Actual picture of me comparing the performance of my fantastic new architecture compared to a vanilla decoder transformer when I do a proper normalization of flops and memory.
4:00 AM · Jun 21, 2026 · 1.8K Views
Meta FAIR's François Fleuret jokes about new model architectures failing against vanilla transformers once normalized for FLOPs and memory
Story Overview
François Fleuret posted a meme showing his own stunned reaction after a new architecture he developed underperformed a basic decoder transformer once FLOPs and memory were properly equalized, turning the moment into a self-deprecating joke about how hard it remains to beat the established baseline.
531322.3K
Open Question
Tuned baselines resist easy replacement
A quick reply noted that transformer++ variants stay exceptionally strong when fairly normalized, so any claimed gains must clear that high bar before they count.
FYI
Even top researchers find the bar exhausting
Fleuret replied that the exercise wears him out, an admission that proper apples-to-apples tests often flatten ambitious new designs without revealing what the next leap should actually be.
Sentiment
Positive users look forward to solutions for transformer issues from the new architecture, while negative users dismiss the effort as exhausting and mock it as weak.
Pos
25.0%
Neg
75.0%
4 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS458LIKES4
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
@francoisfleuret tuned transformer++ is an insanely strong baseline
François Fleuret@francoisfleuret
Actual picture of me comparing the performance of my fantastic new architecture compared to a vanilla decoder transformer when I do a proper normalization of flops and memory.
1hViews 458Likes 4Bookmarks 0
REPLIES1
kache@yacineMTB
@francoisfleuret Give up
François Fleuret@francoisfleuret
Actual picture of me comparing the performance of my fantastic new architecture compared to a vanilla decoder transformer when I do a proper normalization of flops and memory.
39mViews 377Likes 3Bookmarks 0
François Fleuret@francoisfleuret
@teortaxesTex TBH it's exhausting.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
@francoisfleuret tuned transformer++ is an insanely strong baseline
1hViews 125Likes 3Bookmarks 0
WuBu ⪋ WaefreBeorn 🇺🇸 👑@waefrebeorn
@yacineMTB @francoisfleuret this you rn? being broke?
nah i am the broke one!
35mViews 8
J@parkhjaey
@francoisfleuret Your architecture’s got it skipping leg day, clearly.
18mViews 5
Militant Hitchhiker ♥@MilitantAI
@francoisfleuret I've been looking forward to what comes out of your effort on this. I've been working on some solutions to various transformer issues myself. It'll be interesting to see which direction you took.
14mViews 3
Avery@wveriy
@francoisfleuret The classic baseline problem.
13mViews 2