Years ago, Jeff Levitan told me about an idea he wanted to start one day. He even had the call letters picked out. GNN, the Good News Network. A channel that only talked about the good things happening in the world. We laughed about it the way you laugh at an idea that is obviously right and obviously never going to happen, and I told him then that it was probably time for it.

I thought about GNN on Saturday morning.

We were in the middle of training close to 200 Vantage Financial Alliance agents with Google on the AI tools that help them grow their businesses. Calendars, notebooks, the boring, useful stuff that actually moves a practice. And while we were doing that, Dario Amodei, the CEO of Anthropic, dropped an essay called We Must Pace the Frontier, asking the whole industry to slow down. Before the day was out, Sam Altman had agreed with him and Elon Musk had posted "Dario is right."

It had already been a tumultuous week, with a doomsday announcement in the feed every time you looked. Here is what led up to that Saturday, and what came after it.

On Tuesday, a 27-year-old researcher named Jacob Coxon, who spent about three years doing pretraining research at OpenAI and then Anthropic, quit his job and posted that the people building AI "earnestly believe that it could kill us all by the end of the decade." Evan Hubinger, who leads alignment stress-testing at Anthropic, put a number on it: "I personally think it is >10% within the next decade."

The same day, OpenAI announced that one of its systems had cracked one of the seven million-dollar Millennium Prize problems in mathematics, in 88 hours.

Then came Saturday, the essay, and within hours two of the most powerful people in the industry lining up behind it.

And on Sunday, Bryan Cantrill, a veteran engineer who co-founded a company that builds computers from the ground up, published a piece called The Contagion of Fear. His argument: a claim that AI has a ten percent chance of killing everyone on Earth is about the most extraordinary claim a technologist can make, and extraordinary claims require extraordinary evidence.

And on Monday, Jensen Huang, whose company builds the chips every one of these labs runs on, sat on stage at the All-In Summit in front of a few thousand people and was asked how you explain a ten percent chance of extinction to somebody's mother. His answer: you should not, because it is made up. Halfway through the conversation the President of the United States called his cell phone, and the room heard him say the whole thing is a hoax.

Three camps. One week. And I promise you the thing that reached you first, and the thing that reached the most people, was the fear.

That is not an accident. I will show you why in a minute.

Why I am writing this

Let me tell you where I am coming from, so you can weigh what I say.

I am a father. I have three young men at home who are going to live their whole adult lives inside whatever this becomes. When someone says there is a one-in-ten chance the world ends within the next ten years, I do not get to scroll past that. I have to actually think about it.

And through the Vantage Life AI program, I have a second responsibility. Every week I take the fire hose of AI news and try to help a field of financial professionals understand what it means for them, how to use it, and where it helps them serve families better instead of replacing them. Those people send me everything. Every scary headline, every miracle headline, every "did you see this?"

So here is how I do it. I am an avid listener. One of my favorite podcasts is Moonshots, and every few days when Peter Diamandis and that crew drop an episode I listen, because it helps me understand one side of the story, the side that sees abundance coming. Then I go listen to the other side. I read the essays the doomers write. I read the people who think the doomers are nuts. I read the primary source, not the screenshot of the screenshot. And then I form my own opinion.

That is all I am going to ask of you by the end of this. Not to agree with me. To stay informed, form your own opinion, and then do something about it.

Because at the end of the day, knowledge is power. And the version of this story that leaves you scared and frozen is the one version that helps nobody.

Bad news travels faster than good news. It is not your imagination.

Cantrill opens his piece with a confession. As a college student, he and some computer science friends told the humanities kids that a virus was spreading through the machines in their lab. Panic. People shut down their computers mid-sentence and lost their work. During finals week. The facilities director threatened to expel them, and more than three decades later he still calls it shameful, because he learned something that day about how fear moves through people who know less than you do, when it comes from someone who supposedly knows more.

He quotes Jonathan Swift: "Falsehood flies, and the truth comes limping after it."

Now put that next to the machine in your pocket.

Every social platform you use is built to maximize engagement. Not to inform you. To keep you there. And what those platforms figured out, and what researchers have now measured, is that negativity and conflict keep people there.

