Midhun Harikumar

333 posts

Midhun Harikumar banner
Midhun Harikumar

Midhun Harikumar

@HarikumarMidhun

ML Lead and Manager @morphic | Parenting Lead @Home. I write here https://t.co/q94zWlWetE

San Jose, California Katılım Eylül 2019
139 Takip Edilen111 Takipçiler
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@Suhail Use the same process for papers and research as I use for X. Bookmark everything you see into snippets (obsidian clipper is great). Claude summarizes and organizes the topics to review at the allocated time. Once topics get to certain critical mass I review it all together.
English
0
0
0
31
Suhail
Suhail@Suhail·
The pace of frontier AI progress is moving so fast that you could spend an entire day on X reading every interesting thing people publish and get absolutely nothing done. Somehow you have put the blinders on and build.
English
11
3
33
2.3K
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@nikunj Would a more realistic tactic be. "Never trust 'but' listen to anyone that gives advice in absolutes". Averaging advice over many people (users, mentors) has been easier for me than keeping track of a confidence score per person.
English
1
0
0
213
Nikunj Kothari
Nikunj Kothari@nikunj·
Never trust (or listen) to anyone that gives advice in absolutes.. Not in this era. Not at this time. Everything is changing and rearranging so freaking quickly. Sure, some things are evergreen - focus on retention, run a good business and don’t commit fraud. But a lot of the old priors are turning out to be garbage aka don’t boil the ocean, do just one thing, figure out your wedge. PS: I’m sure someone will dig up one of my tweets where I have done this and fine, I was just wrong
English
13
2
104
5.1K
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@paraschopra Don’t prompt engineer use something like GEPA to improve prompts based on what’s important to your task.
English
0
0
2
306
Paras Chopra
Paras Chopra@paraschopra·
it’s easy to dismiss something as mere “prompt engineering” but the right prompt can distill years and decades of highly non-trivial domain knowledge. in fact, as models keep getting smarter, their inability to get something done would likely be attributable to bad prompts.
English
31
12
252
8.1K
Harsh Gupta
Harsh Gupta@hargup13·
Paras Chopra single handedly raised the bar of AI Research in India with @lossfunk . Great use of capital from Wingify's exit. Thank you!!
English
21
54
1.2K
42.4K
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@DominiqueCAPaul Placing your kill switch right under robot workspace is a nice way to break a finger or worse.
English
0
0
2
90
Dominique Paul
Dominique Paul@DominiqueCAPaul·
So using relative joint angles with Pi0.5 isn’t going that well yet. I think it’s some bug in the inference loop. Need to think of a debugging visualization tomorrow.
English
15
3
136
13.1K
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@vijayshekhar Dude why ? You do understand that people look up to you and your words have weight ?
English
0
0
0
89
Vijay Shekhar Sharma
Vijay Shekhar Sharma@vijayshekhar·
Search for the press briefing with Marco Rubio; he was filling in for Karoline Leavitt. It is such a great Q&A. He is such a real, fun ans great human being.
English
3
0
33
6.8K
Caleb
Caleb@caleb_friesen·
AI companies have been rushing to strap egocentric data collection headsets to the heads of Indian factory workers. But is selling this data to foreign buyers legal? @alysha_lobo is conducting a session on the legal implications of data collection/AI in Indian factories 👇
Caleb tweet media
AML@alysha_lobo

🚨: You asked for this and we are finally doing it. This 𝐜𝐥𝐨𝐬𝐞𝐝 𝐕𝐈𝐑𝐓𝐔𝐀𝐋 𝐫𝐨𝐨𝐦 is for 𝐅𝐎𝐔𝐍𝐃𝐄𝐑𝐒 𝐚𝐧𝐝 𝐎𝐏𝐄𝐑𝐀𝐓𝐎𝐑𝐒 🇮🇳 𝐎𝐍𝐋𝐘 who have been dabbling or working in the space of collecting 𝐄𝐠𝐨 𝐂𝐞𝐧𝐭𝐫𝐢𝐜 𝐃𝐚𝐭𝐚 𝐒𝐞𝐭𝐬 from workers in 𝐈𝐧𝐝𝐢𝐚𝐧 𝐟𝐚𝐜𝐭𝐨𝐫𝐢𝐞𝐬/𝐦𝐚𝐧𝐮𝐟𝐚𝐜𝐭𝐮𝐫𝐢𝐧𝐠 𝐬𝐢𝐭𝐞𝐬 to sell to humanoid companies as training material. The session will be delivered by a 𝐋𝐞𝐠𝐚𝐥 𝐭𝐞𝐚𝐦 that has already worked on this matter. We hope to help you understand the compliance and legal challenges around collecting data via factory workers in India 🇮🇳 Link to register and more details in the comment below. 𝐍𝐎𝐓𝐄: • This is for 𝐈𝐧𝐝𝐢𝐚𝐧 Founders and Operators ONLY • We have limited seats so please bear with us. • After registration, you will receive an email from Mosaix Network so please check your Spam folders too. • 𝐕𝐂𝐬 𝐚𝐧𝐝 𝐂𝐚𝐩𝐢𝐭𝐚𝐥 𝐀𝐥𝐥𝐨𝐜𝐚𝐭𝐨𝐫𝐬 please DO NOT apply as we have a separate session for you.

