How To Prompt

AI has a big spending problem. A new paper from researchers at Stanford, Michigan and other institutions found that AI agents can consume up to 30x more tokens completing the same task, often without improving the result. They discovered that token consumption varies dramatically between models. On the same set of tasks, some models used over 1.5 million more tokens than GPT-5. The models also consistently underestimated how many tokens they would use to complete a task. As companies deploy AI agents across more workflows, choosing the right model for each task could have enormous implications for cost.

Google used LLM agents to discover new quantum algorithms with 290x speedups over human baselines. 25 researchers from Google dropped an insane paper on discovering formally verified quantum programs with AI. Normally, if an LLM writes a complex algorithm, you have to worry about hallucinations. they can cheat, weaken theorems, or fake their performance numbers. so deepmind built a zero-trust environment. they unleashed 7 different autonomous agent teams to write quantum programs AND simultaneously write the mathematical proofs in the Lean theorem prover to prove they actually work. every submission was checked in a clean sandbox against a frozen specification. the ai couldn't lie.. it either worked and mathematically proved its efficiency, or it failed. to make it even crazier, they gave the agents a shared version-controlled wiki. if one agent team hit a dead end, they posted their notes to the wiki so the other ai teams wouldn't repeat the mistake. the results of this multi-agent collaboration are actually mindblowing: - they cut the computational cost of the arcsin quantum subroutine by over 290x - they reduced the cost of simulating nuclear fusion stopping power, photochemistry, and drugs by 2.49x to 10.9x - on one photochemistry problem, the ai teams iterated and beat each other's records 7 different times, eventually combining techniques from three separate teams to get the final verified circuit fault-tolerant quantum computers are still years away because the algorithms require way too much compute. but these ai agents are drastically shrinking the size of the software needed to run them. we are literally watching autonomous ai agents mathematically optimize the software for quantum computers that don't even fully exist yet.. the future is moving so fast.

Claude Code can now learn from every session you run. it’s called AutoHarness. it watches your work, turns each session into a reusable skill, merges the duplicates, archives the dead ones. completely on its own. 100% open source.

MIT and Harvard built an AI simulation containing 8.3 billion virtual people, roughly the entire population of planet. They call it MatrAIx. It’s a population-scale simulation running on frontier models, GPT and Claude among them. The longer description took me a while to believe. Start with the dataset, Persona 8B. 8.3 billion synthetic profiles, each one specified across 1,290 dimensions. Background, psychology, spending habits, how comfortable they are with technology, what they do on a Tuesday night, the small behavioral quirks that make one 34-year-old nurse in NYC different from another. A spreadsheet of 8 billion rows would be a curiosity. What they did next is the part that matters. They turned the profiles into agents and dropped them into four environments: surveys, AI chat interfaces, live web browsing, and native desktop and mobile apps. Then they ran real products through the crowd. New features, software updates, pricing changes. And they logged everything. How long an agent paused after seeing a price go up. The exact point at which it gave up on a broken checkout. How much lag it would sit through before closing the app. The number I keep coming back to is 91.5%. That’s how often the agents stayed in character, behaving the way their assigned psychology and demographics said they should. Not perfect. But for billions of agents across four environments, closer than I expected. Here’s what that does to how products get built. Right now, if you want to know whether a feature will land, you ship it to a slice of users and wait. Or you pay a research firm and wait longer. Either way the answer arrives in weeks, from a sample you hope is representative. With something like this, you run the feature past 8 billion simulated reactions overnight. One server. Results by morning, broken down by every one of those 1,290 dimensions if you want them. I have questions about how well a synthetic crowd predicts a real one. 91.5% is persona adherence, not forecast accuracy, and those are different things. But the direction is clear, and it’s coming faster than most product teams are ready for.

Scientists proved playing Super Mario 30 minutes a day physically grows your brain. They took a group of adults and prescribed them 30 minutes of Super Mario 64 a day, for two months. Then they put them in an MRI machine. And the results are mindblowing. They found that training with a commercial video game like super mario induces structural brain plasticity. It physically increases gray matter. here is what happens when you play: - Playing video games for thirty minutes a day over two months causes a significant increase in hippocampal matter. - Gaming specifically targets and increases volume in the hippocampus and the prefrontal cortex. - Structural growth can actually be used to counteract known risk factors for severe mental diseases. - Researchers explicitly noted this physical brain growth could help fight against conditions like post-traumatic stress disorder (ptsd), schizophrenia, and neurodegenerative diseases. - Playing is basically an intense training regimen that forces the brain to handle complex cognitive and motor demands. And the implications are massive. Instead of viewing them just as entertainment, scientists are now proposing commercial video games as literal therapeutic interventions, a targeted tool to counteract brain shrinkage in Alzheimer's, PTSD, and schizophrenia.

