Kevin Tang retweeted
There’s a hidden dagger in almost every PC game from 1995-2005 that murders ARM processors.
It’s kind of an accident, and it all has to do with floating points.
x87 was Intel’s floating point extension to x86. I won’t get into all the historical drama, but Intel decided that all internal float math would be at 80-bit extended precision to reduce rounding errors.
Of course, today no one in their right mind uses 80-bit; standard 64-bit doubles are just too convenient. Modern ARM64 doesn’t have any 80-bit hardware!
This creates a nasty problem for emulation. How much x87 arithmetic can you replace with say, native ARM64 floating point (losing precision!) without breaking behavior?
It’s not so simple; dropping 16-bits of precision can cause funny physics glitches and such. But boy, if you need all 80-bits, the software fallback REALLY stinks.
With all the decomp/recomp hype lately, porting OG Xbox and PC games to mobile / ARM devices is extra tricky. Not impossible, but you’d be amazed at how much x87 code is in old games!
Kevin Tang retweeted
Oh my goodness this has to be right.
Instead of generating a random 256-bit number they generated a random number between 0 and 256.
That’s how they got 4, 66, 155, 156, 230 and 237.
Kevin Tang retweeted
That demo involved a non trivial chance of embarrassing failure! 🫡
I’m only going for the baseline model. At this point in my life, rocket bunny hops are out of my comfort zone.
Kevin Tang retweeted
Amazon dot com’s stock price has now fallen 95% from its high on December 19, 1999.
In just 21 months, this once-promising tech darling has fallen in the dustbin with all the other tech bubble busts.
Kevin Tang retweeted
If you own a 3D printer, you need this tool on your radar right now.
SNAP3D turns a single 2D image into fully assemblable, multi-part 3D prints with built-in peg-and-socket joints!
No glue. No screws. Just print, snap, and build.
it cuts the model into semantic parts, optimizes the connectors, and runs physics tests to guarantee stability.
Feed it a photo, get real-world snap-fit parts ready for your print bed
lucytuan.github.io/SNAP3D/
Kevin Tang retweeted
Recap:
China eBay seller took my money and didn’t ship.
eBay 30 day guarantee window ran out.
I engaged in tactics to get my server that many said would not work.
Today:
The server has arrived.
Don’t you ever doubt my understanding of human nature again lol
Turns out Gigabyte G242-Z11 (for the NEC Vector Engines) wasn’t stuck in customs, China seller tried to not ship and gave fake DHL number.
They ignored my messages and it’s after eBay 30 day guarantee.
So I bought more cheap NICs from them and said if you don’t ship my server I’m reporting all of it. They say they are shipping lol.
Kevin Tang retweeted
Replying to @_KevinTang
Bruh, I showed all my friends this. I want to play. You're the man
Kevin Tang retweeted
@_KevinTang brought Super Smash Bros Melee into mixed reality on Meta Quest, with Astra’s help
your coffee table becomes the stage
players can even hang off the edge
a familiar game becomes something you can place in your room
xitter.cf/_KevinTang/status/2097…
Kevin Tang retweeted
people are using GPT-6 Astra to make robots juggle, play Portal, and put Smash Bros on their coffee table
i went through the demos
8 builds worth seeing and the people behind them 🧵
Kevin Tang retweeted
If you want to time travel to the future, check out chatjimmy.ai/
It generates 15K tokens per second and it absolutely helped me wrap my head around what 2027/8 is going to feel like
Kevin Tang retweeted
ngl, doing software development in the age of (good) LLMs feels far more like caring for a bonsai tree than doing engineering
Kevin Tang retweeted
Introducing Netherite!
Minecraft 1.11.2, rewritten from scratch in C and CUDA, bit-verified against the real game.
This is one trained agent playing in it - then the renderer it actually trains through - then 7,200 live worlds stepping in lockstep on one GPU.
Kevin Tang retweeted
Introducing ACT-2 Preview
The first robotics model to unify broad generalization with high reliability. A single fine-tuning example can teach Memo a new behavior that generalizes.
Zero shot, real unseen homes, 99% success rate.




































