"Reflection API" is a sonnet 3.5 wrapper with prompt. And they are currently disguising it by filtering out the string 'claude'. teddit.net/r/LocalLLaMA/comm…
Joseph @RealJosephus
Sep 7
We now know that "Reflection Llama 3.1 70B" was nothing more than a LoRA trained directly on benchmark test sets, built on top of Llama 3.0. Those who were fooled lacked some basic common sense. teddit.net/r/LocalLLaMA/comm…

Sep 8, 2024 · 10:10 PM UTC

54
68
76
1,025
they see this and remove the filtering, but it is still sonnet-3.5
4
1
128
I've flagged around 100 X accounts that were part of the Reflection hype. About 10 took themselves off the list. The rest are either pushing the hype, bots, or people who are still in the dark about the whole thing. x.com/i/lists/18329249236666…
4
3
37
Replying to @RealJosephus
how do we know it's sonnet (expected) and not just trained on excessive opus outputs (unexpected)
4
51
easy: tokenizer won't lie
intervitens
 @intervitens
11h
The reflection model on their API is just prefilled Claude 3. (All of the models are limited to 10 tokens, temperature 0, top_k 1)
2
2
3
160
more replies
They are keeping eyes on our reactions and making quick 'fix'
4
76
man is busy 'cooking' right now:
indexclone @indexclone1980
10h
Replying to @johnlu0x @RealJosephus
They replaced it with got 4o now. God knows what else they will do
3
24
Replying to @RealJosephus
It’s not Claude. That’s silly. If they are filtering out the tag ( or the model is trained to suppress it ) it’s because the fine tuned on Claude data.
1
12
yep, it's not claude, but 4o now bro is hardworking
/ @gazorp5
10h
Replying to @xiaoqianWX @indexclone1980 @johnlu0x @RealJosephus
Nice find.
2
40
more replies
Replying to @RealJosephus
shameless. they took us for fools
1
16
time to make a fool-list on X, many still believed
2
19
Load more