Brian Elbert

Anthropic onboarding day: Michael Scott introducing Karpathy like he just signed Wemby in free agency.

Mark Cuban says AI will split every company into 2 categories within 3 years "When it's all said and done over the next 3 years, there's going to be two types of companies. Those who are great at AI and those who went out of business" "The biggest challenge for a company is going to be for the CEO to make that decision that we're going to have to blow up a lot of what we do to recreate our company" "We hear a lot of stories about huge public companies that spend a lot of money to try to implement AI and it doesn't really get them a return on investment." "Because you already are running your business the way you've always run it." "The things you've considered outsourcing, your customer call centers, AI agents are perfect for. The stuff you would give to an intern, AI agents are perfect for" "But to really take advantage of AI you have to reformulate your business completely to build it on AI"

Ken Griffin went home on a Friday "fairly depressed" after watching AI agents at Citadel do work that used to take teams of PhDs in finance months to complete. Done in days. His words: "These are not mid-tier white collar jobs. These are extraordinarily high skilled jobs being automated by agentic AI." This is the head of one of the most successful hedge funds in history saying the people he pays seven figures to analyze markets and structure deals are being replaced by software that works in hours instead of months. Not theoretically. In his own office. Right now. The Coatue deck we covered earlier this week called agents "the biggest unlock" in AI. Griffin just confirmed it from the buy side. The shift from copilots to agents is not a future event. It is already happening at the highest levels of finance.

“I am going to probably use $300M of Anthropic this year at Salesforce.” - Marc Benioff “ These coding agents are awesome. Anthropic is awesome. Coding, everything's going to be cheaper to make, it's more efficient. I can do things that I just could not do before. I can go faster than ever before. I can implement my software and sell it at the same time. I've never been able to do that before. Today, I have humans, agents, and headless platforms all interoperating, never before. So the opportunity for my own company and the efficiency that I have in my own company, in service and support, in distribution and marketing, across the board, is unprecedented. What I can do for our customers, unprecedented. And, to that point, my gosh, have you seen Anthropic? It is a rocket ship that will not stop.”

Fun interactive science app ideas | Part 3 Played around with generating 3D biological structures and made an app to explore them interactively UI Design GPT Images 2 Code Gemini 3.1 Pro More demos ↓

Claude Code just got an update nobody is talking about. And it replaced my entire video crew. You type one simple sentence. Claude writes the script. It uses ElevenLabs to clone your voice. It calls Remotion to build the scenes. It grabs HeyGen to make your avatar talk. Ten minutes later, you have a finished video. You never even turn on a camera. This changes how we make content forever.

Anthropic just shipped sleep into agents. When you sleep, your hippocampus replays the day's neural sequences to the cortex during 150-220 Hz bursts called sharp-wave ripples. The replay runs about 20x faster than the original experience. A 10-second sequence gets compressed to roughly 500 milliseconds. Wilson and McNaughton showed this in rats in 1994. You ran this algorithm last night on whatever you did yesterday, whether you wanted to or not. The replay does two things at once. It extracts statistical patterns: what mattered, what generalizes, which sequences predicted reward. And it reorganizes the memory trace from hippocampus-dependent storage into neocortex, which is why old memories survive hippocampal damage but recent ones don't. Disrupt sharp-wave ripples in a rat with optogenetics and the rat fails the next day's task. The replay is causal, not correlational. Most "agent memory" today is a search engine. Past sessions get embedded, you retrieve relevant chunks at the next call. That works for facts. It does not extract patterns and it does not reorganize the trace. Which is why agents plateau. The memory volume keeps growing while real capability flatlines. Dreaming reviews past sessions, extracts patterns, curates memories. That is the brain's actual three-step algorithm. They called it dreaming because dreaming is what the algorithm does, in roughly the same order, for roughly the same reason. Agents that dream between sessions will compound. The ones still running on raw context window will hit the same ceiling humans hit when they pull all-nighters.

Kurzweil set 2029 for AGI. 2045 for the singularity. The question isn't whether AI will transform life science commercial operations. It's whether your infrastructure is structured to benefit when it does.

Most life science companies still run 2015 playbooks. The gap between first movers and everyone else isn't growing linearly. It's growing exponentially. Every quarter you wait, the cost of catching up increases — because the tools on the other side keep getting better, faster.

Read Kurzweil's "The Singularity Is Near" a decade ago. Felt like science fiction. Just finished the sequel. It reads like a status report. Here's what exponential AI growth actually looks like when you use these tools every day. A thread.

Andrej Karpathy (@karpathy), OpenAI co-founder, ex-Tesla AI, "vibe coding" creator. In just 4 mins, he explains why Claude Skills, MCP servers, and AI agents are past the hype and are now the new baseline for building. Worth every second ↓

