A very relevant example being the J-space work by Anthropic: www.anthropic.com/research/glo... (I'm not saying we should buy Anthropic's conclusions here, only pointing this out for the interpretability techniques).
A global workspace in language models \ Anthropic anthropic.com
AI Weekly's analysis
→
- Anthropic says Claude has a 'J-space' of dozens of concepts, under a tenth of neural activity, that mediates multi-step reasoning.
- Swapping 'spider' for 'ant' inside the J-space changed Claude's leg-count answer from 8 to 6, demonstrating a causal role.
- A 'J-lens' tool surfaced silent words like 'fake', 'fictional' and 'manipulation' during deception tests, pointing at safety uses.
Read full analysis →
"Demis Hassabis is leaving his role as CEO of Google DeepMind to be the unit's Chairman. Chief scientist Jeff Dean and another Google AI executive are leaving to start their own company, which Google will invest in." wat www.axios.com/2026/08/05/g...
axios.com
View on Bluesky ·
♥ 10
↻ 0
↩ 1
·
6 from the directory shared this ·
33d ago
I don't even think this is exaggeration re "do". E.g. from Inie et al. (emphasis mine): "[...] examples of language that ascribe false capabilities to the model such as reasoning, problem solving, and the *active use of anything else*." (also "capabilities" itself) firstmonday…
firstmonday.org
As usual, for further arguments for why none of this has an obvious answer as many would like to claim, I'd point to Schwitzgebel: arxiv.org/abs/2510.09858 Chalmers (e.g.): www.youtube.com/watch?v=bskf... Shevlin: www.polytropolis.com/p/contra-chi...
AI and Consciousness arxiv.org
AI Weekly's analysis
→
- Philosopher Eric Schwitzgebel argues we will soon build systems judged conscious by some mainstream theories and not others, with no way to settle it.
- The paper surveys Global Workspace, Higher-Order and Integrated Information theories alongside the Turing Test and Chinese Room arguments.
- Its blunt verdict: none of the standard arguments for or against AI consciousness takes us far.
Read full analysis →
How so? They don't seem to mind: www-cdn.anthropic.com/files/4zrzov...
www-cdn.anthropic.com
Many don't see how much AI has changed recently. Coding agents aren't just chatbots. We can’t deal with the challenges of AI if we don’t understand it. Here's a post to provide some evidence (Claude doing small ML experiments) and form intuitions. davidpreichert.substack.com/p…
If you haven’t recently used Claude Code*, you might not understand where AI is at davidpreichert.substack.com
View on Bluesky ·
♥ 83
↻ 8
↩ 7
·
3 from the directory shared this ·
35d ago
Incidentally, I used to work a bit on getting synchronisation effects into neural nets... this was meant to model spikes, but either way came down to passing continuous complex numbers around (if still on a discrete grid): arxiv.org/pdf/1312.6115
arxiv.org
Why would that not also be ML though? The idea of learning-to-learn(-at-test-time) has been around for a while in ML. E.g. en.wikipedia.org/wiki/Meta-le... and indeed arxiv.org/pdf/2005.14165 is arguably also framed that way.
arxiv.org
Nice! The Zork playing agents were Sonnet 5 (and Haiku 4.5) actually, only the researcher agents were Fable. But there's a lot of setup details that can matter here. You can find (the researcher agent's) write up here if you're interested: github.com/davidpreiche... ...
zork-experiment/REPORT.md at main · davidpreichert/zork-experiment github.com
... would have told us right away that rare = not many in circulation, etc. Or see this Guardian article: www.theguardian.com/technology/2... First picture is a 400 year-old manuscript, which has nothing to do at all with the actual books in question...
‘More than just objects’: Australian booksellers raise alarm over ‘horrific’ destruction of rare titles to feed AI theguardian.com
As usual, for further arguments for why none of this has an obvious answer as many would like to claim, I'd point to Schwitzgebel: arxiv.org/abs/2510.09858 Chalmers (e.g.): www.youtube.com/watch?v=bskf... Shevlin: www.polytropolis.com/p/contra-chi...
David Chalmers: Could a Large Language Model be Conscious? youtube.com
