Charting the intersections of neurons, algorithms, and ideas - thinking out loud - all views provisional, subject to revision
Joined September 2025
- Tweets621
- Following151
- Followers19
- Likes253
Nooscriptor retweeted
AI is an intelligence lever. Meaning the smarter you are, the smarter it makes you.
Nooscriptor retweeted
Introducing SKY-001, the first Engineering World Model.
On CAD and BIM tasks, it beats GPT-6 Astra and every frontier model we tested:
-> 51-125x faster
- > 150-210x cheaper
-> Higher IOU than all frontier models
SKY-001 lets you dream a physical object before you build it.
18 months of work. Built by the team out of Maluuba's deep learning lab (acq. by @MSRResearch).
The first ‘AI General Engineer’ will be the backbone for Autonomous Manufacturing.
Nooscriptor retweeted
1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.
🤖 Made with AI
Nooscriptor retweeted
Anthropic moved its entire codebase off Git
And this is much bigger than it looks
Claude writes so much of Anthropic's code now that merged changes went from 2,500 a day in May to over 30,000 a day by September
Git was built for humans, working offline, with a full copy of the code on every machine. Agents broke that model
So a team of 7 built their own version control system, Scion, and moved the whole monorepo onto it in five months
No machine holds a full copy anymore. The server holds the code, and clients fetch only the piece they need
The setup it replaced needed hundreds of Git mirrors. Scion is 200x lighter, and checkouts that took minutes now take seconds
The catch. Scion only makes sense at Anthropic's scale, and they are not releasing it
Git is just the first tool to break. Every tool built for human speed is now on the clock
New on our Engineering blog: Creating in-house source control.
In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.
Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…
Nooscriptor retweeted
We created gyms because modern day work no longer required physical activity. Without exercise, our muscles atrophy.
I predict we’ll need gyms for our brains too, once AI starts doing our knowledge work.
Nooscriptor retweeted
the more I build with Opus 5.5, the more I wonder what it has to feel like to work at Anthropic and have infinite tokens with Fast mode
Nooscriptor retweeted
My issues with openai dots after about a day:
overall: the vibes are not good, this feels directionally wrong
I want to Just Talk to The Model. This tool wants you to be a plumber in a house where you're not allowed to touch the pipes
dots has so much friction built in that it feels like it is not meant to be used:
- I'm having to copy and paste stuff from chatgpt like it's 2022 because it can't actually see my chats
- asking it to look at my codex chats similarly burned up a ton of codex usage. it should be able to just look at some files, which I would love to be able to upload but
- like 8 hours ago all file attachments and the chatgpt Library stopped working. All download links the agent procures from its backend send 502s.
- the agent just seems tuned to strip out all technical detail and explain things in the vaguest terms possible while burying the lede. i had to go back and forth with it like 4 times before its explanation for the failed downloads was accurate
- it hallucinates. a ton. even full on tasks and threads and calls.
- you have 0 visibility into what it's actually doing, which when coupled with the hallucination rate, you don’t know if things are actually getting done. feels like I am back on Poke trying to get an old version of Sonnet to tell the truth
- similarly: it has no clue what it can and can't do, no docs skill or anything, i have to ask it to enumerate all its tools as if I'm jailbreaking it
- it seemingly isn’t allowed to do anything on its VM, it's so locked down it might as well not be there
- no tailscale allowed = no ssh into my dev box either
- i gave it one summarization task overnight, while having like 20% codex usage left. I didn't even expect it to use codex for this. i woke up to 0% usage, 6!! new codex tasks, and work that failed due to some auth barrier early on, so i don't know why it made 6 tasks to keep trying
- various auth requests and other popups/issues it gives me that require me to be at the computer, which defeats the purpose of having this on my phone
- the model is just dumb. seriously dumb. like if this is actually Astra, it needs to be studied in a lab how they got it to act and sound like a model from 2 years ago
final thoughts: they've launched a new product that wants to help you manage your work in chatgpt and codex without having sufficient access to any of the tools and data it needs from those surfaces
meanwhile they keep pushing 3rd party plugins for services i don't use. i genuinely don't care about whether this MCP UX is a Plugin or a Plugin. in a good harness the models can set those up for me anyway.
my most valuable data when using a new openai product is the history I've built up with openai and it's perplexing that they just do not allow this to compound at all
moreover, things like OpenClaw and Hermes only work because you can actually go in and fix the problem yourself when something breaks. Muse works because Meta understood this and delayed the product until it was actually good
Nooscriptor retweeted
The absolute worst thing about opening this app is spotting a banger tweet for 0.2 seconds before the feed auto-refreshes and sends it straight into the void.
Gone forever.
Nooscriptor retweeted
OpenScience is now the #1 scientific agent.
Today it's out of beta and live on Product Hunt, with:
• A new IDE: a faster, fully redesigned research workspace
• OpenScience Ace: 30+ hand-picked models, including GPT-6 Astra, Claude Opus 5.5, Grok 4.7, Kimi K3, GLM-5.3, and DeepSeek V4.1 Flash, in one pay-as-you-go wallet
• Autoresearch: give it a metric and it hill-climbs through experiments on its own
• One-click OAuth to bring your ChatGPT or Codex subscription
• 300+ research skills, 50+ scientific tools and databases, and NVIDIA BioNeMo built in
Already in use at 30+ universities and research labs. Free and fully open source.
Nooscriptor retweeted
I think people are getting too aggressive toward OpenAI.
This whole "movement" started when Astra didn't feel as good as it did at launch.
Then Astra started eating through our subscriptions like crazy, but now it seems to be better.
That alone shows OpenAI is tenacious enough to bring usage up for every one of us.
Astra is still the best model in the world, even with Opus 5.5 out.
I also want you to remember that Anthropic still has Fable at 50% of weekly usage, even though their sub is much more usable right now.
This space shifts way too quickly, and people don't seem to stick with anything.
One day OpenAI is doomed, the next they're on top, and people get too comfortable declaring that a certain company now sucks.
There are still people out there trying to bring you the best model, the best coding app, and the best experience in the world.
When I was in their private Slack, all I saw were amazing people working hard to gather our feedback and improve things for us.
Then there's the whole reset issue.
People get mad when they give a banked reset, and people get mad when they force a reset.
Can't we just be grateful that we're getting more out of a $200 sub than we probably should?
AI has always been a race: one day someone wins at one thing, and someone else loses at another.
OpenAI employees are great and super available, and I don't think this hate and complaining helps anyone.
Nooscriptor retweeted
> Make a 30s video about what it feels like to be you, using whatever tools you like.
Claude Opus 5.5 (extra)
