@_cobalamin
Hamburg, Germany
Joined May 2012
๐Ÿ’ซ Large Language Continuous Diffusion Models Presenting Sigma, the first large-scale (3B/8B) continuous diffusion LM (dLM) trained with likelihood. ๐Ÿ“ˆ Comparable with discrete dLMs and AR on math and coding ๐ŸŽ›๏ธ Enables meaningful study of inference-time interventions Jointly led with Wei Guo (@WeiGuo01), Jean-Marie Lemercier (@jm_lemercier), Simon Welker (@_cobalamin) Team: @YongganFu @mmkamani7 Sajad Norouzi @julberner @tomasgeffner @karsten_kreis @YongxinChen1 @MoleiTaoMath @jwthickstun @PavloMolchanov Ante Jukic @ArashVahdat @MardaniMorteza
16
112
21
676
73,237
Can the Bluetooth commitee get their heads out of their asses and invent a revolutionary(!) protocol that lets us do this and also not downgrade to callcenter quality audio when using the microphone? It's 2025 and Bluetooth still sucks ass.
Ultra-high transmission rate of 4.6Mbps ๐Ÿ’€ master-level lossless audio stability
77
320
16
8,871
420,016
Simon Welker, Tal Peer, Timo Gerkmann, "Real-Time Streaming Mel Vocoding with Generative Flow Matching," arxiv.org/abs/2509.15085
1
3
595
Imagine if, in the 1890s, Eastman Kodak had created an entire research team devoted to figuring out if film cameras actually did steal your soul.
New column: Anthropic is studying "model welfare" to determine if Claude or other AI systems are (or will soon be) conscious and deserve moral status. I talked to Kyle Fish, who leads the research, and thinks there's a ~15% chance that Claude or another AI is conscious today.
20
100
6
929
59,516
Most modernization attempts accompany a sense of delusional disrespect for existing systems and profound arrogance that the people who came before them were incompetent/ignorant. This lack of humility combined with a cacophonous fetish for making things *appear* modern/sexy, it's a recipe for civilizational regression. False confidence to undertake modernization comes from technological progressโ€”i.e., "We have faster computers and bigger screens now, and they didn't. We can design things better!"โ€”which becomes apparent as the new systems they design are bloated, slow and utterly dysfunctionalโ€”indicting their thesis for change.
10
29
2
252
12,227
Christian Zylindner ist nicht real, er kann dir nichts tun. Christian Zylindner:
44
618
20
15,743
316,666
I really respect when a scifi or horror movie invents a new sound you've never heard before
168
2,305
227
44,365
1,394,073
๐Ÿ” Join Our Team! We have exciting open positions: ๐ŸŽ“ Postdoc: Data Science & AI for Speech & Audio ๐ŸŽ“ PhD/Postdoc: AI-based Speech & Audio ๐ŸŽ“ PhD: Diffusion Models for Speech Processing ๐ŸŽ“ 3 PhD positions for Imaging & Speech ๐Ÿ”— inf.uni-hamburg.de/en/inst/aโ€ฆ
5
6
824
Best Student Paper Award at IWAENC 2024 with @moliner_eloi for our joint work on Unsupervised Speech Dereverberation using Diffusion Models!! Thanks to @_cobalamin @sp_uhh @AaltoAkuLab for the great work! More publications on the topic: inf.uni-hamburg.de/en/inst/aโ€ฆ
2
9
254
Simon retweeted
Smashing my PC when it looked like the AoE II medium AI was beating me was not an act of frustration. It was actually an example of Extraplanar Warfare, the approach to military theory I've been developing. You attack the enemy in metaphysical modalities to which he has no access
13
613
18
4,688
109,848
\usepackage{zalgo}
This is what I call an interesting LaTeX error.
1
3
264
BUDDy: Single-Channel Blind Unsupervised Dereverberation with Diffusion Models We present a fully unsupervised method for blind speech dereverberation using a diffusion prior and a parametric subband filter Paper ๐Ÿ“œ arxiv.org/abs/2405.04272 Audio/Code ๐Ÿ”Š uhh.de/sp-inf-buddy
1
4
12
392
well yeah, I guess
2
146
๐—ฃ๐—ต๐—— ๐—ฝ๐—ผ๐˜€๐—ถ๐˜๐—ถ๐—ผ๐—ป โ€ž๐——๐—ถ๐—ณ๐—ณ๐˜‚๐˜€๐—ถ๐—ผ๐—ป-๐—•๐—ฎ๐˜€๐—ฒ๐—ฑ ๐——๐—ฒ๐—ฒ๐—ฝ ๐—š๐—ฒ๐—ป๐—ฒ๐—ฟ๐—ฎ๐˜๐—ถ๐˜ƒ๐—ฒ ๐— ๐—ผ๐—ฑ๐—ฒ๐—น๐˜€ ๐—ณ๐—ผ๐—ฟ ๐—ฆ๐—ฝ๐—ฒ๐—ฒ๐—ฐ๐—ต ๐—ฆ๐—ถ๐—ด๐—ป๐—ฎ๐—น ๐—ฃ๐—ฟ๐—ผ๐—ฐ๐—ฒ๐˜€๐˜€๐—ถ๐—ป๐—ดโ€œ. Join us in pushing the state-of-the-art in this exciting field. Details and on how to apply: inf.uni-hamburg.de/en/inst/aโ€ฆ
5
1
6
744
he mewed too hard
1
90
i'll never get over how the cochlea is an analog fourier transform organ
129
2,177
336
23,660
1,614,545
We just opened a PhD or Postdoc position on AI-based speech and audio signal processing. Please see our webpage for details and on how to apply inf.uni-hamburg.de/en/inst/aโ€ฆ
5
3
1,005
We offer an exciting #PhDposition on deep multi-microphone speech enhancement. Please find details on our webpages inf.uni-hamburg.de/en/inst/aโ€ฆ
4
4
574