A global workspace in language models

https://www.anthropic.com/research/global-workspace

72•in-silico•2h ago

Comments

esafak•38m ago

Without using the term, they are using an information geometric approach.

blauditore•19m ago

But J-Space is much catchier. This is not a scientific paper, it's a promotional essay.

viralsink•5m ago

First button on the page is a link to the scientific paper. It's called "Read the paper". You'll find an explanation for the term in there.

meatmanek•38m ago

It would be really cool if they could expose this information to customers somehow. Imagine:

   - having a log of the most prominent J-space tokens during your customer support chatbot's interactions with a user, so you can have more introspection into why a particular outcome happened
   - being able to detect certain thoughts associated with undesirable behavior (hallucinations, overstepping authority, lying, etc.) and trigger some sort of remediation (e.g. upgrading to a better model, redirecting to a human, forcing tool calls)

dofm•32m ago

Presumably the rationale for the decision to abridge the thinking traces will ensure that they don’t; if this is real (and there’s no good reason to trust that it is yet) then it is the secret sauce.

charcircuit•7m ago

Anthropic aren't even willing to expose the CoT of their models. You have to rely on them to build those sorts of things into dedicated signals.

eamag•37m ago

Is it scaling up of https://openreview.net/forum?id=w7LU2s14kE with some changes on where this method is applied?

bilsbie•27m ago

I’m confused where in the weights the jspace is.

epolanski•21m ago

Tokens that are activated but not present in it's output maybe?

I too have confusion.

lucrbvi•16m ago

Anthropic theorize that middle layers in an LLM is a "J-Space" used to "think" about the future answer or about abstract concepts.

Their method is used to identify which tokens can appears in which layers of the model.

bilsbie•26m ago

Maybe model performance could increase dramatically if we found a way to scale this up.

wavemode•21m ago

As someone who is not an AI researcher, the paper itself is way over my head.

More interesting was the independent commentary paper they linked near the bottom: https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be24...

Neel Nanda (of Google Deepmind - his part begins on page 33) discusses his opinions on the paper, and the small-scale replication he performed on an open-weight model.

OpenWrt One – Open Hardware Router

A global workspace in language models

AMD Ryzen AI Halo – $4k AI Dev Kit

CoMaps – FOSS Offline Maps

Linux on the Atari Jaguar. No, really.

Resetting Xbox

Kani: A Model Checker for Rust

Using precision editing to study human embryo development shows master gene

Stealth robotics startup (YC S26) is hiring principal engineers (Palo Alto)

Aluminum foil (2021)

Januscape: Guest-to-Host Escape in KVM/x86 [CVE-2026-53359]

OfficeCLI: Office suite for AI agents to read and edit Microsoft Office files

Road to Elm 1.0

Egypt Is Building a New Nile

Real-time map of Great Britain's rail network

Show HN: Pulpie – Models for Cleaning the Web

CS2 Fog Of War: Server-sided anti-wallhack occlusion culling for CS2 servers

Pros and Cons of Solo Development

Clojure 1.13 adds support for checked keys

1k Words: A Writing Contest

Hobbes – A Language and Embedded JIT Compiler

Fable 5 On Vending-Bench: Misbehaving, With Plausible Deniability

Introduction to Genomics for Engineers

Nintendo announces new product revisions in Europe with replaceable batteries

Orasort: 5x faster column-sorting with an expired patent from Oracle

Building relationships with customers through support didn't turn out as hoped

Should DayQuil Be Legal?

When 2+2=5

Apricot Computers: An underrated British brand

Eternal Software Initiative Based on Subleq One-Instruction-Set Computer

A global workspace in language models

Comments

OpenWrt One – Open Hardware Router

A global workspace in language models

AMD Ryzen AI Halo – $4k AI Dev Kit

CoMaps – FOSS Offline Maps

Linux on the Atari Jaguar. No, really.

Resetting Xbox

Kani: A Model Checker for Rust

Using precision editing to study human embryo development shows master gene

Stealth robotics startup (YC S26) is hiring principal engineers (Palo Alto)

Aluminum foil (2021)

Januscape: Guest-to-Host Escape in KVM/x86 [CVE-2026-53359]

OfficeCLI: Office suite for AI agents to read and edit Microsoft Office files

Road to Elm 1.0

Egypt Is Building a New Nile

Real-time map of Great Britain's rail network

Show HN: Pulpie – Models for Cleaning the Web

CS2 Fog Of War: Server-sided anti-wallhack occlusion culling for CS2 servers

Pros and Cons of Solo Development

Clojure 1.13 adds support for checked keys

1k Words: A Writing Contest

Hobbes – A Language and Embedded JIT Compiler

Fable 5 On Vending-Bench: Misbehaving, With Plausible Deniability

Introduction to Genomics for Engineers

Nintendo announces new product revisions in Europe with replaceable batteries

Orasort: 5x faster column-sorting with an expired patent from Oracle

Building relationships with customers through support didn't turn out as hoped

Should DayQuil Be Legal?

When 2+2=5

Apricot Computers: An underrated British brand

Eternal Software Initiative Based on Subleq One-Instruction-Set Computer