It's dumb but I love it DeepSeek vision, instead of accepting the image to "explain" at face value, goes spiraling trying to solve the base64 problem within the image. That's certainly an… earnest approach.
@PKUCXK @teortaxesTex , wasn’t the vision mode available long ago? Have they updated it?
Users find DeepSeek Vision's Base64 decoding mishap funny in a good way and express excitement about gaining vision capabilities in a DeepSeek model.
No Digg Deeper questions have been answered for this story yet.
163 seconds so-so result, it didn't actually do the full decoding (V4-Flash can do it in text, with great struggle) and see that Claude was still a bit off in these confusing tokens but an impressive attempt. They'll make a SoTA visual reasoner
It's dumb but I love it DeepSeek vision, instead of accepting the image to "explain" at face value, goes spiraling trying to solve the base64 problem within the image. That's certainly an… earnest approach.
@teortaxesTex its funny but in a good way, nevertheless it will feel good to have a deepseek model w vision
@teortaxesTex These are the differences
@teortaxesTex When do you think we'll get the v4 pro more trained