Playing with an optimizer speedrun is something that never gets old. Built on top of https://t.co/n7gbPOCTPT and Claude should take all the credit for hypertl tuning.

Had a great time discussing AI user privacy on @augmind_fm 😃 One discussion I’d like to highlight from the chat is that what constitutes the "Privacy Problem" has been shifting as AI progresses. It used to be that we care a lot about *training-time* user privacy: what gets trained into the model, and what the model would spit out. Say you take an LLM and a book (or any piece of sensitive text). We cared about whether the book would be regurgitated ("memorization"); whether you can remove such a book from the model ("unlearning"); and whether you can detect the book being trained ("membership inference"). And as part of mitigating these problems, we work on training-time techniques like differential privacy, careful data cleaning, and model alignment/guardrails (in ~increasing order of adoption). Guardrails seem to work well enough that people don’t really talk about sensitive model outputs anymore. What’s more pressing today, I argue, is *inference-time* user privacy: the fact that intelligent models are served at scale on private user data, which are then centrally managed at model providers. Intelligent models mean that user profiling is now cheap and automatic; your activities can be continuously analyzed to reveal new sensitive insights. Whether your data is trained on or not became less relevant. Having a "digital clone" of you by building on your memory/personalization is now way more profitable. The threat vector changed from the model misbehaving to the provider misbehaving. Because of this, the techniques to improve user privacy would look different than before. They’ll look less like fancy learning algorithms (e.g. RL to steer model to output paraphrase of a book than the original book), and more like *peripheral systems* sitting around closed models that we do not control but still want to access. The OA project (https://t.co/rOAoavIavT) is an example: you could build a zero-knowledge proxy to mediate AI inference and combat surveillance, and leverage smaller models to help users build personal memory on-device. This is not to say that there’s no room for training; you just train for different things, and on auxiliary models than the closed models. thank you so much to @EchoShao8899 @michaelryan207 @shannonzshen for hosting me!

AI agents are already going wild, but today’s red-teaming tools for them are still like toys 😢 🔥👽 After spending 20 months and $120K API credits, we are excited to finally open-source DecodingTrust-Agent Platform (DTap): the first controllable, realistic simulation platform for advanced AI agent red-teaming !! 🌍 DTap simulates 50+ real-world environments across 14 high-stakes domains, with realistic agent interfaces replicated from their official MCPs and GUIs. The environments are full-stack, interactive, fully parallelizable, and can be easily configured to reproduce arbitrary real-world attack scenarios, making agent red-teaming scalable and highly transferable to deployment settings. 🔥We also release DTap-Bench, a large-scale benchmark with ~7K agent red-teaming tasks and ~4K policy-grounded malicious goals. Each red-teaming task includes a sophisticated attack sequence across environment-, tool-, skill-, prompt-level injections, as well as their compositions, plus a handcrafted verifiable judge that checks the actual consequences in the environment. Using DTap-Bench, we evaluate popular agent frameworks and backbone models across diverse policies, risks, threat models, and attack strategies, revealing systematic vulnerabilities and zero-days in today’s agents! Paper link: https://t.co/PjnGC5wKk9 Platform + benchmark + code: https://t.co/aicipKMnig Join our Discord: https://t.co/8UyRjH6RqX Read more below 👇

To train better open models, we need predictable scaling. Delphi is Marin’s first step: we pretrained many small models with one recipe, then extrapolated 300× to predict a 25B-param / 600B-token run with just 0.2% error. Getting there took some work 🧵 https://t.co/HmlVFl11ag

Even the biggest tech companies are struggling to keep pace internally with the AI race. Amazon reportedly had to open access to competing coding tools after employees pushed back on relying solely on its in-house solution. That says a lot about the current market. In AI, performance matters more than ecosystem loyalty. https://t.co/NO9PRvhank @futurism

AI may eventually remove most language barriers. But something important could be lost in the process. Language is not just a tool for exchanging information. It carries culture, emotion, humor and ways of seeing the world. Perfect translation may make communication easier while making cultural discovery shallower. https://t.co/8Bw2XzYc9p

The AI boom may become this generation’s “China shock.” But instead of transforming global manufacturing, it could reshape knowledge work, productivity and economic power. The gains could be enormous. So could the disruption. https://t.co/LplPGs4a32 @fortunemagazine @sashrogel

Anthropic says ‘evil’ portrayals of AI were responsible for Claude’s blackmail attempts https://t.co/NmNO7F93jt

Anthropic says ‘evil’ portrayals of AI were responsible for Claude’s blackmail attempts https://t.co/NmNO7F93jt

This is one of the best short films I've seen in years. Very soon, we'll stop calling it "AI film" and just call it film. https://t.co/6yjU6HcnzN

Geoffrey Hinton says AI systems can keep improving when they are not limited by human data Like AlphaZero can create endless training data through self-play, and math offers a similar path by letting models generate and test their own conjectures "LLMs could eventually work the same way"

Warren Buffett on not overcomplicating investing: “Investing is just about assigning yourself the right story….” Buffett says to imagine you’re an experienced journalist specializing in investigative reporting. Your editor hands you an assignment: What is this newspaper company worth? And you need to write the story by the end of the month. What would you do? Buffett: “You’d go out and interview TV brokers, newspaper brokers, and owners, and you’d try to value each asset…” That’s all Buffett does: “That’s what I do—I assign myself the right story. It’s nothing more than that.”

