The reason top tech companies often hire foreign-born & first-generation engineers over “native” Americans isn’t because of an innate American IQ deficit (a lazy & wrong explanation). A key part of it comes down to the c-word: culture. Tough questions demand tough answers & if we’re really serious about fixing the problem, we have to confront the TRUTH: Our American culture has venerated mediocrity over excellence for way too long (at least since the 90s and likely longer). That doesn’t start in college, it starts YOUNG. A culture that celebrates the prom queen over the math olympiad champ, or the jock over the valedictorian, will not produce the best engineers. A culture that venerates Cory from “Boy Meets World,” or Zach & Slater over Screech in “Saved by the Bell,” or ‘Stefan’ over Steve Urkel in “Family Matters,” will not produce the best engineers. (Fact: I know *multiple* sets of immigrant parents in the 90s who actively limited how much their kids could watch those TV shows precisely because they promoted mediocrity…and their kids went on to become wildly successful STEM graduates). More movies like Whiplash, fewer reruns of “Friends.” More math tutoring, fewer sleepovers. More weekend science competitions, fewer Saturday morning cartoons. More books, less TV. More creating, less “chillin.” More extracurriculars, less “hanging out at the mall.” Most normal American parents look skeptically at “those kinds of parents.” More normal American kids view such “those kinds of kids” with scorn. If you grow up aspiring to normalcy, normalcy is what you will achieve. Now close your eyes & visualize which families you knew in the 90s (or even now) who raise their kids according to one model versus the other. Be brutally honest. “Normalcy” doesn’t cut it in a hyper-competitive global market for technical talent. And if we pretend like it does, we’ll have our asses handed to us by China. This can be our Sputnik moment. We’ve awaken from slumber before & we can do it again. Trump’s election hopefully marks the beginning of a new golden era in America, but only if our culture fully wakes up. A culture that once again prioritizes achievement over normalcy; excellence over mediocrity; nerdiness over conformity; hard work over laziness. That’s the work we have cut out for us, rather than wallowing in victimhood & just wishing (or legislating) alternative hiring practices into existence. I’m confident we can do it. 🇺🇸 🇺🇸

It’s a Brave New World

Democratic agenda will pause AI, and let the world catch up 😀

The RFD4-Proteina throughput on @tenstorrent QuietBox 2 is incredible. The image shows throughput and price of a cluster of 8 x QuietBox 2 compared to a state-of-the-art GPU server. It's easy to build a QuietBox 2 cluster for TT-Bio, because all models in TT-Bio run on a single chip. We’ll make computational drug design accessible to anyone.

some fun slides from my upcoming talk @aiDotEngineer very excited to share how we ship crazy amounts of code (30x'd) without reading it, yet perf is as good if not better then when we launched there's a lot more to improve of course, and the ceiling is far from being reached, but given the relentless shipping speed and our high bar for polish, i think that our agent friendly architecture bets are paying off at taming the slop fun fact: ~230 people landed at least 1 PR to grok bot in october!

You can get into clinic for <$10m. I wrote a playbook on this last year shelbyann.substack.com/p/a-p…

1/18 🧵A thread about the evolution of four generative AI models for biomolecular design (Proteína, La-Proteína, Proteína-Complexa and RFD4-Proteina), which makes me think we're moving toward fully programmable protein design 🔥

It’s been a while! Like every other biotech nerd, I’ve been feeling quite context-bloated as of late, and writing is the only fix I know. So here's a long one on AI and clin dev - if you run trials for a living, I’d appreciate any criticism and feedback :)

Very excited to introduce Paperclip Datasets! Curate your own datasets from millions of papers in minutes. Just describe the dataset you want to build! As example, we release the largest dataset of 20k half-life human measurements curated from 2400+ drugs💊 🖇️ Try it: paperclip.gxl.ai/datasets 📝 Blog: gxl.ai/blog/introducing-pape… 📊 Half-life dataset: huggingface.co/datasets/gxl-…

Solve bottleneck in light 💡 speed and Giga scale

Guess my election-interfering space lasers didn’t work this time 🤷‍♂️

I finished building the interactive 3D book on GCSE and IGCSE Physics Electricity with Opus 5.5. It now has all seven lessons, from charge and current to mains and the plug. Publishing soon on my website.

Large-scale RL has a problem: weight updates from trainers to samplers are slow and expensive We’re sharing weight teleportation: faster weight updates than existing intra-cluster techniques with internet speeds, enabling samplers anywhere (maybe even space!) Weight teleportation is so efficient (<0.23s sampler pause time, ~3s non-blocking compression + prepare), the bottleneck ends up being the GPU hardware’s decompression ASIC We unlock these gains by using GPUs for massively parallel compression of model weight changes, engineering a heterogeneous decompression and patching pipeline, along with other optimizations and overlaps. These techniques are co-designed with hardware-native features in NVIDIA’s Blackwell generation of chips We’ve open-sourced weight teleportation, so everyone can use it today

RFD4-Proteina now runs locally on a MacBook. Introducing BIOS’s Mac port: generate candidate protein binders and screen their predicted structures locally. Built with MLX and custom Metal kernels. • 2.2× faster generation than stock PyTorch-MPS at its default precision • 20 designs generated and screened → 10 ranked candidates in about 12 minutes on GLP-1R (M5 Pro) We believe more researchers should be able to run AI experiments on their own hardware. Test ideas locally, keep your research private, and scale to NVIDIA GPUs when you’re ready. After setup, the workflow runs offline. Watch it run below. ↓

The total difficulty of cooling in space is vastly harder than on the ground But 99.9999% of that challenge is Starship Once you have the ability to get huge amounts of mass into space cheaply, it is genuinely easier to cool things in space Your cooling system is passive Just a giant piece of metal that is a) light w/ b) high thermal conductivity Aluminum is a good candidate And then you have a working fluid that just circulates forever inside the metal structure Ammonia is historical industry standard for radiators in space You need a fluid where the critical point is optimized around the temperature of the chips The working fluid goes from liquid phase to gas phase near the chips, and this change in density drives the hotter gas to pump heat through a loop The high thermal conductivity of aluminum rips heat out of the fluid and black body radiation ejects heat into space Black body radiation goes like temperature to the 4th power so the hotter you can run chips and hotter you can get the working fluid the better you can shed heat It's honestly a way simpler system than on the ground

Cool, it saves money 💴 and time ⏱️

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…

You are fully retarded