Luv Kothari

@paulg Go Illini!!

I bet few people there know it, but YC wouldn't exist without UIUC. My father went there from England on a Fullbright, and came back determined to move to America.

@paulg Twitter is also turning into TikTok with most posts now with a video.

There is something deeply wrong with Twitter. It has always been rough here, but in the past year it has become an even worse kind of nasty. Do you think maybe it's time to try to turn things around, Elon?

Enterprise AI agents need context—not scraping scripts. Cortex Knowledge Extensions are now GA, bringing verified, licensed content from @AP, @WashingtonPost, @StackOverflow, and more into your Snowflake-powered AI. 📚🧠 Use Snowflake Intelligence—now in Public Preview—to connect this content to agents or prompt it directly through Cortex AI APIs using your LLM of choice. No hacks. No headaches. Just answers you can trust: bit.ly/46OQ8Ck

@levie This is spot on!!

AI Agents are *the* big topic for enterprises right now. In most conversations with IT leaders, this is the main thing they are figuring out strategies for. There is a lot of excitement and momentum, and equally a realistic sense of the work that’s ahead. Here are some of the major topics that come up when companies start to think about deploying AI Agents at scale: 1. Agent interoperability: enterprises don’t have their data in just one environment. Salesforce may have their CRM records, Box their documents, Workday their HR data, or ServiceNow their HR and IT workflows; and plenty of Agentic use-cases will span these systems. As an industry we will need to design more interaction patterns for how AI Agents talk to each other and exchange data in the future. 2. Over-permissioned systems: lots of enterprises deal with the fact that software has “overshared” information over the years. This is fine when a human can’t possibly find everything across the tools, but a huge liability when Agents can get access to nearly anything, instantly, and return them to a user. Software products will have to take permissions and access controls more seriously, and we may need new AI Agent permission paradigms to help customers deal with this. 3. Data needs to be in an AI-ready environment: decades of technology being adopted in an enterprise means decades of systems that have important data but are not in environments that AI Agents can easily talk to. There will need to be a continued modernization push to modern, cloud environments, as retrofitting these systems will almost certainly not work. 4. Compliance: given we’re insanely early in the adoption of AI, most industries still haven’t figured out their official stance on where AI can provide suggestions or make decisions. Most regulated industries (like healthcare, life sciences, or financial services) are still in the early innings of developing shared standards for this, and some will need regulatory clarity to be able to do far more. 5. AI Agents executing the work: for a while the standard has been to have a human in the loop when AI is invoked in a workflow. But this becomes increasingly impractical for all steps of the work as workflows become more agentic. Companies will have to go through their own processes for setting their own policies on how much and when you can hand off to an AI Agent. 6. AI model quality: we continue to see rapid breakthroughs from the latest AI models on performance and capabilities, but for certain high value workflows we’re still not 100% there. For instance, you wouldn’t want to vibe code huge pieces of software for a production system, or rely only on AI for mission critical decisions fully just yet. This means we still need constant progress in the model space to get us there. 7. Change management: this will probably realistically take the longest. You can’t take a process that’s been run for decades or a century and expect it to radically transform overnight. In particular, most companies still are working what the best order is of deploying AI to get the biggest impact relative to the change required. 8. Business models of AI Agents: at the moment AI Agents still have a variety of business models being tested. Customers can somewhat feel that we’re early as an industry in getting to the ultimate pricing model of agents, and more stability here will likely go a long way. In all, enterprises are rapidly trying to figure out this space, but the tech industry as well will need to continue to make major progress on these topics to accelerate the transformation. Exciting times ahead!

@maxbranzburg Let’s go, India!!