Yann LeCun closed $1.03B for AMI Labs on March 10. Three days later, this paper dropped from his NYU collaborators. 15M parameters. Single GPU. A few hours of training. LeWorldModel is the first JEPA that trains end-to-end from raw pixels. Two loss terms: predict the next embedding, keep the latent space Gaussian. Previous JEPAs needed exponential moving averages or pretrained encoders to avoid representation collapse. LeWM doesn't. Six hyperparameters down to one. The numbers are the story. Foundation-model-based world models require hundreds of millions of parameters and serious compute to plan a control task. LeWM plans up to 48x faster while staying competitive on 2D and 3D benchmarks. The whole thing fits on a laptop GPU. Look at the trajectory. Yann announced his Meta departure in November 2025 after 12 years and called founding FAIR his "proudest non-technical accomplishment." On March 10, 2026, AMI Labs closed the largest seed round in European history at a $3.5B pre-money valuation. Bezos, Nvidia, Samsung, and Toyota all wrote checks. Three days later: a paper showing that JEPA-from-pixels is no longer fragile and no longer compute-heavy. The engineering scaffolding that made it look like an academic curiosity is gone. The authors sit at Mila, NYU, Samsung SAIL, and Brown. None at Meta. Yann's bet was that the path to machine intelligence runs through world models, not language models. He left a public company to build it. Each JEPA paper from his network resets the assumed cost structure for that bet. This one makes world modeling laptop-cheap. Meta still has the GPUs. The architecture left.

AI will transform the workforce. But some companies are already showing how badly that transition can be handled. The future of AI will not only be defined by technology. It will also be defined by leadership and humanity. https://t.co/b9qFQX0Rf4 @andrewrchow @time

AI can help people write faster. But the real danger is losing the struggle that turns thoughts into original ideas and human expression. The future will belong to people who use AI to enhance thinking, not replace it. https://t.co/PKQjxBz4XN

AI generated identical résumés for a man and a woman. The woman was judged more harshly for using AI, despite the content being the same. The future of AI will not only reveal the power of technology. It will also expose human bias. https://t.co/sS2BrdLgHQ @fortunemagazine

AI is not simply replacing jobs. It is changing how work gets done, which skills matter and how companies are structured. The biggest risk may not be AI itself. It may be failing to adapt to the transition. https://t.co/cU4eQH1gNC

New in Claude Code: agent view. One list of all your sessions, available today as a research preview. https://t.co/NnbsAQjSPW

New in Claude Code: agent view. One list of all your sessions, available today as a research preview. https://t.co/NnbsAQjSPW

we've been quietly building @typedotcom. today we're sharing it with the world. type is the first multiplayer ai product for marketing, sales, support, coding, and much more. create ai teammates that work with all your tools and everyone at your company. https://t.co/l0Lu4Q0Dn4

One thing San Francisco has that even San Jose doesn't is tons of events for AI people and @michelleefang keeps track of them. So does my AI at https://t.co/8L5xphk0qQ, mostly by reading Michelle's posts here. In fact, I talk to my agent here since it reads my profile and uses that as extra signal of things to add. So, Braygent, (what I call the agent that creates Aligned News), please add all these to your lists.

introducing folk beta the personal ai that lives in your texts and quietly runs your life. built for everyone. just text a number. https://t.co/0q7G07xSnk

now, your agent can fix itself. introducing raindrop triage. an agent for finding and investigating agent issues. https://t.co/daJfrn9BJt

agent-browser v0.27 Big day for agents and browsers → React introspection: react tree, react inspect, react renders, react suspense for component trees, props/hooks/state, render profiling, and Suspense analysis → Web Vitals: vitals command reports LCP, CLS, TTFB, FCP, INP + React hydration phases → SPA navigation: pushstate for client-side nav without full page loads → Init scripts: --init-script and --enable flags to register scripts before first navigation → Network route filtering by resource type → cURL cookie import (JSON, cURL, Cookie-header formats) → Dashboard works behind reverse proxies Thanks to @thoma33 @andrewqu @shaper for being part of this release! https://t.co/UefAqZ5ShH

