Maxime Allard

Excited to share what Nithin Parsan and I have been working on at YC! At Reticular (YC F24), we're pioneering the future of explainable AI for biology. Check out the post and our first technical blogpost linked in the comments to learn more about what we're focused on.

82

6 Comments

#RAG DOES NOT REDUCE HALLUCINATIONS IN #LLMS: Think I've called this out a few times, but the validation is nice - https://lnkd.in/eAi6zxW2. The article makes two well established points: 1) "Hallucination is a problem of reasoning and not relevance! Any amount of relevant text fed through RAG to a machine will retain the original perplexity and entropy in the system to hallucinate independent of the text"; and 2) "The initial challenge arises from the retrieval phase’s reliance on semantic similarity, which is quantified by inner product spaces or cosine similarities. This metric, while capturing semantic closeness, does not differentiate between factually accurate and inaccurate information." This should be clear to everyone, but the industry still shills flimsy RAG scaffolding anyway (shovels to gold miners).

104

34 Comments

One year ago, we started Epsilla (YC S23) with a big dream. Today, as we celebrate our first birthday, I want to share the story that brought us here. Many people ask why I left my job as a senior director at TigerGraph, took a huge pay cut, and founded my own company. Since I was a kid in elementary school, I've watched the story of Bill Gates building the Microsoft empire, and dreamed of one day building my own company. This dream grew stronger when I came to the US in 2014 pursuing my master's degree at Cornell, and applied to Y Combinator with my roommate, Hao Ma, with a simple anonymous social app idea. Even though it didn’t work out then, it planted the seed in my heart. During my time at Cornell, I met 4 other entrepreneurial students, Landice Gao, Xiaoyan Wu, Jiaqi Su, and Kejia Tao, and started a company focused on indoor navigation. I wished it could have gone further, but as an international student, I needed a job that could sponsor my working visa to stay in the US. So I joined TigerGraph at 2015, a pre-series A startup at that time. Since day 1, I was trying my best to grow as fast as I can. At TigerGraph, I learned a lot from industry leaders like Amol Ghoting, Li Chen, Mingxi Wu, Ruoming Jin, Yu Xu, Like Gao, Tim Zhang, Songting Chen, Shuo Yang, and many other talented colleagues. Most importantly, I met my future co-founders, Jing Qin and Eric Y., and our future advisor and angel investor, Jay (JieBing) Yu, PhD. When ChatGPT launched in November 2022, it blew my mind. I thought such advanced technology would appear in 2045 (according to "The Singularity Is Near"), but it was already here. However, I quickly realized that large language models (LLMs) didn’t have memory and their knowledge was fixed. Our friend Weimo Liu mentioned vector database to us, and I saw the huge potential of combining AI with data. This led to the birth of Epsilla, aiming to build the data and knowledge foundation for LLMs. Becoming a CEO from a technical background has been tough, especially for an introvert like me. I am so proud of my growth so far. In the past year, I’ve talked to more people than I have in my entire life. I’m grateful for all the interesting people I’ve met around the world, which wouldn’t have happened without starting this journey. Looking ahead, I’m very dedicated to the success of Epsilla. Our journey is just beginning, and I’m excited to see what the future holds for us. Thank you to everyone who has been part of this journey—our customers, partners, team, investors, and supporters. Your belief in our vision keeps us going every day. Here’s to one year of Epsilla, and many more to come! 🚀✨

105

32 Comments

Kolmogorov–Arnold Networks seem like a paradigm shift in ML architectures. Curious to see how they evolve. "While MLPs have fixed activation functions on nodes ("neurons"), KANs have learnable activation functions on edges ("weights"). KANs have no linear weights at all -- every weight parameter is replaced by a univariate function parametrized as a spline. " https://lnkd.in/gdPk_qCQ

29

Teaching SQL to anyone is a pain. Teaching complex SQL to bench scientists is a supreme sacrifice of human productivity. In my imagination, I see a lot of red. It's time murdered without a pang of regret. Many interfaces in biopharma research are beginning to use LLMs. A prime example is text-to-SQL conversion. One of our customers wanted to not have to teach complex SQL to their team. The database schema underlying their tool is fairly complex. More than 30 tables. Lots of complex joins to be done. As one would expect it is hard to teach bench scientists to do complex joins over tables to get any answer. Turns out with a little bit of fine-tuning and nuanced UX design 60-70% of queries could be done just using text based prompts. One does need an understanding of the queries that will be done (easy to get) and some fine-tuning of the relevant LLM. Coming Thursday, I am hosting a webinar around using LLM in real-world R&D settings. Join us for similar use-cases and a technical deep-dive.

113

11 Comments

Want to know why Google is losing the AI race? They are a company of an absolutely mind-blowing tech with terrible product managers. Just let the business/program folks run the engineering, cut the fluff and red-tape and get faster to market please. Using Gemini just does not feel right and integrating to Google is a friction one needs to overcome. Literally they invented transformers and 80% of the tech innovation in AI and still losing. DeepMind has the best AI researchers compared to any other foundational AI giants, hands down. I am a big believer in Google (due to tech), but they really need to do something about their product managers.

21

10 Comments

Synthflow AI recently raised a chunky seed round of $7.4M. Their pitch is that their no-code platform makes it easier than ever for SMEs (which usually have low tech capabilities) to deploy voice agents for a multitude of use cases. And they do have a pretty good flow builder. At Hooman Labs, we too are striving to make voice agents as simple as possible to deploy. But we believe that while self serve is important to enable users to play around with voice agents, it is more critical to build a suite of plug and play agents (including necessary integrations) which don't need a lot of tinkering by the user. To do this, a voice agent provider needs to go deep into a vertical and understand the workflows of their user, end to end. We're doing this for e-commerce and building agents which solve calling needs of brands across the life cycle of the customer, from pre-sales to post. Some of the use cases we're working on are consultative selling, delivery updates, abandoned cart conversion, return requests and warranty claims. If you run an e-commerce brand and are looking to automate your calling function, would love to talk to you and understand how we can help. Tarun Rathore #voiceai #conversationalcommerce #ecommerce

31

The Open Medical LLM Leaderboard on Hugging Face dropped last Friday (thanks Geoff Kwitko for the heads up), and it aims to track, rank and evaluate the performance of large language models (LLMs) on medical question answering tasks (see link below). It evaluates LLMs across a diverse array of medical datasets, including MedQA (USMLE), PubMedQA, MedMCQA, and subsets of MMLU related to medicine and biology. The leaderboard offers a comprehensive assessment of each model's medical knowledge and question answering capabilities. The datasets cover various aspects of medicine such as general medical knowledge, clinical knowledge, anatomy, genetics, and more. They contain multiple-choice and open-ended questions that require medical reasoning and understanding. More details on the datasets can be found in the "LLM Benchmarks Details" section in the link. Ps. Good to know that the GPT-4 model is leading the way (it is what I use on the Medical Coding and Documentation agent). Digital Human Assistants #llmhealthcare

7

LLMs and Optimization: An Unbeatable Combo!

19

Approach to Evaluating AI Projects I see two major problems with AI startups nowadays: 1. There are too many of them, and they often resemble one another. 2. Every startup claims to integrate AI into their thesis, making it unclear what actual AI technology they possess. How does our fund tackle this? We have identified that a project must meet at least one of the four criteria below to have the potential to build something truly significant, rather than just raising a seed round on the current hype. Criteria: 1. Unique proprietary technology. This is very rare, akin to a new ChatGPT. The main value lies in the technology, and no one else has it yet. 2. Unique dataset on which the technology (not necessarily unique) is trained. This could be a publicly available model, but it is fed with a unique dataset. For example, taking all the designs created in AutoCAD and feeding them to a neural network, enabling it to create new and unique products or structures. 3. Team with unique expertise. This is straightforward - a brilliant team with extensive AI experience, preferably not their first project. 4. Non-market growth conditions. For example, a project that already has a customer base for the product, secured contracts, development grants, etc. What do you think of this approach? Share your thoughts in the comments 👇

7

4 Comments

Using experimental data to improve performance of generative AI models can be a powerful way to improve protein fitness over rounds of design. This is why we now support fine-tuning our AI models using assay data such as but not limited to thermostability, enzyme activity, and antibody binding affinity. Get in touch to learn more on how our software recommends point mutations for property optimization based on experimental results!

32

Been heads down for a while, but figured I would post about something super exciting. We have been working on how to create real time experiences and a really important new feature just came out yesterday... When working on real time experiences, its all about latency. The problem with latency stems from all sorts of different angles, especially when you deliver an experience live within a meeting that has a ton of latency components. One of the longest tails in latency is your AI... and thats why this release is so critical. Anthropic released a way to cache your big prompts (and ours are huge). By caching it, you save it in memory thus significantly reducing the time to process a big question by 80%. This is HUGE! In order to deliver a realistic real time experience, you need to be at sub 1 second respond rate, and this is finally a reality. That means you can have a sophisticated conversation or reaction from an AI agent in real time to fundamentally facilitate and run a meeting. The next frontier is here and its only getting faster everyday. Get ready for the next evolution of product and can't wait to show the next generation of what we have been working on. #ai #anthropic #aiprojectmanager

20

"Are you an AI company?" might be one of the most expensive questions in biotech right now. Forcing this narrative can lead to biotech companies picking sides and making costly mistakes that lead them to over or under-invest in the ways they advance their R&D and programs. Rather than picking a side ('traditional biotech' versus 'AI company'), the sharpest leaders and most competitive companies will find ways to weave in elements from both angles. And, over time, the conversations will hopefully evolve away from "which one are you?" and more towards whether companies are covering their blind spots and leveraging every edge they can be, in their race to getting drugs to market. https://lnkd.in/dkCtnDtK

53

2 Comments

Next up in the series of posts about starting a startup for non-technical founders: 2. Finding a co-founder How do you even find one in the first place? For anyone who's tried, it's exceptionally hard. Let's start with the framework. You want to find: (a) Someone that has complimentary skills. Think one plus one equals three. (b) Someone equally passionate about the same problem. If you’re not equally excited about the challenge you’re tackling, it’s going to be hard to keep the momentum going. (c) Someone that is available - Timing. You both need to be available and ready to commit at the same time. After you go down this list, you might be lucky to know one or two people in your immediate network to work with. In terms of what you can do to navigate this stuff, there were 2 things that I thought would give me the best shot. One was immersing myself in a founder-dense environment like SF. Another was co-founder dating. The best matching platform I’ve seen is from YC. They have tons of profiles and make it simple to meet a lot of people quickly. None of this is easy, but all of it is possible. I will be sharing learnings from the process that Tennison C. and I went through (very different; I met 8 new people vs Tennison met 80 new people)—hope they’re helpful for anyone experiencing this.  --- Yesterday, we covered "Things you can’t say as a non-technical founder." For the remainder of the week, I will write about 3. Importance of a deep trial period 4. Investing in your co-founder relationship

26

4 Comments

Can LLMs act as Doctors? Meet AMIE, a chatbot developed by Google that's making waves in the medical field. It was trained on a huge dataset of medical conversations and questions about symptoms and diagnoses. It simulates doctor-patient talks to improve its diagnostic skills. Users chat with AMIE about their symptoms, and it asks questions and provides possible diagnoses. In tests, AMIE had a 90% accuracy rate, better than the 77% for human doctors. Source- https://lnkd.in/g_z9Yq5v #AI #healthcare

48

6 Comments

Wiz - $0 to $23B in 4 Short Years + The Actual Reason Why Alphabet Inc. is Happy to Make it Happen Alphabet is likely buying Wiz for $23B - an insane amount of value created in just 4 years since its founding in 2020. For context, it took OpenAI 7+ years to get to $20B. Alphabet buying Wiz at $23B puts a fantastic comp out there for investors to underwrite to and justify insane valuations (kinda kidding, kinda not). Currently, Wiz is over $500M in ARR (~46x) but plan to get to $1B in ARR by the end of '25, which would put it at a bit more reasonable 23x, but still a steep price and certainly an indication of conviction in the strategic value of Wiz. So why could Alphabet spend such an insane amount to buy Wiz at such a hefty price? Because the actual golden goose here is winning market share in the ~$300B/yr Cloud Services market that is growing at an insane 21% YoY. The means nearly $60B in new TAM will be up for the taking the next 12 months alone.. Today, Google Cloud (11%) is a distant 3rd in market share behind Amazon Web Services (AWS) (31%) and Microsoft Azure (25%), but it is growing the fastest at 28% YoY. If buying Wiz can help Google accelerate their growth and continue to gain share, they'll be able to capture a bigger slice of that fast growing Cloud Services TAM. And with a market that big growing that fast, there is almost no price too high. Every marginal point of market share is worth at least $15B of EV (assuming 5x). Alphabet currently trades at 5.5x revenue so assuming they can get a marginal $5B of ARR out of the deal over the next few years, the economics make sense almost immediately. Google Cloud is also Alphabet Inc.'s best hedge against AI's potential disruption to their core search business as more people turn to generative AI instead of that trusty search bar to answer their questions. This will be Google's biggest acquisition ever by nearly 2x, and it will be a massive, much-needed replenishment of funds into the venture ecosystem (Index Ventures Sequoia Capital Greenoaks, Insight Partners, Andreessen Horowitz) Fun aside - Wiz has some wildly notable angels on the cap table like Howard Schultz and Bernard Arnault. If this deal goes through it's going to be a huge win for a bunch of founders, investors, and employees and an overall great mark for the early-stage technology ecosystem! All of this is assuming Lina Khan lets it happen!

83

11 Comments