India has boundless potential for crypto adoption and innovation It's a country of builders who are often leapfrogging the US and Europe into the next era - something I've seen firsthand as they skipped credit cards to go straight to digital payments Time to bring India onchain

Congratulations to @Cisco for launching AI Defense today! We worked with Cisco to build the guardrail models that power this end-to-end solution to safeguard enterprise AI transformation. blogs.cisco.com/news/you-cant-…

As we launch into 2025, our Public Sector team reflects on our work over the past year. We focused on strengthening national security by equipping U.S. agencies with the best-in-class commercial AI technologies. Highlights include: ✅Deepening ties to Missouri’s geospatial and defense tech industry ✅Work with the U.S. government to develop and deploy new AI capabilities, including a T&E framework for @DeptofDefense and novel perimeter security technology with @DIU_x and @usairforce ✅Collaborations with @anduriltech, @CISAgov, & Holos ✅Product launches like Defense Llama and major upgrades to Scale Donovan Read more 👉 scl.ai/public-sector-…

Want to design better AI agents? Take notes from code writing systems. Techniques include — Multi-agent — Tool choice — Underlying model — Diff format — Innovative Signals — Code retrieval + knowledge graphs — LSP — Fault localization Let's dive deeper with real examples: 1/10

📣 Introducing Llama 3.2: Lightweight models for edge devices, vision models and more! What’s new? • Llama 3.2 1B & 3B models deliver state-of-the-art capabilities for their class for several on-device use cases — with support for @Arm, @MediaTek & @Qualcomm on day one. • Llama 3.2 11B & 90B vision models deliver performance competitive with leading closed models — and can be used as drop-in replacements for Llama 3.1 8B & 70B. • New Llama Guard models to support multimodal use cases and edge deployments. • The first official distro of Llama Stack simplifies and supercharges the way developers & enterprises can build around Llama to support agentic applications and more. Details in the full announcement ➡️ go.fb.me/229ug4 Download Llama 3.2 models ➡️ go.fb.me/w63yfd These models are available to download now directly from Meta and @HuggingFace — and will be available across offerings from 25+ partners that are rolling out starting today, including @accenture, @awscloud, @AMD, @azure, @Databricks, @Dell, @Deloitte, @FireworksAI_HQ, @GoogleCloud, @GroqInc, @IBMwatsonx, @Infosys, @Intel, @kaggle, @NVIDIA, @OracleCloud, @PwC, @scale_AI, @snowflakeDB, @togethercompute and more. With Llama 3.2 we’re making it possible to run Llama in even more places, with even more flexible capabilities. We’ve said it before and we’ll say it again: open source AI is how we ensure that these innovations reflect the global community they’re built for and benefit everyone. We’re continuing our drive to make open source the standard with Llama 3.2.

Let's say you’re a pre-revenue tech startup with a shiny new product that’s almost ready for primetime. You’ve landed a couple of hot leads that could make or break your credibility—and get investors pumped for your next round. Here’s where I’d focus my energy right now: 1. Value Prop Check Nail down your sales pitch, but be ready to adapt as real-world feedback rolls in and shows you where your message hits—or misses. You won't get it right the first time, so plan to iterate. 2. Scrappy Sales Team Assemble a small, agile team that includes you (the CEO or Founder) and one or two sharp folks from engineering with a decent customer UI. At this stage, it's as much about being good learners than about being good closers. 3. Test and Tweak Sales Process Start with a basic sales playbook, but treat it like a lab experiment—track everything, learn fast, and adjust on the fly. *Always* ask tons of questions first before you pitch your product. 4. Find Your First Believers Go after the early adopters who feel the pain your product solves, and dig deep into their world to refine your ideal customer profile. 5. Pilot with Partners Lock in some pilot deals for low initial dollars that let you test your product in the wild—think of it as R&D with potential revenue. Always charge something to test willingness to pay. 6. Fast Feedback Loops Set up quick and dirty feedback loops between your customers and the product team—iterate like your runway depends on it. Learn and adapt. 7. Insights Over Instincts Make every customer interaction a chance to gather intel—use it to tweak your product and pitch until you’ve got a winner. Focus on these 7 key areas, execute relentlessly, and stay adaptable. Nail this phase and you'll transform from pre-revenue to market-ready before you know it.

Google DeepMind have produced a program that in a certain sense has achieved a silver-medal peformance at this year's International Mathematical Olympiad. 🧵 deepmind.google/discover/blog/…

Unlock the secrets to mastering generative AI with our AI Readiness Report. Based on a survey of over 1,800 experts, discover how leading organizations are fine-tuning models, tackling data challenges, and implementing robust evaluation frameworks to drive AI success.

@coinbase seriously..

