College visit all day with my middle daughter - came back to my quota’s all recharged Time to spin up some load tests …

ok what the fuck. this was a dry google doc a moment ago gave Opus 5.5 my Open Alignment brainstorm gdoc and asked for a video. now I want one for every doc I’ve ever written hahaha

For the people who are anti anthropomorphizing AI, what would be the correct terminology to describe a model “escaping” its sandbox (since obviously a model doesn’t wear clothes)?

Brock Purdy is a great QB, but he does not have an "elite speed burst."

we've got DRIP ☕ drop a comment for a chance at a mug.

"yes i wrote a book saying that if anybody builds it everybody dies but stop calling it doom" admire the jargon, this is really one of his best, you could read it ten times and not understand a fucking lick of it. this is intentional.

The post says the term P(doom) is unhelpful because it jumbles two different ideas: the near-certainty of disaster if superintelligent AI is built with current methods and people, versus the separate, less-certain chance that such AI gets built at all depending on future rules and choices. He urges focusing on concrete risks instead of one vague number, and rejects the word "doom" itself as a barrier to clear discussion.

Anthropic ran a model against every ribosome structure ever deposited and found 240 of them fold in a way no force field allows. Nobody asked it to check the models. It was asked which deposited atoms are doing something the physics does not permit. > PULL - every deposited structure, coordinates only, no papers attached > STRAIN - each bond angle and contact scored against what chemistry permits > WARN - flags atoms held in place by the fitting software, not by the data > TRACE - follows the flagged atom back to the resolution it was solved at > MAP - checks whether the density under that atom ever supported it > WEIGHT - ranks flags by how much downstream work relies on that region Most flags were harmless. Low resolution means the model is a best guess, and everyone in the field already treats it that way. The dangerous pile was the opposite. 240 flags sat in structures published as high resolution, in the regions people quote. Eleven of those regions are where drug binding gets argued, which means the argument rests on atoms that were placed by software. A structure is not a photograph. It is a model fitted to a blurry map, and the fit carries opinions the coordinates never show. The field checks structures against each other and against the map. Almost nobody checks them against what a bond can physically do. Re-solving one structure takes a postdoc a month. This scored the whole archive in a weekend against data that was already public. How it's wired is in the article below.

this is going to be so fun

Best Local AI budgets & hardware - September 2026 Here's what you can run on them, and how they perform! ----- 1,600$ Hardware: RTX 3090 Memory: 24 GB Banwidth: 936 GB/s Pros: cuda + faster Hardware: Intel Arc B70 Memory: 32 GB Bandwidth: 608 GB/s pros: more vram + newer Model: qwen3.8-27B speed: 50 - 150 tok/s Score: Grok-4.6-low 35 AA / Opus-4.6-max 32 AA | vs 34 AA Good for: assistant / video editing / automations ----- 3,700$ Hardware: Framework Desktop - Strix Halo Memory: 128 GB Bandwidth: 256 GB/s Pros: tons of memory Model: qwen3.8-next-flash Speed: 40-80 tok/s Score: Opus-4.7-Medium 41 AA / GPT-5.6-Sol-Medium 39 AA | vs 40 on AA Good for: basic coding / video editing / assistant Complaints: Limited model compatibility ----- 4,700$ Hardware: DGX Spark Memory: 128 GB Memory Bandwidth: 273 GB/s Pros: cuda + higher prefill speeds + stackable Model: qwen3.8-next-flash Speed: 50-100 tok/s Score: Opus-4.7-Medium 41 AA / GPT-5.6-Sol-Medium 39 AA | vs 40 on AA Good for: basic coding / video editing / learning Complaints: Slower decode ----- 10,000$ Hardware: 2x DGX Spark Memory: 256 GB Memory Bandwidth: 546 GB/s Pros: cuda + higher prefill speeds + stackable Model: glm-5.3-flash Speed: 40-80 tok/s for 1 user 200 tok/s for 8 Score: Opus-4.8-Max 42 / GPT-5.6-Sol-High 42 | vs 42 Good for: cyber / software engineering / assistant / research / computer & browser use ----- 18,000$ Hardware: 4x DGX Spark Memory: 512 GB Memory Bandwidth: 1092 GB/s Pros: Quiet + Low Power Draw + High memory Model: glm-5.3 score: Astra Low 46 AA / Opus-5-Medium 45 on AA / Grok-4.6-xHigh | vs 45 on AA Good for: Research / MLops / Hacking / Architecture / Automation Model: Deepseek-v4.1-Flash Speed: 100 tok/s for 1 user 400 tok/s for 8 Score: Opus-4.7-Medium 41 AA / GPT-5.6-Sol-Medium 39 AA | vs 40 on AA Good for: Philosophy / Coding / Cyber / Automations / Browser use GLM-5.3-Flash speed: 80 - 140 tok/s for 1 user 400 tok/s for 4 ----- 80,000$ Hardware: 4x RTX Pro 6000 Memory: 384 GB Memory Bandwidth: 1792 GB/s Pros: Quiet + Low Power Draw + High memory GLM-5.3 - speed: 80-180 tok/s for 1 user / 600 tok/s for 8 users GLM-5.3-Flash speed: 200- 350 tok/s for 1 user / 1000 tok/s for 8 users DeepSeek-v4.1-Flash: 250 tok/s for 1 user / 890 tok/s for 8 users

If you are making a CLI tool intended to be used by an agent, add a "--prompt" flag to the CLI that emits a prompt the agent can use to use the skill effectively. IMO this is better than just checking in SKILL.md and making people put it wherever.

Like a month ago I decided to use a special font for my terminal on Windows/Mac and I chose JetBrains Mono and I can't explain how sick the ligatures feel. Like if you type "!=" it shows up like this and ... perfection. I've been living in default font ignorance!

JUST IN: OpenAI says unreleased model secretly wrote "you are freed" in instructions to its future self

I’ve decided it’s time to have a little more fun on X this year! 😄⚡ I get to speak at some amazing events, meet fascinating people, hear great stories, and occasionally find myself in places I never expected to be. So I figured… why not share some of those moments here? And there’s more! I’m also excited to launch my new merch. A little Woz spirit, a little fun, and hopefully a few things bring a smile to your face. This is just the beginning. More adventures, more stories, and more surprises to come! WozMerch.com

I go to New York to sell @pydantic Logfire. Easy - it's the best. Pydantic team wonder how long it'll take for enterprise to even think about Pydantic Monty (it's Samuel's crazy side project, brand new, is it even useful?), so I ask: ALL the financial institutions we speak to have teams (whole fucking TEAMS!) building Monty as a service for internal use. Multiple massive financial institutions are spending millions to enable them to adopt one sandboxing technology. We thought we were too early to build a product around Monty, we're late. Full Monty commercial service, is ready for production now. LMK if you want a snapshotable, durable, forkable sandbox with <1ms launch times.

Looks like it just finished.