Technology

Model can be downloaded on Huggingface: https://huggingface.co/IntervitensInc/pangu-pro-moe-model

In modern LLM applications like RAG and Agents, the model is constantly fed new context. For example, in RAG, we retrieve relevant documents and stuff them into the prompt.

The issue is that this dynamically retrieved context doesn't always appear at the beginning of the input sequence. Traditional KV caching only reuses a "common prefix," so if the new information isn't at the very start, the cache hit rate plummets, and your GPU ends up recomputing the same things over and over.

CacheBlend changes the game by allowing for the reuse of pre-computed KV caches regardless of their position in the input sequence.

This makes it possible to achieve a 100% KV Cache hit rate in applications like RAG. The performance gains are significant:

• Faster Time-To-First-Token (TTFT): Get your initial response much quicker.
• More Throughput: Serve significantly more users with the same hardware.
• Almost lossless Output Quality: All of this is achieved with little degradation in the model's generation quality.

CacheBlend works by intelligently handling the two main challenges of reusing non-prefix caches:

• Positional Encoding Update: It efficiently updates positional encodings to ensure the model always knows the correct position of each token, even when we're stitching together cached and new data.
• Selective Attention Recalculation: Instead of recomputing everything, it strategically recalculates only the minimal cross-attention needed between the new and cached chunks to maintain perfect generation quality.

An interactive CacheBlend demo is available at: https://github.com/LMCache/LMCache-Examples/tree/main/demo-rag-blending

https://archive.ph/lC5R0

https://archive.ph/xbx3M

We are constantly fed a version of AI that looks, sounds and acts suspiciously like us. It speaks in polished sentences, mimics emotions, expresses curiosity, claims to feel compassion, even dabbles in what it calls creativity.

But what we call AI today is nothing more than a statistical machine: a digital parrot regurgitating patterns mined from oceans of human data (the situation hasn’t changed much since it was discussed here five years ago). When it writes an answer to a question, it literally just guesses which letter and word will come next in a sequence – based on the data it’s been trained on.

This means AI has no understanding. No consciousness. No knowledge in any real, human sense. Just pure probability-driven, engineered brilliance — nothing more, and nothing less.

So why is a real “thinking” AI likely impossible? Because it’s bodiless. It has no senses, no flesh, no nerves, no pain, no pleasure. It doesn’t hunger, desire or fear. And because there is no cognition — not a shred — there’s a fundamental gap between the data it consumes (data born out of human feelings and experience) and what it can do with them.

Philosopher David Chalmers calls the mysterious mechanism underlying the relationship between our physical body and consciousness the “hard problem of consciousness”. Eminent scientists have recently hypothesised that consciousness actually emerges from the integration of internal, mental states with sensory representations (such as changes in heart rate, sweating and much more).

Given the paramount importance of the human senses and emotion for consciousness to “happen”, there is a profound and probably irreconcilable disconnect between general AI, the machine, and consciousness, a human phenomenon.

https://archive.ph/Fapar

https://archive.ph/Q9FEn

https://github.com/VectorSpaceLab/OmniGen2