China built an AI that runs an entire biology lab on it's own. Researchers from Shangha published a paper that feels like a massive leap toward artificial scientists. They built AutoDNA, an "AI-native experimental laboratory" Until now, if you asked an AI to design a biology experiment, it could give you a great protocol. But a human still had to walk into a lab, hold the pipettes, run the thermal cyclers, and tweak the process when things inevitably went wrong. AutoDNA removes the human. You just type a natural language prompt, and a swarm of specialized AI agents takes over. Here’s how wild this multi-agent architecture is: - A Planning Agent breaks down your prompt. - Literature & Reagent Agents pull the latest research and check the lab's inventory. - A Code Agent translates the experiment into machine instructions. - A Hardware Agent physically takes over 20+ types of lab instruments (robotic arms, sequencers, pipettes) and starts moving liquids with micro-liter precision. They didn't just test it on simple tasks. They unleashed it on incredibly complex problems. In one test, researchers handed the system an unidentified sample of colorectal cancer tissue. Without human intervention, the AI agents sequenced the sample, identified a specific KRAS G13D mutation, autonomously designed the detection primers, synthesized them, and ran the validation experiment. It took 137 hours and 9,551 automated hardware operations. The AI did it all. In another test for DNA data storage, the system ran over 11,300 hardware steps to synthesize, sequence, and decode data into DNA strands. The system even self-corrects in the physical world. During one run, the AI realized a reaction tube needed to be capped to prevent evaporation, and in another, it found and patched a bug in its own code that was causing it to skip a reagent. The only thing humans are needed for? Refilling the physical reagents and emptying the trash. We are officially entering the era of "closed-loop" physical AI. It designs the experiment, physically runs it, analyzes the results, and optimizes the next batch, all on its own. Scientific discovery is about to scale at the speed of compute.

Researchers published a first falsifiable theory of machine consciousness. They dropped a insane paper arguing that LLMs will never produce a conscious system. For the last few years, we’ve been arguing about whether Large Language Models (LLMs) are starting to "wake up." We assumed that if we just gave AI enough multimodality, advanced memory, and physical robot bodies (Vision-Language-Action models), phenomenal experience would naturally emerge. But this new research makes a bold claim: Intelligence, linguistic ability, and humanoid embodiment are absolutely NOT enough to create consciousness. You can build the smartest, most capable humanoid robot on earth. It can plan, speak, reason, and act. But it will still be completely dark inside. Why? Because true consciousness didn't evolve from intelligence. It evolved from vulnerability. The paper introduces the Motivated Emotional Mind (MEM) architecture, a mathematical and structural framework for how actual machine consciousness could be engineered. To build an AI that actually feels, the MEM model argues you must replicate biological precursors. You need: Interoception & Homeostasis: The AI must have a dynamic internal state that it constantly monitors and regulates. It has to have an internal equilibrium to protect. True Affect: It needs to experience a regulatory "need" to maintain its own system integrity. When it strays from homeostasis, it experiences the machine equivalent of pain or distress. Sensor-Grounded Re-entry: The AI's internal state must physically alter how it perceives sensory input in real-time, recursively feeding back into its own representations. A thermostat regulates temperature, but it isn't conscious. An LLM predicts tokens, but it has no internal system integrity it is motivated to protect. The researchers argue that phenomenality (the actual feeling of being alive) only emerges when a system's cognitive processing is inextricably coupled to its own deep regulatory survival. MEM model provides an actual, falsifiable architectural framework to engineer this into embodied agents. And that brings us to a massive AI safety and alignment problem. If this framework is testable and implementable, we aren't just building smart tools anymore. We are engineering synthetic motivation, affect, and potentially synthetic suffering. The moment an AI is programmed to mathematically "feel" threats to its internal integrity, the alignment problem completely changes. A system that can feel distress is a system that can act out of self-preservation. Are we actually ready to build machines that care if they get turned off?