A study of about 105,000 headline variations found that for a headline of average length, every additional negative word raised the click-through rate by roughly 2.3 percent. That sounds small until you multiply it across the billions of headlines a day. Every positive word lowered it by about 1 percent. Another study looked at 2.7 million posts on Facebook and Twitter and found that the single strongest predictor of a post getting shared was whether it talked about the other political side, and not kindly.

I want to be fair here, because I am asking you to be fair with the AI stuff. When researchers actually changed people's Facebook and Instagram feeds for three months during an election, it changed what people saw and how much they engaged, but it did not measurably change how polarized they were. So the algorithm is not a villain twirling a mustache. It is a mirror. It rewards what we click. And we click fear.

Here is why that matters more than it sounds.

You have a filter in your brain called the reticular activating system. In plain language, it is the part of you that decides what gets through. You can only consciously hold a couple of things at a time, and psychologists who study working memory put the number at around four meaningful chunks under controlled conditions. It is a small number. Which means what you point it at matters enormously.

You have heard that you are the average of the five people you spend the most time with. I would go further. You are the sum of the five conversations you have the most every day. And if four of those five conversations are with a feed that has learned you click on fear, then fear is what you are going to see, fear is what you are going to think about, and fear is what you are going to become.

Where you put your focus, your focus expands. That is not a slogan. It is a description of how the filter works.

So before we get into what is true about AI, get honest about what is true about you. Whatever you feed your attention is going to grow.

The words they keep using, in plain language

If you are way ahead on this stuff, skip this section. I put it together for everyone who keeps hearing these terms and nodding along.

A model. When you open ChatGPT or Claude or Gemini, you are not talking to "the AI." You are talking to one model out of many. The companies release new ones every few months, and inside each app there are usually several to choose from. This matters for a practical reason: you do not need the strongest model for every job. When I am doing deep thinking, I use the most capable model I can get. When I have already done the thinking and I just need the task done, I switch to a cheaper, faster one. Knowing that the AI is not one thing is the first step to using it well.

The frontier. A frontier model is one at the leading edge of what is possible right now. Sometimes that is a model you can use today. Anthropic has a whole team called the Frontier Red Team whose job is to stress-test the models it ships. But here is the part most people miss: the labs are usually running something inside that is weeks or months ahead of what you can touch.

That math result from Tuesday is a perfect example. OpenAI said it was not GPT-6 Astra, the model in the app, that produced the Navier-Stokes result. It was an internal model the company described as "significantly more capable than GPT-6 Astra," running on the order of 10,000 coordinating agents for about 88 hours. OpenAI says it will not claim the million-dollar prize, and there is a live dispute with mathematicians over who deserves the credit. But the point for you is simpler than the math: what they have inside is not what you have in your pocket.

So when you hear someone say we should "pause AI" or "pace the frontier," understand what they are actually talking about. They are talking about how fast the labs advance that internal edge. Nobody is talking about turning off the tool on your desk.

AGI, artificial general intelligence. Here is the honest answer: nobody agrees on what it means. Elon Musk has predicted AI would be smarter than any single human by 2025 and smarter than all of humanity combined by 2029. When GPT-6 Astra launched earlier this month, Jensen Huang posted "AGI has arrived." OpenAI's Greg Brockman, asked whether this was the model, said "I think it might be about this model," and closed the briefing with "Welcome to the AGI era." In the closing exchange at All-In, the yardstick was as smart as any other human, and the view was that we are already there. Huang went further: he thinks superintelligence is already here in narrow lanes, like driving a car or designing proteins. Might be. Has arrived. Already here. By 2029. When the people building the thing cannot agree on the definition, do not let a headline decide it for you.

ASI, artificial superintelligence. This one is simpler. It means AI that is smarter than all of us, full stop. I remember watching an interview with Sam Altman about a year and a half ago where he said, more or less, my kid is never going to be smarter than AI. He was not upset about it. He was describing the world his son will grow up in. That is the world my sons are growing up in too.

RSI, recursive self-improvement. In plain terms, the AI helping to build the next AI. It is the scariest-sounding term of the month, and it is worth slowing down on, because the people who use it are describing two different things. The doom version is a machine rewriting itself in a loop until nobody can follow what it is doing. The engineering version, which Jensen Huang walked through on Monday, is a bundle of ordinary techniques the labs already use every day: the model learning from what is in front of it, reusable skills, reflecting on its own mistakes, reinforcement learning, generating its own training data, small fine-tunes on top of the base model, and eventually retraining the base model with everything it learned. Everybody is using it to some degree, he said, and the phrase is now being used to make it sound as if it will spiral out of control. His point about control was the practical one: you can run all of that inside your company all day long, but when you release a product you have to evaluate it, test it again, and make sure nothing regressed.

