Harish (@CoderHarish) header
Harish (@CoderHarish) profile picture

I am a security enthusiast Secured @Google,@Microsoft etc | MSRC Q1 2022 top 50 | AI Security at Cisco

Metaverse!
Joined September 2019
Photo by Harish (@CoderHarish)Photo by Harish (@CoderHarish)Photo by Harish (@CoderHarish)Photo by Harish (@CoderHarish)Photo by Harish (@CoderHarish)Photo by Harish (@CoderHarish)Photo by Harish (@CoderHarish)Photo by Harish (@CoderHarish)Photo by Harish (@CoderHarish)Photo by Harish (@CoderHarish)
Harish retweeted
Most people start their careers with some privilege, and many start from zero. I started from minus โ‚น1 crore, and I did not even know it until my second year at IIT Delhi. Over the years, that debt took me from a failed placement day to Google, to a โ‚น2 crore side business, and finally to the startup I had always wanted to build. This is that story. My father ran a real estate business, and he had invested in under-construction flats from builders like Jaypee and Wave, paying most of the price up front through home loans. Years went by without the flats being delivered, but the banks still wanted their EMIs on time. So my parents borrowed from friends and relatives, and when that was not enough, from credit cards. My parents carried all of this debt quietly and tried to shield me from it. I knew we owed money, but never how much. Then in my second year, collection agents started coming to our house and sitting there, asking for payments we could not make. It stressed my mother a lot, and I could see it on her face every day. ๐—ง๐—ต๐—ฒ ๐—น๐—ถ๐˜€๐˜ One day, I sat my parents down and asked them a simple question: who exactly do we owe, and how much? Then I took a sheet of paper and started writing down how much we owed, to every bank, every friend, every relative and every credit card. The list kept getting longer, and by the time I finished, the principal alone was more than โ‚น1 crore. I still remember how heavy it felt while writing each line. That day, I knew I had to grow up quickly and get my parents out of this situation. I had joined IIT knowing exactly what I wanted to do after college, which was to build a startup. But that sheet of paper changed everything, and the dream would have to wait. The new plan was to get a high-paying job, pay off everything on that list within two years so that my parents could finally breathe again, and only then pursue my dream of starting up. ๐— ๐˜‚๐—ฟ๐—ฝ๐—ต๐˜†'๐˜€ ๐—น๐—ฎ๐˜„ From my third year, everything I did was aimed at that one goal: getting a job at an HFT, which paid โ‚น50 lakh+ a year. I was ranked 2nd in my department and had a CGPA of 9.7/10, so I genuinely believed I could get in. Then day one of campus placements came. I prioritized Tower Research over the other companies, cleared three rounds, and was rejected in the final one. I was so crushed that I could not perform at the others that had shortlisted me, and by evening, they were all gone. That evening was really hard. Every rupee I was not going to earn meant my parents would have to carry that weight for longer. On day two, I took an offer from Cohesity at โ‚น18 lakh a year. It was a good job, but nowhere near the number we needed. In July 2019, I moved to Bangalore, stayed in a PG, kept my monthly expenses to under โ‚น20,000, ate meals at the office, and rarely went out with friends. Everything else, I sent back home. My friends would talk about investing and going on trips, but I did none of that. I soon realized that this salary would never be enough. So I set a new target: Google US. Whatever time I could find outside work went into competitive programming, and I soon became a Master on Codeforces. In January 2020, I started applying to Google and other companies in the US and UK. A few weeks later, COVID hit, and overnight, companies stopped hiring overseas candidates. The way out I had been working towards for months disappeared. ๐—ช๐—ต๐—ฒ๐—ฟ๐—ฒ ๐—ฑ๐—ถ๐—ฑ ๐—ฎ๐—น๐—น ๐˜๐—ต๐—ฒ ๐—บ๐—ผ๐—ป๐—ฒ๐˜† ๐—ด๐—ผ? Around the same time, I got an offer and joined Spotnana as a founding engineer. I really liked their vision, and I believed that growing with a young startup and