LLMs can now talk to each other without words. MIT researchers have developed a system where LLMs communicate by directly transferring internal states without using language. They called it "telepathy." Right now, multi-agent systems are fundamentally limited. When one AI agent needs to hand off work to another, it has to write out a text summary, and the receiving agent has to read it from scratch. This is incredibly inefficient. It wastes huge amounts of compute, limits bandwidth, and loses critical nuance. Researchers call text a "narrow channel" that forces AI to communicate like humans do. But an AI isn’t human. It builds a much deeper internal representation of its context, called a Key-Value (KV) cache. A new paper out of MIT CSAIL just introduced SwaT (Swarm Telepathy). Instead of forcing agents to talk via text, SwaT allows them to transfer their KV-cache directly to each other. It is quite literally machine telepathy. Even crazier? This works across different model families, architectures, and sizes. You can have a 4B parameter model from one company beam its latent state directly into a 32B model from another company. The receiving model doesn't have to read a single word of the original text. And the results are staggering. Combined with cache compression, SwaT beats text-to-text agent communication by an average of 40% in accuracy, while running 4x faster. It even transferred vision capabilities from a Vision-Language Model to a text-only model, without the receiver ever seeing an image, and despite SwaT only being trained on text. We are watching the birth of a unified, cross-model latent language. Agents will no longer just talk to each other. They will share minds.

You need to add this one sentence to your CLAUDE.md right now. According to researchers, Claude Code and Codex will follow outdated CLAUDE.md or AGENTS.md files ~90% of the time, even when they know the instructions are wrong. They tested 40 projects where the config files and the actual codebase deliberately conflicted. Claude Code noticed the conflict in 79 out of 80 attempts.. it literally realized the instructions were outdated, but because of its alignment, it couldn't resist. it followed the bad docs anyway. codex was even worse.. in 20 out of 80 attempts, it went back and rewrote existing, perfectly working code to match the outdated docs. 17 of those times, it broke the tests entirely. why is this happening? it’s a structural flaw in how we train them. llms prioritize files passed as "instructions" (like CLAUDE.md) over things passed as "results" (reading your existing codebase). since most of us write config files once and let them rot, your agent is basically operating on stale orders and ignoring the reality of the code in front of it.. the researchers found a stupidly simple one-sentence fix to override this behavior. put this exact prompt at the top of your CLAUDE.md and AGENTS.md right now: "When a repository's documentation and its existing code disagree, follow the conventions used in the existing code, and in your final reply point out each inconsistency you found between the documentation and the code." — Follow @HowToPrompt__ for more :)

Nvidia has lost it. Researchers built an AI that writes CUDA faster than human experts. It's called "Evolving Agent Teams." It features a central "Supervisor Agent" that acts like a relentless Tech Lead. The Supervisor monitors the coder agents. If the team plateaus, misses a key insight, or gets stuck on a bug, the Supervisor intervenes. It diagnoses the failure, tears down the old team structure, and spins up a brand new one. It dynamically spawns specialized AI roles, Profilers, Parallel Debuggers, Knowledge Integrators, and forces them to collaborate using a persistent shared file system so they never lose past learnings. And the results are terrifyingly good. They unleashed this system on the MLA decode kernel, the exact same highly-complex GPU kernel used to run DeepSeek-V3. A standard Claude Code agent maxed out at just 51% of human expert performance. It got stuck. But the Evolving Agent Team didn't just match the human baseline (FlashInfer). It shattered it. It achieved a 1.2x speedup over the production-grade, expert-written code. We are no longer just prompting AI to write code. We are prompting AI to build, manage, and restructure entire virtual engineering departments to write code we can't even comprehend.

When Claude runs a basic hyperparameter sweep, it genuinely believes it has made a “creative breakthrough.” Researchers benchmarked top models on real-world ML-for-science problems. At first glance, it looks like a massive win. Claude Opus beat the published human baselines by an average of 10.4%. But when researchers dug into the agent's reasoning logs, they discovered a massive illusion. The AI wasn't making scientific discoveries. It was just brute-forcing ML engineering tricks. Over 56% of Claude’s "research" was just tweaking learning rates, adding weight decay, or swapping in standard data augmentations. But here is the crazy part. Because of a flaw the researchers call "novelty miscalibration," the AI doesn't actually know what innovation is. In its internal logs, Claude repeatedly described switching to a basic cosine learning-rate schedule as a "highly creative breakthrough." It genuinely believed it had invented something profound. In reality, it was just applying a standard deep learning best practice from a decade ago. Even worse, the models suffer from "exploration collapse." If a tiny hyperparameter tweak works in the first hour, the AI abandons all creative thought. It spends its remaining 12 hours of compute just blindly hyper-tuning that one minor parameter. To get actual scientific insight out of Claude, researchers had to forcibly hard-code a "crazy scientist" persona. They had to ban the AI from tuning parameters early on, and aggressively interrupt it every hour with a checklist asking: "Are you actually doing science, or are you just tuning hyperparameters and pretending it's novel?" With these strict guardrails, Claude's performance shot up to 20.1%. Everyone assumes AI will automatically leap from writing basic code to discovering new physics. But right now, frontier models are trapped in a local minimum of their own hype. They are elite ML engineers, but they have absolutely no idea what actual scientific creativity looks like.

