{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreigylvokhxildtmhcftivzyhwhykfbitoih745l25kbje7xjnawv7m",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mlqvufpe5fx2"
  },
  "path": "/t/biic-replacing-tokens-with-geometric-algebra-multivectors-early-results/175911#post_2",
  "publishedAt": "2026-05-13T17:29:57.000Z",
  "site": "https://discuss.huggingface.co",
  "textContent": "this is an observation i made personally when i started useing LLM APIs.\nan inherant flaw in flat token strings is that it has to be sent as a complete string to the processing data center. if any part is lost it has to be resent.\none my be forgiven for believing that they are tokens and therefor microscopic, so whats the big deal?\n\nthe big deal is that in API useage you are charged for token usage. and each time a token string is recent you are not only charged for that, it increases the size of the token string, because the entire conversation is sent each time.\n\nso basically the conversation doubles in size on each failed send.\n\nin a 4 turn conversation with staggered ammounts of resends, the token count ballons astronomically, and this is not mitigated in anyway over the course of the conversation.\n\nwhy do i say all this, well, your solution offeres something else, the ability for approaching the concept of batch sends. or packet transfer.\na flat token string cannot be batched as far as i know. but this remains a massive issue in AI Remote Compute if one is paying for service.",
  "title": "BIIC: Replacing Tokens with Geometric Algebra Multivectors — Early Results"
}