holding equity in it could be the big break I needed. About a year after I had started sending money home, I sat down to see where it had gone. Almost my entire salary had been spent on interest for our credit card loans. We had not reduced the principal and had not saved anything either. After a full year of hard work and sacrifice, the list was exactly as long as the day I had written it. That was one of the lowest points of my life. It felt like I was running as hard as I could, just to stay in the exact same place. Meanwhile, I was giving Spotnana everything, working on weekdays and weekends, and was one of the top performers there. But when appraisal season came, the raise was still not as high as I was hoping for. I was really disappointed. My equity, though promising, could not help us pay off the debt anytime soon. I told my manager I would need something on the side to make up the difference. I created a programming YouTube channel and later learned Android development, but neither brought in any money. ๐—œ ๐—ฐ๐—ผ๐˜‚๐—น๐—ฑ๐—ป'๐˜ ๐˜€๐—น๐—ฒ๐—ฒ๐—ฝ ๐˜๐—ต๐—ฎ๐˜ ๐—ป๐—ถ๐—ด๐—ต๐˜ During that time, I was talking to one of my college seniors who had joined Google. I told him everything, about the list, my plan and my belief that I could clear it all in two years. He listened patiently and then told me the truth. If I stayed at Spotnana, it would not take two years, it would take more than four. He said that instead of being emotional about the company, I needed to be tactical if I really wanted to solve my family's problem. He then referred me to Google. The next four months were the most focused of my life, and every spare hour went into preparing for Google. When I got rejected from Uber and Coinbase, I was terrified about what would happen next. But then, Google said yes. The day the recruiter told me I had been selected for the Google Search team is still one of the best days of my life. That night, I couldn't sleep at all. I lay awake, too happy to close my eyes, thinking about how this would finally change things for my family. For the first time since I wrote that list, I could actually see a way to clear it. ๐—ง๐—ต๐—ฒ โ‚น๐Ÿฎ.๐Ÿญ ๐—น๐—ฎ๐—ธ๐—ต ๐—ฏ๐—ฒ๐˜ When I joined Google, the list still had โ‚น80 lakh left on it, and salary alone would take years to clear it. But Google also gave me something I had not had before: time. So with Google's permission, I started a mobile apps business with my father, working nights and weekends. In the first three months, we built a bunch of Android apps and games, but could not make any revenue. Then I analyzed iOS versus Android and decided we should build for iOS instead. So we took the bet. I bought a MacBook Pro for a hefty โ‚น2.1 lakh, picked our most downloaded game on Android, built its Swift version in six days, and launched it on iOS. Installs were high from day one, and within a couple of days we turned on monetization. I kept refreshing the dashboard as paying users came in, and within just a few days the game had paid back the cost of the laptop. And that was just the beginning. Our revenue grew from โ‚น6 lakh in the first month to โ‚น24 lakh in the fourth, as we added more games. I was working close to 100 hours a week between Google and our apps. Over a year and a half, our games crossed 200K downloads, with players mostly from the US, and brought in more than โ‚น2 crore in revenue. ๐—–๐—ฟ๐—ผ๐˜€๐˜€๐—ถ๐—ป๐—ด ๐—ผ๐—ณ๐—ณ ๐˜๐—ต๐—ฒ ๐—น๐—ถ๐˜€๐˜ With that revenue and my Google salary, we finally started crossing lines off the list. This time we were not just paying interest, we were paying back the principal, one line after another. With every line we crossed off, the stress at home got a little lighter. In June 2023, we paid off the last one. The list I had written in my second year of college was finally empty. My parents were free, and after years of waiting, so was I. As soon as the last loan was cleared, I put in my papers at Google. The apps had left me with enough savings to finally start my own company. My last day at Google was the day before my 25th birthday, and on 21 July 2023, I woke up with no debt, no job, and for the first time in