Aditya Vivek Thota

One day my friend asked, “Which is better, the iPhone or the Google Pixel?”. He was a long-time Android user and wanted to have an experience iPhone. Both of us brainstormed which was better and I being an iPhone user tried to explain why I bought it and how the Apple ecosystem is helping me in my coding, researching, writing, and business things. This morning I read a paper titled “Reward Steering with Evolutionary Heuristics for Decoding-time Alignment”. As I was reading two things dawned on me: 1. LLMs are sometimes very irritating because even if you give the right prompts they just scam you. 2. Why cannot they brainstorm ideas before generating a response as I did with my friend by asking us follow-up questions to know our preference? (They do only asked to do) What is the issue at hand? Traditional user preference optimization methods such as RLHF, DDO, etc., interfere with the model’s original training knowledge when they align themselves with the user preference. This is good but not all the time. Sometimes we expect a different style, format, and tonality in the responses. The authors in this paper came up DARWIN framework that utilizes exploration and exploitation methods: o Exploration: The authors imply instruction mutations, where initial instructions are modified to create diverse instructions. These guide the model to explore various potential responses. o Exploitation: Periodic replacement is performed where the low-rewarded generations are replaced with the high-rewarded ones to produce well-aligned responses. Essentially, DARWIN separates the preference modeling from the main LLM. This means the model's core capabilities remain intact, and alignment is achieved during the response generation phase. This approach prevents the alternation of the parameter. Here is a step-by-step walkthrough. 1. Input Prompt: "Which is better, iPhone or Google Pixel?" 2. Archiving: The input prompt is stored in a database called archive. 3. Instruction Mutation: Generate multiple mutated versions of the input prompt — “What are the pros and cons of the iPhone compared to the Google Pixel?", "Which phone has better performance, iPhone or Google Pixel?" etc. 4. Response Generation: Response is generated using every prompt and stored in the archive. 5. Evaluate Responses: A reward model evaluates the prompts. 6. Response Replacement: Responses with low scores are removed with high scores in the archive. 7. Iterative Mutation and Replacement: This iterative process continues until a stopping criterion is met, such as a set number of iterations or convergence in the quality of responses. 8. Final prompt “What are the long-term benefits of choosing an iPhone over a Google Pixel?” with its respective response. I think the next iteration of LLM must bring these capabilities, this will help us in long-term planning, researching, brainstorming, learning complex subjects, and assisting us in our everyday lives. #LLM

7

4 Comments

Are you a product leader or founder in SF Bay Area or Bengaluru? If yes, you don't want to miss this post, period. 🚀 Bits-n-Atoms, our flagship mixer for founders, operators is live in Bengaluru this month and our 6th edition of GrabChai is back to the Bay Our previous editions have had folks from Meta, Google, OpenAI, Snap, Swiggy, Flipkart, Uber, Microsoft and others come in and share their journey 🗓️ May 23rd - Bay Area & May 25th - Bengaluru 📬 Invite Only What’s on the agenda? 🤝 Meet product leaders, startup founders, and fellow growth folks 💬 Get inspired and gain new insights. 🍴 It's all about great food and even better conversations. Follow The Product Folks so you don’t miss future ones! Check the comments section for RSVP links, see you there 👋

7

most thrilling & exciting 30 days of my professional journey 👇🏼 together with Shlok (the crazy guy amongst the 2) - we spent the first 7 days solving the problem without writing a single line of code, we’re not building an ai wrapper but solving a real consumer problem - built a toy version of the product in 15 days (it worked really well, wasn’t a toy to be honest lol but i like to call it that way) - interacted with 500+ people to validate our thesis, got some learnings & implemented features accordingly - got rejections from 50+ VCs telling us that we’re too early, get some traction & other bs (we knew that we needed just one to trust us & we’re lucky to have the most amazing folks backing us) we’ve been executing things at crazy speed & we believe in failing faster & iterating with feedback our goal is to build a 100 people / $100 billion ai-first company from India (a realistic version of the 1person-$1billion company proposed by sam altman, founder of openai) looking for some exceptional designers, engineers & MiSfiTs to help us achieve that we’ve already received applications from some great folks at Meta, Google, Microsoft, Adobe which is kinda overwhelming for us too but a guiding principle of our hiring is that we don’t care about your qualifications/titles, all we care about is your ownership in what you do a few roles we’ll be aggressively hiring for  - senior ml, backend & devops engineers  - ui designers with a strong visual sense - product & design engineers for business opportunities, please reach out to Shlok ps. this would be the best place for you if you want to build a really deeeep-tech consumer product & have loads of fun at a hacker house!

