I'm excited to share that I will be starting my lab at NTU Singapore. It is bittersweet to be moving halfway around the world. On one hand, I'm excited for a new adventure. I have spent my entire adult life chasing this dream, and it's finally happening. On the other hand, I still love America and believe in all of the good that it stands for. My parents moved here to give me the gift of the American Dream and I never thought I would experience what it's like to be an immigrant to a new country myself. I can only say that life has a way of offering blessings in disguise, and I’m grateful for this opportunity in every way.

Some thoughts on Colossal Biosciences CEO Ben Lamm's new spinout WARHOUND: 1) The basic pipeline is that they're going to SCNT a few of the best elite military, police, and first-responder dogs. Think Dolly the Sheep. Here's it's mainly Belgian Malinois, German Shepherd, or a mix. 2) This is downstream of Colossal's acquisition of ViaGen Pets, a startup that offered to clone your pet for $50k. I don't think that market was every going to be big enough. But ViaGen probably gives them an in house SCNT tech for dogs that's better than current state-of-the-art. For example, published SCNT success rates are low, maybe 2-10%. If they can routinely get 10% live birth per transfer, then this can work reasonably well, since you can transfer 10-20 embryos per surrogate. 3) The genetic enhancements are unnecessary. If you have a great working dog, it's not going to have recessive hip dysplasia or cancer risk. What would you change about Secretariat? If you've got canine MJ, are you really trying to make the Monstars? The same applies to healthy adult humans -- there is virtually no genetic edit that I could make that would majorly enhance your child. Most of the traits you care about are highly polygeneic. 4) This is going to be the ultimate nature vs nuture test for behavior and cognition. If you make 100s of genetic twins (a better term than clones), do they have the same trainability, temperament, and instincts? How much variation from developmental noise is there really? 5) These dogs are going to be like Monsanto seeds - only available for purchase from the source. You won't be able to breed them because they're not inbred, and so any F2 is going to risk picking up undesirable recessives. 6) Picture below is a myostatin mutant, perhaps the most striking single mutation one could engineer. It is common in racing dogs, but not likely to be relevant for working dogs.

Inside the NBA still has that juice. I hope all you brainrotted gen-Zs can manage to buckle down for six minutes to watch greatness unfold

Who he play for?? These guys are a national treasure

This is super cool. I remember playing the 24 game as a kid growing up (二十四点). Many Asian Americans of my generation will remember it. Flip four random playing cards. Use addition, subtraction, multiplication, and division to combine all four cards to make 24. You must use all 4 cards. By the time you get good, you can instantly buzz in as soon as the cards are flipped, so apparently the math camp kids invented a harder version, make 163 with 6 cards. Scott is showing off a version here where you make 2025 with 8. It seems like total magic for the uninitiated. But the trick is you factor 2025 in your head to 25 x 81. Then you just have to find 2 ways to make 5 and 2 ways to make 9 from the remaining cards. For weirder numbers like 2027, you reduce it to a nicer number like 2025 and find an extra 2 for the ending. These things feel superhuman, but there's a trick to it. It brings me great joy that these math genius kids are now billionaire CEOs thanks to America. Our parents generation did not have this same opportunity. AI and tech is what made this possible. For the world of the 90s and early 00s that we grew up in, I will never forget another kid just like Scott from my high school, Mathcounts National Champion in 2008, 3 years before Scott. He didn't get the chance to live this dream, and he should have been at Harvard too, but this was before SFFA and asian math nerds were still underpriced relative to our private school peers. Anyways, that's a story for another day.

Just curious are voting machines pwned? Sooner or later right??

There's an underrated reason why AI's improving cyber-hacking abilities won't lead to the imminent collapse of the internet or global power grids, etc: it's actually quite hard to monetize a new vulnerability. Take the example of the recent French grocery store hack that ruined millions of dollars of produce by raising the temperature on hundreds of supermarket freezers. If you were a cybercriminal in possession of this hack, how would you actually extract a payday from it? Let's say you contact the grocery store CEO and demand a ransom. Now you face a thorny set of problems that are arguably more difficult than hacking a temperature controller. 1) How do you contact the grocery store CEO while shielding your identity? 2) How do you prove that you can actually cause damage? If you execute the hack, now the clock's ticking on the other side finding a patch. If you tell them it involves the freezers, they can minimize damage by moving produce out. 3) If they agree to pay you, how do you actually get your hands on the money? Sure, crypto—but you, random sysadmin guy working from home, how do you get a foreign bank account to launder that money and deliver it to you in a usable form? 4) Okay, let's say we want to fix the credibility issue by causing damage, making a name for ourselves, and then the next time we make a ransom demand, they'll know we mean business. Well, this is a terrible plan because you've now made yourself a major target for law enforcement. They will come after you, and the consequences of a major hack still involve rotting in a jail cell for the rest of your days. In reality, no one has taken credit for the French grocery store hack, no ransom demand has been reported, and there's no evidence that anyone has made a single penny off of it. Similarly, with this week's PHP vulnerability, the immediate damage has largely been the exposure of sensitive data and the massive costs of patching vulnerable infrastructure. But has anyone actually made billions of dollars from these hacks? At this point, the number of hacking incidents that have leaked our personal information onto the internet is too many to count. Anything that's leaked or sold quickly gets devalued as the same information becomes available to everyone else. Can you sell the vulnerability itself? Maybe, but only if you know the right people in the criminal underworld. Plus, a vuln is a depreciating asset. Once you use it on a single target, it's only a matter of time before it gets patched everywhere. There just aren't that many ways to make a huge amount of money off a vulnerability, even if discovering new ones becomes easier than ever. This is also why ransomware is such an effective business model for cybercrime. It has three rather unique properties: 1) It provides proof of compromise. 2) The damage is potentially reversible. Since encrypted files can be restored, the victim has a massive an incentive to pay you. 3) Patching the vulnerability doesn't undo the damage. By the time you deliver the ransom message, the attack has already happened. Even if they fix the original flaw, they are doomed by the math of encryption. In other words, ransomware solves the central economic problem of cybercrime: you have something the victim needs, you can prove you have it, and they can pay you to get it back. Very few other destructive hacks have this combination of properties. So what's the endgame of our new security situation, where AI makes discovering and exploiting vulnerabilities dramatically easier? 1) Maybe we get more ransomware attacks. 2) Maybe someone discovers a new BUSINESS MODEL for hacking that makes previously unprofitable attacks lucrative. This seems unlikely. 3) But widespread, indiscriminate destruction of infrastructure remains unlikely and everyone needs to just calm down. Just becuase a technological capability exists, doesn't automatically mean that it will be used. Just because AIs will eventually achieve superhuman capability, doesn't mean they will automatically WANT to kill us all.

