Cyrus Shaoul

Linxin Gu great article. thanks for sharing. Pls let us know folks in Boston area can help with your journey

5

3 Comments

Excellent work on this topic. Also wildly dangerous LLM scaling. https://lnkd.in/gYEiMj9c

30

2 Comments

Some VC investors believe that brick and mortar - referring to a physical presence of some kind - should be avoided at all costs. Yet, the track record of clinic and hybrid care businesses in health care indicates otherwise. ==> One Medical (sold for $3.9 billion to Amazon), Oak Street (bought by CVS for $10.6 billion), Lifestance (IPO at $7.5 billion market cap), and PE-backed Premise Health (acquired by OMERS Private Equity for $1.1 billion) If a clinic is indicated as the optimal way to deliver care for that patient population, there are creative ways to build and operate them. Full article below with advice and strategies on how to build and operate hybrid clinics from my teammate Anarghya Vardhana from Maveron, Amir Dan Rubin from OneMedical Group, Steven Eidelman from Modern Animal and more!

13

1 Comment

Science-fiction is inherently speculative. Author George #Orwell of ‘#1984’ predicted concepts such as doublethink, little earbuds and society’s obsession with the black mirror screens that dominate our lives — and according to the BBC, novelist and short story author JG Ballard, of “High-Rise” and “Crash” fame, has predicted the current tool and cultural fixation of #GenerativeAI in the modern society — only he didn’t just speculate them within a story, he is actually one of the first use-cases of a generative AI used in a creative context — Ballard was so fascinated with the growing technological advancement of computers that he used one to compose poems, essentially the precursor to what we have now with ChatGPT. Today, the debate about the creative use of generative AI and its value on creative industries is an extremely fraught and controversial topic, but if controlled in a space that doesn’t replace human labor or try to equate itself to traditional art forms, it is interesting to see what computers can do and how they are creatively limited in comparison. Ballard’s two poems, The Yellow Back Novels and Machine Gun City are generated, and both include notes on how Ballard achieved producing them. Regardless of quality (Ballard himself probably wouldn’t describe these two poems as his best work), it is interesting to see primitive versions of the technology that we utilize on the daily today, and how these moral conversations that concern them would continuously evolve.

1

1 Comment

A fun summer activity with your kids (and for adults too): websim.ai is powered by Anthropic's Claude 3.5 Sonnet and creates an infinitive multiverse of a simulated browser environment. The more creative you are with it (hint: with the URLs), the more it plays with you (and makes you lol)! Our son, Levi, has created some really fun pages with it. Warning: parody only - if you aren't a fan of TV shows like "Rick and Morty", this might offend you. I think this shows us all how the future of entertainment is going to evolve to infinitely personalized content, especially as video models like OpenAI's Sora become super inexpensive as the inevitable technology curve of compute prices falling while rising compute (and solar and/or fusion energy) plays out. Read "The Singularity is Nearer", Ray Kurzweil's new book (launched yesterday), to think more about how that curve will undoubtedly fall, as it always does and has. Tri-Nation Quantum Fusion Investment Portal: https://lnkd.in/gU9b9v-D AI Evolution Timeline: https://lnkd.in/gDHxAWyy Quantum Computer Solves P vs NP Problem, Mathematicians Retire En Masse - news from Dimension X-23 on Rick's Interdimensional Cable News: https://lnkd.in/gj8UKj-p

6

A new artificial intelligence system, called OpenScholar, is promising to rewrite the rules for how researchers access, evaluate, and synthesize scientific literature. At OpenScholar’s core is a retrieval-augmented language model that taps into a datastore of more than 45 million open-access academic papers. When a researcher asks a question, OpenScholar doesn’t merely generate a response from pre-trained knowledge, as models like GPT-4o often do. Instead, it actively retrieves relevant papers, synthesizes their findings, and generates an answer grounded in those sources. One particularly damning finding involved GPT-4o’s tendency to generate fabricated citations—hallucinations, in AI parlance. When tasked with answering biomedical research questions, GPT-4o cited nonexistent papers in more than 90% of cases. OpenScholar, by contrast, remained firmly anchored in verifiable sources. In expert evaluations, OpenScholar’s answers were preferred over human-written responses 70% of the time, but the remaining 30% highlighted areas where the model fell short—such as failing to cite foundational papers or selecting less representative studies. These limitations underscore a broader truth: AI tools like OpenScholar are meant to augment, not replace, human expertise... Source: https://lnkd.in/engfpdKj Image credits to @VentureBeat. #artificialintelligence #scholar #research #science #academia

