ReadBench evaluates bypassing text tokenization by rendering long-context documents as images for vision-language models
The technique builds on Google's pixel-only CLIPPO model design
501.7K2171.3K195.1K
“When frontier labs suddenly cut costs, or say "we found a way to dramatically cut inference memory!!" This is what they found. The secret sauce. Vision always wins! Of course, my very capable ex-colleagues who now work on Gemini and Claude already found out years ago: https://x.com/giffmana/status/2073318749273231394/photo/1 https://twitter.com/intcyberdigest/status/2073200409439568031”
GiffmanaTECH#72Story Brief
ReadBench evaluates bypassing text tokenization by rendering long-context documents as images for vision-language models is circulating through @giffmana, with the visible post framing the story around The technique builds on Google's pixel-only CLIPPO model design. So far, the public evidence is narrow, so the safest read is what the post says and what remains unverified. The useful read for readers is therefore limited but clear: what was shown or claimed, who put it in front of the conversation, and what would need more confirmation before anyone treats it as settled.
“@suchenzang The trick is to choose the winning team long before anyone knows it's the winning team and then later keep rubbing it in with "i told you so"s. ezpz”
GiffmanaTECH#72