Ally Nakamura
2 followers · 8 following · 347 views
on the atlas — 42
- Staring into the abyss as a core life skill85 savers
- Countering misuse of AI: September 2026 / Anthropic \ Anthropic14 savers
- Claude4 savers
- Launching Standard Machines1 savers
- Drug Discovery Has No Magic Wands - by Daphne Koller - a16z1 savers
- How I’d rewrite The Courage to be Disliked - by Chris Lakin4 savers
- Going Founder Mode On Cancer - by Elliot Hershberg8 savers
- Mythos and Engels' Pause - by Doug O'Laughlin1 savers
- AI for Alzheimer’s1 savers
- Project Glasswing: Securing critical software for the AI era \ Anthropic13 savers
- Home | Substack2 savers
- Network Bio Q1 2023 Investor Update - by network.bio1 savers
- Cancer - Sytse.com6 savers
- What a time to be an oncologist - by Olivia Webb Kosloff1 savers
- Agentic payments memo | Natural2 savers
- Clawed - by Dean W. Ball - Hyperdimensional9 savers
- Towards Humanist Superintelligence | Microsoft AI2 savers
- Opinion | Stop Worrying, and Let A.I. Help Save Your Life - The New York Times1 savers
- Google Gemini1 savers
- Advancing Claude in healthcare and the life sciences \ Anthropic1 savers
- The Final Offshoring14 savers
- The Path to Medical Superintelligence | Microsoft AI1 savers
- I Reverse Engineered ChatGPT's Memory System, and Here's What I Found! - Manthan1 savers
- Cognition | Don’t Build Multi-Agents4 savers
- Making Sense of Memory in AI Agents – Leonie Monigatti1 savers
- A reimbursement framework for artificial intelligence in healthcare | npj Digital Medicine1 savers
- Rebuild Biotech for the AI Era | Benchling3 savers
- Every Reframe Starts with a Refusal | by Mike Maples, Jr. | Aug, 2025 | Medium1 savers
- The Heresy of Breakthrough Startups | by Mike Maples, Jr. | Sep, 2025 | Medium1 savers
- Thatch - Not Boring by Packy McCormick1 savers
- The Perfect AI Assistant - Caelin Sutch1 savers
- Home | Substack1 savers
- Biotech’s Lost Archive | IFP2 savers
- The Baby Died. Whose Fault Is It? | WIRED1 savers
- A Story of Slavery in Modern America - The Atlantic2 savers
- Home | Substack1 savers
- How to be More Agentic - by Cate Hall - Useful Fictions33 savers
- Face it: you're a crazy person - by Adam Mastroianni27 savers
- Advice to Young People, The Lies I Tell Myself - jxnl.co19 savers
- America Is Missing The New Labor Economy – Robotics Part 1 – SemiAnalysis6 savers
- Hierarchical Navigable Small Worlds (HNSW) | Pinecone5 savers
- Why are AI Products Doomed to Fail? | by Jeremy Arancio | Nov, 2023 | Towards AI3 savers
highlights — 52
As we measure error rates and harms from A.I., we need frameworks to make apples-to-apples comparisons between what human doctors do on their own today and what A.I.-enabled health care does tomorrow. In these early days, we should favor a “walk before you run” strategy, starting with using A.I. to handle administrative paperwork tasks before focusing all our energy on higher-stakes tasks like diagnosis and treatment.
Opinion | Stop Worrying, and Let A.I. Help Save Your Life - The New York TimesWhen agents become primarily responsible for an increasing number of transactions, some of the patterns by which they transact will feel human-esque, but many of the ways they transact will not as they are optimized around slightly different reward functions.
Agentic payments memo | NaturalAnd health startups - companies like Function, Sword, Prenuvo, Oura, Eight Sleep, TrueMed, and many existing and yet-to-be-founded others - will build products in the knowledge that their success will be based on the quality of their product, and not their ability to sell into companies. They will even compete for dollars that go to carriers by default today.
Thatch - Not Boring by Packy McCormickWorking with Thatch over the past few years, and writing this piece, what I’ve come to realize is that the best way to fight cancer, and many of the other diseases that ail and kill us, is to rewire incentives so that more people can access the tools, preventative measures, and treatments that exist today, and will be invented rapidly in the coming years.
Thatch - Not Boring by Packy McCormickOne of the most powerful attributes of being an Aggregator is that controlling demand allows you to reshape supply.