We’re excited to see @cohere launching Command-R today! Enterprises can test and evaluate Cohere’s models on Scale GenAI Platform, incorporate these models into their RAG pipelines, and use these models to build custom GenAI apps built on their own data 💪

Scale GenAI Platform is now available on the @Azure Marketplace! Azure customers can purchase Scale GP through the Azure Marketplace, leveraging their existing @Microsoft Azure Consumption Commitment (MACC). 👉 scl.ai/sgp-on-azure 👉 scale.com/genai-platform

We are excited that @scale_AI has now added Cohere’s Command model and Rerank technology to their GenAI platform to customize business applications with fine-tuning and use RAG for accurate responses. This move will help unlock the powerful potential of AI for enterprise customers.

📣Today, we're launching a major update to the Scale GenAI Platform, the full-stack platform to transform your data into customized enterprise-ready Generative AI applications. Any use case, any model, anywhere 🧵 scale.com/blog/genai-pla…

Being at the 95th percentile in fitness ≈ being 30 years younger. (Image via @andrewchen.)

Here's a list of my favorite LLM papers I read this month: 1/ Zephyr LLM - a 7B parameter model with competitive performance to ChatGPT on AlpacaEval; applies distilled supervised fine-tuning to improve task accuracy and distilled direct performance optimization on AI feedback data to better align the model; shows performance comparable to 70B-parameter chat models aligned with human feedback. arxiv.org/abs/2310.16944 2/ LLMs Meet New Knowledge - presents a benchmark to assess LLMs' abilities in knowledge understanding, differentiation, and association; benchmark results show arxiv.org/abs/2310.14820 3/ Llemma - an LLM for mathematics which is based on continued pretraining from Code Llama on the Proof-Pile-2 dataset; the dataset involves scientific paper, web data containing mathematics, and mathematical code; Llemma outperforms open base models and the unreleased Minerva on the MATH benchmark; the model is released, including dataset and code to replicate experiments. arxiv.org/abs/2310.10631 4/ LLMs for Software Engineering - a comprehensive survey of LLMs for software engineering, including open research and technical challenges. arxiv.org/abs/2310.03533 5/ Self-RAG - presents a new retrieval-augmented framework that enhances an LM’s quality and factuality through retrieval and self-reflection; trains an LM that adaptively retrieves passages on demand, and generates and reflects on the passages and its own generations using special reflection tokens; it significantly outperforms SoTA LLMs. arxiv.org/abs/2310.11511 6/ Instruct-Retro - introduces Retro 48B, the largest LLM pretrained with retrieval; continues pretraining a 43B parameter GPT model on an additional 100B tokens by retrieving from 1.2T tokens. arxiv.org/abs/2310.07713 7/ Overview of Factuality in LLMs - a survey of factuality in LLMs providing insights into how to evaluate factuality in LLMs and how to enhance it. arxiv.org/abs/2310.07521 8/ LLMs Represent Space and Time - discovers that LLMs learn linear representations of space and time across multiple scales; the representations are robust to prompt variations and unified across different entity types; demonstrate that LLMs acquire fundamental structured knowledge such as space and time, claiming that language models learn beyond superficial statistics, but literal world models. arxiv.org/abs/2310.02207 9/ StreamingLLM - a framework that enables efficient streaming LLMs with attention sinks, a phenomenon where the KV states of initial tokens will largely recover the performance of window attention; the emergence of the attention sink is due to strong attention scores towards the initial tokens; this approach enables LLMs trained with finite length attention windows to generalize to infinite sequence length without any additional fine-tuning. arxiv.org/abs/2309.17453 10/ Retrieval meets Long Context LLMs - compares retrieval augmentation and long-context windows for downstream tasks to investigate if the methods can be combined to get the best of both worlds; an LLM with a 4K context window using simple RAG can achieve comparable performance to a fine-tuned LLM with 16K context; retrieval can significantly improve the performance of LLMs regardless of their extended context window sizes; a retrieval-augmented LLaMA2-70B with a 32K context window outperforms GPT-3.5-turbo-16k on seven long context tasks including question answering and query-based summarization. arxiv.org/abs/2310.03025 You can find more interesting papers for this and past months here: github.com/dair-ai/ML-Pap…

Good intro/overview to LLM retrieval (RAG) from @c_viernickel and Bryan Zhu on Scale's blog — it's not RAG vs. fine-tuning, it's tuning for RAG; for interesting tasks you want both: