🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

How comes people are not hysterical about this? arxiv.org/abs/2410.01131

🎙️✨ Join me and the researchers behind NatureLM-audio (Marius Miron and David Robinson) as we present a live technical walkthrough on Nov 21 at 5pm GMT (12pm US Eastern / 9am US Pacific)! Interested? Head to our Discord community for more information: discord.com/invite/H2Y532a…

16th highest scored paper at ICLR 2025 with 3(!), 8, 10, 10, 10 tldr: they scale sparse autoencoders to GPT4 and show that interpretability techniques used on toy models can work on larger models too (hmm i wonder who these people who have access to GPT4 activations are!)

In arxiv.org/abs/2303.00848, Durk Kingma and Ruiqi Gao had suggested that noise augmentation could be used to make other likelihood-based models optimise perceptually weighted losses, like diffusion models do. So cool to see this working well in practice!

✨ Train language models directly on continuous data - without tokenization ✨ We propose an easy way to train GPT-style autoregressive models on continuous data, without error accumulation. We test it on audio 🔊, but this method can easily work with other modalities 🎆 👇🧵

🚨 PhD Alert! 🚨 We’re seeking 2 motivated PhD students to explore the evolutionary origins of language through cutting-edge field research with wild orangutans. 🌍🐒 Please RT 📍 Based @WarwickPsych, w/ fieldwork in Sumatra (Indonesia) 📅 Apply now! 👉findaphd.com/phds/project/w…

this neurips is really going to be remembered as the "end of pretraining" neurips notes from doctor Noam Brown's talk on scaling test time compute today (thank you Hattie Zhou for organizing)

Paper alert––*Awarded best paper* at NeurIPS workshop on Foundation Model Interventions! 🧵👇 We analyze the (in)abilities of SAEs by relating them to the field of disentangled rep. learning, where limitations of AE based interpretability protocols have been well established!🤯

so happy for this!

🔥Excited to introduce RINS - a technique that boosts model performance by recursively applying early layers during inference without increasing model size or training compute flops! Not only does it significantly improve LMs, but also multimodal systems like SigLIP. (1/N)

after 10 years I switched to uv from conda. meanwhile my environment's dependencies have been resolved successfully :D

Large Language Diffusion Models Introduces LLaDA-8B, a large language diffusion model that pretrained on 2.3 trillion tokens using 0.13 million H800 GPU hours, followed by SFT on 4.5 million pairs. LLaDA 8B surpasses Llama-2 7B on nearly all 15 standard zero/few-shot learning

happy to announce that Biodenoising was accepted at ICASSP 2025. this is essentially the equivalent of speech enhancement for non-human vocalizations. it can be easily used in Python with pip install biodenoising.

ESP co-founder, Aza Raskin spoke with Kenneth Cukier on the Babbage podcast by The Economist, where he shared how we're leveraging AI to decode animal communication and working toward a future of interspecies understanding 🌍 economist.com/podcasts/2025/…

I am at ICASSP and will be presenting biodenoising on Friday morning.happy to talk to people interested in bioacoustics,cross-domain representation transfer or simply curious about our work at ESP.we released a new version of biodenoising including self-training on your own data

AVES is now pip-installable 🎉 This self-supervised, transformer-based model is pretrained on large-scale animal vocalization datasets & thanks to our incredible engineering team is now more accessible than ever–ready to run with just a single command. 🔗bit.ly/44uTbhN

I'll be at #ICLR2025 in Singapore this Saturday for poster session 6, co-presenting our NatureLM-audio project! If you're around and interested in AI for bioacoustics, ecology, or related areas, would love to meet up — feel free to reach out!

📢 We've open-sourced NatureLM-audio, the first audio-language foundation model for #bioacoustics. Trained on large-scale animal vocalization, human speech & music datasets, the model enables zero-shot classification, detection & querying across diverse species & environments 👇🏽

daré una charla por videoconferencia mañana a las 2PM hora de Colombia sobre cómo usar inteligencia artificial para decodificar el comportamiento animal Earth Species Project (ESP)

excited to finally share on arxiv what we've known for a while now: All Embedding Models Learn The Same Thing embeddings from different models are SO similar that we can map between them based on structure alone. without *any* paired data feels like magic, but it's real:🧵