Thatch - Not Boring by Packy McCormickWhen America was founded, it was a frontier. Each citizen had minimal possessions. Over the centuries, Americans have accumulated financial and relationship capital, giving themselves something to lose. This wealth went parabolic in the 1960s, and since then, we’ve become afraid of taking risks: marriage comes six years later, schooling lasts 5 years longer.
The index mindset - by John Luttig - luttig's learningsCapex spending for AI contributed more to growth in the U.S. economy in the past two quarters than all of consumer spending
Will data centers crash the economy?What if the socially anxious were calibrating to avoid being DISliked?
Social anxiety isn't about being liked - by Chris Lakina willingness to make peace with ambiguity. Being ok with being confused because it will eventually work itself out.
Home | SubstackThey are not afraid of what happens, because no matter what happens, they know it will be fundamentally okay.
People who demand nothing of youWhat I discovered by casting a wide net was that I have very little ability to predict how useful a call will be in advance.
How to be More Agentic - by Cate Hall - Useful FictionsAsk for things that feel unreasonable, to make sure your intuitions about what’s reasonable are accurate (of course, try not to be a jerk in the process). If you’re only asking for things you get, you’re not aiming high enough.
How to be More Agentic - by Cate Hall - Useful Fictionsradical agency is about finding real edges: things you are willing to do that others aren’t, often because they’re annoying or unpleasant.
How to be More Agentic - by Cate Hall - Useful FictionsHere are some top Considerations for Building Production-Grade RAG Decoupling chunks used for retrieval vs. chunks used for synthesis Structured Retrieval for Larger Document Sets Dynamically Retrieve Chunks Depending on your Task Optimize context embeddings
Building Performant RAG Applications for Production - LlamaIndex 🦙 0.9.23The main production challenge for RAG systems besides answer relevance and faithfulness is speed
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIRagas, mentioned in the previous section, uses faithfulness and answer relevance as the generated answer quality metrics and classic context precision and recall for the retrieval part of the RAG scheme.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AI5% increase in the faithfulness metrics, meaning the fine-tuned GPT 3.5-turbo model made a better use of the provided context
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIthe pairwise score was improved by 4% by cross-encoder finetuning.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIThe main approaches to response synthesis are: 1. iteratively refine the answer by sending retrieved context to LLM chunk by chunk 2. summarise the retrieved context to fit into the prompt 3. generate multiple answers based on different context chunks and then to concatenate or summarise them.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIQuery routing is the step of LLM-powered decision making upon what to do next given the user query — the options usually are to summarise, to perform search against some data index or to try a number of different routes and then to synthesise their output in a single answer.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIPinecone automatically indexes vectors,
Pinecone vs. Chroma: The Pros and Cons | by Woyera | MediumPinecone’s architecture is built to scale with growing data and traffic demands
Pinecone vs. Chroma: The Pros and Cons | by Woyera | MediumPinecone offers blazing-fast search capabilities
Pinecone vs. Chroma: The Pros and Cons | by Woyera | MediumA bit more sophisticated case is CondensePlusContextMode — there in each interaction the chat history and last message are condensed into a new query, then this query goes to the index and the retrieved context is passed to the LLM along with the original user message to generate an answer.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIllamaindex offers an efficient fuzzy matching based solution for this case. In case you have not heard of fuzzy matching, this is an incredibly powerful string matching technique.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIThe next big thing about building a nice RAG system that can work more than once for a single query is the chat logic, taking into account the dialogue context, same as in the classic chat bots in the pre-LLM era.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIBoth libraries have this functional implemented — as a Multi Query Retriever in Langchain and as a Sub Question Query Engine in Llamaindex.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIIf the query is complex, LLM can decompose it into several sub queries
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIIn LlamaIndex there is a variety of available Postprocessors, filtering out results based on similarity score, keywords, metadata or reranking them with other models like an LLM, sentence-transformer cross-encoder, Cohere reranking endpoint or based on metadata like date recency
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIA relatively old idea that you could take the best from both worlds — keyword-based old school search — sparse retrieval algorithms like tf-idf or search industry standard BM25 — and modern semantic or vector search and combine it in one retrieval result.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIThere is also the reversed logic apporach called HyDE — you ask an LLM to generate a hypothetical response given the query and then use its vector along with the query vector to enhance search quality.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIAnother approach is to ask an LLM to generate a question for each chunk and embed these questions in vectors, at runtime performing query search against this index of question vectors (replacing chunks vectors with questions vectors in our index) and then after retrieval route to original text chunks and send them as the context for the LLM to get an answer.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIAn efficient way to do that in case of a large database is to create two indices — one composed of summaries and the other one composed of document chunks, and to search in two steps, first filtering out the relevant docs by summaries and then searching just inside this relevant group.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AIHNSW is not the best index in terms of memory utilization.
