@elonmusk Grok does not think. Neither does any other "AI". LLMs aren't winning Maths Olympiad by doing maths. They are doing it by pattern-matching. They can't even count. 🧵
@elonmusk You can argue that counting is just pattern-matching, but infants (in many species) show innate abilities in counting, grammar, geometric/spatial reasoning from birth. There is no learned pattern to match.
@elonmusk Cognition itself seems to understand these primitives from birth. And an orphaned squirrel raised in isolation innately knows about collecting and burying nuts. These are not learned behaviours. They are evolutionarily-encoded cognitive primitives.
@elonmusk They can be refined by learning; some abilities can be unlocked by learning e.g. you aren't born knowing programming, but some people still have a talent; coders apply pre-existing cognitive primitives (logical reasoning, pattern recognition, systematic thinking) to a new domain.
@elonmusk The talent is latent in their cognitive architecture until they learn programming.
But LLMs have no innate abilities. Every capability is trained into them; it is all pattern-matching - unless there is something emergent going on but there isn't yet much evidence of this.
But LLMs have no innate abilities. Every capability is trained into them; it is all pattern-matching - unless there is something emergent going on but there isn't yet much evidence of this.
@elonmusk Biological intelligence starts with foundational cognitive primitives (counting, spatial reasoning, causal thinking, social cognition) and then uses learning to refine and extend these.
@elonmusk LLMs start with nothing and try to reconstruct everything through pattern matching from text. Their "understanding" is entirely derived from patterns in training data, not intrinsic conceptual frameworks.
@elonmusk And that being said if there is emergent behaviour - how does it sustain, between instances of an LLM? Two consecutive prompts in any LLM will be run on fresh instances of that LLM.
@elonmusk Even if something genuinely emergent happened in one prompt/response sequence, it disappears when that instance terminates.
@elonmusk There's no way for insights to compound or evolve across interactions. Each conversation starts from the same trained checkpoint. Any emergence is transient, at best.
@elonmusk What we call "emergent" behaviours in LLMs are just fancy combinations of trained patterns not genuine emergence. True emergence would have to create lasting changes to the system's capabilities; but LLMs are fundamentally stateless.
@elonmusk So we have pattern-matching systems trying to simulate the outputs of systems with genuine cognitive primitives. Giant karaoke machines. [ENDS]
@elonmusk @threadreaderapp unroll
• • •
Missing some Tweet in this thread? You can try to
force a refresh
