Performance Engineer. Tokyo based 🇯🇵
Tokyo
Joined September 2022
- Tweets1K
- Following432
- Followers2.4K
- Likes1.2K
Sasha retweeted
America has added an estimated 10GW of AI data-centre capacity over the past two years.
Meanwhile, the sorry state of Britain's national grid means that our flagship 50MW project has faced so many delays that Microsoft has pulled out. Nscale is one of two major British neoclouds, but both are choosing to IPO in America. Every month that passes, we fall further behind.
This technology is going to determine the winners and losers of the 21st century. It looks like Britain's elites have chosen their side.
Sasha retweeted
Western liberals imported the third world into France and other countries in the name of diversity and compassion. They ignored fundamental questions of assimilation and compatibility. They were told it would lead to this and they didn't listen.
I pray that France, a beautiful country of amazing people, will stop this madness before more of its landmarks burn.
And I hope Americans are paying attention. This is the future congressional Democrats seek for your country.
Sasha retweeted
Lucky coincidence the labs are only successfully cracking esoteric math problems and not the problems where a solution would give the winner an asymmetric commerical advantage.
yeah I've had like 5 existential crises ytd on this very extrapolation
just spend more time touching grass
Sasha retweeted
Bradford and Birmingham alone are costing us nearly £6 BILLION a year in benefits. This simply is not sustainable.
Sasha retweeted
Birmingham is ranked 1 out of total spending for benefits as a local authority at £4.1bn
The benefits bill per resident is £3,521 which has increased by over 25% in just two years!
benefitsdata.uk/mybill?local… – at Birmingham, England
Sasha retweeted
We are through the looking glass now. Police have put an entire British town under a Section 34 dispersal order (meaning they can even order residents to leave and not return) just so they can import a boat full of illegal immigrants.
Sasha retweeted
some strong feels as I try to navigate this.
1. i f**king love LLMs for generating code, reviewing blog posts, and doing a million other things that make me more productive.
2. i f**king detest reading LLM-generated material from colleagues.
(1/3)
Sasha retweeted
Guy who wrote Ruby on Rails:
Performance is important actually
Eventually I imagine models will have enough intelligence and prowess to imbue good style, design and iterate until the simplest / most compressed solution is found (or humans are so out of the loop, that the latter doesn't matter).
But I reckon in the intermediary; it's going to be very difficult for engineers to manage - offloading all work / thought to agents will result in total context loss, coupled with creeping complexity, slop & boilerplate: productivity may hit an asymptote soon and even decline.
Incentive structures are currently as such, that this may be increasingly difficult to mitigate.
Sasha retweeted
Lazy work used to mean too little output.
Now, with AI, it often means too much and more work for everyone else.
@tobi call it "slop grenades."
A "Slop Grenade" is when you let AI produce the work and pass it on without adding any value (including checking it).
Someone else has to wade through it, catch the mistakes, and clean up the mess.
You save time and look productive but someone else pays for it.
My third conversation with Shopify co-founder and CEO @tobi.
0:00 How Shopify Uses AI
7:18 River: Shopify's Internal AI
8:55 How to Encourage Osmosis Learning
10:52 AI Dreaming and Self-Reflection
11:53 How to Use AI for Strategic Decision Making
14:11 The One Thing AI Cannot Do
16:04 What AI is Making Worse at Shopify
19:46 Predictions: Where AI is Headed Next
21:55 The Future of AI-Powered Software
24:40 Will CEOs Be Replaced with AI?
27:54 Can Superintelligence Be Controlled?
31:13 Critical Skills in AI Age
34:22 Why Complex Solutions are Usually Wrong
36:33 Conditions Needed for True Intuition
38:02 The Best Path Doesn't Have Instant Feedback
44:51 How Affirmations Can Shift Your Behavior
50:44 The Inobvious Thing Hurting Companies
52:37 Relationship Between Beauty and Creation
56:23 How SpaceX Moves Forward By Subtraction
1:00:24 Why Companies Need Refounding Events
1:01:50 Books as Cheat Codes
1:02:39 Three Books to Change Your Thinking
Enjoy!
(Includes paid promotions.)
For the record, multiple simd registers + fused FMA gets us to ~15% of theoretical upper bound in terms of GFLOPs on the host machine for a single core. Still some way to go.
Also, its already naturally following that tiling lends itself nicely for multiple cores.
Single tiling provides about a 30% increase in perf. Lower than I had from theoretical perspective. Tile sizes targets L1 cache size.
Probably need to do something similar for SIMD registers & reduce some of the tiling bookkeeping - L1 cache misses reduce by 4x, yet instructions per FLOP increase by ~same amount, likely some inefficiencies, although when measured against control (TM,TN,TK=N) suggests bookkeeping cost less than 1%.
