Compute, Context and Concierge Service* You can have all the compute & context in the world, but if you don’t know how to make it actually useful in your business, which most do not, you just waste a boatload of bits and atoms.

Momentum is oxygen for startups. When you have it, you'll do anything possible to bottle it up, make sure everyone on your team knows how precious it is, and preserve it at all costs. When you're missing it, you only have one goal. Recapture it. You'll do things that don't scale. You'll will wins into existence. You'll do anything, literally anything to get it back.

It's official. @tenex_labs is hosting the goddamn BATTLE OF CHAMPIONS. An invite-only hackathon in August featuring NYCs best builders, solving real, hard problems from big, complex enterprises. 150 builders. 24 hours. 1 winner. Applications are officially open. And limited sponsorship slots are officially open. Head to the official site:

Favorite new AI phrase: benevolent prompt injection Definition: The act of assembling a packet of context automatically, across the agent environments our engineers use. Managed startup hooks inject a project's rules, requirements, process, and current state before a new task begins, so a new session does not depend on a user starting with the right prompt. Killer concept by @tenex_labs' engineers.

4 years ago, @ArmanHezarkhani and I had a conversation about creating the Presidential Brief for everyone on earth. The Presidential Brief is a daily summary of high-level, all-source information and analysis on national security issues produced for the president. What would that look like but for you? A 1-of-1 newsletter for every person on earth. The concept has always been appealing, but pre-AI it wasn't feasible to create killer customized daily newsletters for anyone. Now it's not only possible, but inevitable. The /daily-brief skill builds you a personalized daily newsletter that combines your external interests with all of your internal context (email, calendar, slack, etc). Try it out. It's really freakin sick.

It's official. @tenex_labs is hosting the goddamn BATTLE OF CHAMPIONS. An invite-only hackathon in August featuring NYCs best builders, solving real, hard problems from big, complex enterprises. 150 builders. 24 hours. 1 winner. Applications are officially open. And limited sponsorship slots are officially open. Head to the official site:

Favorite new AI phrase: benevolent prompt injection Definition: The act of assembling a packet of context automatically, across the agent environments our engineers use. Managed startup hooks inject a project's rules, requirements, process, and current state before a new task begins, so a new session does not depend on a user starting with the right prompt. Killer concept by @tenex_labs' engineers.

4 years ago, @ArmanHezarkhani and I had a conversation about creating the Presidential Brief for everyone on earth. The Presidential Brief is a daily summary of high-level, all-source information and analysis on national security issues produced for the president. What would that look like but for you? A 1-of-1 newsletter for every person on earth. The concept has always been appealing, but pre-AI it wasn't feasible to create killer customized daily newsletters for anyone. Now it's not only possible, but inevitable. The /daily-brief skill builds you a personalized daily newsletter that combines your external interests with all of your internal context (email, calendar, slack, etc). Try it out. It's really freakin sick.

going from good-to-great on the non-verifiables (writing, marketing, design) is going to be exponentially harder then going from non-existent to world-class on the verifiables (coding, finance, support).

Token costs are gonna continue to plummet. This is why having a robust marketplace for models, closed and opensource is so so important. The way AI leads to the future we all want, with abundance and access is through healthy competitive markets, limited regulatory capture, and attacking bad actors at the source not the model layer.

Really important piece to read. It seems obvious that every company is going to have to create private benchmarks for tasks and there's going to be some platform that allows companies to create those benchmarks. Right now everyone's business is a fleet of vehicles (workflows) going to the gas station to get gas (model), but no one has the label on the side of their door saying whether they need unleaded, premium, or supreme. So they unnecessarily use premium. And the only information they have on what gas to use are generic standards (arena ai, SWEbench) saying all SUVs should use premium, all pick up trucks should use diesel, etc. Prob 12-18 months out from this having a huge market because most enterprises are earlier in their AI journey then necessitates this precision, but feels like the right time to get started solving this problem.