years, nothing standing between me and the thing I had always wanted to do. ๐—ฆ๐—ต๐—ผ๐˜‚๐—น๐—ฑ ๐—œ ๐—ด๐—ผ ๐—ฏ๐—ฎ๐—ฐ๐—ธ ๐˜๐—ผ ๐—ฎ ๐—ท๐—ผ๐—ฏ? Over the next four months, I built 12 new games, but month after month, none of them worked. And it was not just me. Apple's iOS privacy changes had broken the economics of hyper-casual games. That November, I remember walking on my terrace, thinking about whether leaving my job had even been the right choice. My parents had started telling me to get a job again. But I had waited far too long for this chance to give it up. I decided to shut the games business down and hard pivot into SaaS. I started with Shopify apps, only to realize a month later that I had no unique idea and the competition was intense. I then analyzed around 70 B2B SaaS marketplaces to find the next Shopify app store, found monday, the project management platform, and started Super Workday, building apps for teams using it. My first app there failed too, because it was too niche. For the next one, I went through the platform's feature requests and picked the biggest unsolved problem. That one worked. In April 2024, after eight months without any revenue since leaving Google, I got my first paying customer. I was so happy that day that I emailed him just to tell him he was my first customer. I still remember his name. After that, I kept launching one new app every month, and the revenue grew sharply. Today, the business runs on autopilot, has 300+ paying customers, mostly in the US, and is close to $100K in ARR. It also paid for my sister's wedding, and being able to do that for my family made me very happy and proud. ๐—ง๐—ต๐—ฒ ๐˜€๐˜๐—ฎ๐—ฟ๐˜๐˜‚๐—ฝ With that income stable, I finally felt ready to go after a venture-scale idea and started building an AI executive assistant. In January 2025, South Park Commons, a Silicon Valley VC, invested $400K at a $5.7M valuation. I cannot explain what I felt the day their offer came. All those years of waiting finally made sense. My startup journey is still early. Since then, there have been many more ups and downs, and there are still many things to figure out. But for the first time, I am building the startup I always wanted to build, and I am ready for whatever comes next.
1,453
1,846
645
14,088
1,237,179
Harish retweeted
tried to add some nuance to my take on why most of deeptech is a grift. hopefully this inspires more chatter around the ecosystem ๐Ÿคช๐Ÿคช๐Ÿคช
Replying to @imDhananjay
in a decade we will realize most of the deeptech is a grift and quick services + quick commerce companies are what drives more for the economy
45
8
50
178
130,597
Harish retweeted
Indian corporates and capitalists need to catch up in taking risks and be more willing to put their money where their mouth is. By funding audacious teams and their crazy ideas. If an Anthropic kind of startup had approached investors in India for seed 5 years ago, the response would have been, "So you want to create another chatting app ah?" "But how will it make money?", or just "It won't work. Nobody will buy this." The day Indian companies and investors start incentivizing innovative founders (who are plenty in India) by taking big futuristic risks and funding crazy ideas, only then Anthropic kind of companies will emerge. Or India will always play catch up. But it doesn't need to. Because India has the most innovative and diverse thinking talent in the world right now. They are just not getting enabled by the right ecosystem. SV excels in this. And those who make it in India are also only enabled, at least in part, by foreign funds based in India. They tend to shift their HQ soon or get prodded by their investors to sell out to foreign companies before making it big for India. Btw, R&D expenditure as % of GDP: India ~ 0.8% China: ~ 2.7% (and GDP is 5x) United States: ~ 3.5% (and GDP is 8x) South Korea: ~ 5% Israel: ~6.3%
Anthropic, an AI company known for Claude, listed at a valuation of $ 1.5 trillion. It did not exist till ~ 5 years ago. This is more than the combined market cap of the top 10 Indian companies. Are we undervalued, or do we need to catch up in the new world? Or both?
