Video scraping is only one of many new tips doable when the newest giant language fashions (LLMs), similar to Google’s Gemini and GPT-4o, are literally “multimodal” fashions, permitting audio, video, picture, and textual content enter. These fashions translate any multimedia enter into tokens (chunks of knowledge), which they use to make predictions about which tokens ought to come subsequent in a sequence.
A time period like “token prediction mannequin” (TPM) may be extra correct than “LLM” these days for AI fashions with multimodal inputs and outputs, however a generalized various time period hasn’t actually taken off but. However it doesn’t matter what you name it, having an AI mannequin that may take video inputs has attention-grabbing implications, each good and doubtlessly dangerous.
Breaking down enter limitations
Willison is much from the primary individual to feed video into AI fashions to realize attention-grabbing outcomes (extra on that beneath, and this is a 2015 paper that makes use of the “video scraping” time period), however as quickly as Gemini launched its video enter functionality, he started to experiment with it in earnest.
In February, Willison demonstrated one other early utility of AI video scraping on his weblog, the place he took a seven-second video of the books on his bookshelves, then received Gemini 1.5 Professional to extract all the ebook titles it noticed within the video and put them in a structured, or organized, record.
Changing unstructured information into structured information is essential to Willison, as a result of he is additionally a data journalist. Willison has created instruments for information journalists previously, such because the Datasette project, which lets anybody publish information as an interactive web site.
To each information journalist’s frustration, some sources of knowledge show immune to scraping (capturing information for evaluation) as a result of how the information is formatted, saved, or introduced. In these circumstances, Willison delights within the potential for AI video scraping as a result of it bypasses these conventional limitations to information extraction.