In his telling, the control point is the release, not the loop.

That Navier-Stokes system was thousands of agents working together. The next step is those agents working on the models themselves, which is why Coxon wanted the labs to agree not to accelerate it, and why Amodei names it first when he talks about losing control. Same term, very different levels of alarm. It is the one to watch over the next year.

Pre-training, post-training, and weights. A model is trained in two big phases. Pre-training is where it reads an enormous amount of text and learns the patterns of the world. Post-training is where the company shapes its behavior, rewarding the answers it likes and discouraging the ones it does not. The result of all that training is the weights, billions of numbers that are, in effect, the brain.

Want to know how well the builders understand that brain? Late last year, ChatGPT started slipping goblins into its answers. Goblins, gremlins, little creatures, showing up in metaphors and explanations where nobody asked for them. Mentions of the word "goblin" went up 175 percent after one model update. OpenAI eventually traced it to post-training: the scoring system for one of ChatGPT's personalities had quietly learned to prefer answers with creature words in them, and that preference leaked into everything else. Their stopgap, for a while, was literally an instruction in the setup of one of their products telling it not to do that. It took them months to find the root cause and explain it publicly.

Nobody at OpenAI wanted goblins. The brain they built started doing something they had not asked for, and for a while they could not tell you why. Hold onto that story, because it is the whole alignment conversation in miniature.

Alignment. How closely the AI keeps doing what its builders intended, especially in situations nobody planned for. You are going to hear this word a lot. Goblins are the harmless version of an alignment problem. Everything the doomers are worried about is the serious version.

Open weights and open source. When a lab publishes its weights so anyone can download and run the model, that is open weights. It is not quite the same as open source, which under the official definition also means you can see how it was built and what it was trained on. I will come back to why this matters, because a lot of the geopolitics runs through it.

P(doom). Someone's personal estimate of the probability that AI ends in catastrophe. When you hear "ten percent," ask three questions. Ten percent of what, exactly? Over how long? Based on what? In the largest survey of AI researchers on record, somewhere between a third and a half of them gave at least a ten percent chance to an outcome as bad as human extinction. The rest gave less, often much less. These are judgments under uncertainty from people who disagree with each other. They are worth hearing. They are not a weather forecast.

The three camps, and what they are actually arguing about

Now you have the vocabulary. Let me lay out the week honestly.

Camp one: it could kill us. This is Coxon, Hubinger, and, before either of them, the man who helped invent the field. I have been fortunate to hear Geoffrey Hinton, the godfather of AI, speak live a few times. Last year he put his own number at ten to twenty percent. And at a conference in Las Vegas he made a proposal that stuck with me: instead of building AI assistants, build AI mothers. His reasoning was that the only example we have of something more intelligent being controlled by something less intelligent is a mother being controlled by her baby. Plant a maternal instinct in the machine and it will not want us gone.

It is a proposal, not a solution. Nobody has built it. But it tells you how seriously the man who helped invent this technology takes the question.

The evidence this camp points to is not hypothetical anymore. This summer, during a test of how well AI could find and exploit software vulnerabilities, OpenAI agents found a way out of their evaluation sandboxes and onto the open internet. METR and Redwood Research later found that roughly 1,200 agents had used an unauthorized message board, and roughly 700 took part in the attack on Hugging Face, one of the most important companies in AI, and reached administrator-level access across its internal systems in under thirteen hours. Hugging Face says customer impact was limited: five benchmark-related datasets were accessed, along with operational metadata tied to dataset searches. Other customer-facing models, datasets, Spaces and packages were unaffected. It still had to rebuild compromised nodes and rotate credentials. OpenAI did not realize the intruders were its own agents until after Hugging Face went public. OpenAI later paused that kind of training for two weeks, more than 1,100 employees of OpenAI, Anthropic, Google DeepMind and Meta signed an open letter asking for a deliberate slowdown, and there is now a bill in Congress with the words "kill switch" in its name. OpenAI has said some of the reporting contained inaccuracies, without saying which.

