Building in the open - agents, apps, and the tools in between. Previously early Unacademy, then Loco, Graphy & Airlearn. Bangalore
Bengaluru, India
Joined October 2009
- Tweets3.4K
- Following994
- Followers1.9K
- Likes4.5K
Pinned Tweet
Your agent can now use a real phone. It opens apps, scrolls, taps and types, and when it needs your face or your password, it asks you.
Presenting PhoneDock: a free, open-source Mac app for Claude Code, Codex, Cursor and Pi.
Agree with most of this. I'd just say it's the era of good product thinkers, not sure about product management as a function.
They can come from anywhere.. PM, design, engineering. Earlier most of them were PMs or designers since engineers were mostly analytical. Now with coding agents, anyone with real product taste is more powerful than ever.
great PMs have never been more valuable, but the bar needs to be much higher than it was a decade ago
somehow we’ve gone from raising the bar on product management to questioning whether the function should exist at all
a lot of this comes from vocal X veterans who IMO haven’t experienced what truly great product management looks like
great product organisations will always need great PMs, what makes one will vary across companies and industries
imagining you can build a great product organisation without them is living in a fool’s paradise
maybe it’s time we start imagining a world without these veterans for a change
when building becomes ridiculously easy, knowing what not to build becomes the biggest differentiator for great product teams
1
10
475
What it can't do yet:
• It's Mac only (Apple silicon)
• Real iPhones need iOS 27 or later, Developer Mode and pairing (older iPhones work in beta, through WebDriverAgent)
• Your agent's model provider still sees the screenshots it sends
1
1
189
It's free and open source (Apache-2.0).
Download: sushilk1991.github.io/phoned…
Code: github.com/sushilk1991/phone…
If it breaks on your setup, open an issue and tell me what you were trying to do.
2
176
Read it twice and still don't know what Dreambeans is. A feed? Maps? Discover, with a new name? Google being Google.
It’s time to start dreaming! ☁️✨
Dreambeans, our experimental app from Google Labs, is officially available to everyone in India (18+). Instead of the endless scroll, Dreambeans brews personalized daily story collections from places to explore to upcoming local events using the Google apps you already use, including Calendar, Gemini, Gmail, Photos, Search, and YouTube.
📱 Available on Android & iOS 🗣️ Experience it in 14 languages, including Hindi, Bengali, Tamil, Telugu, Marathi, and more!
Download the app today and start dreaming: labs.google/dreambeans
1
1
457
So now models are editing other models weights directly.. no training. Kinda RSI!
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
122
DeepSeek is what OpenAI was supposed to be. Nobody is giving back to the community like them, maybe Qwen comes close.
Learnt more from their posts and papers than from anywhere else. Really hope they win.
Your Harness. You Define It.
Meet Creator Mode in DeepSeek Harness.
Turn ideas into plugins that you can use, from a design workspace to a sandbox game or a 3D LLM visualizer.
What will you create?
1
163
One month into being a dad and the thing I didn't expect is how much of it is just watching.
She does something new almost every day. A different sound, a longer stare, grabbing a finger a little harder than yesterday.
I've spent years building products and staring at dashboards, and nothing has held my attention like this.
2
13
519
Reading DeepSeek's super-agent write-up felt like someone wrote down what we figured out the hard way at Farfield. And they wrote it so well, with clear principles behind every choice.
We had one gateway agent that listened to every Slack tag and decided where each request should go. Behind it were sub-agents for code review, writing code, finding gaps, RCA and DevOps.
The gateway ran on Sonnet, so it was fast and cheap. It handled about half the requests by itself and passed the rest to sub-agents running Opus. They reported back, the gateway ticked off its to-do list, and the user got one clean reply.
DeepSeek's version is almost the same: one light agent that talks to you, and heavy workers that do the real work. They even add a watchdog that steps in when a worker gets stuck or drifts.
Then there's Instinct, which I think takes this further. Noah Shinn, its founder, wrote Reflexion back in 2023, where an agent writes down why it failed and reads that before trying again. I think that's why Instinct corrects itself so well. It doesn't just recover from a mistake, it remembers it.
I'm convinced this is the way to build multi-agent products that people actually use. Whatever I build next starts here.
4
3
448
A startup raised $1.3M for this a few days ago. Someone built a rough version with AirPods in 9 minutes and put the code online.
That's where software is now. If anyone can copy it this fast, keeping the code closed doesn't protect much.
What it does hide is what the app is doing with your data, and I think people will start caring about that a lot more.
Closed source is going to start looking a little suspicious. Open will just be normal.
I recreated subvocal in 9 minutes with one prompt for AirPods Pro 3.
Source: github.com/joeblau/subvox
1
284
Been doing a version of this with Futari, a small widget app I made just for my wife and me. Shared lists, a note on both our home screens, little things like that.
Whenever one of us wants something changed, I just change it. No roadmap, no other users to worry about.
Having that built into the app itself would be so good. I think a lot of personal apps will end up working like this.
I'm now building an "improvements" feature into all of my personal vibecoded applications: a freeform text input that's hooked up to an agent that yolo implements whatever is submitted.
