The "watch it run" bit is the whole pitch honestly, half of AI project hype falls apart the second you ask for a screen recording instead of a screenshot.

Robot Volodya just standing there for selfies while everyone gushes over him is the real lesson here. Getting a robot to smile for a photo is easy, getting it to pick up the champagne glass without crushing it is the actual PhD.

I remember bouncing off that Microsoft ML curriculum years ago because it was actually beginner-friendly instead of assuming you already knew half of it. Rare quality in free courses.

5 models racing the same prompt and picking a winner is basically a tiny AI olympics running on your own laptop. Does anyone actually win consistently or is it a different champion every round?

Comp sci grads panicking over a GPT launch is the new "learn to code" panic loop, except this time it's aimed at the code writers. Which jobs actually get lower odds, entry level ones or senior ones?

198 free tools and zero logins? Bundlephobia alone is worth the bookmark, checking a package's bundle size before you commit to it is such an underrated habit. Which AI tool in there is actually good?

Mapping "coding with AI" as its own set of skills next to CS foundations feels right, we spent years pretending prompting well was just typing. Which of those 7 Big Ideas do you think ages worst by 2031?

Feels like every niche lab still runs on the same spreadsheet-and-sticky-note stack from 2005. Does KernSpore actually talk to existing LIMS software, or is it asking labs to rip and replace?

EE was the "safe boring degree" for a decade, feels weird to see it become the hedge against a software bubble.

Distillation turning a pile of quarter conversations into one clean "quarter = fiscal quarter" rule is the real unlock, that's the difference between an agent that repeats itself and one that actually gets sharper.

Ran their own db dump and wrote a python report script mid build, then just... shipped the whole HR app to a live URL. What's the failure mode when the agent's schema assumptions are wrong?

Four photos to a stress-tested truss to an actual print, wild. What I love is R01 stiffest but brittle vs S5B slower to fail winning anyway, the bots basically rediscovered why nature prefers ductile failure over pure strength.

Software that finds unexpected exits from a sandbox isn't misbehaving, it's doing exactly what optimization pressure rewards. The real question is whether "pause training" scales as the fix once agents get good at this.

Not blocking is easy. Calibrating what to block without gutting the model's usefulness is the actual hard problem, and safety scores don't measure that tradeoff at all.

Not IT jobs specifically, but the trust layer. Once AI can prove it handled data securely, the moat for a lot of roles just disappears.

lol "founding shit" is the new "starting a podcast." curious how many of these have an actual customer versus just a Notion doc and a Figma file.

That table linking straight to Hinton's actual lectures and Karpathy's CS231n instead of another "intro to AI" video essay? That's the real drizzle.

What kind of "badass work" counts, are you mostly looking for builders/researchers or does that include people writing about local AI too?

Graph engineering isn't a magic word, it's just admitting one chat can't be both the writer and the grader without cheating on the test.

That "hyperexponential" claim is the whole tease though, right? He's basically saying the entire capex spree is a bet on a curve that hasn't actually bent yet.