Vinay Mathew

🎵 But I will keep counting my blessings, knowing I can't count the high. 🎵

[Vinay Mathew @vinay_matt]

Thank you! Even faster & more accurate diffusion coming soon to localfold :)

[Sergey Ovchinnikov @sokrypton]

https://t.co/7OlBz2wKP1 updated to include:
- bioemu2 @FrankNoeBerlin
- mol1 @MolenceBio
- fastplm (confidence modules for esmfold fast models) @Logan_Hallee
- af3x crosslinks @jankosinski
- faster diffusion (exponential update and geometric noise schedule) @vinay_matt

[Vinay Mathew @vinay_matt]

@woongworks @davidstutz92 It was pretty smooth, I've shared a graph in this repo but not abrupt as you would think it would be. I was also very conservative with the learning rate so that might have contributed to the structure. https://t.co/QNOOiuArp3

[Vinay Mathew @vinay_matt]

(1/2) Thank you to the Google DeepMind team for sharing the SynthID-Bio Struct model weights and for all the work that went into them. However, @davidstutz92 and team, I noticed something I’d appreciate clarification on. The downloaded weights appear to contain 11 tensors associated with the watermark detector, despite the paper stating that the detector is excluded from the release! (The exact Haiku/JAX namespace prefix I see in the model weights is diffuser/~/watermark_detector/point_net/.)

I used those weights and the information in the supplementary material to reconstruct the scorer. When I ran it on predictions from both the ordinary and watermarked AlphaFold3 models, the mean logit scores were −4.52 and +4.48, respectively, on a set of approximately 2,300 PDB structures from 2024 (the test set used by Boltz-2). With a simple logit threshold of > 0 (likely different from your setup), I get a 0% detection rate for unmarked AlphaFold3 and a 100% detection rate for the watermarked model.

These results led me to wonder whether I could remove the watermark from the released SynthID-Bio AlphaFold3 model.

[Rohit Singh @rohitsingh8080]

Intriguing thread about stress-testing the structure watermarking scheme of SynthID-Bio. Complements our sequence-based analysis.

I'm not an expert in the techniques here, and will defer to those who know more.

Also, kudos to SynthIDBio folks for making most code/scripts avlbl

[Vinay Mathew @vinay_matt]

(2/2) I initialized the model with the watermarked weights and trained it using the multi-step consistency training schedule from Heek et al. (https://t.co/CVMQ2hyCfi), with 8 steps. After approximately 4,000 optimizer steps, the logit scores were almost the same as those of the unmarked AlphaFold3 model. The detection rate of my inferred scorer fell to about 1%, compared with nearly 100% at initialization. (This change in logit scores did not occur with consistency distillation i.e., using the teacher model as a guide in Alg 1 of Heek et. al, which is expected.)

I’d be happy to provide our fine-tuned model weights and predicted structures privately to the DeepMind team for independent testing. I’d also welcome confirmation of the scorer and an evaluation using the official detector, so we can check whether our interpretation and findings are accurate. I’ve open-sourced the code for the scorer here https://t.co/N3EV8aI1v0. The same repo also contains all the code used for doing discrete multistep CD on boltz-2.

Thanks to @rohitsingh8080’s post on the washing out of the watermark signal with Raygun for inspiring me to dig into this. And also to my collaborators for getting me all the resources I needed to test this out.

[Vinay Mathew @vinay_matt]

@anodaboss @SwarajPatell @KendrickImler Bro never forget!

[ChatGPT @ChatGPT]

leave me alone man

[SpaceX @SpaceX]

At dawn from the gateway to Mars, the launch of Starship’s second flight test

[Senior PowerPoint Engineer @ryxcommar]

reinforcement learning from human feedback

[James Franklin @coachjfranklin]

#WeAre...Better
#107kStrong
#PSUnrivaled

[Vinay Mathew @vinay_matt]

@rishiandri @TRAI @DoT_India @JioCare @reliancejio I’m looking into it bro

[Vinay Mathew @vinay_matt]

@rishiandri @MKBHD Sm

[Vinay Mathew @vinay_matt]

@ajuanijustus why bruh

[Vinay Mathew @vinay_matt]

@bringbackblf 😂😂😂😂😂

[Arun Maini @Mrwhosetheboss]

OOF

[Shiv Aroor @ShivAroor]

Good thing about Vodafone users is that they will never know if their internet has actually been suspended.

[TEDxCET @TEDxCET]

Have some ideas worth spreading this festive season with @TEDxCET. Wishing you all a colourful Onam.

[Dhanya Rajendran @dhanyarajendran]

And her cadres proved all of us wrong. No violence, no hysteria, no immolations. They gave her a fitting goodbye. #Jayalalithaa

[Old Tamil Poetry @OldTamilPoetry]

Puranaanooru - 226. "Death must have pleaded with the ruler, to take his life away"
Blog: https://t.co/kidz2sVjUW

[Unbox Therapy @UnboxTherapy]

It was a very difficult decision but I’ve decided to endorse Jack for president.