How much of what a neural network learns is complexity introduced by the language we give it?
“Breaking the Token Ceiling,” from researchers at Meta and UW, explores how changing the representation changes model efficiency and performance.
arxiv.org/abs/2609.12303
1/n
This connects directly to our work at Heurai.
With AMAR, our proprietary kernel language for meaning, we’re asking: if meaning and its operations are explicit in the language itself, how much less does a model need to rediscover?
2/n
1
1
4
heurai retweeted
Replying to @GoogleResearch @IamEmily2050
heurai retweeted
Replying to @marcba
LLMs hallucinate.
hallucination is a type of thinking.
reasoning is another type.
LLMs satisfy a subset of thinking, not the whole
heurai retweeted
Replying to @heurai_ai
@heurai_ai already has continual learning. no manual retraining. no “training” at all, actually. (not llm) it queries and captures knowledge and forms representations/concepts as the need for it appears
Replying to @iandcosmos
That’s level 3 or 4. Continual learning is not a requirement, AI models are already better than 99% of humans and getting better 10x faster than humans even if it’s not continuous yet.
heurai retweeted
an LLM would say the color is blue. @heurai_ai’s architecture reasons logically that the user has not provided enough supporting detail ( where, when, etc) to give a near perfect answer
this is gonna kill openai and claude
like the ease of use of ollama but want the speed of llamacpp?
we made a script to serve ollama models through llamacpp. free + opensource
github.com/heurai/ollama_to_…