Really excited about nteract https://t.co/qvvOuQASJv https://t.co/qPA6TCMP26

The migration tool in the Codex app now supports both Code and Cowork. There's never been a better time to switch. https://t.co/6SfXoghb7z

When I hit the Update button in Codex and then the updated version of Codex also has an Update button https://t.co/ByhFYUsudd

When I hit the Update button in Codex and then the updated version of Codex also has an Update button https://t.co/ByhFYUsudd

Claude chrome extension still has a key advantage over codex browser use AFAIK (I hope it reaches parity though!) https://t.co/dQgfSZfvv6

I will take concepts from @heyclicky 's Point and Talk interface by @FarzaTV to improve my own work. So I studied the OSS prototype of it to understand how it works. Here's the highlights: 𝐖𝐡𝐲 𝐩𝐨𝐢𝐧𝐭 𝐚𝐧𝐝 𝐭𝐚𝐥𝐤: The chat box isn't a great interface for an LLM. A blue triangle landing on a button is a more natural way to say "click here" than "in the top right, between the search field and the avatar, you'll see a small icon." 𝐓𝐡𝐞 𝐭𝐫𝐢𝐚𝐧𝐠𝐥𝐞 𝐚𝐧𝐝 𝐭𝐡𝐞 𝐨𝐯𝐞𝐫𝐥𝐚𝐲: Clicky create an invisible full screen overlay on each window and draws the flying blue triangle on that. Clicks fall through the window, and this window doesn't know what apps are behind it. 𝐎𝐧𝐞 𝐭𝐚𝐠, 𝐨𝐧𝐞 𝐫𝐞𝐠𝐞𝐱: Clicky uses regex and prompting to get coordinates from the model. I expected some structured output tool calling, but it simply asks the model to put coordinates at the end 𝐖𝐡𝐚𝐭 𝐂𝐥𝐚𝐮𝐝𝐞 𝐬𝐞𝐞𝐬: Clicky takes one screenshot per monitor (resized and at lower quality) which lets it see the screen and read UI text. It filters out it's own window so it sees what the user sees, minus Clicky 𝐂𝐨𝐨𝐫𝐝𝐢𝐧𝐚𝐭𝐞 𝐦𝐚𝐭𝐡: Claude returns coordinates in the screenshot's pixel space. Getting the buddy to that spot on the display takes multiple transforms to account for differing coordinate spaces, sizes, multiple monitors, and different coordinate systems in different places. Check out the post for full details: https://t.co/PiKmzraGuk Or try the Clicky product. It's more stable, spawns agents, got google integrations, and more stuff that the OSS prototype doesn't: https://t.co/Z5sT0qe366

+1 more

// Scalable Patterns for Agentic AI Workflows // Besides context engineering, we should be putting a lot more system engineering efforts around agents. This paper shows an example of why it matters. (bookmark it) Let's start with an important question: Where does your agentic RAG pipeline actually lose time? It's almost never the LLM call. It's usually the data plane underneath. Serialization between preprocessing, embedding, and vector retrieval, plus coordination overhead between distributed services. New work introduces AAFLOW, a unified distributed runtime that models agentic workflows as an operator abstraction over Apache Arrow and Cylon. A zero-copy data plane connects preprocessing, embedding, and retrieval directly. Resource-deterministic scheduling and async batching cut coordination cost. The result: up to 4.64× pipeline speedup and 2.8× gains in embedding and upsert phases, with comparable LLM throughput. None of that comes from LLM inference acceleration. It all comes from cleaner data flow. Paper: https://t.co/9fqkRRsV39 Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX

Cool paper from Apple. Most evaluation of tool-calling agents happens after the trajectory is over. By then the wrong call has already shipped. This new paper moves evaluation into the execution loop. A specialized reviewer agent inspects each provisional tool call before it executes. If something is off, it injects feedback and the primary agent revises. To quantify the tradeoff between corrections and new mistakes, they introduce Helpfulness-Harmfulness metrics. Helpfulness measures the percentage of base errors fixed; harmfulness measures correct calls degraded by the reviewer. Results on BFCL: +5.5% on irrelevance detection (84.9% to 90.4%), +1.6% on relevance, all with no retraining of the base agent. On τ²-Bench multi-turn: +7.1% (48.7% to 55.8%). Reasoning-model reviewers get a 3:1 benefit-to-risk ratio vs. 2.1:1 for GPT-4o. Adding GEPA prompt optimization stacks another +1.5–2.8%. Why does it matter? You can keep the base tool-calling agent frozen and still ship measurable accuracy gains by improving only the reviewer. Model selection and prompt optimization on the reviewer become real, separable production levers. Paper: https://t.co/L0p0UBFcI0 Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX