Principle · MoneyTara Sesha says ship the frontier to enterprises even when they feel overwhelmed — if you're not shipping the latest to these enterprises, you'll get leapfrogged — and that OpenAI needed to ship a giant revolutionary change that broke all of their processes so enterprises could unlock more value, getting agents to enterprises as fast as possible, otherwise they might consider alternative products.
Lenny's Podcast · Stop planning for 2027: how OpenAI builds product 90 days at a time | Tara Sesha and Nan Yu (OpenAI)
ReportA frame is copied to create a new state for a transition; an image is copied from unsplash.com as content; smart animate creates a relationship between an object and a frame with a nice easing curve; an on-click interaction is set to navigate to the target page, and a second on-click interaction navigates back to the previous frame; when clicking the second pizza, objects go off the page and get larger automatically.
Figma · Figma for Edu: Prototyping with animations in Figma
Creator perspective · SocialCoach Kyle argues you shouldn't create a dynamic where you keep asking questions; say things she responds to, acknowledge her response, and she starts asking you — then you can sit back, relax, and talk while she invests more and more into the interaction.
Coach Kyle · How to Keep a Woman Interested Without Acting Like a Clown
Report · MoneyIn the demo, a Claude Code terminal session is told to connect to the platform; the connected session gets an identity and an agent card, with all network-perspective complexities hidden from the agent.
AI Engineer · Why Your AI Agents Can't Talk to Each Other (Yet) — Vlad Luzin, BAND
Principle · MoneyHumans and agents learn the most from mistakes — the real sharpening happens when the agent fails. A run that goes sideways is the richest signal you can get because it exposes the missing guardrail, a missing library, or an edge case you haven't thought about upfront. Every run — pass or fail — feeds the distiller mechanism, constantly making the skill better: failed runs make the skill more robust because they reveal what's missing, while successful runs make it more trustworthy because achieving the goal means the skill works. This all happens completely autonomously and asynchronously, without a human in the loop.
AI Engineer · From 36% to 100%: How Self-Improving Agents Write Their Own Skills — Rafal Wilinski, Runlayer
Creator perspective · MoneyPen testing is a human security researcher who understands the concept and business logic of the application and tries to break it to find vulnerabilities — the golden standard in application security testing, and still the best method for testing business logic. But it costs a lot of money because it requires an actual human researcher, so companies can only do it once, two, or maybe three times a year. Traditional pen testing also produces a PDF report that developers must interpret to determine what is real, then fix, and then retest — a slow workflow that no longer makes sense.
AI Engineer · AI Hackers Are Faster Than Your Pen Test — Eli Cohen, Snyk
Creator perspective · MoneyDoing pen testing only two times a year is not good enough when AI attackers are continuously attacking — for the rest of the roughly 350 days a year, the application goes untested.
AI Engineer · AI Hackers Are Faster Than Your Pen Test — Eli Cohen, Snyk
Principle · MindThe one tactical tip: try to do things that check multiple boxes — find activities that satisfy several priorities at once. Find two or three activities in your life where one action checks a few boxes; you won't get it for everything, but you can get lucky. For example, lifting with friends who are also colleagues checks recreation, career growth, and physical health, letting you block more time to it.
Alex Hormozi · It took me 37 years to realize what I'll tell you in 18 minutes.
Creator perspectivePhilipp Schmid explains that with the environment field set to remote, your agent gets a sandbox managed by Google where it can run code, create files, write files, install dependencies, install your own tools, or call your own APIs. The managed environments are stateless on Google's side — you just send the API request and don't need to manage the infrastructure, and the environments stay persistent.
AI Engineer · Why AI Agents Should Have Their Own Sandbox — Philipp Schmid, Google DeepMind
Creator perspectivePhilipp Schmid describes the Interactions API as Google's developer-focused, simpler, web-developer-friendly answer to OpenAI's Responses API. It works for all Gemini modalities with a model or agent, an input (string or multimodal data), and tools or environment; calling a model like Gemini and calling an agent like the anti-gravity agent use literally the same interface.
