Who's Who of AI

Lucas Beyer

630 trust @giffmana · 146,896 followers
Why they matter

Tracked through public AI activity and peer connections inside the directory.

AI signals
6
past 30d
Sources
3
distinct domains
Discussions
0
past 30d
Latest signal
3h ago
View every signal from Lucas Beyer →
What they're sharing

Articles & links

RT @AIatMeta: 🔗 Download Muse Glimmer on @huggingface: https://t.co/s7Lzb8MqCG 🔗 Read the technical blog: https://t.co/X6htFnhRbc 🔗 Find…

meta-models (Meta Inc.) huggingface.co
AI Weekly's analysis →
  • Meta released Muse Glimmer on August 10, 2026, a 30B multimodal model under Apache 2.0 aimed at local, on-device agent workloads.
  • Q4_K_M GGUF quantization and a DFlash speculative decoding drafter fit the model into a 24 GB or 32 GB consumer GPU envelope.
  • Meta reports category-best scores on MCP Atlas (75.5), SWE-Bench Pro (51.2), AIME 2026 (94.7), and Charxiv Reasoning (78.8).
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 4 from the directory shared this · 13d ago

https://t.co/nr447ojdFs

Revisiting Self-Supervised Visual Representation Learning arxiv.org
AI Weekly's analysis →
  • Kolesnikov, Zhai and Beyer revisit self-supervised visual representation learning and argue CNN architecture choice deserves attention equal to the pretext task.
  • The authors report standard CNN design recipes from supervised learning do not always translate to self-supervised representation learning.
  • The team claims their study outperforms previously published state-of-the-art self-supervised results 'by a large margin,' though the abstract lists no per-benchmark numbers.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 3h ago

@gabriberton @Nik__V__ The dream? My friends at Google derisked this exactly a decade ago: https://t.co/iReVZecPwX And pretty much around the time you describe in the thread when we were "looking for scalable supervision" we also discussed this exact thing (we had streetview a…

PlaNet - Photo Geolocation with Convolutional Neural Networks arxiv.org
AI Weekly's analysis →
  • PlaNet reframes photo geolocation as classification over thousands of multi-scale geographic cells rather than image retrieval.
  • The convolutional network is trained on millions of geotagged images and integrates cues like weather, vegetation and architecture.
  • Combining the model with an LSTM over full photo albums yields a 50% performance improvement over single-image inference.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 2d ago

Not always! There are situations where flops and wall time genuinely are not interchangeable, and it's not just a "missing cracked cuda"; not all flops are equal. Here i had one irl, from FlexiViT paper: https://t.co/AEsF2rWWVn Context: the efficiency misnomer https://t.co/wnZ…

arxiv.org
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 7d ago

@BlackHC If you change architecture, don't put params on X though. flops or walltime. https://t.co/Sl5dOJn3m5

arxiv.org
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 8d ago

It turns out there is a form where you can register your interest in a linux version of the codex app, also select the distro. I assume besides the twitter noise, they may also use this to gauge interest. If you want this, consider registering interest: https://t.co/UnqGGceCIZ…

openai.com
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 62d ago