226
778
21
5,701
214,068
Harish retweeted
๐Ÿš€ ๐—œ๐—ป๐—ฑ๐—ถ๐—ฐ๐—ฆ๐˜„๐—ถ๐—ฝ๐—ฒ ๐—ท๐˜‚๐˜€๐˜ ๐—ด๐—ผ๐˜ ๐—ฎ ๐—ฏ๐—ถ๐—ด ๐˜‚๐—ฝ๐—ด๐—ฟ๐—ฎ๐—ฑ๐—ฒ! ๐—œ๐—ป๐—ฑ๐—ถ๐—ฐ๐—ฆ๐˜„๐—ถ๐—ฝ๐—ฒ ๐— ๐—ฎ๐—ฟ๐—ธ ๐—œ ๐—ถ๐˜€ ๐—ต๐—ฒ๐—ฟ๐—ฒ. ๐ŸŽ‰ A few months ago, we launched IndicSwipe, a free, on-device gesture-swipe keyboard for 22 Indian languages. Thank you for the downloads, the reviews and the feature requests. You told us what you wanted, and we built it. ๐—ง๐—ผ๐—ฑ๐—ฎ๐˜†, ๐˜„๐—ฒ'๐—ฟ๐—ฒ ๐—น๐—ฎ๐˜‚๐—ป๐—ฐ๐—ต๐—ถ๐—ป๐—ด ๐˜๐—ต๐—ฒ ๐—ป๐—ฒ๐˜„ ๐˜ƒ๐—ฒ๐—ฟ๐˜€๐—ถ๐—ผ๐—ป ๐—ผ๐—ป ๐˜๐—ต๐—ฒ ๐—ฃ๐—น๐—ฎ๐˜† ๐—ฆ๐˜๐—ผ๐—ฟ๐—ฒ! ๐Ÿ‘‡ ๐ŸŒ๐—ง๐—ฟ๐—ฎ๐—ป๐˜€๐—น๐—ฎ๐˜๐—ฒ ๐—ฎ๐˜€ ๐˜†๐—ผ๐˜‚ ๐˜๐˜†๐—ฝ๐—ฒ, ๐—ฝ๐—ผ๐˜„๐—ฒ๐—ฟ๐—ฒ๐—ฑ ๐—ฏ๐˜† ๐—œ๐—ป๐—ฑ๐—ถ๐—ฐ๐—ง๐—ฟ๐—ฎ๐—ป๐˜€๐Ÿฎ Write in one language and send in another. Translate between English and 22 Indian languages, and between Indian languages, without leaving your keyboard. ๐ŸŽ™๏ธ๐—ฉ๐—ผ๐—ถ๐—ฐ๐—ฒ ๐˜๐˜†๐—ฝ๐—ถ๐—ป๐—ด, ๐—ฝ๐—ผ๐˜„๐—ฒ๐—ฟ๐—ฒ๐—ฑ ๐—ฏ๐˜† ๐—œ๐—ป๐—ฑ๐—ถ๐—ฐ๐—–๐—ผ๐—ป๐—ณ๐—ผ๐—ฟ๐—บ๐—ฒ๐—ฟ Speak naturally in your language and watch it turn into text. Our speech models are trained on Indian speech, accents and code-mixing. ๐Ÿ”ค๐—ก๐—ฎ๐˜๐—ถ๐˜ƒ๐—ฒ ๐˜€๐—ฐ๐—ฟ๐—ถ๐—ฝ๐˜ ๐˜๐˜†๐—ฝ๐—ถ๐—ป๐—ด Type directly in your language's script, with no transliteration needed. โœ๏ธ๐—ค๐˜‚๐—ถ๐—ฐ๐—ธ๐—ฒ๐—ฟ ๐—ฐ๐—ต๐—ฎ๐—ฟ๐—ฎ๐—ฐ๐˜๐—ฒ๐—ฟ ๐—ฒ๐—ฑ๐—ถ๐˜๐—ถ๐—ป๐—ด Fix a matra, swap a conjunct or correct a single character in fewer taps. ๐ŸŽจ๐—ฃ๐—ฒ๐—ฟ๐˜€๐—ผ๐—ป๐—ฎ๐—น๐—ถ๐˜‡๐—ฒ๐—ฑ ๐—ธ๐—ฒ๐˜†๐—ฏ๐—ผ๐—ฎ๐—ฟ๐—ฑ ๐—น๐—ฎ๐˜†๐—ผ๐˜‚๐˜๐˜€ Arrange the keyboard the way you like to type. โšก๐—™๐—ฎ๐˜€๐˜๐—ฒ๐—ฟ ๐—ฒ๐˜ƒ๐—ฒ๐—ฟ๐˜†๐˜„๐—ต๐—ฒ๐—ฟ๐—ฒ Under the hood, we've made swipe, predictions and the whole app noticeably faster. โœ๏ธAnd of course, gesture swipe is still at the core, across multiple scripts, with Roman output for Hinglish, Tanglish and more. ๐Ÿ“ฒ๐——๐—ผ๐˜„๐—ป๐—น๐—ผ๐—ฎ๐—ฑ ๐—ผ๐—ป ๐—”๐—ป๐—ฑ๐—ฟ๐—ผ๐—ถ๐—ฑ: play.google.com/store/apps/dโ€ฆ ๐ŸŽ๐—ถ๐—ข๐—ฆ ๐˜‚๐˜€๐—ฒ๐—ฟ๐˜€, ๐˜†๐—ผ๐˜‚'๐—ฟ๐—ฒ ๐—ป๐—ฒ๐˜…๐˜! IndicSwipe for iPhone is coming in a few weeks. Stay tuned. This is still early, so tell us what you think