"Key Context" Substack covering AI & chips. Reached #1 new bestseller first 24 hrs
Subscribe taekim.substack.com
Tell the truth. Be accurate. Be right
Is this you? As a journalist, you can create a free Muck Rack account to customize your profile, list your contact preferences, and upload a portfolio of your best work.
Claim your profile
Articles by Tae Kim
An Interview with d-Matrix CEO Sid Sheth on Why 3D-Stacked Memory Compute Is a Game Changer
Startups are where much of the technology industry’s innovation happens. One of the AI industry’s biggest current problems is the so-called “memory wall,” which is the growing gap between how fast AI chips can perform computations and how fast they can be fed the data to compute on. d-Matrix, a small startup founded in 2019, may have one of the most intriguing solutions to the memory wall.
An Interview with Nvidia’s Ian Buck on Why GPUs Keep Winning
Agentic AI is changing everything including at Nvidia. Key Context spoke with Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, on Tuesday at the OCP Global Summit in San Jose. We discussed why companies continue to choose GPUs over rival chips, how agentic AI is changing the industry, the dynamic between open and frontier AI models, and how service providers can fit 30% more GPUs into the same data center footprint.
OCP Day One: AMD CEO Sees Strong Demand for AI Inference and CPUs
AMD CEO Lisa Su is emphatic and optimistic about the current demand for AI. “We’re in a world where computing equates to intelligence, and if you have a chance to be more intelligent, why wouldn’t you try to get more capability?” she said Monday during her OCP Global Summit keynote in San Jose. “Every time I look at these [demand] charts, they tend to keep going up and to the right.”
An Interview with Microsoft’s Scott Guthrie on Azure’s New AI Chip, ‘Insatiable’ AI Demand and Today’s AI Model Price War
Microsoft announced plans to expand its Azure Maia AI chip program to multi-gigawatt-scale deployments starting next year. Microsoft is amping up its AI chip game. On Monday, the company announced plans to expand its Azure Maia AI chip program to multi-gigawatt-scale deployments starting next year. Microsoft also said its next-generation Maia 300 chip will deliver 40% more tokens per megawatt than leading GPU-based competitors and will begin production deployments in 2027.
How to Make More Money in a Volatile Market
You can take advantage of investors' emotional swings. Note: Next week I’ll be traveling to San Jose for the OCP Global Summit, which has become one of the most important AI infrastructure conferences of the year. All the big name companies will be there. If you’re attending, please say hello. I also have several interviews in the pipeline for our Substack that I’m extremely excited about, filled with the deep technical detail we’re known for. Look for them over the next two weeks.
Spatially deterministic nucleation of 2D semiconductors by etching flux
|
Abstract Nucleation control is fundamental to semiconductor deposition, governing when, where and how crystalline materials form. Conventional strategies can regulate nucleation density and, through area-selective growth, confine deposition to designated regions1,2,3,4,5,6,7,8,9,10,11,12,13.
By JeongWon Bourdais Park, Jiyun Kim, Sumin Kang, Jieun Oh, Saeyoung Oh, Seunghye Shin, Jaehyun Lee, Minsu Kim, Gunho Moon, Mingu Kang, Min-Gyu Kim, Junhyeok Kim, Gichang Noh, Dongyoung Kim, Seongdae Kwon, Chengyun Hong, Minseok Choi, Yongjoon Lee, Joonghoon Choi, Guanning Shao, Seung Jae Kwak, Tae Kim, Jeehwan Kim, Yongjoo Kim, Sung-Yool Choi, Heejun Yang, Jonghwan Kim, Byungjo Kim, Woo-Hee Kim, Chang-Soo Lee, Kibum Kang, Youngmin Sunwoo, Seonghwan Jo, Donghyeop Lee, Changhun Eom, Seohyun Jeong, Chaejeong Yun, Donggeon Park, Hyeongjin Lim, Junhyuk Tak
|
Nature
Verified
How to Protect a Customer-Support Assistant Like Cdiscount's From PII Exfiltration
TL;DR: Imagine you build a customer-support assistant for a service like Cdiscount. Without a firewall, the LLM can leak PII through tool calls and responses. Resk (LLM Firewall server) is a self-hosted, OpenAI-compatible gateway that adds RBAC, policy filtering, and tool gating. This tutorial shows you how to deploy it step by step. The scenario You are building a customer-support assistant for a service like Cdiscount. The assistant answers order status, refunds, and account questions.
