vish

6.8K posts

vish banner
vish

vish

@vishctx

•⁠ ⁠research scientist at @meta •⁠ ⁠rust developer writing c •⁠ ⁠work on Linux kernel •⁠ ⁠dm me if you are building on ai agents or infra!

bay area Katılım Eylül 2015
2.7K Takip Edilen1.2K Takipçiler
alli
alli@sonofalli·
everyone say happy birthday to @brexton !!!
alli tweet media
English
13
0
63
2.5K
vish
vish@vishctx·
It is also the fact that some organizations don’t need the development velocity that Claude Code brings. It’s not better than an average engineer in terms of skill, and the only thing it offers is speed compared to a developer manually typing. Non tech companies usually don’t ship a lot of features per year, neither do their customers have capability to absorb/grok those changes even if shipped.
English
0
0
1
80
Arnav Gupta
Arnav Gupta@championswimmer·
@ponnappa It is not just supervisor ignorance but also supervisor capability gap. Not every manager of an engineer can replace that guy simply with Claude Code + coffee + sleepless nights themselves.
English
2
0
33
1.8K
vish
vish@vishctx·
@ivanburazin There’s multiple subreddits of people just browsing Zillow or Redfin listings. They’re never going to buy, bud, view or even email asking for details on any of them, but just look at houses! It was wild when I discovered it!
English
0
0
9
849
Ivan Burazin
Ivan Burazin@ivanburazin·
I recently heard an a16z partner on a podcast say "The average person doesn't want to save time. They want to waste time." It kinda fucked my mind. I never understood this. I don't use YouTube browse features. Don't scroll feeds. Don't go down rabbit holes. Just what do I need → get it → leave Zero tolerance for wasted time. But it seems that's not how 99% of humans work. They want to scroll through shopping sites. They want to click random videos. They want the browse interface. Text-field-only UIs won't dominate because only Sama and Elon types prioritize pure efficiency. Everyone else wants to scroll, browse, and waste their time. Entertainment disguised as productivity. All the more reason why AI won't kill traditional interfaces.
English
65
37
752
212.4K
vish
vish@vishctx·
@jonchu @ankrgyl I launch my company. Anthropic creates a skill next week. 😎
English
0
0
0
40
vish
vish@vishctx·
@ivanleomk Hmu if you need help understanding it!!
English
0
0
0
163
vish
vish@vishctx·
@khoomeik Priced in where is the question! It still hasn’t been priced into a lot of Cloud contracts or api billing. I’d assume over next few years all of the contracts will see significant bumps. Hertzner bumped their prices by 30-40%!
English
0
0
2
528
Rohan Pandey
Rohan Pandey@khoomeik·
sooo is CPU demand due to scaling RL envs and inference sandboxes priced in yet?
English
29
6
603
64.2K
vish
vish@vishctx·
We talk a lot about "serverless," but we’re about to get a hard reminder that the "server" part still has a price tag. Hetzner won't be the only ones increasing prices by 40%!
English
0
0
1
283
vish
vish@vishctx·
@jonchu Every engineer is a QA engineer because all you’re doing is reviewing AI code! 🧐
English
0
1
4
442
Jon Chu // Khosla Ventures
Ironically, I think QA engineers became a lot more valuable today. Whereas a couple years ago, it was a shrinking profession
English
20
4
148
27.6K
vish
vish@vishctx·
@brexton You can use Claude to build yourself unlimited inference.
English
0
0
2
89
brexton
brexton@brexton·
I waste an absurd amount of inference just asking AI to explain things to me like I'm an idiot
English
3
0
17
901
vish
vish@vishctx·
@axboe @_Felipe Nothing says 'refined taste' like a 250MB hydration layer for a site that could have been a 4KB static HTML file. True craftsmanship.
English
0
0
1
49
Jens Axboe
Jens Axboe@axboe·
@_Felipe Agree. Honestly I take great offense in statements such as these. That’s like saying with generative AI, taste is now important in arts/music. It wasn’t before?!
English
3
0
7
255
Tijana Minić
Tijana Minić@tiksiminic·
long distance relationships suck because you think you moved past the times where you have to see your husband off to battle until you have to see your boyfriend off to the airport
English
1
0
6
397
vish
vish@vishctx·
vish tweet media
IT Unprofessional@it_unprofession

