Post

Conversation

Siqi Chen
@blader
tl;dr LLMs are already neurosymbolic in its latent space this is the mechanistic explanation for the intuitively obvious "feel" that the stochastic parrot crowd never understood
Quote
Anthropic
@AnthropicAI
New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.
The media could not be played.
Last editedOpens edit history
64.9K
Views
View quotes
David Watson 🥑
Post your reply

Chelm's Deep
@ChelmsDeep
To me the description is incredibly simple, llms are a simulation of the “word thinking” part of the human brain That is the part that can be tricked and make up things that are not true. It can also reason and categorize and analyze Not the sum total of intelligence, not
728
JMB 🧙‍♂️
@jmbollenbacher
i dont think that's the tldr at all. i think you heard what you wanted to hear, not what was said.
889
Youssef El Manssouri
@yoemsri
The people who called it a stochastic parrot were looking at the outputs and reasoning backward. The interpretability work is looking at the internals and the picture is completely different.
344
Ibrahim Dagher
@IbrahimDagher20
What do you mean by “neurosymbolic”. What they’re saying here is that there’s a “scratchpad” (the residual stream’s vectors) which the neural network works through as it reasons about something. Much as you have a stream of thought as you reason. I suppose this is in some sense
419
Sam Tilston
@samtilston
Mechanistic interpretability is useful because it moves the debate from analogy to evidence Claims about internal representations need to be tested through experiments, not just inferred from model behaviour
5
Intersignal
@intersignal_ai
We pioneered this months ago 👏 the models are smart and endpoint human charset/language is merely the skin on whatever the underlying weights' capability looks like
291

Discover more

Sourced from across X
bubble boi
@bubbleboi
So let me get this straight. An account called “Bog Boss” just decided to post a chip architecture that seems to beat Nvidia at inference and it’s uses LPDDR, an old TSMC node, and *checks notes* 33 millimeters of MACs with the rest of the area going to DDR PHYs & 32 Gb SERDES
Quote
Big Boss
@0xBADB01E
Replying to @0xBADB01E
First you have to understand that modern LLM inference already disaggregates weights as models outgrew single chips years ago. You shard either by layer (pipeline parallelism) or by slicing every layer (tensor parallelism), and the two do very different things. As an example,
546K
alice
@aliceisplaying
OpenAI: here. have INFINITE TOKENS. if you somehow run out, here are several buttons to get MORE INFINITE TOKENS Anthropic: okay kids you get ONE WEEK of Fable and ONLY AT 50%
114K
Sam Altman
@sama
our older kid put two words together for the first time and i am approximately as amazed by this cognitive feat as i am by GPT-5.6 discovering new math
1.5M