Shayon Mukherjee

243 posts

Shayon Mukherjee banner
Shayon Mukherjee

Shayon Mukherjee

@shayonj

Everything, everywhere - always in progress. Do less, well. 🏃‍♂️👨🏽‍🍳💻

Cambridge, MA Katılım Ekim 2009
701 Takip Edilen1.1K Takipçiler
Shayon Mukherjee
Shayon Mukherjee@shayonj·
@fatih @WHOOP So much this. I recently got one, the onboarding was frustrating tbh. Took me a good 15 min to figure out their strap situation. Every UI is trying to sell something to me, jarring. This is after $250/yr. Finally, I decided to cancel which in itself is a horrible experience.
English
0
0
0
422
Fatih Arslan
Fatih Arslan@fatih·
I canceled my @WHOOP subscription, and the only reason was that they made it extremely hard to cancel. It’s not possible via their app, and even if you find their obscure link, they don’t allow you to cancel unless you answer 9/10 questions. Feels extremely deceiving. Better to avoid paying companies that have zero respect to their users.
English
129
86
3.2K
361.4K
Shayon Mukherjee
Shayon Mukherjee@shayonj·
To be led by a fool is to be led by the opportunists who control the fool. Octavia Butler.
English
0
0
3
161
Shayon Mukherjee retweetledi
Glenn Greenwald
Glenn Greenwald@ggreenwald·
Liberating the Iranian people by bombing their oil refineries, poisoning their air, destroying their fresh water supplies, exploding their elementary schools, and now "obliterating" their power plants. Who wouldn't be grateful for this?
Rapid Response 47@RapidResponse47

English
1.6K
15.1K
58.5K
1.4M
Shayon Mukherjee retweetledi
Rasmus Jarlov
Rasmus Jarlov@RasmusJarlov·
I have a genius idea: Dont bomb Iran. That will keep the straight of Hormuz open. It was open and no ships were attacked by Iran before that started.
English
2.5K
6.9K
49.1K
1.3M
James Cowling
James Cowling@jamesacowling·
Solid take. Start with the wrong architecture you can get past the point of no return in a day now. You used to have to spend months of suffering to get there, usually learning something in the process.
modi@modisulak

@jamesacowling the compounding is the scary part. you used to have time to realize you were building wrong. now you've shipped the whole thing in a weekend before anyone notices.

English
4
1
77
6.3K
Shayon Mukherjee retweetledi
James Cowling
James Cowling@jamesacowling·
Whatever your take on this, writing code has always been tremendously easy compared to designing/maintaining large systems. The best engineers in the world are the best because of this stuff. Currently LLMs will not save you from bad architectural decisions. You need to constrain them to do the right thing.
David Cramer@zeeg

im fully convinced that LLMs are not an actual net productivity boost (today) they remove the barrier to get started, but they create increasingly complex software which does not appear to be maintainable so far, in my situations, they appear to slow down long term velocity

English
30
40
485
44.4K
Siqi Chen
Siqi Chen@blader·
@charles_irl psa: knowledge cutoff time is august 2025 - pre opus 4.5, pre gpt 5.2
Siqi Chen tweet media
English
6
2
82
6.2K
Charles 🎉 Frye
Charles 🎉 Frye@charles_irl·
We find that the adoption of Cursor leads to a statistically significant, large, but transient increase in project-level development velocity, along with a substantial and persistent increase in static analysis warnings and code complexity. arxiv.org/abs/2511.04427
English
22
33
590
100.6K
Shayon Mukherjee
Shayon Mukherjee@shayonj·
Cloud Hypervisor now supports sub 100ms VM snapshot restores using userfaultfd 🚀
Shayon Mukherjee tweet media
English
1
0
3
252
Shayon Mukherjee
Shayon Mukherjee@shayonj·
This. I think there are enough primitives that sandboxes are “solved”. Even though it’s somewhat difficult tech to pull off at scale and security. However, without a good DX and ease of coordination, everything else falls flat on its face. Including 60ms boot times.
Ashton Chew@iamashtonchew

1 billion is no surprise. We build our RL eval environments on Modal. Many sandbox platforms offer fast isolated containers. Modal offers that plus the most seamless coordination and orchestration primitives in Python. Sandboxes are a dime a dozen, coordination around them is the hard part.

English
1
0
1
581
Vasek Mlejnsky
Vasek Mlejnsky@mlejva·
We're introducing 𝗔𝘂𝘁𝗼 𝗥𝗲𝘀𝘂𝗺𝗲 𝗳𝗼𝗿 𝗦𝗮𝗻𝗱𝗯𝗼𝘅𝗲𝘀 Many agents don’t need a sandbox running all the time, but when they do need it, it should just work, whether it was paused or not. Auto resume handles this 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆. A paused sandbox wakes up when activity arrives. Perfect for use cases such as 🔸 agent runs on request 🔸 active dev previews on inbound traffic Great ship by @badphilosopher!
Vasek Mlejnsky tweet mediaVasek Mlejnsky tweet media
English
25
7
175
21.4K
zhi
zhi@zhipan01·
@ivanburazin What if branching requires write actions like sending an email. You can’t send 5 emails.
English
3
0
0
252
Ivan Burazin
Ivan Burazin@ivanburazin·
When an agent hits a decision point, it forks into 5 parallel branches exploring different solutions. - Branches 1, 3, 4 work; 2 and 5 fail - Snapshot the working ones - Tear down the failures - Fork 5 more branches from each winner That's how agents solve problems. And the infra has to support it natively.
English
4
0
40
4.7K
Shayon Mukherjee retweetledi
Arcee.ai
Arcee.ai@arcee_ai·
Trinity Large Preview just crossed 2T tokens in its first 35 days on @OpenRouter. We are grateful to our launch partners and core apps driving this traffic at @cline, @kilocode, @opencode, and @openclaw. Growth at this scale requires world-class infrastructure, and we couldn’t serve this volume without the continued support and reliability of our partners at @PrimeIntellect and @modal. Here's to the next 2T 🥂
Arcee.ai tweet media
English
5
16
118
8.8K
Shayon Mukherjee
Shayon Mukherjee@shayonj·
LLMs are so humble. Estimates 1 day of work, cranks it out in 5 mins on the next prompt.
Shayon Mukherjee tweet media
English
1
0
3
313
Shayon Mukherjee
Shayon Mukherjee@shayonj·
With nvproxy, GPU ioctls still reach host driver paths, so GPU-driver bugs remain part of the risk model? gVisor reduces a lot of surface, but not all GPU-driver risk right Which makes me think you need policy enforcement and proxy control for secrets
English
0
0
1
229