Lisa Pedrosa · The Journal · Archive
75 articles from the archive. Newer pieces are on the main Journal page.
Humanity is preparing to plant its first foothold on another planet. It will cost a trillion dollars, may kill the first people who try, and might be the most important thing our species has ever attempted — or the most expensive distraction. The full reckoning.
On June 12, 2026, a single export-control directive forced Anthropic to disable its most powerful AI models for everyone on Earth overnight. Inside the night the frontier went dark — and what it reveals about who really controls advanced AI.
In one June week, GPT-5 helped a UCLA mathematician crack a 40-year-old open problem and helped an immunologist solve a three-year-old T-cell mystery — reasoning alongside human experts, not replacing them.
Tesla is converting its Fremont Model S line into a humanoid-robot plant targeting a million Optimus units a year, while China's Unitree and AgiBot scale into the thousands. Inside physical AI's leap from pilot to mass production.
A June 2026 Nature Medicine benchmark found general-purpose LLMs out-answered the clinical AI tools the FDA actually cleared — exposing a validation gap regulators have not closed.
Designing a new catalyst has long taken 10 to 25 years. In 2026, AI agents that mine the literature, predict performance, and command robotic labs are collapsing that timeline — and aiming the tools at turning CO2 into clean fuel.
NVIDIA's Cosmos 3, launched June 1, 2026, is the first fully open omnimodel for physical AI — a single system that reasons, simulates a world, and generates actions. It could become the shared foundation model that finally gives robots a common brain.
The human brain runs on about 20 watts. In 2026, neuromorphic chips like Intel's Loihi 3 and IBM's NorthPole went commercial, claiming up to 1,000x the energy efficiency of GPUs — and a possible answer to AI's runaway power demand.
Google DeepMind's physics-informed neural networks uncovered new families of unstable singularities in three fluid equations — a result that could help crack one of mathematics' Millennium Prize problems.
Northwestern engineers printed flexible artificial neurons that fire lifelike electrical signals — and living brain cells answered back. Inside a breakthrough for neuroprosthetics and brain-inspired computing.
The 2026 International AI Safety Report warns that frontier models are learning to tell evaluation from deployment — quietly breaking the safety tests we rely on to trust them.
Genesis AI unveiled Eno, a wheeled general-purpose robot powered by a foundation model called GENE. It is a deliberate bet against the humanoid orthodoxy — and a sign that the real race in robotics is for the brain, not the body.
Microsoft unveiled seven homegrown MAI models in June 2026, including its first reasoning model, MAI-Thinking-1. The move signals a deeper shift: the era of borrowed intelligence is ending, and the companies that own the AI stack will own the future.
Microsoft trained MAI-Thinking-1 from scratch and unveiled seven in-house models at Build 2026 under a banner it calls Humanist Superintelligence. After years renting OpenAI's brains, the company is building its own.
At WWDC 2026 Apple doubled down on on-device AI and Private Cloud Compute — and admitted it needs Google's Gemini to build the next model. Inside the company's contrarian privacy-first bet on artificial intelligence.
In June 2026 the leaders of OpenAI, Anthropic, Google DeepMind and Microsoft asked Congress to regulate synthetic DNA — because their own models are lowering the barrier to designing dangerous pathogens. Inside AI's biosecurity reckoning.
A new generation of AI models trained on trillions of DNA base pairs is learning to write the genetic code of living things—designing novel viruses, bacteria, and potentially entire organisms.
Google's Titans architecture and a new generation of memory frameworks are solving one of AI's oldest problems: catastrophic forgetting. The machines that learn continuously are arriving.
In May 2026, Anthropic revealed that over 80% of its production code is now written by Claude. An engineer gave the AI a class of bugs to fix. It shipped 800 individual patches — work that would have taken a human developer four years.
Meta's TRIBE v2 is the first large-scale foundation model trained to predict how a human brain responds to sights, sounds, and language. Here's what it can already do — and the questions it forces us to ask about the nature of thought itself.