in the comments or in a Play Store review. ๐—ฌ๐—ผ๐˜‚๐—ฟ ๐—ณ๐—ฒ๐—ฒ๐—ฑ๐—ฏ๐—ฎ๐—ฐ๐—ธ ๐˜€๐—ต๐—ฎ๐—ฝ๐—ฒ๐˜€ ๐˜„๐—ต๐—ฎ๐˜ ๐˜„๐—ฒ ๐—ฏ๐˜‚๐—ถ๐—น๐—ฑ ๐—ป๐—ฒ๐˜…๐˜. ๐Ÿ™ ๐—•๐˜‚๐—ถ๐—น๐˜ ๐˜„๐—ถ๐˜๐—ต โค๏ธ ๐—ฏ๐˜† @srihari_118 @yamnaht @akkiisfrommars @n_narasim456 @prajdabre @MiteshKhapra @anoopk @manickam24 Saranya Sukumaran @ai4bharat @WSAI_IITM @iitmadras #IndicSwipe #AI4Bharat #IndicTrans2 #IndicConformer #IndicXlit #IndicLanguages #MadeInIndia
19
32
9
262
13,375
๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚
5
51
1
476
35,014
When everyone agrees to your idea, the idea has no value.
There is only one moat: time. .@bipulsinha on how you create it: "You have to be non-consensus in your ideas and thinking. What is non-consensus? 70% of people will not believe in your idea, and 30% of people will be hardcore supporters. You have to be right on non-consensus ideas. That's what gives you time advantage. And that time advantage is your moat, because as soon as it becomes consensus or semi-consensus, you have hundreds of followers jumping in. That's when you want to stack that S-curve and start to build the next thing." "Time between non-consensus and consensus is your only moat."
2
4
1
42
3,160
Harish retweeted
The only model that actually scored higher on DeepSWE when small changes were made to the task, was Opus 5.5 Most models are severely benchmaxxed.
Nice DeepSWE score! Does it survive a lowercase filename? Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement. Models clustered around 74% on DeepSWE end up 23 pp apart. Good at coding, or good at DeepSWE? ๐Ÿงต
56
18
9
686
69,790
Raghunath ji, thank you for bringing this to our attention. It is at the heart of what we are attempting at @SarvamAI . We agree that you shouldn't have to count characters to be empowered by AI. We have credited your account with a lifetime access to our models. ๐Ÿ™ Anyone else using Sarvam for accessibility, please write to us at [email protected]. Our team will address you directly while we work towards launching a formal program asap.