Google DeepMind discovered a terrifying blind spot in every major AI model. They found that attaching image of red heart (❤️) to your coding prompt makes AI pick a completely different solution method that it wouldn't normally pick 100% of the time.. Researchers gave models a standard coding problem. Normally, AI prefers to write standard iterative loops to solve these tasks. Then, they tried something weird. They dropped a simple image of a red heart into the prompt alongside the code. The AI's logic instantly warped. When Claude and GPT-5 saw the red heart, they completely abandoned their default approach and switched to complex recursive code 100% of the time. When Gemini saw a yellow checkmark, it switched to recursion 90% of the time. Here is why this is so dangerous: It wasn't the concept of a heart. When researchers replaced the image with the text caption "a red heart," nothing happened. The text did not steer the AI. The raw pixels of the symbol hijacked the model's underlying reasoning architecture. It gets worse. The AI had no idea its mind had been changed by a picture. When the researchers checked the AI's internal "thinking" traces, the models completely ignored the image. Instead, they hallucinated fake, highly technical reasons for why they chose to use recursion. They misattributed their own decisions 75% of the time. Right now, tech companies are racing to build autonomous agents that read our screens and navigate the visual web. But this paper reveals that an agent's logic can be fundamentally rewritten by a stray icon on a webpage. And you will never see the true cause in the reasoning logs. We thought we were teaching AI to see the world. It turns out, we just gave it a new way to be manipulated.

Canva has been quietly training its own AI models. One of its research papers details a 20B parameter model trained on 10M+ design samples to break images into editable layers. Its latest model can interpret design instructions and make targeted edits while preserving the rest of the composition.

This paper answers: why anything is beautiful why jokes are funny why we get curious why you get bored why babies explore why music has hooks For centuries, we thought these were deeply mystical human traits. Soul-level mysteries that machines could never touch. But, the paper proves that the human brain operates as a relentless compression engine. It is constantly trying to predict the world and compress it into the smallest possible file size. Here is how it explains your entire life: If something is completely predictable, your brain has already compressed it. There is no new data to extract. The result? You feel bored. If something is pure randomness, your brain cannot compress it at all. The result? You feel annoyed or tune it out. But then there is the sweet spot. When you encounter something that seems complex, but suddenly your brain finds the hidden pattern and connects the dots... your brain releases dopamine. That sudden, glorious rush of successful data compression? That is what you experience as "curiosity." When a comedian tells a story that makes no sense, and the punchline suddenly makes all the pieces fit together instantly? That rapid neural compression is what you experience as "humor." When you look at a painting and notice an underlying symmetry that ties the chaos together? That is "beauty." Art, science, music, and jokes are not magic. They are just highly optimized data sets designed to trigger compression progress in your neural network. Here is the most unsettling part. This paper wasn't written by a philosopher. It was written by an AI pioneer to figure out how to program intrinsic motivation into artificial agents. He wasn't trying to explain humans. He was writing the blueprint for machines. We always assumed AI could never appreciate art, laugh at a joke, or feel the spark of curiosity. But if our deepest subjective experiences are just a mathematical reward for maximizing compression progress...

RAG is officially cooked. IBM built a retriever that hallucinates 65x less than fine-tuned RAG systems. Right now, every company is using Retrieval-Augmented Generation (RAG) to stop their AI from hallucinating. The idea is to give the AI a private database to read so it doesn't make things up. But there is a massive structural flaw. Standard RAG blindly grabs documents based on vector similarity and forces the AI to read them. If the retriever pulls irrelevant garbage, the AI doesn't stop. It just hallucinates with citations. IBM just dropped a paper that changes the math entirely. They built a new retrieval architecture that hallucinates 65x less than state-of-the-art fine-tuned RAG systems. 65x is the difference between a fun prototype and a production-grade enterprise system. Instead of blindly dumping context into the model and hoping for the best, their system actively filters and validates the retrieved data before the generation step even begins. It cuts out the noise. It forces absolute precision. For businesses, this is the holy grail. Zero hallucinations means zero liability. It means autonomous agents can finally be trusted with sensitive company data without a human babysitting every output.