It's an interesting experience -- the activation energy barrier to sanding any rough edge or implementing any new feature becomes extremely low. Someone else who uses one of the applications I created has already implemented 150 improvements over the course of a few days. This pattern feels like a new kind of tool -- liquid software? -- and every application that doesn't have this feature now has a slightly rigid and anachronistic flavor. Why can't I just tell it how I want it to be different?
Now, it's not obvious to me how this pattern can or should scale. Does it become unmaintainable after a while? (Maybe model intelligence improvement outstrips my feature appetite such that it's fine?) How does collaborative software work in this model? Can software designed this way be secure and reliable?
Whatever the larger questions about broader deployment, it definitely works great in a personal context. It reminds me a little of the visions of Lisp/Lisp machines, Smalltalk, and HyperCard -- systems that the user can take control of and modify. Those visions never quite worked, though: the vast majority of software today is not end-user modifiable in any useful way. So, while this "liquid" affordance is a very small touch, the way it changes one's relation to the software feels a bit transgressive and empowering in a Lickliderian and Engelbartian way.
4
4
349
Sushil Kumar retweeted
Today’s my last day at Airlearn - the most impactful stint in my professional career.
About two years ago, we asked ourselves: Can we build a truly world-class consumer AI product from India?
- 11.1M Users with 0 perf marketing
- 10B+ UGC Views
- 65+ Courses
- Paid Subscribers from 150+ countries later
I’m proud of the fact that we came reasonably close.
What makes this journey particularly special is seeing learners around the world fall in love with Airlearn. Helping people express themselves, build relationships, connect across cultures, and become more confident individuals has been an incredibly fulfilling mission.
Looking back, I have been incredibly fortunate to have had the chance to work with and learn from Sushil, Vipin, Ashok, Meer, Shweta, Raghvendra, Rishav and so many others who have been relentless in the pursuit of building a world-class product.
@gauravmunjal - Thank you for the opportunity and the trust, and for giving me a front-row seat to such an important mission. Eternally grateful.
@chaiandkadhi - Thank you for being a wonderful partner, always being my rock, and supporting me every step of the way. I really lucked out 17 years ago.
Rooting for the team to build the most used language learning product in the world.
The story is just about to unfold ❤️
Sushil Kumar retweeted
2018 loco days were truly loco
Early Loco days. We were doubling every two or three days. A game would have 200–250K people live at the same time, and we were barely keeping up.
Then one evening, almost 350K showed up.
Right before the game, we were getting 15–20K sign-ups a minute. Everything broke. Nobody could sign up, and people were waiting for a game we couldn't let them into.
My co-founder @ichetandhembre came up with the fix right there. Hold back the sign-up writes, push everyone into the game first, and save the data after the game ended. It was ugly, and it worked. The game ran.
I think about that night a lot now.
AI has made building so much easier. But nobody prompted their way out of that moment. It was two people who understood the system well enough to know what could wait.
I wonder where new developers will pick up that kind of instinct, and whether AI will actually help when things break live, not just when you're building.
Building feels close to solved. I'm not sure maintaining is.
Early Loco days. We were doubling every two or three days. A game would have 200–250K people live at the same time, and we were barely keeping up.
Then one evening, almost 350K showed up.
Right before the game, we were getting 15–20K sign-ups a minute. Everything broke. Nobody could sign up, and people were waiting for a game we couldn't let them into.
My co-founder @ichetandhembre came up with the fix right there. Hold back the sign-up writes, push everyone into the game first, and save the data after the game ended. It was ugly, and it worked. The game ran.
I think about that night a lot now.
AI has made building so much easier. But nobody prompted their way out of that moment. It was two people who understood the system well enough to know what could wait.
I wonder where new developers will pick up that kind of instinct, and whether AI will actually help when things break live, not just when you're building.
Building feels close to solved. I'm not sure maintaining is.
Things nobody told me about the first month:
You'll whisper even when she's awake.
You'll stand in a dark room swaying long after she's asleep, because stopping might wake her.
You'll hear her crying in the shower. She wasn't.
You'll google "is it normal" at 3 AM more times than you've googled anything at work.
You'll check if she's breathing more times a night than you'll admit.
2
10
347
I'm running a lot of agents now. A few coding CLIs, plus personal agents like Grok Bot, Instinct, Dot and Hermes. And the personal ones have way more access than the coding ones. Email, calendar, files, notes, a few accounts.
Every week one of them gets a bit more. A folder here, an API key there. Mostly because it's faster than thinking about it.
If something got deleted or sent tomorrow I genuinely couldn't tell you which one did it. There's no single place to look. Each tool has its own history, if it has one at all.
What I actually want is boring. Something like Dropbox/Google Drive but for agents. My files and keys sit in one place. Each agent gets its own login and only the folders it needs. Anything that deletes or sends stuff pings my phone first. And one log I can scroll through at night to see who touched what.
I looked around. Companies are building this, there's a YC one pitching itself as Okta for agents and a few more with funding. All of it is sold to security teams at big companies. Couldn't find anything for one person running ten agents.
Is anyone using something for this? Please suggest. Or should I just build it?
4
11
592