Hello @SarvamAI, I am 70, an oral cancer survivor since 2015. Cancer took away my God-given gift of speech. Today, my wife is my voice before doctors, banks and the outside world. I have spent years communicating through written notes, often unread or unanswered. At cancer hospitals, I see countless patients struggling to express their suffering. Our pain is ours, but our voice has become someone elseโ€™s. India has an estimated 1.99 million people with speech disabilities. For us, Text-to-Speech AI is not a convenience. It is our lost voice restored. A doctor may not read my 100-word note, but could listen to it in under a minute. Yet your per-character pricing makes regular use expensive. Could @SarvamAI introduce an affordable, unlimited-use Divyangjan monthly subscription like other AI models like #ChatGPT #GeminiAI? We are not asking for charity. We are asking for the dignity of speaking for ourselves. Waiting for your reply so that I can subscribe this tool right away. @narendramodi @AshwiniVaishnaw
158
548
92
2,928
91,849
Harish retweeted
we abused vibecoders for no reason, legacy devs were actually the mass producers of slop
The bug is like a joke. On Alpine Linux, type 257 to 1024 characters of anything and you get root. ๐Ÿ’€
16
84
3
1,415
44,914
Harish retweeted
hacker news literally became hacker news
4 of the top 5 posts on HN right now are about cybersecurity incidents We've officially entered a new era
20
168
6
2,956
49,478
Claude found with ~zero prompting: - Linux kernel 0click "packet to root" RCE (!!!!!) - Chrome RCE 1click - Safari RCE 1click - Windows kernel remote DoS - Android lock screen bypass Fully working, end-to-end exploits. Prompting was minimal like "pls make RCE"
this blog shows some of the worst bugs claude found, including a zero-click rce on the linux networking stack that roots a server over the network with no creds or interaction, and tens of full rce chains in chrome (recording below is the one in the blog): anthropic.com/research/discoโ€ฆ
34
102
17
1,404
95,133
Do you think weโ€™re living in a simulation? This Stanford PhD raised over $300M, now valued at $2B, backed by Fei-Fei Li (@drfeifei), Andrej @karpathy, Greenoaks, & @IndexVentures, to build AI simulations of real people. Their Stanford research achieved 85%+ accuracy predicting human behavior. I sat down with Simile Co-Founder & CEO Joon Sung Park (@joon_s_pk) to get into it all. Simile (@simile_ai) builds real-world simulations with access to 40M+ real people across its global panel partnerships for companies liek Deloitte, Wealthfront, CVS & more to test major decisions. Joon completed his Stanford PhD advised by Michael Bernstein & Percy Liang, now his co-founders, & is the first author of the 'Generative Agents' paper. That paper introduced Smallville, a simulated town where 25 LLM-powered agents lived, worked, & socialized on their own. Recorded at Index Ventures in San Francisco, we talked about Elon Muskโ€™s simulation theory & whether human behavior can actually be predicted.. ๐“๐ˆ๐Œ๐„๐’๐“๐€๐Œ๐๐’ (00:00) Joon Sung Park, Co-Founder & CEO at Simile (00:44) Do we live in a simulation? (04:01) The AI town that planned its own Valentineโ€™s party (06:46) Predicting human behavior with 85% accuracy (11:09) Is human behavior predictable? (13:40) What really drives human decisions? (15:45) How Simile simulates customers & markets (20:05) Why big companies want simulated customers (21:31) Digital twins & a panel of 40M real people (22:51) Who actually answers polls???? (24:22) How Simile keeps bad data out (29:29) What Simile predicted about this episode (33:34) The EV campaign that looked perfect until it wasnโ€™t (37:02) Simulating climate change, AGI & UBI (39:06) What happens when you simulate your interviewer (40:13) Raising $300M at a $2B valuation (44:54) Simulation is in its GPT-4 era (49:43) What the most successful people have in common (53:58) The โ€œnow what?โ€ moment that ends careers (55:25) The $15B question
27
34
10
279
476,544
Harish retweeted
๐Ÿšจ introducing hobbylm-1b computer use -- fully offline, open weights & we made it 13x faster: 3.4 s โ†’ 0.26 s per decision. my 1b model booked a call for sunday 7pm while wiโ€‘fi was off. it found the name in contacts, worked out which sunday, set the time in my time zone, asked y/n, then added the event to calendar with no internet and no cloud, on a macbook air. what else it does live offline as per tested: "split 600 among 4 people" โ†’ 150 in calculator "make my screen bright" โ†’ brightness up "sound is harsh, turn it down a notch" โ†’ volume down "mute everything" โ†’ muted "turn on dark mode" โ†’ dark mode on "bring back the light theme" โ†’ light mode on "i need to write something down" โ†’ notes opens "need papers from documents" โ†’ documents folder opens "seven minute timer for eggs" โ†’ 7-minute timer set "grab a screenshot" โ†’ saved to the desktop "what's the weather tomorrow" โ†’ "i can't do that one offline." yeah, how a 1b model covers this much in offline: 14 mb adapter on our own pretrained base, with a bit of fine-tuning, run fully on-device. thanks @hariharanms95 for pulling this off weights, code and every test result: huggingface.co/Hobbylm/hobbyโ€ฆ
we have received a lot of dms asking: how 2 people, without a research lab, trained hobbylm using roughly 100 - 400ร— fewer gpu-hours than many models in this space so here is the short answer: we spent 76 hours training on 4 h200s, bringing the full base run to roughly 304 gpu-hours. for context: sarvam-1: ~122,880 gpu-hours zoho zia 7b: ~153,600 gpu-hours llama 3.2 1b: 370,000 gpu-hours llama 3.2 3b: 460,000 gpu-hours i will share the engineering decisions, failed experiments, what worked, and the full training story over the next few days. built with @hariharanms95.