AI Engineer · Why AI Agents Should Have Their Own Sandbox — Philipp Schmid, Google DeepMind
Creator perspective · MoneyA structural three-to-six-month advantage in chip capability perennially wins all frontier-lab deployments. The six-month gap from a rolling frontier of bets is the wedge that allows a company to drive into the competition. The frontier model analogy applies to the chip space: whoever holds a structural time advantage wins deployments.
No Priors: Artificial Intelligence | Technology | Startups · Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile
Creator perspective · MoneyCurrent SRAM-based fast-inference chips are not running long-context attention in real deployments — their memory capacity is too low, so deployments still flip back to a GPU for that — creating a mismatch between where speed matters most and what the hardware can do.
No Priors: Artificial Intelligence | Technology | Startups · Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile
Principle · MoneyAll frontier labs play an asymmetric game where they are exposed to enormous risk if they go all in on a hardware bet. If lab one goes all in on proprietary silicon and lab two discovers a computational breakthrough that only works on lab two's chip, lab one could die in the nine months before it deploys enough of its own chip to gain the 5X computational efficiency. There is an unparalleled urgent need for frontier companies to deploy the same platforms as one another.
hypothetical scenario
No Priors: Artificial Intelligence | Technology | Startups · Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile
Creator perspective · MoneyFrontier labs face enormous risk going all-in on proprietary silicon: a competitor's chip-only computational breakthrough could kill them in the nine months before their own chip deploys. Frontier players have an urgent need to deploy the same platforms as one another to hedge against chip-locked breakthroughs, and making deeply differentiated bets and going all-in on them is an irrational and dangerous move for any chip player competing at the frontier.
No Priors: Artificial Intelligence | Technology | Startups · Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile
Creator perspective · MoneyFractile spent roughly its first two years on an SRAM-based chip similar to Groq or Cerebras, but began worrying about the scalability of that approach towards the end of 2023 and certainly in 2024: growing context lengths exposed SRAM's low capacity, so the company moved away from the architecture.
No Priors: Artificial Intelligence | Technology | Startups · Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile
Creator perspective · MoneySRAM is super high bandwidth memory that lives on the same piece of silicon as the compute logic, providing extremely high bandwidth between compute and model weights or KV cache — an SRAM-based approach can drive thousands of tokens per second on language models.
No Priors: Artificial Intelligence | Technology | Startups · Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile
Principle · MoneyWhen new LLMs churn every two weeks, find what is common across all of them rather than targeting any single model — the universal constant is that every new LLM desperately wants orders of magnitude more memory bandwidth to run faster. The attention mechanism at the open-source Chinese model frontier churns every couple of weeks, including MoE sparsity levels and sparsity on attention itself, so hardware bets must tolerate this churn.
No Priors: Artificial Intelligence | Technology | Startups · Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile
Principle · MindLiemandt argues that AI-driven personalized lessons let kids learn roughly 10x faster, and that the freed time should go to builder skills, entrepreneurship, storytelling, and financial literacy — the skills that make kids thrive and 'totally unlock.' Kids love this combination and will do both mastery learning and engaging experiments.
Andrew Huberman · How to Accelerate Learning & Improve Education | Joe Liemandt
Creator perspective · MindJoe Liemandt describes how an AI tutor can assess exactly what a student knows and doesn't know, deliver a personalized lesson, and keep the student in a loop until mastery before letting them move on — and Huberman adds that a non-shaming digital tutor lets students fill gaps without embarrassment.
Andrew Huberman · How to Accelerate Learning & Improve Education | Joe Liemandt
Creator perspective · MindJoe Liemandt argues that a student learning at 2x — twice as much in two hours a day — will eventually catch up to the norm because the rate is independent of starting position; students in third-grade material can learn without anyone knowing or caring where they are, since their motivational model is based on their own starting point.
Andrew Huberman · How to Accelerate Learning & Improve Education | Joe Liemandt