I think 'next token prediction' isn't really accurate, as it's also next sentence prediction and next thought prediction. Sure, those outouts are made of tokens - but that's too reductionist and hides these models' true capabilities.
It's going to be a busy week at TC-39 this week... I've got three proposals on the table I'll be trying to get advanced.
First is TypedArray concat github.com/tc39/proposa...
const u8_1 = new Uint8Array([1,2,3]);
const u8_2 = new Uint8Array([4,5,6]);
const u8_3 = Uint8Array.concat([u8_1, u8_2]);
Memory transfer time. Just goes to show there's no moat and no real lock in.
I've heard feedback on the @markpub.at lexicon! Thanks to @bmann.ca @blaine.bsky.social @thisismissem.social and others who have given really useful comments. I've published a new version for comment that is still very simple, but has more room for flexibility. Let me know what you think! markpub.at
Right, basically Macbeth.
And kind of ironic 😁
it's beginning:
github.com/tldraw/tldra...
Almost missed it, but there's a new mercury diffusion model! It seems to beat Haiku 4.5 while being quite a bit faster. I don't think the first one was open source (?), and the API prices aren't fantastic imo, but it isn't *expensive*, while being really fast (1k tk/s)
I think it's impressive because React is a well thought out UI architecture pattern where a component is just a function transforming state to UI primitives. It scales well because it's composable.
https://lckhd.eu/MiHLUj
There's a big gap between building an AI agent in a notebook and running one in production. The article below goes over the full path, from #LangGraph graph 🧵
Very cool. I was just thinking that something like this could be used for a federated trusted age verification system, where the data isn't held centrally. It would be good enough proof for age from a trusted source for DIDs where regulations required it.
Yeah, that would democratize the data and analytics - a good thing. But I'm still thinking about what the business model should be for the atmosphere, if it's not advertising and attention. Maybe there doesn't need to be one, and affiliation $ to converted premium content is enough.
Data brokers pay a lot to closed systems for data, with some analytics. With an open system the investment won't flow back to the platform.
Yes, but doesn't everyone have access to the firehose? How do the AT protocol platforms benefit financially?
Interop is the future, once you taste it you can't go back to walled gardens
Also - it was really inspiring & fun to work with @margin.at - happy to be building the Atmosphere's annotation layer together ;)
We have moved our ATProto handle to eurosky.social, and you can too! For European data storage, and to demonstrate how a decentralized, open social network looks like in practice. We are testing personal data server (PDS) transfer this week, sign up if you want in www.eurosky.tech/register #eurosky
ATProto is now in IETF standardization process. BOF stage passed at IETF 124 Dublin. Working group formation is next.
datatracker.ietf.org/wg/atp/about/
Wow, but why... All that effort to scale vertically and deal with contention. Surely they have a half decent distributed data lake for training data.
This is amazing....
FastMCP 3.0 is here!
This is our largest and most ambitious release ever.
We've re-architected the framework to support the next generation of MCP applications.
It's time to move fast and make things.
🧵
@lenscrafters.bsky.social (if that's you) - your website has a bug so that you cannot book online appointments. The phone number validation seems to be broken.
You may have noticed that search on the VS Code website has gotten a lot quicker lately.
In our latest blogpost, João Moreno breaks down the engineering process behind docfind, a search engine we built that runs entirely in your browser using WebAssembly: code.visualstudio.com/blogs/2026/0...
atproto ecosystem growth really strengthens my belief that this project is critical to reclaiming the internet we deserved. the protocol offers the interoperability and extensibility that Web 2.0 promised but failed to deliver
Holy shit.
I'm only halfway through this video but it is already one of the best pieces of science communication I have ever witnessed.
Truly amazing even if you are only peripherally interested in microchips -- just from a physics / chemistry / engineering perspective.
Hello atmosphere ! Over the last few weeks, a bunch of us have been collaborating on a new set of lexicons for publishing, indexing, and building out social features across long-form reading and writing apps ! 🧵1/7
Love it...
Astronauts staring back at Earth. Earth is labelled "software development". The first astronaut says "with AI it's no longer about writing code". The second astronaut points a gun at the back of the first astronaut and says "never was".