296

63 Comments

I've been going to many events and mixers where LLM and GenAI discussions are the norm. And invariably there is talk of Indian language LLMs. I love it. But I wonder, how many folks asking about all of this used Koo? How many posted content in different languages? how many checked out multilingual Topics that Koo developed more than two years ago? Where were all the ardent Indian language LLM aficionados when Koo was breaking out in Brazil? Koo did a lot of work on Indian language tech. KooBERT was trained on all Indian language Koos with AI4Bharat and open sourced, Topics in all supported languages was an industry first consumer app feature, NER, Classification models, use of multilingual embeddings, vector databases (before vector dbs became popular) and so much more. There are so many stories to tell and learnings to share. I will be starting to record podcasts to share some of these experiences from the trenches. Stay tuned. #koo

48

2 Comments

Startups don't need 'big brains' - they need 'big hearts.' Hearts that can withstand the breathtaking ride, hearts that have experienced the noise and emerged bolder. A Harvard MBA or a stint at Google might look great, but what doesn't mean someone knows how to roll up their sleeves and get their hands dirty. Well, I don't know which one I fall into but I am the one who can roll up their sleeves and get their hands dirty. Real entrepreneurship is chaotic. And requires more than just big brains and a polished resume. It needs grit. #career #entrepreneurship #Coaching #Leadership #Linkedin ♻️ Reshare if you found this helpful. If you’re stuck somewhere and need personalised coaching, book a call with me at https://lnkd.in/ggxZjxtQ

22

3 Comments

Very well put and pertinent to today's fluid, volatile and dynamic world

3

Ever wondered why your Indian language searches fall flat? Uncover the hidden 'Search Problem' through my blog on Srujanee: https://lnkd.in/grQTeedY Read, share, and join the conversation! Mention some interesting experiences that you have undergone while searching on the internet using your native language search queries in the comments below. #Srujanee #MultilingualSearch

28

‘India set to have world’s largest developer community,' says GitHub CEO Thomas Dohmke And that's precisely the inspiration behind TechX! We're all about creating an ecosystem of *Top Indian Engineers Globally*. We're incredibly bullish on the amazing outcomes that Indians, as a strong tech community, can achieve. Picture this: a global network bursting with Indian engineering talent, buzzing with innovation, and set to take the world stage by storm. That’s TechX for you! Imagine the brilliance of India’s best and brightest engineers, connected across continents, collaborating on groundbreaking projects, and pushing the boundaries of what’s possible. With TechX, we’re not just building a community; we’re creating a powerhouse of innovation that spans the globe. This is the place where top Indian engineers unite, thrive, and lead the charge in shaping the future of technology. The future is bright, and with the collective strength of Indian engineering talent worldwide, it's going to be nothing short of extraordinary. Cheers to the great times ahead! 🍻

16

Who would invest in founders who have wasted $50M? Koo, an Indian microblogging platform, recently shut shop after raising $50M in capital. Everyone has their views on why Koo failed. IMO, the biggest reason behind its failure was a lack of PMF (Product Market Fit). Koo spent much money on marketing and didn’t build a product people wanted. It kept harping on nationalism, a Twitter for Indians, made in #India. This approach failed. However, this doesn’t mean Koo’s founders aren’t smart or intelligent. I wrote in an earlier post that we shouldn’t be surprised if Koo’s founders raise fresh capital for a new startup. To many of us, this might seem grossly unfair. Millions of smart, hardworking founders are struggling to raise capital while investors are backing failed founders who have wasted millions of dollars. How is that fair? It confirms the widely held belief that the #venturecapital industry is broken and corrupt where a few people with the right connections get all the capital while others don’t see a dime. It shows that the VC industry is a closed group, a cabal where who you know matters more than anything. While these are just complaints, most people who criticise VCs don’t understand how startup investing works. VCs welcome failures. They embrace them because they know failures are inevitable. When VCs invest in 20 #startups, they know 15 will fail. The goal is to invest in startups that can give them a 100x return. Failures are part of the game. They are unavoidable. The aim is to back founders who can build such 100x return companies. Hence, failures are not looked down upon in the VC industry. It’s similar to the American culture where society appreciates the people who take risks, and it easily accepts personal bankruptcies that result from this risk-taking approach. There’s no taboo in failing in the USA and the VC industry. I’m sure Koo’s founders have learnt from their mistakes. It’s an expensive mistake and not everyone is so lucky to have this luxury to spend $50M on their education. ;) However, most VCs will back Koo’s #founders because these founders know how to raise capital and they know how to not build and scale tech companies. Their past failures have improved their chances of future success. In the VC industry, Failure is Good.