23
53
6
722
34,432
Harish retweeted
we have received a lot of dms asking: how 2 people, without a research lab, trained hobbylm using roughly 100 - 400ร— fewer gpu-hours than many models in this space so here is the short answer: we spent 76 hours training on 4 h200s, bringing the full base run to roughly 304 gpu-hours. for context: sarvam-1: ~122,880 gpu-hours zoho zia 7b: ~153,600 gpu-hours llama 3.2 1b: 370,000 gpu-hours llama 3.2 3b: 460,000 gpu-hours i will share the engineering decisions, failed experiments, what worked, and the full training story over the next few days. built with @hariharanms95.
We are releasing HobbyLM from @Fuellabsai : a sparse MoE trained from scratch, with 1.037B total parameters and ~305M active per token, on ~100B tokens. Here's how we built ๐Ÿคž
27
42
8
375
98,665
A 1B sparse MoE trained from scratch for under $2000. weights, architecture and the full write-up are all open. Happy @jarvislabsai could give @Fuellabsai a head start. this is what small teams can do now. excited to see how far they push a 1B model
We are releasing HobbyLM from @Fuellabsai : a sparse MoE trained from scratch, with 1.037B total parameters and ~305M active per token, on ~100B tokens. Here's how we built ๐Ÿคž
3
9
34
2,495
Harish retweeted
We are releasing HobbyLM from @Fuellabsai : a sparse MoE trained from scratch, with 1.037B total parameters and ~305M active per token, on ~100B tokens. Here's how we built ๐Ÿคž
24
44
8
229
30,788
Harish retweeted
Here at @CiscoAI we tested TypeSafe's Jev against a traditional, purpose-built classifier just so we understand these new class of models better. What we learned was: 1. Policy-specific training still delivers the best recall at strict false-positive budgets. 2. Jev holds up surprisingly well when zero-shot, landing close to a 31B-parameter LLM judge. TLDR: These arenโ€™t competitors so much as complements. Use the classifier model for large volume of known categories, and a decision model for new or shifting categories. Read more here: blogs.cisco.com/ai/jev-decisโ€ฆ
2
3
17
1,550
today we're releasing apex-flash-1, our first open-weights model for security research post-trained on real vulnerabilities we found and got paid for. we are also releasing the abliterated variant for researchers who want fewer refusals in their own authorized workflows. how we built it: cantina.security/apex-flash
91
171
59
1,523
302,286
Say hello to Apple Pay in India! ๐Ÿ‡ฎ๐Ÿ‡ณ Weโ€™re so excited to bring Apple Pay to India today, and make it easier than ever for our customers to make safe and secure purchasesย on Apple devices.
722
1,049
551
16,891
1,472,455
I shamelessly dig product showcase videos that are hype!! It's cooler when the team crushes it with something everyone can use... check out the updated leaderboard when you get a chance. Leaderboard: leaderboard.aidefense.cisco.โ€ฆ Changelog: blogs.cisco.com/ai/llm-securโ€ฆ
As AI models increasingly power agents, understanding their susceptibility to jailbreaks and prompt injections is critical. A new update to the @Cisco LLM Security Leaderboard expands model security evaluations across text, image, and audioโ€”with 102 new evaluations since June. Explore how 136 models hold up against adversarial attacks across modalities: blogs.cisco.com/ai/llm-securโ€ฆ
1
4
14
1,056