Best explainer on Kimi K3 i've read. It walks you through how the model works & the elegant innovation behind it: - K3 is the biggest open model anyone has released: 2.8 trillion parameters total, though only 104 billion of them do the work on any given word. - The problem Moonshot went after is memory. Normally a model keeps notes on every word it has read, and that pile grows with every token, which is why long conversations get slow and expensive. - K3 mostly stops the memory pile from growing. Three out of every four layers use a new mechanism called Kimi Delta Attention, which keeps a fixed-size working memory and edits it as it goes, overwriting what's stale instead of hoarding everything. The fourth layer keeps a compressed record of each word so exact details are still recoverable when they matter. - They proved this on a small model first. A 48B test version used up to 75% less memory and ran about 4× faster at long context, while matching or beating the conventional design on the benchmarks they reported. - Training leaned hard on long jobs including coding, browsing, research, visual work, agent sessions running hundreds or thousands of tool calls in a row. - 1 Million cached input tokens costs $0.30. An agent can pull the same repo, the same docs, the same tool definitions back into context over and over without the bill getting stupid. Moonshot credits that to the memory design working alongside their serving stack. - The biggest takeaway: context window size is the main character, but what matters is what a model compresses, what it forgets, and how it gets exact information back.

Couldn't help myself. Forward deployed engineer or AI engineer sounds dry as cardboard, so I asked my team to come up with the most creative job titles describing what they do... The current list: - Transformer tamer - LLM whisperer - Hallucination exorcist - Token daddy - Weights guy - Human out of the loop - Latent freak - Gradient guru - RL gym junkie - Inference dealer - Nothing-but-the-harness - Benevolent prompt injector - The metaharness - Context window's context window Who can top this?

If you think you're AI Native, here's a good test: When you're heading to bed tonight, are you asking yourself: "What agents can I be running over night to get a head start on my day tomorrow" If that's an honest thought, you're among the top .01% of AI users right now.

Best explainer on Kimi K3 i've read. It walks you through how the model works & the elegant innovation behind it: - K3 is the biggest open model anyone has released: 2.8 trillion parameters total, though only 104 billion of them do the work on any given word. - The problem Moonshot went after is memory. Normally a model keeps notes on every word it has read, and that pile grows with every token, which is why long conversations get slow and expensive. - K3 mostly stops the memory pile from growing. Three out of every four layers use a new mechanism called Kimi Delta Attention, which keeps a fixed-size working memory and edits it as it goes, overwriting what's stale instead of hoarding everything. The fourth layer keeps a compressed record of each word so exact details are still recoverable when they matter. - They proved this on a small model first. A 48B test version used up to 75% less memory and ran about 4× faster at long context, while matching or beating the conventional design on the benchmarks they reported. - Training leaned hard on long jobs including coding, browsing, research, visual work, agent sessions running hundreds or thousands of tool calls in a row. - 1 Million cached input tokens costs $0.30. An agent can pull the same repo, the same docs, the same tool definitions back into context over and over without the bill getting stupid. Moonshot credits that to the memory design working alongside their serving stack. - The biggest takeaway: context window size is the main character, but what matters is what a model compresses, what it forgets, and how it gets exact information back.

I'm either going to look like an absolute idiot or a fucking genius. i believe that most people are multi-dimensional & don't want to do one thing at work all day long. marketers want to build software. engineers want to create content. product people want to learn to market. at the same time, companies are (or should be) fighting for distribution more than ever before. it's one of the few remaining moats in business & will continue to get harder as AI drives abundance, while attention remains finite. it's why i've introduced a sort of hybrid role at @tenex_labs that I haven't seen at other companies. it's like the professional version of a cyborg or centaur. engineers can be creators. creators can build software. PMs can be teachers & trainers. all of this to say, I think companies should start leaning into their people being jack-of-some-trades, especially in the interest of earning the attention of the internet. if you want to be a professional centaur, shoot me a dm. some examples: - an AI engineering looking to build software & create content. - a product designer looking to build with AI & teach about it. - a full-stack video producer/editor looking to build with & create content around AI. - an AI-native consultant working with clients & looking to write deep essays about this wave.

I’m reworking my content diet. Less content junk food that’s bite-sized and short-lived. More deep analysis that will stand the test of time. With that, who does the best long form writing about AI on the internet?