226

41 Comments

The Unicorn Hunt: How I Landed a Founding Engineer for a Harvard-Stanford Health-tech Startup in 1 introduction The catch? The candidate had to be based in India while the founders were in the US. Also this is the first engineer for their team After months of searching and working with agencies, the founders were ready to throw in the towel. But not me - I was determined to find their perfect match. Here's how I did it: Step 1: Crack the Founders' Code Before I could even think about finding the right candidate, I had to understand the founders on a whole new level. I dove deep into their psyche, personality, and non-negotiables. It was like cracking the Da Vinci Code, but with a startup twist. Step 2: Uncover the Vision Next up, I needed to know the why and how behind their startup. I asked all the tough questions - short-term goals, long-term dreams, the whole nine yards. It was like a game of 20 Questions, but with higher stakes. Step 3: Define the Unicorn Persona The job description was just the tip of the iceberg. I had to figure out the exact persona they needed - think ex-founders, CTOs, and founding engineers. These are the high-agency individuals who can make magic happen. Step 4: Conquer the Distance Since the founders were based in the US, they had no clue about the talent landscape or salary benchmarks in India. I took them on a crash course, educating them on the local market. It was like a crash course in international relations, but with a tech twist. After all this research and preparation, I went on the hunt for the perfect candidate. And guess what? I found a unicorn! Sure, luck played a part, but my intel gave me a serious edge. I introduced a candidate that was a match made in startup heaven. The founders had to stretch their budget, but the value this person brought to the team was off the charts. They were blown away, and so was I. So there you have it, folks, even with an ocean between you. If I can do it, so can you. Happy hunting! #Recruiting #Startups #FoundingEngineer #TalentAcquisition #RemoteWork #InternationalRecruiting #startupjobs #Techjobs #remotejobs

20

Just saw Zomato's new feature. Here’s what I think about it 👇 (For context: They launched a feature where we can see the total order count of the day.) This update received some criticism for adding adding an admin-side internal widget to the consumer-side. I’m not privy to Zomato’s internal workings but it could be a listed company gimmick, to play and showcase their scale or maybe just an experiment to improve customer trust and create social proof. I personally think it’s a harmless update, the one occasional feature that’s fun, interesting, cool and may eventually die. Does every product feature mandatorily needs to move metrics? Let me know what do you think? #zomato #product #productupdate

27

1 Comment

The company's search boss, Prabhakar Raghavan, told employees at an all-hands meeting in March that Google plans to build teams closer to users in key markets, including India and Brazil, where labor is cheaper than in the U.S. Seems like SkillsFuture is leveling up in India and Brazil; also means the Commoditization of typical "Core" roles could be performed beyond the US, with much lower cost. Now, my question, what can our next-generation Singaporean fresh graduates do, than those countries mentioned yet to? Also goes to many staff with the business-as-usual SG&A roles, how to keep them from moving into Thailand and Malaysia?

52

23 Comments

What if the next billion internet users prefer to use the internet in their mother tongue? Explore the cultural, economic, & social impact of Indian languages online in my latest blog post about "Why Indian Languages Matter on the Internet" on Srujanee: https://lnkd.in/gkxc-3Tq Join the conversation & help build a more inclusive digital world! #IndianLanguagesMatter #BeyondEnglish

39

2 Comments

As the elections in India have started, we can will see the social media buffering with the invoking and criticising terms or queries which can create an unexpected impact on elections so the meta has blocked the certain specific election terms or keywords which can be typed in meta ai. Does this shows the fear of misuse of ai??. What do u think does the ai will influence the elections in future in respect to this measure taken by meta??. Comment your thoughts on this. #meta #ai #artificialintelligence #machinelearning #metaai #chatgpt

1

2 Comments

📣 I agreed to do a "webinar" for PANIIT USA this Saturday morning on "Demystifying LLMs." The event is run by and for IIT alumni, but is open for anyone. You need to register at the link below to get the webinar link. 👉 https://lnkd.in/gkgh4AXh Abstract: The main aim is to help you develop the right mental picture of LLMs–as n-gram models on steroids, trained on web-scale corpus of our collective (sub)conscious, doing approximate retrieval in the context of the prompt. Armed with this picture, we will also look at the many myths/misunderstandings surrounding LLMs–especially in regards their factuality, whether they can ever be made to overcome hallucinations, and whether LLMs are AI-Complete (in that they can also do planning, reasoning etc. that are typically associated with System 2 competencies in humans). Time permitting, and in Q&A, we will also look at broader ethical/societal implications. This column in The Hill will serve as a good extended abstract: https://bit.ly/4b6VYy0

43

1 Comment

Koo shutdown news got everyone talking about it. Huge respect for Koo Founders; they are my fellow Blumiers! 💜 I remember writing a post about Web 2.0 Social Networking Tools in 2006! Yes, in 2006! I thought to share here 👇 Social networking tools targeting a particular social segment might not work out in the future. For instance, a student having Facebook during graduation, LinkedIn for building business contacts, and Bebo for something else…will this be feasible in the long run? I have one Hi5 account, created a Bebo when a friend requested it, left Orkut and continued using LinkedIn. For GenZs, Hi5, Orkit, Bebo were social network of the past decades! :) I was one of Spoke's key designers and developers for Spoke (it is dead as of 2024), the first of the few social networking-based tools (though Spoke targets enterprise networking and solutions like SalesForce and others). Spoke - https://lnkd.in/g3u7kmPY I was involved in the Spoke first MVP launch in 2002. LinkedIn has targeted the professional community successfully. Friendster couldn’t sustain its success after Myspace arrived; however, Myspace had a good strategy of allowing various communities to coexist in the same social space. In a single Myspace, a convergence of various social groups happened. Taking this argument further, many social networking tools are definite bubble-burst types. Note: This is from my blog post from July 2006 👇 https://lnkd.in/gGqDqvdx

6

1 Comment