66

34 Comments

RAG ... a necessary but *not* sufficient element of reliable NIR ... "Neural Information Retrieval". As revolutionary as LLMs are, NLU is not solved. The old NLU benchmarks are obsolete, but the challenge of genuinely *understanding* and fully responding to specific user intent remains before us. And many of the old tools will find new applications in a new synthesis. Call it Neural Information Retrieval. https://lnkd.in/eD72UBQ4

18

1 Comment

On the latest Vital Signs, Nikhil Krishnan and I had a ton of fun chatting with Dr. Atul Butte. Dr. Butte is a professor at UCSF, Chief Data Scientist of the UC Health System, and a co-founder of numerous healthcare startups. In our conversation, Dr. Butte shared his takes on healthcare AI opportunities, challenges, hype, and more. A few highlights: ⚡LLM use cases, data, and growth LLMs are already helping to solve the “drudgery” aspects of medicine, such as automatically drafting SOAP notes and patient messages. Dr. Butte is excited about future applications of LLMs that can help with the “cognitive” aspects: with difficult medical problems an LLM might be able to stitch together disparate pieces of information to generate a potentially novel solution. He also highlights that prompt engineering will become the next “coding language” as so much of leveraging LLMs is knowing how to ask them the right questions. Additionally, Dr. Butte discusses how since 2016, UCSF has been working to deidentify its patient notes so that it can query them to answer clinical questions and ultimately enhance patient care. ⚡The need for patient-facing AI As patients are increasingly able to access their own personal health data in nicely structured formats (e.g., EHRs, wearables), they desperately need more tools to help them interpret and make decisions using that data. Dr. Butte understands that these patient-facing solutions are challenging to get funded but he emphasizes that patients need to receive the benefits of AI sooner. ⚡Evaluating healthcare AI solutions In order to evaluate AI solutions, we often put human specialists – such as medical professionals – on the same tasks. Understanding that specialists often disagree with each other, a significant challenge arises in figuring out how we can evaluate AI considering we can’t just “average” all the specialist opinions. Dr. Butte points out that there still remains a massive gap in AI evaluation tools. Additionally, there are so many organizations that want to have a say in how healthcare AI is regulated that the landscape is uncertain. An awesome conversation with an expert on all things AI meets healthcare. Listen to the full episode below: Spotify: https://spoti.fi/3Wr3643 Apple: https://apple.co/4acwgan

82

8 Comments

Legislative processes are boring and a lot of people don't pay attention to laws until they get passed, so I'll apologize in advance for talking about a California bill currently in committee. It's called SB 1047. The bill recently passed the California Senate. To become law, it needs to pass through committee next week, get a vote of support in the Assembly this summer, and then be signed by Newsom. The problem is that SB 1047 is a deeply flawed bill that could kill the AI industry in California. Jeremy Howard, a guy I respect a lot in AI, has written about it here: https://lnkd.in/gcQy6T_g Here are his very good points: - The definition of “covered model” is overly broad and could encompass a wide range of open-source models. - The requirements for developers are overly burdensome and could discourage open-source development. - The potential for civil penalties for AI developers could have a chilling effect on research and innovation. - The ability to levy fees could create a barrier to entry for small businesses. There are better ways to support the responsible development of AI. I hope the tech community will rally to oppose this bill and create something more reasonable in its place.

6

Anthropic releases Fascinating paper diving deeper into model inner workings. The paper studies how neurons in language models respond to specific features and how this changes as the models get bigger. Key Terms: Monosemantic Neurons:** Neurons that respond to one specific feature (i.e. a specific word). Polysemantic Neurons:** Neurons that respond to multiple features. Main Points: 1. Detection: Create methods to identify neurons that only respond to one feature. 2. Inhibition: Develop ways to stop neurons from only responding to one feature during training. Why It Matters: Reducing monosemantic neurons can make AI models more flexible and understandable, leading to better overall performance. Paper terms Neuron: A basic unit in a neural network that processes information. Language Models: AI models that understand and generate human language. Feature: A specific characteristic or element that a neuron responds to. This approach can lead to more reliable and easier-to-understand AI models. Kelly Cohen Josette Riep #artificialintelligence #llm