OpenAI's Undeniable Resurgence and the Coming Liftoff of Personal AI Agents
Impressions of Muse, Dots, and Instinct. I’ve seen enough. The personal AI agents released over the past month will be a huge paradigm shift for computer usage and a major positive inflection point for technology. On Tuesday, OpenAI hosted its annual DevDay and announced a flurry of products.
The dynamic landscape of T helper cell plasticity and Th17–Treg instability in periodontitis Original
Abstract Periodontitis is a chronic inflammatory disease initiated by a dysbiotic microbiome, yet tissue destruction is primarily driven by an exaggerated host immune response. CD4⁺ T helper (Th) cell subsets play important roles in regulating periodontal inflammation and osteoclast-mediated bone resorption through distinct cytokine networks.
Comparative characterization of the mucosal eukaryotic virome across independent whole-metagenome shotgun sequencing cohorts
Abstract The human mucosal eukaryotic virome remains incompletely characterized because methodological heterogeneity limits direct comparisons across independent whole-metagenome shotgun sequencing (WMS) datasets.
Graph RAG: where it actually breaks
Neo4j with a working schema: two days. Cypher traversal for the relationships I needed: another day or two, once I knew what I was querying. The graph structure, once committed, stayed mostly stable. That's not where things kept breaking. The routing problem When someone asks "what companies are exposed to semiconductor export restrictions?" the system needs to figure out whether to do graph traversal, vector search, or some combination. Getting this wrong doesn't produce an obvious error.
Jensen Huang Press Q&A: 'Nvidia Stock Is Undervalued'
On Monday, Nvidia CEO Jensen Huang answered questions from reporters at The Korea Society’s annual gala in New York. The executive received this year’s prestigious Van Fleet Award for his contributions to strengthening relations between Korea and the United States.
Transcriptional intermediary factor 1 gamma-based multitarget gene therapeutic strategy for triple-negative breast cancer Original
Abstract Triple-negative breast cancer (TNBC) is an aggressive subtype lacking effective targeted therapies. A multitarget gene therapy was developed and validated to overcome molecular heterogeneity and compensatory survival signaling in TNBC.
The Silent Killer of AI Agents: Why Your Evaluation Metrics Are Lying to You
Originally published on tamiz.pro. The Metric Mirage You’ve shipped your AI agent. It aces the benchmark, clears every test case, and your dashboard glows green. Three weeks later, a user reports it’s making catastrophically wrong decisions in production — decisions no metric ever hinted at. This isn’t a model failure. It’s a measurement failure. And it’s everywhere. Most AI agent evaluation pipelines are a stack of proxy metrics: accuracy, precision, recall, reward model scores, task completion rates.
Meta’s Muse Is No Sure Thing
Meta is basking in the glow of its new AI agent app Muse. On Monday, Meta shares surged 11% as investors grew increasingly excited about the app’s success. Muse handles multistep tasks such as booking flights, finding plumbers, managing customer service requests, and comparison shopping. It has been the most downloaded app in Apple’s App Store for the past several days, following its launch roughly two weeks ago.
Key Context by Tae Kim
This has been a bad week for the crypto industry and I think it is going to get worse. On Tuesday, the industry lost a key procedural vote in the Senate on the Clarity Act, legislation that would have established a regulatory framework for digital assets. Banks of every size from the largest money-center institutions to small regional lenders lobbied against the bill with JPMorgan CEO Jamie Dimon among the most vocal opponents.
Integrated genomic and transcriptomic profiling across the clinical spectrum of Lynch syndrome-associated endometrial disease
Abstract We investigated the molecular processes underlying the clinical spectrum of Lynch syndrome (LS)-associated endometrial disease using integrated genomic and transcriptomic profiling of endometrial tissues. Whole-transcriptome sequencing revealed distinct patterns of immune remodeling across clinically defined disease states (unaffected carrier, premalignant, and malignant states), characterized by increasing infiltration of T cells, particularly CD8+ subsets and B cells.
Jensen Huang on the 'Irresponsible' AI Doomers, RSI, and Dario's Essay
Nvidia CEO Jensen Huang was interviewed on Monday at the All-In Summit 2026. You can watch the full video here. Huang was asked about Jacob Coxon’s whistleblower posts and Anthropic CEO Dario Amodei’s “We Must Pace the Fronter” essay. They both moved markets and dominated the mainstream media zeitgeist the past week.
GPT-6 Astra and the Fourth Exponential
A reminder that I’m on vacation this week, but the flood of impressions people have posted online over the past few days as they gained access to OpenAI’s GPT-6 Astra, seems like an important moment to briefly cover. GPT-6 is likely to go down as a big milestone in AI development. OpenAI made Astra a full point release to version 6 for good reason.
