Abhinav Upadhyay

Well, Focusrite’s mythically garbage ability to write a driver that can stomach Windows 11 has triggered a new voice conversion model training run already. None diffusion based voice conversion methods like sovits and rvc are even more sensitive to this, where they need a very clean input file, but to deal with at all is very annoying. Basically, with even relatively small quality problems with the input sound, you can get what I’ll call a voice conversion fail - that is, the result is not similarly as bad as in the input, but dramatically worse. With a diffusion model, there really is no excuse not to be robust as hell against that. So a new training run with a decent percent chance of background noise, reverb, dropout, clicks, and 5 other ways to crap on audio. You do it all on CPU in the data loader workers and it doesn’t even cost anything. Bonus, make the source and target speaker the same and you have a dope speech enhancement model as well.

1

Found And Played This Trust And Safety Game On Stanford Internet Observatory - T&S Consortium... A Very Cool Way To Learn About Various Decision A T&S Team Would Probably Make, And Testing Oneself On Personal Decision One Would Make If On Such Team Try Your Hands On This Game Here https://lnkd.in/dfY6WaQ3 #trustandsafety #onlinesafety

6

🚀 Ready to elevate your LLM-powered applications? 🚀 Discover how Chaitanya AI can help you achieve your goals with advanced AI techniques. 📞 Schedule a call: https://lnkd.in/dC8st677 📚 Explore our case studies: [Chaitanya AI Case Study](https://lnkd.in/dkuCXAqJ) Let's drive innovation together! #AI #LLM #MachineLearning #NLP #ChaitanyaAI

10

👉 Indian language RAG with Cohere multilingual embeddings and Anthropic Claude 3 on Amazon Bedrock Media and entertainment companies serve multilingual audiences with a wide range of content catering to diverse audience segments. These enterprises have access to massive amounts of data collected over their many years of operations. Much of this data is unstructured text and images. Conventional approaches to analyzing unstructured data for generating new content rely on the use of keyword or synonym matching. These approaches don’t capture the full semantic context of a document, making them less effective for users’ search, content creation, and several other downstream tasks. Read more... 👇 #amazonbedrock

5

❗ Interference in Indian Elections, powered by AI ❗ Yesterday, OpenAI announced that they detected and blocked covert operations to influence Indian Elections. Foreign firms used AI-generated content favouring one side, spread through bots on social media, to sway public opinion. Similar operations for other parts of the world were also blocked. Bots have been around for a while but AI has closed the automation loop. It now enables the mass production of unique, persuasive, and highly tailored content that's far harder for platforms to detect. The report mentions techniques used for detecting/blocking foul actors: defensive model designs, using another gpt to moderate content, API access control, domain monitoring. And the techniques used by attackers to avoid detection: rotating the topics for every API call, using multiple keys and ofc proxy servers, vpns. Interestingly, attackers also used gpt to debug the code for their bot scripts and automations, which was detected using query analysis. While OpenAI should be applauded for it's elaborate safety measures, there are other uncensored models available for use. This attack was reported by a service used by the attackers; not the platforms the content was posted and consumed on, those using some different service would have probably gone undetected. What do you think about these risks? Should platforms be responsible for detecting and mitigating such threats? How can we design AI models that are inherently resistant to misuse? #AI #Disinformation #Elections #Ethics #OpenAI #TechForGood #Democracy

2

Is Protobuf the future default for everything? Timothy Chen & I chatted with Buf CTO Akshay Shah about the future of Protobuf, their new Kafka-compatible streaming layer, and everything he learned scaling Uber that led him to build Buf. This episode covered a pain point I've repeatedly experienced while working at Snyk & see at many other companies: managing cross-team migrations and building a productive internal API ecosystem. Here's a great example: You have two services (A and B), and A depends on B. How do you modify an API or Event in A without requiring B to do a complicated lock-step migration? These are the challenges in any at-scale organization where a project that should take 30-60 days turns into months as you try to navigate alignment and delivery. They're mundane papercuts that turn into hundreds of hours of lost productivity. I love what Buf are doing to help make this a problem of the past.

48

3 Comments

Here's an even better explanation of why LLM's can't count characters. Understanding how these models work is key to using them effectively.

4

1 Comment

🚨 Preparing for Fault Proofs on Base Mainnet 🚨 Exciting changes are on the horizon for Base Mainnet, with fault proofs expected to go live on 30th October! 🗓️ Fault proofs are a crucial step towards decentralization, allowing for a more open and community-driven approach to validating L2 state. Here’s how they’ll enhance the network: 🔍 Permissionless Output Proposals: In traditional L2s, only centralized proposers could submit output roots. With fault proofs, anyone can propose output roots for Base's state, breaking down centralization barriers. ⚔️ Permissionless Challenges: Anyone can now challenge a claim if it appears faulty or fraudulent, enhancing security and trust. These changes mean that anyone can withdraw funds from Base to L1 without depending on centralized actors. 🌐💸 What's Changing for Withdrawals: Withdrawals before the upgrade: Must wait the standard 7-day challenge period before finalization. Withdrawals during/after the upgrade: Will be proven using the Fault Proof system, with a general 7-day finalization period (potential delays if challenged). Note: bridge.base.org now redirects to Superbridge and other "Superchain bridges" for managing deposits and withdrawals. Check out our docs for more details. 🌍🚀 These updates are setting the stage for a more decentralized, secure, and community-governed Base Mainnet. Let’s embrace this exciting step towards decentralization! #Blockchain #DeFi #L2 #Decentralization #Crypto #Ethereum #BaseMainnet #Bridging #Layer2 #Web3 #SmartContracts #FaultProofs #DeFiSecurity #BlockchainScaling #EthereumEcosystem 

3

The truth is too multifaceted to be contained in a five-line summary. - Stanley Kubrick, filmmaker I experienced Kubrick's frustration while working on my story on Sarvam AI, a venture from India that has built a large language model (LLM) with 2 billion parameters. In contrast to the Big Tech LLMs that are proficient in English, Sarvam's Gen AI offerings are available in 10 Indian languages. It has built its LLM to serve large enterprises in India, for whom technology is still a cost (not an investment). Hence, Sarvam's pricing is affordable, like AI voice agents at Re 1/ minute. And its APIs are economically priced too. That's my ~five-line summary, but it doesn't even begin to describe the excitement I experienced while unravelling how the LLM was built, with its genesis in AI4Bhārat, IIT Madras. At The Arc, we serve our readers by first talking about the business fundamentals of a start-up's premise. So, Part 1 of my story is on the Re 1 price proposition. https://lnkd.in/gaWRsgSy But to even gauge how this is possible from a cost perspective, I had to dig deeper into the technical nous of building the LLM. That is Part 2, which is coming soon. For now, soak it in: a venture from India has built an LLM based on our inherent multilingual strengths and diversity, rather than ape the west with another English-proficient LLM. I got goosebumps when I reflected on what Sarvam AI's founders have managed to put together in the past three years. Look forward to your comments and feedback on my story, (and suggestions when you would like Part 2 of the Sarvam report). 😊

10

3 Comments

Reddit's metadata store handles 100,000 req/sec at p99 of 17 ms 🤯 I dissected the blog and compiled the learnings into a quick 7-minute video and covered their entire architecture, trade-offs, and the data partitioning strategy. give it a watch - youtu.be/b9zh0SFbkqM This is the 53rd video in the real-world system design series where I dissected engineering blogs and made them easier to understand. ⚡ Subscribe to the channel and follow along; I keep things no-fluff. #AsliEngineering #SystemDesign

670

5 Comments

Imagine if technical book publishers increased the royalty rate, to say 50% and reduced the size of the teams publishing the books. Technical books would actually make sense to write at that royalty rate. Sell a 100 dollar book @50/book is 50K for only 1k books. This could be an actual job for many talented authors. Instead, the technical book publishing industry is disrupting itself because it is illogical to give a publisher 95% of the royalty when you can do it completely yourself and actually get a decent reward for a year of your life. Greed and inefficiency has consequences.

7

Come on, #Google! Don't replicate the mistakes from #WebRTC! While working on #HTTP3 I understand that Google #Chrome does not trust user-added Certificate Authorities for #QUIC. Of course it makes sense if Google plans to make changes to QUIC while deploying it. But it's been since 2019... https://lnkd.in/eVWuxK4J This ecosystem deserves #openness.

16

Wondering what top research got published this last week? 📑🤖🤔 So essentially, 1. "AI Model StyleFeatureEditor gives highly detailed image editing results!" 2. "We have a Universal Quantum Chemistry dataset to find new drugs now!" 3. "Stylebreeder clusters diverse art styles from 6.8M images on Artbreeder, enhancing personalized AI-driven artistic expression." 4. "LongRAG addresses the limitations of traditional RAG by using long retrieval units" 5. "Language model tuned to avoid toxic language in English will also avoid it in other languages!" 6. "YOUDREAM Auto-Generates Perfect 3D Creatures from Text!" From cutting-edge research on Artificial Intelligence to advances in Machine Learning, Computer Vision, and Neuroscience, we distill the latest technical papers into bite-sized insights you can apply today! For more subscribe at https://lnkd.in/gJ_wJxAC

21

1 Comment

Thrilled to support 💪 Perplexity return citations param - https://lnkd.in/g7kkztxX (+4 more updates 👇) ⚡️ Azure - return response headers ✅ Streaming - break out of infinite model loops 🎉 Instructor - fix completion_cost calc 👀 [PR] logging by API key - https://lnkd.in/gAG5YKQa

6