8

1 Comment

Re-reading this from 2017 (a good 7.5 years ago) and way pre-ChatGPT event. What has changed, and what is now clear ? From the original post: ‘’🧬Biological evolution has taken us to the point we enjoy today over millions of years. 💡Technological evolution is different; moves in exponential manner. 🛠️Technological evolution has made it possible to educate more people, increase life span and significantly reduce the number of people under the poverty line. 🦾Technological evolution is now moving us to a new era. The Era of the AI. Man and intelligent machine will coexist and merge. More than 50% of current jobs will be lost. Are we prepared for this ? I doubt that we are at this point but we’ve got to try. What are the key questions we should be asking ourselves and our world leaders and government representatives ?’’ #ai #tech #genai #superintelligence #humanity #ethics

14

2 Comments

This analysis, at least anecdotally makes sense to me. In many situations, infringe and holdout is the least painful option for large infringers. The system enables it. - The plaintiff must allocate the time and $$ to assert, taking time and $$ away from selling and developing products. Even if they engage litigation financiers they are limiting their upside potential. - The infringer continues to infringe while employing stall tactics or simply relying on the painfully slow, overburdened legal system. - The cost benefit analysis for infringers - "infringe and pay lawyers" typically outweighs the cost of design around or license. - Enhanced damages and injunctions are rare. - Litigation insurance products reduce risk all around. "The Marcum report shows that there is indeed something wrong with the patent system in the United States but it’s not excessive litigation or patent trolls. The real problem is the inability of patent owners to enforce their rights against infringers. Given the difficulty of obtaining injunctive relief and the limited enhancement of damages, it is unsurprising that predatory infringement has become a standard business practice when, at worst, all an infringer can be expected to pay is a reasonable royalty and, at best, nothing at all."

5

François Chollet is a prominent French software engineer and artificial intelligence (AI) researcher, currently serving as a Senior Staff Engineer at Google. Chollet is best known for creating Keras, a widely-used deep learning library that was released in 2015. In 2019, he devised ARC. The Abstraction and Reasoning Corpus (ARC) is a benchmark designed to measure AI's ability to acquire skills and progress towards human-level intelligence. Learn more about it in this guest post with Jurgen Gravestein of Teaching Computers How to Talk. We need to continue to devise new benchmarks for our reasoning and agentic machines and tools. The ARC prize and context relating to AGI is of particular interest to me. Read more and check out some of the links in this article. Including videos by Machine Learning Street Talk (MLST) and Dwarkesh Patel.

20

In the LLM world, I'm seeing more and more people moving to the Mistral models as the goto base models for fine-tuning. What's the secret sauce here ? What makes Mistral so good ? Know one on the outside knows for sure, but I have a guess as to how to detect these kinds of great models. In this post, I conjecture why Mistral 7b works so well–based on an analysis with the open-source weightwatcher tool, and drawing upon the Didier Sornette’s theory of Dragon Kings. A Dragon King is a special , mechanistic outlier that drives the dynamics when certain systems exhibit a power law distribution. Now I doubt that the Mistral team looks at their training this way--but you can. You can use the weightwatcher tool to look at models during or after training to see how well each layer is converging, and adjust your hyperparameters and layer sizes accordingly. --- WeightWatcher is a one-of-a-kind must-have tool for anyone training, deploying, or monitoring Deep Neural Networks (DNNs). Reach out to get a sneak peak at the upcoming WeightWatcher Pro--the SAAS service you need to evaluate you fine-tuned LLMs

83

12 Comments

I'll do even a braver poll than last, but only because of this Eric Topol, MD article in Science Magazine :) Multimodal large AI models will be used in medical forecasting (disease risk, onset, trajectory) A) now B) soon but at population level only C) sooner than in diagnostics D) later than in diagnostics E) never https://lnkd.in/gGBHSptv

7

I am very excited to share the pre-print study on Atropos Health ChatRWD beta results(link to paper below). ChatRWD beta outperformed ChatGPT 3.5, Google Gemini 1.5 Pro. and Anthropic Claude 3 on answers to clinical questions when reviewed by independent physicians. OpenEvidence also outperformed the Big Tech LLMs by a wide margin. Most important to us, ChatRWD performed highest on Physician trust and Best Answer and was able to answer questions asked 94% of the time. As the Gen AI cycle continues to move at a rapid pace we firmly believe that "Trust" and "Accuracy" are going to be critical, and that publishing your performance on both is table stakes. It is also clear to me that in the quest to be able to answer 100% of questions with Clinical grade accuracy training LLMs solely on the literature will not be enough. We need to produce new evidence in the form of high-quality transparent studies using appropriate methodology. We are adding beta users post this announcement, so please sign up to try it, and we will have more to come on full launch this summer. Exciting times and hope everyone has a great July 4th. https://lnkd.in/es5gWK84 #generativeAI #LLM #Realworldevidence

225

13 Comments