English
4
5
80
5.4K
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
You’re making a commonly inflated assumption here . CTC is yearly so your google friend has a CTC of Rs30.75L not 62L . I’ve hired contract employees before and they don’t get any benefits and it’s higher chance to get layed off and >90% chance your are working on non critical functions.
English
1
0
4
2.8K
Devaansh Bhandari
Devaansh Bhandari@ThisIsBhandari·
My friend has a ₹62L CTC at Google. Another friend is a Remote Contractor making ~₹36L ($38k). On paper → almost 2x higher compensation (₹62L vs ₹36L) In the bank → Remote dev wins by a mile. Here’s the math 👇
English
49
31
1.1K
363.5K
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@chris_j_paxton Agreed , spot was a blip but managing to be a near zero revenue research lab for more than a decade changing hands ever so often and being responsible for a majority of robot viral videos is a no coincidence . Good thing to come from BD .
English
0
0
1
26
Chris Paxton
Chris Paxton@chris_j_paxton·
@HarikumarMidhun as i've said other places in this thread i was honestly super bearish on them until last year; i've never been a fan of spot (terrible for mobile manipulation) and the old atlas was behind the times. but they've changed tactics really successfully
English
1
0
0
67
Lee Sharkey
Lee Sharkey@leedsharkey·
My team at @GoodfireAI has been cooking up a new way to do interpretability: decompose a language model’s weights, not its activations. Our decomposition natively handles attention (!) and behaves less like a lookup table and more like a generalizing algorithm. (1/6)
English
34
191
1.5K
233.4K
Gowthami
Gowthami@gowthami_s·
@arnie_hacker I certainly am. It starts at 10pm and it doesn’t end till 2. Sometimes it’s Claude’s dumbness and sometimes it’s greed when free GPUs are just there. 🫩
English
3
0
18
2K
Arnie Ramesh
Arnie Ramesh@arnie_hacker·
Are ML researchers constantly sleep deprived because they're trying to submit all their jobs before they sleep
English
39
43
847
72.4K
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@nikunj @_inception_ai Anything <200k context that you need ultra fast responses to. Parsers, decision trees, live user interactions. Performs better that all small LLM's and competitive with gemini-flash on quality. Much faster.
English
0
0
0
59
Nikunj Kothari
Nikunj Kothari@nikunj·
It’s criminal how cheap and how good Gemini Flash is.. that too with 1M context windows and structured outputs. Probably, my most used model in production workloads. Separately their new live voice model is mindblowingly good if you haven’t tried it yet!
English
37
11
293
20.6K
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@willbitsky @oyhsu @JacobZietek TRI found chain-of-thought doesn't help robot policies. But π0.7's whole thing is chaining language sub-tasks and visual subgoals to condition actions... which is basically CoT? Maybe the issue isn't CoT itself ?
English
0
0
0
59
will bitsky
will bitsky@willbitsky·
beyond excited for the engineering + deployment phases of robotics good time to reread @oyhsu + @JacobZietek blog posts
will bitsky tweet mediawill bitsky tweet media
English
4
4
77
7.5K
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@paulg Largely overlooked fact is the bay area weather that allows one to work out of a garage most of the year without additional heating or cooling .
English
0
0
5
56
Midhun Harikumar
Midhun Harikumar@HarikumarMidhun·
@thegautamkamath Switching off notifications is an option but the only real value i've seen is all day tracking of health and pretending to be a secret agent when I voice message my wife.
English
0
0
1
98
Gautam Kamath
Gautam Kamath@thegautamkamath·
Recently replaced the battery for my (dumb) watch. I get the appeal of a smartwatch, but I prefer not to be bombarded with the sum total of human knowledge and information when I just wanted to know the time.
English
4
0
18
3.8K

Keşfet

@Suhail @nikunj @paraschopra @PrivateThilak @hargup13 @lossfunk @DominiqueCAPaul @vijayshekhar