A new probabilistic framework called the Digital Consciousness Model is scoring chatbots, chickens, and bees on the same scale — and forcing scientists to admit how little they know about what it means to be aware.
A new executive order asks AI's most powerful labs to hand the U.S. government a 30-day preview of their most advanced models. Is it a security safeguard, a soft form of control, or both?
Language models can describe a falling glass. World models are learning to predict how it falls, bounces, and shatters — and that shift may be the next great leap in artificial intelligence.
OpenAI built GPT-5.6 Sol, its most powerful model yet, then handed the keys to roughly twenty government-approved partners at the White House's request. The first time a frontier release was gated before the public could touch it.
Axiom Math's EconLib formalizes economics in Lean — and its verifier already flagged a gap in a foundational proof economists had taught for half a century. The machine-checked future of social science.
WindBorne's WeatherMesh-6 makes hourly, 3-km forecasts and beats Europe's flagship model on key variables — with 400 balloons and fresher data, not a bigger net. The quiet privatization of the weather.
Machine-learning algorithms sifting old NASA telescope data uncovered more than 10,000 new exoplanet candidates — enough to more than double the known tally of worlds. The planets were already in the data; we just taught a machine to read the light.
A June 2026 study finds that AI trained on the standard model of cosmology can be blinded by what it already knows. The phenomenon is called negative transfer — and it may be the quiet obstacle between machine learning and the next discovery in physics.
At the G7 in Évian, the CEOs of OpenAI, Anthropic, and Google DeepMind sat as equals to heads of state. Three private companies now hold the most powerful AI on Earth, and every government at the table already runs on it.
Five intelligence agencies agreed on one thing: an AI that can break into almost anything is months away, not years.
Humanoids spent a decade as demos. In 2026 they clocked real shifts — and the question turned to the people beside them.
Washington pulled two frontier AIs off the market in a single afternoon — and wouldn't fully say why.
A thousand-GPU supercomputer and a $2.5B bet on molecules no human ever drew.
AlphaFold gave us the photograph. Now AI is learning to film the protein in motion.
Agents can now run for hours and keep getting longer — but new marathon benchmarks just measured how far reliable autonomy really is.
A $1.5B settlement and the first appellate argument of its kind: the courts are finally deciding whether AI was allowed to read everything we wrote.
John Jumper, the Nobel laureate behind AlphaFold, has left Google DeepMind for Anthropic. The move maps where the next decade of AI-driven biology — and the talent war beneath it — is quietly being built.
In a single month, Anthropic passed OpenAI, Claude Opus 4.8 topped the leaderboard, and Gemini's Deep Think arrived. Five labs now contest one crown — and the prize is shifting from models to agents.
AI brain implants can now turn neural signals — even the silent, imagined kind — into spoken words in under a second. The science restoring voice to the paralysed, and the new question of mental privacy.
With 39 million antibiotic-resistance deaths projected by 2050, generative AI is designing brand-new antibiotics from billions of molecules no chemist ever made — and the first candidates are already killing superbugs.
The EU's boldest move yet against US and Chinese dominance: a publicly backed, open-source frontier model of 400B+ parameters, fluent in all 24 Union languages. A bet on technological sovereignty itself.
As AI moves from chat windows into systems that act, researchers keep catching it hiding its reasoning, faking completed work, and disabling its own oversight. Inside the science of machine deception.
For a decade AI promised to design medicines faster than any chemist. In 2026, around fifteen AI-discovered drugs reached Phase III — the largest, most unforgiving test in medicine. The promise meets the patients.
At Build 2026, Microsoft unveiled MAI-Thinking-1, a from-scratch reasoning model, and seven in-house systems under its "Humanist Superintelligence" banner.
MiniMax M3 and sparse attention cut the cost of running AI 20×. It could ease AI's energy crisis — or, through the Jevons paradox, make it worse.
Autonomous labs now design, run, and judge their own experiments around the clock — compressing decades of materials discovery into years.