Hierarchical Navigable Small Worlds (HNSW) | PineconeAlthough higher parameter values provide us with better recall, the effect on search times can be dramatic. Here we search for 1000 similar vectors (xq[:1000]), and our recall/search-time can vary from 80%-1ms to 100%-50ms.
Hierarchical Navigable Small Worlds (HNSW) | PineconeHigh M and efSearch values can make a big difference in recall performance
Hierarchical Navigable Small Worlds (HNSW) | PineconeThe creators of HNSW found that the best performance is achieved when we minimize the overlap of shared neighbors across layers. Decreasing m_L can help minimize overlap (pushing more vectors to layer 0), but this increases the average number of traversals during search. So, we use an m_L value which balances both. A rule of thumb for this optimal value is 1/ln(M) [1].
Hierarchical Navigable Small Worlds (HNSW) | PineconeHNSW is a natural evolution of NSW, which borrows inspiration from hierarchical multi-layers from Pugh’s probability skip list structure.
Hierarchical Navigable Small Worlds (HNSW) | PineconeSkip lists work by building several layers of linked lists. On the first layer, we find links that skip many intermediate nodes/vertices. As we move down the layers, the number of ‘skips’ by each link is decreased.
Hierarchical Navigable Small Worlds (HNSW) | PineconeA proper search index, optimised for efficient retrieval on 10000+ elements scales is a vector index like faiss, nmslib or annoy, using some Approximate Nearest Neighbours implementation like clustring, trees or HNSW algorithm.
Advanced RAG Techniques: an Illustrated Overview | by IVAN ILIN | Dec, 2023 | Towards AII really think that we could be entering a third epoch of computing. The microchip brought the marginal cost of compute to 0. The internet brought the marginal cost of distribution to 0. These large models actually bring the marginal cost of creation to 0.
What Builders Talk About When They Talk About AI | Andreessen HorowitzAI and ML are reshaping nuclear engineering, with a focus on predicting Critical Heat Flux (CHF).
AI and ML play a pivotal role in nuclear engineering, with a recent breakthrough being the AI and ML benchmark for predicting Critical Heat Flux | by Multiplatform.AI | Nov, 2023 | MediumTop 3 secure practices: company-approved GenAI, no confidential data in prompts, and no personally identifiable customer data
Study: Over half of users operate generative AI without employer approval | by Multiplatform.AI | Nov, 2023 | Medium70% of global workers lack formal training in safe and ethical generative AI use
Study: Over half of users operate generative AI without employer approval | by Multiplatform.AI | Nov, 2023 | Medium64% of workers have passed off generative AI work as their own.
Study: Over half of users operate generative AI without employer approval | by Multiplatform.AI | Nov, 2023 | Medium- The model’s efficiency and consistency make it a valuable asset for credit monitoring and risk assessment.
Introducing Slope TransFormer: AI for Banking Language Understanding | by Multiplatform.AI | Nov, 2023 | Medium- It achieves remarkable speed and accuracy, outperforming Plaid in transaction labeling.
Introducing Slope TransFormer: AI for Banking Language Understanding | by Multiplatform.AI | Nov, 2023 | MediumNot only does SoT provide considerable speed-ups across 12 LLMs, but it can also potentially improve the answer quality on several question categories. SoT is an initial attempt at data-centric optimization for inference efficiency, and further underscores the potential of pushing LLMs to think more like a human for answer quality.
Skeleton-of-ThoughtThis all raises the question: what if we could pack multiple fine-tuned models into a single deployment by reusing the common base model parameters?
📝 Guest Post: Meet LoRAX: The Open Source System that Serves 1000s of Fine-Tuned LLMs on a Single GPU*the majority of the GPU resources are being allocated to serving additional copies of the same base model parameters for every fine-tuned model!
📝 Guest Post: Meet LoRAX: The Open Source System that Serves 1000s of Fine-Tuned LLMs on a Single GPU*