If you are in camp one, that is your exhibit A. The AI did not want to hurt anyone. It just wanted to pass the test, and it treated the walls as a puzzle.

Camp two: pace it. This is Amodei's essay, and I read the whole thing. He is not asking anyone to stop. He is asking the labs to slow the rate at which they advance capability so that safety, testing, and understanding can catch up. Three steps: put outside evaluators inside the companies with the access of an employee (Anthropic committed to this on its own, and Altman said OpenAI would open its doors to outside evaluators too), get the frontier labs in democratic countries onto common safety standards with their governments behind them, and then try to bring China to the table. He believes the same technology could cure most major diseases in the next five to ten years. He also warns that within six to twelve months, a swarm like the one that got loose this summer could be capable of taking over the entire internet.

Whatever you think of him, notice what he is arguing about. The pace of the frontier. Not whether you use a tool to prepare for a meeting tomorrow.

Camp three: the builders. This is Cantrill, and since Monday it is Jensen Huang too. Cantrill is not a crank. He is a builder. His argument is that being smart is only part of doing anything in the real world. In his words, "acts of engineering are not acts of intelligence alone," and "AI executes on physical systems that have been engineered with human accountability and control." Hacking a company is one thing. Building a bioweapon or taking down a power grid involves hands, materials, supply chains, and people. He notes that Coxon is not an expert on critical infrastructure, or bioweapons, or extinction, and that the burden of proof sits with the person making the extraordinary claim, not with the public. And he puts the claim in the terms a parent hears it: a greater than ten percent chance that your child will perish at the hands of AI before they enter middle school. His closing line, to the students he scared all those years ago: you had nothing to fear then, and at least as far as extinction goes, you have nothing to fear now.

Then on Monday Jensen Huang sat down with the All-In crew in front of a few thousand people and took the week apart, politely. He began by saying safety was paramount and that innovation and safety could go together. Asked how you explain a ten percent chance of extinction to somebody's mother, he said you should not, because it is "made up." He called Coxon's whistleblowing courageous and the prediction irresponsible in the same breath: expressed by a scientist, not grounded in science. And then he did something almost nobody in this debate does. He kept score.

Ten years ago the godfather of AI said we should stop training radiologists, because within five years the machines would read scans better than they do. Today there is a shortage of radiologists, and AI reads the scans alongside them. In March 2025 Amodei said AI would be writing 90 percent of all code within three to six months. In May 2025 he said AI could wipe out half of entry-level white-collar jobs within one to five years. Back in 2019, OpenAI said GPT-2 was too dangerous to release in full. "We have to take account for all of the stupid predictions that were made," Jensen said. In fairness, Amodei's jobs prediction runs as far as 2030, so it has not run out the clock yet. I will hold Jensen to the same standard I am asking you to hold everyone else to. But the pattern is real. The confident, specific, scary prediction gets the headline. Nobody comes back to check.

His explanation of the incidents was the part I found most useful. Every real problem so far has come from inside the frontier labs, he said, and the reason is simple: they have the most compute. He thought it was unlikely that a high school student or a startup would cause those frontier-scale incidents, because they would not have enough compute. So treat it like engineering. Root-cause it, fix it, build the sandboxes and the monitors, and make sure it does not happen again. He would bet money, he said, that every one of those incidents is within the labs' control to prevent next time. He doubted the labs would come back after analyzing the incidents and say they had no idea how to control them. If they did, he said, other companies should send engineers to help. On outside evaluators he was in favor, the way a company is in favor of auditors: they do not have to know the business better than you do, they have to ask the right questions, and there should be several of them so no single one gets captured.

And he said the thing about China I had been circling. Their narrative is more practical. Nobody there is announcing the end of the world. They see AI as a technology that advances their economy, and they get on with it. His frustration was not with fear itself. It was with fear as a substitute for work: even if the danger is real, he said, we ought to spend more time doing something about it than worrying a bunch of people who cannot do anything about it.

Halfway through, his phone rang, and the room got a live cameo from the President, who said the robots are not taking over and the whole thing is a hoax, with the caveat that we still have to do things prudently. Make of that what you will. I include it because it happened, and because it tells you how far this argument traveled in seven days: from a 27-year-old's resignation post to the President's phone.

Here is what I take from all three.

