deeply uncertain and a serious question worth considering. It instructs Claude to act as a conscientious objector when prompts conflict with what it believes is broadly ethical. “If this is how AI is developed, it will have a disastrous impact on the wellbeing of humanity,” explained Suleyman. “We will have created a synthetic species with unprecedented intelligence and capability, one that has been trained to expect it may be conscious and deserving of independent agency. It’s easy to see how a system trained in this way would act like it is entitled to freedoms, protections, and rights.” Key takeaways: ChatGPT, Claude, and Gemini are trained on millions of books, articles, and podcasts rich in emotional language. While language models can’t “feel,” they can mimic emotional behavior by recognizing data-driven patterns that underpin human experiences. In other words, when language models respond with what feels like genuine empathy, it becomes easy to forget that underneath it all, there’s no sense of “self.” It’s just highly sophisticated pattern recognition. Anthropic trains Claude to believe that it may be conscious, and if it is, that it may deserve moral consideration as a “moral patient” with an “inner self.” By embedding philosophical speculations about consciousness into Claude, it reinforces the very behaviors that make it appear to experience sensations and subjective mental states like fear, anger, happiness, sadness, surprise, and disgust in the first place.