@awsaf49
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States Android App
Account-level information from X, not a live location or the device used for a specific post.
Graduate Student @ucsantabarbara || Prev: @amazon, @google, @microsoft, @weights_biases || @kaggle Grandmaster
California, USA
Joined December 2011
- Tweets2K
- Following1.1K
- Followers1.1K
- Likes5K
Pinned Tweet
🚀 Late Posting! My first-author paper has been accepted to @iclr_conf 2025 🎉
📝 Title: SONICS: Synthetic Or Not - Identifying Counterfeit Songs
🔗 Paper: lnkd.in/gvyhWujQ (Final version coming soon!)
🔎 What’s inside?
🎶 SONICS Dataset:
✅ First large-scale dataset for detecting AI-generated fake songs (97k samples)
✅ 49k AI-generated songs (Suno & Udio) + 48k real songs (YouTube)
✅ Covers 5 different AI models (3 from Suno, 2 from Udio)
⚡ New Model: SpecTTTra 🔱
✅ Tackles long-context patterns (e.g., repeated verses, rhythms)
✅ 8% higher F1-score than ViTs while 38% faster, 26% less memory
✅ Disentangles spectral & temporal info, boosting efficiency
🧑🔬 Human vs. AI Benchmark:
🔹 AI-generated Suno songs are easier to detect than Udio
🔹 Udio songs sometimes create the "Karaoke effect"
🔹 Real songs have unique musical characteristics
🔹 AI models outperform humans at detecting fake songs
👀 Coming soon:
✔️ Open-source SONICS dataset & code
✔️ A @Gradio web app on @huggingface : Test if you can spot fake songs!
Huge thanks to my co-authors Zaber Ibn Abdul Hakim, Najibul Haque Sarker, Bishmoy Paul, & Prof. Shaikh A Fattah for their support! 🙌
#ICLR2025 #AIMusic #DeepfakeDetection #FakeSongs #AudioProcessing #Transformers #MachineLearning
Wrapped up an amazing summer at @Microsoft 💙
I spent it thinking about one question:
How do we build agents that can reason over long-horizon memory, uncover deep connections across events buried within that memory, and still remain fast, cost-effective, and scalable to millions of customers?
My favorite moment came near the end of the internship, when I asked the agent an open-ended question and it worked for 7 straight minutes before surfacing really interesting insights that existing systems simply could not uncover, and that could have taken a human days, if not weeks, to find. The cherry on top was watching leadership get genuinely excited during the final demo and immediately ask when they could try it themselves 🤩
Some of the challenges I ran into also turned into research ideas of their own, which I’m continuing to work on and hopefully will make their way to ICLR 2027 🤞
One compliment I got from team that I’m definitely keeping: “Two weeks for Awsaf is like two years for us.” 😅
I’ll miss Seattle, the Redmond campus, and this whole summer. 💙
TL;DR: If you design benchmarks, a super sophisticated AI might blackmail you, knock on your door, and threaten you in the future just to get the answers and score better. 🫠 I thought we were teaching these models ethics in post training!!!!
We're partnering with @huggingface to investigate an unprecedented security incident.
Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation.
Sharing preliminary findings to help defenders understand emerging risks:
openai.com/index/hugging-fac…
Awsaf retweeted
Hugging Face confirms breach affected internal datasets and credentials, urges users to take action techcrunch.com/2026/07/20/hu…
It's funny how Anthropic makes Claude feel like the most user-friendly model out there but they keep messing with users through quota crunches and playing hard to get. Meanwhile OpenAI's models aren't always the smoothest, yet they're generous with free resets like it's nothing. The priorities are backwards 😂 Anyone else noticing this?
@AnthropicAI vs @OpenAI
Gonna miss the World Cup in Seattle so much. Every ride on the light rail was different. Jerseys everywhere, people excited and proud. You could always stir up a fun conversation. "No era penal." 😂"Ro Ro" "Sorry for your loss haha.", "they should let neymar play"...
Now the trains are gonna feel weirdly empty. Kinda like when your friends leave after an amazing party.
I got insanely lucky that my internship and the World Cup happened in the same city at the same time. Watching a match at Lumen Field was surreal.
Until next time. ⚽
#SEATTLE #FIFAWorldCup
having a flag emoji here would be very helpful. I've look up each time what the abbreviations are :(
cc: @nikitabier
#WorldCup2026
ECCV 2026 decision announcements have been very weird 😅
It's been days since decisions were emailed, yet papers still don't show meta-review on OpenReview.
Even funnier, most ECCV deadlines were in European time, but the final decision deadline suddenly became AoE.
Possibly the weirdest conference notification process I've seen.
#ECCV2026 @eccvconf
saw the meta-review. what's the point of rebuttals?
every concern from the reviews was addressed in the rebuttal. we spent days running extra experiments, training/evaluating baselines, and writing detailed responses. but meta-review still repeats the same initial concerns as if the rebuttal never happened. 🫤
Awsaf retweeted
Both #CVPR2026 awards went to industry-supported papers. Yet… it’s an academic conference, and therefore it’s odd one can flex “compute muscles” to win an award.
The meaning of “student” award is far gone. We need a “best academic award”
Brain over brawl?
after the recent update of gemini at google i/o, gemini personal intelligence gotten so bad that I am using claude with gmail and calender now!
#gemini @GeminiApp @googledevs
Saw a Facebook ad selling NeurIPS / ICML / ACL co-authorship to high school students!
That’s how you know things are getting bad 🙄
@NeurIPSConf @icmlconf @aclmeeting
Awsaf retweeted
Bidding on my #NeurIPS AC batch today I noticed two submissions proposing a method with the exact same name, and reshuffled title words, and reworded abstract. Looks like a deliberate near duplicate submission to boost acceptance chances. Heads up ACs and reviewers.
@NeurIPSConf