David Weiss

A large fraction of valuable business work is repeatable and decomposable into steps that don't each require frontier intelligence, and the act of decomposing it is where the durable value concentrates.

@signulll Chat is about knowing things. The marginal cost of knowledge is heading toward zero. But doing… Doing things, that’s where there’s a LOT of value, and promoting tasks isn’t just a response to Claude Co-work, it’s about where the value is coalescing.

did openai demote chat to a second class citizen in the new desktop app? the chats appear under a tasks grouping if you execute them as a task but if you start from chats, it’s treated differently. pretty interesting to see the product crew making very intricate tradeoffs as it navigates from a chat first world to an agent first world.

@alexocheema What's the max? 4?

You can stack Macs too. MacBooks, Mac Minis and Mac Studios have native support for RDMA over Thunderbolt 5 (M3 Ultra, M4 Pro, M4 Max, M5 Pro, M5 Max). The interconnect latency is single digit microseconds so clusters scale ~linearly.

🐧 Linus Torvalds gets angry when people say 99% of our code is written by AI, he told the audience at the Open Source Summit.

Agentic loops are a watered down version of process engineering. You should be doing process engineering. Not so much /goal /loop /ralph etc.

This is absolutely insane. Elon Musk's XAI reportedly spent $40 billion to build their data centers Based on public disclosure of the Anthropic and Google deal, XAI will get paid $26 billion per year to license the compute from these data centers That is a payback period of 18 months for all the data center spend, from just two customers And you still think AI infrastructure capex is a bubble?

Divine Intelligence != Human Intelligence != Artificial Intelligence youtube.com/watch?v=Ts5Z64…

Scaling knowledge gives you static competence. Intelligence gives you adaptability.

The U.S. Senate has just rejected a motion to add the SAVE America Act as part of budget reconciliation on a 48-50 vote. Republicans who voted against it: -Thom Tillis -Lisa Murkowski -Mitch McConnell -Susan Collins

@MarioNawfal Voter ID and proof of citizenship are not radical positions. They are basic legitimacy checks. When 80% of the country supports something this obvious and Washington still kills it, people are not being paranoid when they say the system is protecting itself.

Coding agents are like contractors.

🚀 Gemma 4 12B is here! We partnered with @GoogleDeepMind to bring and optimize their new dense and unifed multimodal model for Apple Silicon. ◈ 12B dense · 256K context ◈ Thinking mode (built-in reasoning) ◈ Vision: dynamic res, OCR, UI + charts ◈ Native audio: ASR + speech translation ◈ Function calling for agents ◈ Text + image + audio, interleaved Runs local. Get started now ⚡ > uv pip install -U mlx-vlm github.com/Blaizzy/mlx-vlm

When you run a fleet of agents, system throughput is gated by the human: initiating, reviewing, confirming, and correcting. The bottleneck shows up to be you.

@DavidSacks What is the single biggest threat to Anthropic or any of the frontier proprietary labs? Open source LLMs. Open source LLMs run on local hardware or even on local to the company data centers. The biggest threat is akin to the PC revolution, where things move from mainframes to personal computers, because at that point, the cost per token, (which is the business model of these proprietary labs) goes to the cost of electricity, and disrupts the business model of selling tokens. And IF models are converging in capacity or staying within a few points of each other, then that signals commoditization and these labs DO NOT want that. (Note how the LLMs are increasingly training for in harness interactions, so that if you don't use their harness, the output is "off distribution") No wonder then, that, Anthropic would suggest that AI is dangerous and must be regulated and guardrails must be implemented, and therefore, what's the conclusion? Kill distributed open source LLMs so that their existential risk (price per token going to zero) is fully nullified.

Blue Origin just vaporized a rocket, a launch pad, and Amazon's entire satellite deployment timeline in nine seconds. NG-4 was supposed to fly June 4 carrying 48 Amazon Leo satellites. That mission was the first of 24 contracted Blue Origin launches Amazon needs to build its Starlink competitor. Amazon has roughly 240 satellites in orbit against an FCC requirement of 1,618 by July 2026. They already filed for a two-year extension because they were falling short. Losing your primary heavy-lift rocket on the pad doesn't help that math. The pad damage is the part people aren't thinking about. New Glenn carries roughly 2.4 million pounds of propellant. The explosion toppled one of LC-36's lightning protection towers. That launch complex took years to build and billions to outfit. You can manufacture a new rocket in months. You cannot rebuild a launch pad in months. The cascade gets worse. Blue Origin's Blue Moon MK1 lunar lander is supposed to launch on New Glenn this fall for NASA's CLPS program. That mission is the pathfinder for Artemis III, which needs Blue Moon MK2 to fly on New Glenn in mid-2027 to land astronauts at the lunar south pole. Every month LC-36 sits damaged pushes Artemis further into the late 2020s. Jeff Bezos has two companies betting on the same rocket. Amazon Leo needs 24 New Glenn launches to close the gap with Starlink. NASA needs New Glenn for Artemis. Both timelines just broke simultaneously, and LC-36 is on fire.

@dexhorthy @walden_yan @tobi @karpathy Don't use an LLM when you can just write a script or some focused tool. Use the right tool for the job.

someone hit me up about the new "claude dynamic workflows" feature, claiming "see, multi-agent works" But really, the launch of this feature proves the exact point that I made back in June of 2025, along with @walden_yan, @tobi, @karpathy, and many others: Deterministic workflows orchestrating small agent loops beats non-deterministic multi-agent or "agent soup" systems every dang time everything is context engineering

SpaceX has almost finished writing V1.0 of an in-house AI training stack in C that exact-maps to 220k GB300s with 800G NICs, making heavy use of pipeline parallelism and getting as close to bare metal as possible. The potential speed improvement vs JAX for large training runs is over an order of magnitude.

I found an interesting thing: Due to a bug in my cog system, before I submit a PR I subject the PR locally to a context free review (meaning the coding agent doesn’t know about any of the main context used to author the PR) locally with the same model that built it (Opus 4.7) and it does this 3 times before creating the PR and paying for Copilot to do a review. Results: I almost never get a comment by Copilot, and in the past I’ve really depended on Copilot to catch things Claude Code or Codex missed. Just goes to show that you can increase code quality if you throw tokens at the problem.

@lennysan mechanicaladvantage.ai

What are your top 3 dream companies to work for right now?

Excited to see how Grok Build performs on this new benchmark: deepswe.datacurve.ai