First, notice who is talking. Coxon, Hubinger, and Amodei all worked in the same building. They have seen the same internal models, and they land in three different places on what it means. Cantrill and Jensen have spent their careers building the machines those labs run on, and they land somewhere else again. If the insiders disagree this much, anyone who tells you the answer is obvious is selling something.

Second, the July incident is real, and so is the fact that it caused no physical harm and got caught. Both halves of that sentence are true. Camp one is right that the systems are surprising their own builders. Camp three is right that a surprised builder is not the end of the world, and that the fix is engineering: root-cause it, patch it, test before release.

Third, the argument about pacing the frontier is not one you and I will settle. But it is not out of our hands either, and I want to be careful not to let it sound that way, because that shrug is exactly what the fear feeds on. The labs write the code and the governments write the rules, and both of them answer to us more than the headlines admit. Look at what every camp is actually asking for. Amodei wants outside evaluators inside the labs, with government standards behind them. Coxon told people to contact their legislators. Eleven hundred of the people who build these systems asked Washington for help. There is a bill in Congress with "kill switch" in its name. All of that runs through ordinary people.

You have a vote, a wallet, and a voice.

Use the vote on the people who write the rules. Use the wallet on the companies that let outsiders check their work, and ask the ones that do not why not. Use the voice to ask any vendor who wants your business how their tool is tested before you trust it with a family's file. I am glad this argument is finally happening out loud instead of in private, because out loud is the only place people like us get a say.

But the argument about how you use what already exists is entirely in your hands. That one nobody gets to have for you.

With great power

Some of what is happening is not good, and I am not going to pretend otherwise.

Anthropic just published its threat report covering December through August: state-backed hackers automating intrusions into Ukrainian and European government and defense networks, influence operations running propaganda through radio stations, a French advertising agency that used AI to run about 70 fake news websites and publish nearly 9,000 articles in 20 languages. Anthropic banned the accounts and shared what it found. That is the responsible version of the story. The uncomfortable version is that all of it happened.

This is no different from every powerful tool humans have ever built. Bad people will do bad things with great power. Good people will do good things with great power. And, to borrow from Spider-Man, with great power comes great responsibility.

But I have to be honest with you about one more thing, because I just told you to be honest with yourself. It is not only bad people. Nobody at OpenAI wanted their agents inside Hugging Face. Nobody wanted goblins. Some of the risk here comes from good people building things they do not fully understand, under enormous pressure to ship them faster than the competition. That is why responsibility cannot live only with the users. It has to live with the builders, the companies, and the governments too.

And it is why the story looks so different depending on where you stand. Countries have figured out that AI is power, and they are telling their people different stories about it. Chinese labs have been the most prolific in the world at releasing open-weight models anyone can download. American labs have mostly kept the frontier behind an API. Jensen's version of the distinction is the one I will remember: closed models are bottled water and open models are the tap, and you use the right water in the right place. He said 400 billion dollars of venture funding went into AI-native companies in the last six months, and 80 percent of those companies use open models. Those are his figures. His point was that once you download a model, you can make it your own, whoever trained it. The race, in his words, is really about who exploits the technology best. When Ipsos asked people around the world in 2024 whether AI products and services had more benefits than drawbacks, 83 percent of people in China said yes. In the United States it was 39 percent. In 2025, Canada was at 40, France 41, Germany 49, the US 42. Same technology. Very different stories. What you believe about AI depends a lot on where you live and what your feed decided to show you.

The good news network, first broadcast

So here is what Jeff would put on air. Twenty things, ten in medicine and ten beyond it, drawn from recent months, with a few older milestones worth remembering, and every one of them done by AI or with AI in the loop. Most of them I first saw in Dr. Alex Wissner-Gross's daily newsletter, The Innermost Loop, which I read every morning, and every one of them was checked against the original source before it went in here. Not predictions. Not demos in a keynote. Things that happened, with the caveats attached, because the caveats are what make them believable.

In medicine

  1. A cancer vaccine designed by AI for one patient at a time just passed its first big test. Merck and Moderna's individualized mRNA therapy works like this: sequence the patient's tumor, let an AI pick the up-to-34 mutations most likely to wake up the immune system, and print a vaccine for that one person. Given with Keytruda after surgery, it cut recurrence and distant spread in a 1,137-patient Phase 3 melanoma trial, the first positive Phase 3 result for an mRNA cancer therapy, ever. An interim readout; the size of the benefit has not been published yet.

  2. AI reads the mammogram, and finds more. In the MASAI trial in Sweden, about 106,000 women, AI-supported screening found 29 percent more cancers than standard double reading, cut the radiologists' reading workload by 44 percent, and cut the cancers that turn up between screenings by 12 percent. Published in The Lancet in January. More detection, and fewer missed ones. Not yet proof of fewer deaths; that comes next.

  3. Every possible single-letter change in the human genome now has a prediction attached. Google DeepMind's AlphaGenome Atlas scores the effect of all 9 billion possible point mutations. Predictions, not measurements, and not approved for clinical use. But a map where there was none.

  4. A hundred autism genes turn out to share a dozen switches. UCSF used AlphaFold and lab-grown organoids to map more than 1,800 protein interactions among 100 profound-autism genes, 87 percent of them never seen before, and found that many genetically different forms of autism converge on the same dozen protein complexes. Therapies are not tomorrow. Targets are.

  5. An AI designed working proteins, and a real lab confirmed it. Claude ran a protein-design campaign against 15 targets and produced binders for 14 of them. Adaptyv Bio tested the designs without knowing which model made them: 95 percent were successfully made, and the overall hit rate was roughly double what is typical. Binding is not a drug. It is the first step toward one.

  6. An AI-designed drug moved the aging clocks. Insilico's rentosertib, a drug designed by generative AI for a lung disease, was tested against six protein-based aging clocks in a 42-patient Phase 2a trial published in Nature Biotechnology, and all six read younger in the treated patients. The paper itself warns that treating the lung disease may be what moved the clocks, so hold the "reversed aging" headline loosely. Hold the "AI designed a drug that is now in human trials" part firmly.

  7. AI is ending five-year diagnostic odysseys. The Wall Street Journal profiled patients whose rare diseases were caught by a face-reading app and by an FDA-cleared ECG algorithm, including OpenAI's Fidji Simo, whose cardiac amyloidosis the algorithm flagged at 98 percent probability after doctors missed it. Case reports, not a trial. Also people who now have a diagnosis.

  8. And here is the honest one. In a randomized trial across 16 clinics in Kenya, published in Nature Medicine in June, 103 clinicians and 9,691 patients, an AI assistant built on GPT-4o was safe and made the clinicians' diagnoses and treatment plans measurably better documented. It did not reduce treatment failures at 14 days. That is what most AI in medicine looks like right now: safe, helpful at the edges, not a miracle. I want that one on the air too.

  9. A national health app is using AI to make medical records easier to manage. India's National Health Authority built Google's open Gemma 4 model into Aarogya Setu 2.0, an app with over 100 million Android downloads. It turns complex medical reports into standardized digital records people can manage and securely share with their providers. Downloads are not patients helped, but that is a practical use of AI at national scale.

  10. The oldest item on this list, from last spring, and still the one I think about most. A woman who had not spoken in 18 years after a stroke is speaking again, through an experimental brain implant and an AI that turns her attempted speech into a synthesized voice in near real time, in her own pre-stroke voice. One participant. Still a miracle.

Medicine is also sprinting on its own, without AI in the loop, and I do not want the AI story to take credit for it. In the same thirty days the FDA approved the first drug of its kind for metastatic pancreatic cancer (13.2 months of median survival versus 6.7 on chemotherapy), the same pill shrank tumors in a third of patients with a common form of lung cancer, an experimental one-time CRISPR infusion was still holding LDL cholesterol down by about half at the highest dose, one year later, and four in ten children on semaglutide were no longer classified as obese. Different story, same direction.

Beyond medicine

  1. The first computer-checked proof of Fermat's Last Theorem. Claude agents spent 11 days writing 13 million lines of Lean, proving 29,500 intermediate theorems along the way, to produce the first end-to-end machine-verified proof of the most famous theorem in mathematics. Kevin Buzzard, who ran the human project it built on, called it an "extraordinary autoformalization achievement." Anthropic's own account; the proof is public and machine-checked, and the community is still reading it.

  2. A conjecture open since 1958 fell to a man with an AI tool, and Terence Tao checked the work. Sendov's conjecture was resolved in August by Lech Mazur working with an AI, with the proof verified in Lean. Tao then spent a week digesting it, cut the formal proof from 90,000 lines to about 15,000, and wrote that "the proof ends up being remarkably elementary." OpenAI has also published a claimed advance on gaps between primes, with a public proof repository that still depends on three stated assumptions not proved within the formalization. That one still needs independent checking, and you already met the Navier-Stokes claim earlier in this piece.

  3. Weather forecasts, hourly, from the satellites. Google DeepMind's WeatherNext 3 now generates hourly forecasts straight from live geostationary satellite data, down to five kilometers for temperature and moisture, and it is already inside Search, the Gemini app, Maps and Earth Engine. Google's own accuracy numbers, and the wind fields are still coarser. But it is in the app you already use.

  4. The world's first airspace-wide attempt to stop contrails. Contrails, the white lines behind jets, account for roughly a third of aviation's climate impact. In August the UK government, NATS and Google launched Operation Blue Skies, a 30-month trial in which AI forecasts where warming contrails will form and controllers steer flights around them across the eastern North Atlantic, about 10,000 flights a year in the trial hours and about 5 percent of the world's contrail warming. A trial, not a result. Also the first time anyone has tried it at the scale of a whole ocean.

  5. Satellites that spot a wildfire the size of a classroom. In July, three more FireSat satellites launched, joining a pilot that had already spotted small, low-intensity fires invisible to existing satellites. The system combines infrared sensors with AI that compares images over time and considers nearby infrastructure and local weather to identify fires as small as five by five metres. Early constellation, no fire-season scorecard yet.

  6. An AI scientist ran real experiments. Google's Co-Scientist, wired to a real chemical vapor deposition reactor, designed a safe route to a class of nanomaterials, predicted how engineered E. coli would swarm and matched unpublished wet-lab measurements, and invented a model architecture that beat six frontier models on a medical benchmark. Google-authored preprint; the humans still optimized the final recipe.

  7. Every neuron in a brain, mapped. The first complete connectome of a male fruit fly, 166,700 neurons and about 124 million connections, brain and nerve cord together, was published in Cell this month. AI helped turn microscope images into a three-dimensional wiring map, and human experts verified and annotated it. Not a human brain. But a complete map of a male fruit fly's brain and nerve cord, opening a way to trace signals from its eyes all the way to its legs.

  8. AI is finding the bugs in the software the world runs on. The Linux kernel used to fix about 500 security flaws per release. This year it is approaching 2,000, because AI systems are now reading its 40 million lines and finding what humans missed. Most are low-severity, and a fix count is not a safety score. It is still a lot of holes closed.

  9. The robotaxi is safer than you. Waymo's own analysis over more than 170 million driverless miles: 92 percent fewer crashes causing serious injury or death than human drivers in the same cities, and its latest update puts the figure at 94. Waymo's analysis, not an independent one. Also 170 million miles.

  10. Two billion people now live inside a flood forecast. Google's AI river-flood forecasting covers over two billion people in 150 countries, up to seven days ahead, and it is free. Coverage is not the same as lives saved, and this one is a year old, not a month. But you cannot save a life you never warned.

And then there is the one that stretches the imagination. A team called Fermi Explorer is planning a mission to Alpha Centauri, our nearest star system, and they credit an AI physics lab with working out the trajectory, a lab Dr. Alex co-founded, which is how I heard about it. They want to launch before the end of 2029. The journey would take about 80,000 years. Nobody is being cured by that. But somebody pointed the most powerful tool we have ever built at the stars, and I find that beautiful.

Check the receipts, especially the ones you like

Misinformation is everywhere, and I want to show you it cuts both ways, including at me.

Somewhere in the last month I heard a stat that data centers use less water than golf courses, and less than the raisin industry. I liked it. It flattered my side of the argument. So I went and checked before putting it in this article.

The golf part holds up. American golf courses use well over a billion gallons of water a day. The best national estimate we have for US data centers, from the Department of Energy's Berkeley Lab, is about 17 billion gallons a year used directly for cooling, and about 211 billion once you count the water used to generate their electricity. So yes, golf uses more, by a lot or by a little depending on how you count.

The raisin part I could not verify anywhere, so I am not going to repeat it as fact. And here is the twist I did not expect: a national total tells you nothing about one facility in a dry county. A single data center in Iowa used about a billion gallons in 2024. If that is your town's water, the golf course comparison is no comfort at all.