Just got a notification from my bank about "suspicious activity." Someone tried to use my card to buy $400 worth of Roblox gift cards. I know exactly who it was. My 10-year-old. I went into his room. He was on his iPad playing Roblox, very focused, trying to look innocent. I said "did you try to buy Robux with my credit card?" He said "no." I showed him my phone with the fraud alert. $400 in attempted charges. He got quiet. Then: "It was an accident." I said how do you accidentally try to spend $400? He said his friend told him how to find saved payment info in the browser. He was "just looking" and then "it kind of just happened." Then he added: "I only tried to get the medium amount though. Not the big one." This kid tried to steal from me but showed restraint about how much to steal. I asked why he didn't go for the $1,000 option if he was already committing fraud. He said "that seemed greedy. And you might notice." So he calculated the optimal amount to steal. Enough Robux to matter. Not enough to trigger immediate attention. He's 10 years old and already understands the concept of staying under detection thresholds. I should be mad. I am mad. But I'm also realizing this kid has better operational security instincts than half my IT team. He committed fraud responsibly. That's somehow worse. I told him he's grounded from Roblox for a month. And from my laptop forever. He said "that's fair." Then he asked: "If I had done $200 would you have noticed?" The answer is probably not. And he knows it. I'm either raising a future cybersecurity expert or a future criminal. Maybe both.

ZXX
0
0
1
658
vish
vish@vishctx·
@ivanburazin I definitely agree with the sooner part, especially after seeing OpenClaw and the likes. What I am curious about is in the long run. Running a headless browser takes much lower resources (cpu, memory etc) than running these device use agents.
English
1
0
1
59
Ivan Burazin
Ivan Burazin@ivanburazin·
@vishctx I used to align with your thoughts, but the truth of the matter is, we will sooner have agents that are capable for computer use than we will get to transferring all knowledge work software to headless mode.
English
1
0
2
108
Ivan Burazin
Ivan Burazin@ivanburazin·
Android sandboxes make a compelling case for themselves. 1/ Many services are mobile native/simply work better on mobile Your AI agents need WhatsApp (messaging), DoorDash (ordering food), banking apps, social media, and almost every other mobile first service 2/ Devs building Android apps need AI coding environments Someone building "Cursor for Android" needs sandboxes that can run Android to test/build apps like these Agents need to operate in every environment humans do: - Linux (most common for dev work) - Windows (for financial services, enterprise tools) - Android (for mobile apps & mobile first services) They can function as digital knowledge workers across all platforms only when they have access to the full stack.
English
13
1
48
3.8K
vish
vish@vishctx·
@sachdh That looks fun!!
English
0
0
2
116
Sachin
Sachin@sachdh·
Let’s talk Agentic Engineering in Hall 2
Sachin tweet media
English
20
1
174
5.8K
vish
vish@vishctx·
Too academic for industry, too industry for academia.
English
1
0
6
354
vish
vish@vishctx·
Building software was never about the code itself, but a sequence of architectural and business decisions expressed through code. The real risk of offloading that final step to an AI agent is losing the steering on how those subtle choices are interpreted. @margaretstorey hits the nail on the head with her post on Cognitive Debt and it serves as a warning that losing the "theory" of our programs turns our systems into black boxes we can't control. margaretstorey.com/blog/2026/02/0…
English
1
0
2
168
Ivan Burazin
Ivan Burazin@ivanburazin·
Everyone assumes sandboxes are just Linux CPU boxes. But if foundation model companies are training AI on financial services work (which happens on Windows), how can you do RL training for those tasks on Linux?
English
2
0
26
3.1K
sid is in sf 🌉
sid is in sf 🌉@sidbing·
life update: im joining @NousResearch to work on agents and RL in the post-training team next month!
sid is in sf 🌉 tweet media
English
68
7
457
30.2K