I think it's notable that in this era of super-viruses and vulns, this dogshit website is still up and running, despite Elon having fired 95% of its engineers

I went to see TikTok streetball star MK (Matt Kiatipis) during his Singapore tour, and I regret to inform you that Gen Alpha is even more cooked than Gen Z. We Twitter-ites remain blissfully unaware of the ecosystem of social media microstars on TikTok and Instagram. MK's shtick is simple: travel the world challenging local hoopers to pickup-style one-on-one games, trash-talk them mercilessly, and turn the spectacle into viral clips for his millions of followers. It's the 30-second clip economy in action. It's a totally different and deflating experience in person. In the clips, you see the highlight reel. In person, MK airballed and bricked about 8 threes in a row and had trouble getting enough lift for his dunks. I don't think he is actually 6'3" standing barefoot at the doctor's office. It's honestly embarrassing to watch the constant clowning and fake confrontations, followed by a 360-degree three-point attempt that airballs. These moments make it clear that the purpose of the event is purely to farm clips for social media, in front of an audience of mostly 12-year-old boys. Even worse, the audience seems to be there for the same reason -- to show their friends that they were there, that they were part of the spectacle. Watching it all play out, I was left with the overwhelming feeling that the online world has already eclipsed the real one for Gen Alpha. It's a generation of humans whose behavior is guided by the invisible hand of the algorithm. It's all so depressing. I guess this is the Gen Z dream now: to become a social media star and make a living performing algorithm-optimized clown shows for 12-year-old boys, who have somehow managed to surpass their idols in brain rot.

Did you know that reptiles have evolved live birth over 100 independent times? It's not that hard -- instead of laying an egg, you just keep it inside of the mother and make the shell thinner to allow for oxygen exchange. So if reptiles can do live birth and birds/dinosaurs can do warm blood, then what's really so special about mammals that allowed them to become the dominant species on Earth?

A little analytics for Seahawks fans: Many things went wrong in the Denver Debacle, but one mistake that really cost us was the overtime kickoff return. Shaheed fumbles the catch, Tyrice Knight gets called for a holding, and they start the drive on the 13 yard line. If they'd started at the 31 yard line, which is their average starting field position after a kickoff this year, then the Broncos would have gotten the interception at the DEN 39 instead of the SEA 43. Good chance they go 3 and out and give the ball back to Seattle. Field position is paramount in overtime, with limited possessions. It's a shame the defense didn't give up a single first down and still gave up points. Can't put it all on Shaheed -- his punt return put the Seahawks in position to tie in regulation, and he's been good overall on returns. This was just the worst possible time to flub one. For the 2026 season as a whole, the Seahawks have been 5th in the league on returns and middle of the pack on coverage. The kickoff coverage unit has regressed from 2025, when it was 2nd in the league. So perhaps we were a bit spoiled as fans. For context, this regression in coverage alone has been worth ~0.6 points per game, which is pretty significant in a league where every edge matters. Let's hope the Seahawks can get back to being elite in all phases, go hawks! @robstaton @hawkblogger

There's a beautiful parallel between this year's Nobel Prize in Medicine for optogenetics and the quest to understand how advanced AI models think and work. For animal brains, we face a real black box problem. It's impossible to record in real time how all of the neurons in a network are behaving, and it's hard to manipulate them one by one to decode their function. Optogenetics is a real breakthrough for systems neuroscience because it allows us to activate a specific clump of neurons, and then observe a change in behavior in the animal. It should be no surprise then that all of these classic neuroscience experiments are easier to do with an LLM. In fact, we have perfect observability of the input prompt, the chain of thought, the output, as well as the "firing patterns" of the entire neural network throughout. If anyone is worried about a rogue agent that can deceive us while plotting against us, one of the key safeguards will be if we can log all of its neural firing and map out exactly what it's thinking at any given moment. I think this is a feasible goal, given the perfect observability of the system. The data streams will be enormous, so we'll need ways to compress and summarize them. But I essentially imagine training a model that can examine these "neural recordings" and reconstruct the original prompt, chain of thought, and output from the recordings alone. This kind of predictive power is what mechanistic interpretability means to me. The complication, though, is that this approach may depend on the underlying model remaining static. At some point, we'll want models that can learn and adapt continuously. If the neural representations themselves are changing over time, how do we ensure that our interpretations can keep up?