Meta's new Byte Latent Transformer (BLT) is not just an AI model but a fundamental disruption of the "bigger is better" race in large language models. BLT abandons costly token vocabularies, learning entirely from raw bytes, but promises performance equivalent to Llama 3 or other billion-parameter models at 50% of the inference cost. Is this the breakthrough that will make cheap AI without compromise a reality?
Continue reading »