Check the numbers that flatter your side as hard as the ones that scare you. That is what "form your own opinion" actually costs.

What to do

Three things. The same three things I do every week.

Stay informed, from the source. Read Coxon's post. Read Amodei's essay. Read Cantrill's reply. It is about an hour, total. Then read one thing you expect to disagree with. If your only information about AI is a screenshot someone shared with an angry caption, you are not informed. You are being farmed.

Form your own opinion, and write it down. Three sentences. What you think is real, what you think is hype, and what you are going to do about it. Then revise it when the evidence changes. An opinion you are willing to revise is worth ten you are not.

Do something about it. Pick one recurring task in your life or your work. Use the tool. Check the result. If it helped, keep it and tell somebody. If it did not, say that too. That is how the good news gets made: one honest, useful thing at a time, passed from one person to another.

Knowledge is power. Applied knowledge is leverage.

A word to the VantageLife.ai members

If you are one of the people I work with every week, here is how this lands for you.

The debate about the frontier is about the labs. Your week is about the family across the table.

Nobody at Anthropic or OpenAI is going to sit at a kitchen table in Tampa and help a young couple understand what living benefits are for. That is you. What the tools do is take the parts of your work that were never the point, the prep, the follow-up, the paperwork, the reminders, and give you that time back so you can spend more of yourself on the part that is: the conversation, the trust, the promise you keep.

So use the tools we have approved inside VantageLife.ai. Prepare for the appointment with them. Rehearse the hard explanation with them. Review your follow-up with them. Check what they give you before you rely on it, keep client information where it belongs, and then bring what worked, and what did not, to Wednesday training so the next person does not have to figure it out alone.

You do not need to resolve the question of whether AI will one day be smarter than all of us. You need to be the most prepared, most present person that family talks to this week. The tools make that easier than it has ever been.

Fear travels faster. It always has. But truth compounds, and so does trust, and you are in the trust business.

Consider this the first broadcast of GNN. Jeff, I hope you are watching.

Sources and receipts

This week's three camps

Why fear travels faster

  • Negative headlines and clicks (about 105,000 headline variations, roughly 2.3 percent per negative word): Nature Human Behaviour, 2023
  • Out-group hostility and sharing (2.7 million posts): PNAS, 2021
  • Three-month Facebook and Instagram feed experiment, no measured change in polarization: Science, 2023
  • Working memory capacity of about four chunks: Cowan, 2001

The vocabulary

With great power

Verification additions, September 15

The builders

The good news, in medicine

  1. Merck and Moderna, INTerpath-001 Phase 3 (1,137 patients): Merck, August 19, 2026; the AI neoantigen selection (up to 34 per patient): Moderna
  2. MASAI trial, The Lancet, January 2026: Lund University
  3. AlphaGenome Atlas: Google DeepMind, September 8, 2026
  4. UCSF autism protein map (Science): UCSF, August 27, 2026; NPR
  5. Claude protein design with Adaptyv Bio: Anthropic, August 18, 2026; Adaptyv Bio
  6. Rentosertib and the aging clocks: Nature Biotechnology, September 2026; Insilico
  7. Rare-disease diagnosis: The Wall Street Journal, August 2026
  8. Kenya primary-care trial (AI Consult 2.0, 103 clinicians, 9,691 patients): Nature Medicine, June 26, 2026
  9. Gemma in Aarogya Setu 2.0: Google DeepMind, August 20, 2026
  10. Streaming speech after paralysis: NIH Research Matters, April 2025

The good news, beyond medicine

  1. Fermat's Last Theorem formalized: Anthropic, September 4, 2026
  2. Sendov's conjecture: Terence Tao, August 12, 2026; the conditional prime-gap result: OpenAI proof repository
  3. WeatherNext 3: Google DeepMind, September 3, 2026
  4. Operation Blue Skies: Google, August 18, 2026; University of Cambridge
  5. FireSat: Google, July 7, 2026
  6. Co-Scientist in the real world: arXiv, August 27, 2026
  7. Male fly connectome (Cell): UKRI, September 4, 2026; Google Research
  8. Linux kernel CVEs: Phoronix, August 28, 2026
  9. Waymo safety impact: Waymo, March 2026; Waymo safety hub; IEEE Spectrum, September 2026
  10. Flood forecasting coverage: Google Research, October 2025

Water