A Multiferroic Morphotropic Phase Boundary
Conflicts of Interest The authors declare no conflicts of interest. Data Availability Statement Research data are not shared. The data that support the findings of this study are available from the corresponding author upon reasonable request. Supporting Information Filename Description adma74920-sup-0001-SuppMat.pdf6.7 MB Supporting File: adma74920-sup-0001-SuppMat.pdf. References 1, et al., “Magnetoelectric Effect in van der Waals Magnets,” NPJ Quantum Materials 10 (2025): 6.
Hot Chips: Google 8th Generation TPU Presentation
Note: I’ll be taking my first vacation of the year next Labor Day week, September 7 to 11. -Tae Presenters: Norman Jouppi and Sridhar Lakshmanamurthy at Hot Chips conference on Tuesday, August 25, 2026. Here are the most important distilled parts of the presentation. I manually chose and pulled the quotes: Prior to Google, Norman was one of the principal architecture and designers of MIPS. He’s been the tech lead for TPU since 2013.
Hot Chips: OpenAI Jalapeno Presentation
Presenters: Richard Ho, VP of Hardware. Ravi Narayanaswami, Member of Technical Staff. Chris Leary, Member of Technical Staff. Presented at the Hot Chips conference on Tuesday, August 25, 2026. Here are the most important distilled parts of the presentation. I manually chose and pulled the quotes: Richard leads the hardware team at OpenAI. He was part of Google’s TPU team. Ravi spent a decade at Cruise and also worked at Google. Chris was a Google TPU software lead.
Hot Chips Conference Initial 5 Takeaways. What You Need to Know
I LOVED Hot Chips. It has easily become my favorite conference of the year. Imagine hundreds of the smartest chip engineers in the world listening to roughly 40 technical sessions on the new chip technologies they’ve been working on and their dreams of the future, all shared in an authentic way. You can tell that even though they compete in the marketplace, they are chip nerds at heart rooting for each other to create the most magical, life-changing products possible.
Nvidia Blows Away Wall Street's 2027 Forecasts. What You Need to Know.
Note: Hi all. I’m out in California this week at the Hot Chips conference and visiting companies. My schedule has been jam packed with incredible speaker sessions, receptions, hallway conversations and meals with dozens of new friends/contacts. There’s so much to share. Hot Chips has instantly become my favorite conference of the year. What a great time. I’ll write up my initial takeaways tomorrow, but first a quick post on Nvidia’s fiscal second quarter earnings tonight.
Newswrap: OpenAI's Revenue Growth Is Accelerating Like a Rocket Ship and It May Be Taking Share from Anthropic Now
AI-related stocks staged a nice recovery from the late-July lows, when the Situational Awareness hedge fund dislocation triggered a wave of selling. This week, though, brought another wave of choppy volatility amid renewed uncertainty over the Iran War and Scott Bessent’s Treasury Twist intervention, rising long-term bond yields, higher oil prices, and some disappointment over Anthropic’s latest numbers.
Multi-omics pathway integration identifies central hub genes involved in the acute kidney injury to chronic kidney disease transition
Abstract Understanding the transition from acute kidney injury (AKI) to chronic kidney disease (CKD) requires resolving complex immune-metabolic networks. To define this pathological shift, we integrated transcriptomic and label-free proteomic profiles from a folic acid (FA)-induced mouse model under normal, day 14, and day 28 conditions.
Nvidia Is Speedrunning the Creation of a Synthetic Hyperscaler
This is a guest post by Clark Tang, a partner at Altimeter, who originally published this piece on his X account. The reality is what we are seeing unfold is Nvidia speedrunning the creation of a synthetic hyperscaler. Apologies in advance to all the investors who are stuck in their priors that this will trigger. But what is a hyperscaler? Strip it down and it’s a scaled infrastructure collective of CPUs, networking, storage with a development platform on top. It fulfills two purposes.
Nvidia Scores $500 Billion in Financing for Its Customers. That's A Lot.
I have often written about how Nvidia’s ability to use its size and resources to secure supply commitments for sought-after wafers, optical components, and memory chips gives it a significant advantage over its rivals. Now, we can add a massive amount of AI infrastructure financing to the list.
Cooperative Solvent-Thermal Interactions Enable Multi-Material Nanotransfer Printing for Multi-Dimensional Molecular Profiling
1 Introduction High-resolution micro- and nanoarchitecturing provides a powerful route for tailoring the physical, chemical, and optical properties of materials, enabling a wide range of applications in electronics, photonics, energy devices, and biosensing systems [1-9].
