more evidence of flash lite with zero thinking budget

April 10, 2026 - 21:21 UTC

4
1
69
free intro to LLMs right here. note the 10x difference in cost and 20x difference in output token count
2
2
38
i agree with your general point but i dislike the vapid "count N letters in this phrase" bench that people use to dismiss LLM capabilities

it's like if i was trying to use a screwdriver to hammer in a nail; it's the wrong way to use the tool!
2
0
2
in the reasoning trace it counts it correctly 4 different times to make sure lol
1
0
27
I’d suggest this is product malpractice but that would require Google to have good product instincts to begin with
1
0
8
I would have to guess that it would be way too expensive to do it right?

But also that's a problem
1
0
0
i've been trying to figure out how to explain that models under these conditions are basically just doing automatic writing.
0
0
3
Flash light instruct with no sandbox
0
0
1