The more your agent remembers, the less it knows. This sounds counterintuitive, but it is actually a direct result of how agent memory is built today. Agent memory inherits the cognitive shape of its store. - A vector DB gives it associative memory to recognize familiar patterns. - A graph gives it relational memory to understand how things connect. Most agents run on the first and skip the second. Here's an example that explains the failure it leads to: Say a study assistant stores three facts about a student in a vector DB: - Mark is in grade 10. - Grade 10 has final exams in March. - The library closes 2 weeks before final exams. Mark asks: "Will the library be open next week?" The vector DB likely returns the first and third facts, because the query mentions Mark and the library. But it skips the middle fact, which links Mark's grade to the exam time, because that fact mentions neither Mark nor the library. It sits in embedding space too far from the query to make it to the retrieved context. So the Agent answers with partial info, or it fills the gap with a plausible guess that sounds right but might be off by weeks. This is not a corner case, but it's actually what real queries look like. Any question that spans two or more hops exceeds what a similarity search can do. Increasing context size and retrieving more context is one solution. But accuracy drops over 30% when the relevant fact sits in the middle of a long context, which is the well-known "lost in the middle" problem. A bigger window is not the same as better memory. It just gives the model more room to miss things. To actually solve this problem, you need to stop treating memory as a single store and start treating it as three complementary layers, each doing a job the others cannot. - Relational: It stores where a fact came from, when it was stored, and who has access. This is the provenance layer. - Vector: It stores what a fact means and what it is semantically similar to. This is the retrieval layer. - Graph: It stores how facts connect, what depends on what, and who relates to whom. This is the reasoning layer. All three are important and complementary: - A vector DB alone gives similarity without relationships. - A graph alone gives relationships without semantic search. - A relational store alone tracks where data came from but cannot reason over it. If you want to see this in practice, Cognee (open-source) implements this approach. It runs an ECL pipeline (Extract, Cognify, Load) that writes into all three stores in a single pass and keeps them synchronized as new data arrives. So the vectors and graph edges are built together during indexing, not glued together later. On top of this, there are two things Cognee does differently from most memory tools: 1) Smarter entity resolution: You can give Cognee a domain vocabulary file, and it uses it to merge duplicate mentions automatically. So "car manufacturer," "automobile maker," and "vehicle producer" collapse into one canonical node instead of being available as three separate entries. 2) Local-first defaults: The default stack runs on a single pip install and stays fully local. You can switch to Postgres and Neo4j for production without changing the API. My co-founder wrote a first-principles walkthrough of agent memory that takes the same problem and works through every layer of the stack, ending in a real working agent built on Cognee. Read it below.

Mark Cuban just described the largest wealth transfer of the AI era. Almost nobody understood what he said. Cuban: “There are 33 million companies in this country. Aren’t going to have AI budgets. Aren’t going to have AI experts.” Not tech startups. The shoe store. The regional trucking outfit. The accounting firm with 12 employees. The businesses that actually run the physical economy. They know AI is coming. They have no idea what to do with it. Cuban: “You’ve got the head of Microsoft saying software is dead because everything’s going to be customized to your unique utilization.” Software is dead. The SaaS era ran on one rule. Build a generic product. Force millions of companies to bend their workflows around it. Charge rent forever. AI ends the contract. The business stops bending to the software. The intelligence bends to the business. But customized by whom. The third-generation manufacturer cannot tell Claude from Gemini. The county hospital is staring at a reactor asking where the light switch is. Cuban: “Who’s going to do it for them?” That question is worth more than the frontier models themselves. Hundreds of billions are being burned to build the foundation. The smartest engineers alive are locked in a bloodbath over who owns the base layer. Let them fight. Let them burn the capital. Let them drive the cost of raw intelligence toward zero. Because the wealth does not collect where the brain is built. It collects where the brain meets the business. Every ambitious kid in college right now thinks survival means a seat at OpenAI or Anthropic. Cuban is staring at the other 99 percent of the economy. Learn the models. Then learn the messy, unglamorous reality of how a 50-person company actually operates. Walk through the door. Understand their problems. Wire the intelligence directly into their revenue. That is not a job title. That is an entire economic class being born. You do not need to build the brain. You need to build the nervous system. The biggest winners of the electricity era were not the engineers who built the generators. They were the ones who walked into dark factories and showed the owners where to plug in. 33 million companies are standing in the dark right now. Silicon Valley is racing to build the god. The fortunes will belong to whoever teaches him a trade.

There’s $1T up for grabs for agent-first startups and this window is WIDE open. Probably 10,000+ niches. How it plays out: 1. Every SaaS company follows salesforce and goes headless within 18 months 2. a new category of "agent-native" startups emerges that treat salesforce, HubSpot, workday etc as dumb backends. the startup IS the agent. the SaaS is just the database. 3. the entire consulting/services industry around enterprise SaaS gets compressed into software. the agent replaces the implementation team. 4. outcome-based pricing becomes default. nobody pays per seat when the "seat" is an agent making 10,000 API calls a minute. you pay when revenue hits your account. 5. the winning founders are ex-operators who understand a vertical workflow cold. the code is the easy part. knowing that a property manager spends 14 hours a week on lease renewals? that's the insight worth $100M. 6. distribution becomes the moat. when anyone can wire agents to APIs, the company with the audience and the brand wins. media + agents is the new SaaS. There’s a rush to incubate live/short form shows. 7. Silicon Valley goes all influencer. Roy lee gets this. Pat Walls gets this. Sam Parr gets this. 8. the first $1B agent-native company in each vertical will look nothing like the SaaS it replaced. smaller team, higher margins, no implementation cost, no churn from bad UX because there is no UX. the fastest path to wealth right now: find an industry that still runs on dashboards, phone calls, and spreadsheets. build the agent-native version. charge per outcome. own the workflow end-to-end. someone reading this right now is going to build a $100M company off this exact shift. tell me about it on the @startupideaspod when you do. Im rooting for you. Less reading, less bookmarking, more building. the last wave rewarded people who built pretty interfaces on top of ugly data. I think this wave rewards people who build smart agents on top of exposed APIs. Or who just build the APIs themselves Here we go