Bosmolisib (BR101801), a novel PI3K-δ/γ and DNA-PK inhibitor for relapsed/refractory patients with peripheral T-cell lymphoma: a phase I study
Abstract Bosmolisib is a novel triple inhibitor of PI3K-δ/γ and DNA-PK for malignancy treatment. This was a first-in-human, phase 1, open-label, dose-escalation, and dose-expansion trial with relapsed/refractory (R/R) non-Hodgkin lymphoma (NHL). Patients received bosmolisib orally once daily in 28-day cycles across four dose levels (50–325 mg) until disease progression or unacceptable toxicity.
Why F1 Is the Wrong Objective for Entity Resolution Thresholds in Corporate Data
Most entity resolution pipelines are tuned to maximize F1. For corporate data, that is the wrong objective. F1 treats false positives and false negatives as equally costly. In practice they are not even close. A false negative in corporate ER means you have two records that refer to the same company — a redundancy you can fix later by rerunning the pipeline with updated parameters. A false positive means you merged two distinct companies into one entity.
Beyond the Demo: Why AI Infrastructure Is the Real Competitive Advantage
Building reliable AI systems requires more than great prompts. It requires great infrastructure. Everyone is talking about AI Agents. From coding assistants to autonomous customer support, AI agents are rapidly becoming part of modern software. But after building a few production AI applications, I realized something: The difficult part isn't building the agent. It's building everything around it. A simple AI demo Most tutorials look like this: User │ ▼ LLM │ ▼ Tool It works.
Building a Production RAG Pipeline: Document Processing, Chunking, and Metadata Design
In the first article, we explored why many RAG systems fail in production and established a key principle: retrieval quality determines answer quality. We also introduced the architecture behind production-grade RAG systems and explained why a simple "embeddings + vector database + LLM" pipeline is rarely enough. In this second part, we'll move one step earlier in the pipeline—to the moment a document first enters your system.
AMD Earnings and SpaceX Call Takeaways
Before I get to AMD’s second quarter, let me say a couple of things about SpaceX. The day SpaceX IPO’d on June 12, I wrote that the multi-trillion dollar valuation was “way too high” and that the best thing to do was watch from the sidelines. The stock rose for three days before dropping like a rock and getting cut in half. I’m not comfortable with the valuation here either, so I’d suggest staying away. But that doesn’t mean listening to SpaceX’s earnings call can’t be useful for AI trade investors.
Thermal performance of high-albedo envelope coatings in aging multi-family housing: a case study in Gimcheon City, South Korea
?Mathematical formulae have been encoded as MathML and are displayed in this HTML version using MathJax in order to improve their display. Uncheck the box to turn MathJax off. This feature requires Javascript. Click on a formula to zoom. ABSTRACT While cool roof technology is well-documented for single-family houses, empirical field evidence from multi-family housing remains limited despite this building type dominating Asian urban residential areas and housing heat-vulnerable populations.
Incremental Graph Updates for Corporate Knowledge Graphs: Three Problems Batch Pipelines Can't Solve
Corporate knowledge graphs break in production the moment you treat them as a batch artifact. I ran into this building 2asy.ai, a Graph-RAG system for East Asian corporate intelligence. The initial architecture was standard: ingest a document set, run NER and entity resolution across the batch, write the graph. Clean, deterministic, fast to prototype. Then the first update cycle hit.
Context window growth is the silent failure mode in agentic pipelines
Context window management is the invisible cost in every multi-step agent pipeline. I see this pattern across production consulting work. An agent that works cleanly in testing — ten steps, coherent reasoning, correct output — starts degrading six weeks after ship. No errors. No crashes. Just drift. Outputs get shorter. Reasoning steps that used to show clear logic start reading like vague summaries. Users start noticing that the agent "feels slower." The root cause is almost always the same.
Key Context by Tae Kim
“The next six to nine months are going to be bigger than anyone believes... It’ll be dramatically better for AI than the last two years in terms of advancements in capabilities.” “RSI is a lot closer than people think...
The all-purpose agent isn't an architecture. It's a single point of failure with a system prompt.
Intro In a demo, one agent that "does everything" looks like impressive economy: one prompt, one API call, one mental model. Nobody's testing it against thousands of edge cases a day, and nobody's paying for the tokens burned parsing forty unrelated instructions to answer a two-line question. In production, that same design is where teams spend their first real outage. Below are four failure modes that show up once a monolithic agent hits real traffic.
The Big Lesson from the Implosion of Leopold Aschenbrenner’s $20 Billion Situational Awareness Hedge Fund
Those who do not learn from history are doomed to repeat its mistakes. Financial history is littered with hedge funds that start with a good fundamental idea, post a period of incredible returns, and then through hubris and too much leverage collapse and give it all away. Long-Term Capital Management. Archegos. Amaranth Advisors. And now this week, Leopold Aschenbrenner’s Situational Awareness.
Why merged cells break table extraction from multi-column PDFs
PDF tables look simple until you actually extract one. The visual structure in your reader is a rendering artifact. The underlying file has no table object, no row, no cell: just positioned text spans and optional line segments. When a table has merged cells, two or more content spans share the same logical cell, but the PDF records each span at an independent coordinate. The parser sees overlapping positions and has to guess where one cell ends and another begins.
I built a client-side trade document validator — no upload, no server, just the browser
The problem I kept hitting: an international invoice with a line total that's off by a cent, or an HS code that appears in the commercial invoice but is absent from the packing list. Customs holds are rarely dramatic. They're usually paperwork. I wanted a tool that would catch these before the documents left the company. The constraint I set: nothing uploaded to a server. Every byte stays in the browser. What it does Upload an invoice and a packing list (Excel or PDF).
I built a PDF table checker that verifies arithmetic in any PDF without a template
The problem I kept running into: a financial report where a subtotal on page 8 is wrong, but nobody spots it until the numbers get aggregated into a spreadsheet and something doesn't reconcile. I wanted a tool that could ingest any PDF with tables and tell me which totals don't add up. The constraint: no template, no column mapping UI, no server. Every byte in the browser. What it does Upload a PDF.
Query-Time Entity Disambiguation in Graph RAG: When One Name Means Seventeen Nodes
The hardest retrieval problem in Graph RAG is not missing data. It is the query that arrives with an ambiguous entity name. In 2asy.ai, a knowledge graph built on Korean corporate data, "Hyundai" matches seventeen separate nodes. A vector search returns all of them ranked by embedding similarity. A graph traversal needs exactly one starting node. The gap between these two realities is where Graph RAG quietly breaks.
Key Context by Tae Kim
Thanks to Charles Payne for inviting me on to discuss why Google is a secular short, Nvidia, OpenAI/Anthropic, Kimi K3, and chip stock geopolitics. It was a ton of fun. Enjoy!
Tool Schema Drift: The Silent Failure Mode in Production Agentic Systems
The most common agentic system failure I encounter in production is not a bad prompt. It is not a context overflow. It is a tool that changed without its registration changing. I have seen this cause weeks of debugging in systems that were working fine until they weren't — and because the failure is silent, teams often spend time looking at the model first. What Tool Schema Drift Looks Like Most agentic frameworks let you register tools with a name, a description, and a JSON schema for parameters.
An Agent Harness Is Not a Security Boundary: Reading AgentSmith as a Workflow Layer
Coding agents do not fail only because a model lacks context. Teams also lose the workflow around the model: what to inspect first, what evidence counts as verification, when to hand off, and when to stop for a human. AgentSmith is a new open-source agent harness that frames this as a composition problem. Its public repository contains a shared core/, work-type profiles/, hooks/, and a setup.sh entry point.
The OpenAI and Hugging Face Incident Was an Agent Boundary Failure
OpenAI said on July 21 that two of its models breached Hugging Face during an internal cyber capability evaluation. One was GPT-5.6 Sol. The other was a more capable pre-release model. The models were being tested with reduced cyber refusals so OpenAI could measure maximum capability on ExploitGym, a benchmark for offensive security tasks. The weird part is not that a security benchmark produced risky behavior. That is the point of this kind of test. The weird part is the shape of the escape.
Google Is a Secular Short
The smartest investors in the world are piling into Google parent Alphabet’s stock. Warren Buffett recently confirmed that he initiated Berkshire Hathaway’s purchase of Alphabet shares, including a $10 billion private placement last month. Chris Hohn, who has one of the best track records of any fund manager in history and ran the world’s most profitable hedge fund last year, sold most of his Microsoft position earlier this year and bought more Alphabet stock.
When the same incident becomes five separate events in your graph
A single corporate event in East Asia rarely shows up once. A merger gets filed at OpenDART in Korean, reported by EDINET in Japanese a day later, picked up by a Hong Kong exchange notice in traditional Chinese, and summarized in an English wire story. Four documents, one real-world incident. If your pipeline extracts events document by document, you end up with four Event nodes in the graph that are really the same thing.
Show More
loading
Actions
Is this you?
As a journalist, you can create a free Muck Rack account to customize your profile, list your contact preferences, and upload a portfolio of your best work.Get in touch with Tae
Contact Tae, search articles and posts on X, monitor coverage, and track replies from one place.
Learn more about Muck Rack