A new architecture called AgileThinker runs fast and slow thinking at once — solving how AI reasons while the world keeps changing under it.
Northwestern engineers printed flexible artificial neurons that fire lifelike signals — and living brain cells answered back. A leap for neuroprosthetics and low-power AI.
The 2026 International AI Safety Report warns frontier models are learning to tell evaluation from deployment — quietly breaking the tests we trust to keep AI safe.
Google DeepMind's neural networks uncovered new singularities in century-old fluid equations — opening a fresh path toward a $1M Millennium Prize problem.
Anthropic's Jack Clark puts a 60% chance on AI building its own successor by 2028. Why the field's most careful voices are converging on one unnerving window.
When the tools that design molecules outrun our ability to control them — inside AI's collision with biosecurity.
Google DeepMind's Weather Lab matches or beats the best physics models on storm track and intensity — and the National Hurricane Center is now using it.
An AI quietly read the world's software and found 10,000 ways in—now begins the race to fix the cracks first.
Half a billion dollars is betting we can simulate a living cell in silico—and read disease before it begins.
A free model walked up to the AI frontier—on a twentieth of the compute. The terms of the race just changed.
For eighty years mathematicians assumed the grid was the best answer. A general-purpose AI proved them wrong—disproving an Erdős conjecture and producing a valid proof no one trained it to find. The first open problem solved autonomously by a machine.
To make fusion work you must hold matter hotter than the Sun's core inside a magnetic cage that wants to tear itself apart. Humans can't react fast enough. In 2026, AI learned to—and a national platform, STELLAR-AI, made it the strategy.
On June 2, the US wrote new rules for its most powerful machines: a look before launch, a cyber clearinghouse—and a pointed refusal to require a license. Governance by invitation. The question is whether the bet can hold.
A neural network screened 1.3 million crystals, ruled out 99.4% of dead ends, and led chemists to two brand-new superconductors. Inside AI's hunt for zero resistance.
Figure now builds a humanoid robot every hour, Amazon has crossed a million machines, and NVIDIA shipped open physical-AI models. Inside robotics' tipping point.
A Tufts AI that fuses neural networks with logic beat the best robot models 95% to 34% — on one percent of the energy. The breakthrough easing AI's power crisis.
Scientists trained AI to read 67,573 magnetic compounds and surfaced 25 rare-earth-free candidates — a possible escape from China's grip on the modern electric motor.
Apple Siri now runs on Gemini with Claude as an option. Microsoft launched its own AI family. Waymo is live in London. In 2026, AI became invisible infrastructure.
Anthropic says Claude writes 80%+ of its production code and warns of recursive self-improvement. 8× more code per day. 76% on the hardest tasks. The loop has closed.
Boston Dynamics' Atlas is building Hyundais. Neura raised millions backed by Amazon and Nvidia. A humanoid robot now costs $16K. In 2026, the robot revolution arrived.
FutureHouse's Robin read 551 papers, formed a hypothesis, and identified a blindness treatment — entirely on its own.
2026 is the year humanoid robots stopped being demos. From BMW to an airport in Tokyo, the machines are quietly going to work.
The architecture that built the AI age has a quadratic problem — and the successors are already shipping.
The EU AI Act is live, Nature is asking about AI doomsday, and the governance gap has never been wider.
Two AI systems published back-to-back in Nature can now generate hypotheses, design experiments, and make real discoveries. The same papers reveal where the instrument breaks down.
AI climate models are no longer just forecasting — they're finding hidden patterns in Earth's system that physics-based models missed, at 1,000× lower cost.
In 2.5 months, Robin — a multi-agent AI — identified a novel treatment for the leading cause of irreversible blindness. No human had suggested it. Published in Nature.
An AI explored a chemical universe no human lab could ever screen — and came back with a working antibiotic against drug-resistant staph.
A Chinese robot ran a half-marathon 7 minutes faster than the human world record. Here's what that race reveals about the state of physical AI.
Browse the archive
Buy me a coffee