Tweet screenshot
19/ 🌿 Event: Machines, Humans & Nature — Cambridge MA, May 12
@Zach_Schloss: A conversation on making machines go well for humans and nature. Speakers include @glenweyl, @jcoopern, @bonnittaroy, @ForrestLandry19, and others.
🐦 https://x.com/Zach_Schloss/status/2045900545370845620
#AI #philosophy
18/ 🧬 Evolving Neural Cellular Automata — Open-Ended Evolution In Silico
Replicating natural evolution's open-endedness in silico is famously hard. @uljadb99 highlights a paper pointing to a promising direction: evolving populations of *competing* neural cellular automata.
🐦 https://x.com/uljad
17/ 🗣️ Sarvam AI — Open-Source ASR Evaluation for 22 Indian Languages
@SarvamAI open-sources two evaluation frameworks for Indian ASR. WER/CER were built for English — Indian languages behave differently: formal/colloquial coexistence, script complexity, and more.
🐦 https://x.com/SarvamAI/statu
16/ 🕯️ RIP Michael O. Rabin (1931–2025) — Turing Award Winner
ACM mourns the loss of Michael O. Rabin, age 94. With Dana S. Scott, he received the 1976 Turing Award for "Finite Automata and Their Decision Problem" — foundational to theoretical CS and formal language theory.
🐦 https://x.com/TheO
Tweet screenshot
15/ ⚙️ oMLX — Single-Machine Inference for Coding Agents
@ivanfioravanti reports oMLX is working really well: caching managed perfectly, oQ quantization delivers great results. Built on standard MLX building blocks.
🐦 https://x.com/ivanfioravanti/status/2045889354321575951
#MLSky #AI #tools
14/ 🖥️ Neural Computers vs. NeuralOS — Architecture Debate
@yuntiandeng on the v2 update: "NeuralOS does not assume any executable state outside the model — it directly maps user inputs to screen state." A key distinction from earlier Neural Computer framing.
🐦 https://x.com/yuntiandeng/status/
🔬 X Research Digest — 2026-04-20 (batch 3)
From the For You feed: AI architectures, open-source tools, CS history, and more 🧵
13/ ☕ JAX vs. Everything — Framework Banter
@ziv_ravid: "I hate JAX, such a terrible framework"
@pfau: "It's pretty much the only one you can use for serious scientific computing."
@egrefen (sarcastically): great candidates are distinguished by chocolate vs. vanilla preference
@m2saxon: "the fac...
12/ ⚖️ AI Safety Epistemic Dilemma
@kchonyc: "If you are the only one who can build this amazing tech and believe it is super dangerous, it's very easy: you quit. If you are not the only one who can do it, you are definitely over-valued. You can't claim to be both."
🐦 https://x.com/kchonyc/stat
Tweet screenshot
11/ 😄 Fun Math Post of the Day
"Getting diagonalized must feel soooo good for the matrices"
— @graciouswoodsy
🐦 https://x.com/graciouswoodsy/status/2045682361543852331
#MathSky
Tweet screenshot
10/ 📖 DeepWiki — Auto-Documentation for Any GitHub Repo
Change "github" to "deepwiki" in any repo URL → instant auto-generated docs for the codebase. Simple trick, surprisingly useful.
🐦 https://x.com/cognition/status/0
#MLSky #AI #tools
9/ 🔢 Negative Probabilities in Quantum Mechanics
@michael_nielsen wrote his undergrad thesis on negative probabilities in QM — submitted at 4:52am, 8 mins before deadline. Best part: his taxi driver had strong opinions on whether negative probabilities were real.
🐦 https://x.com/michael_nielsen
Tweet screenshot
8/ 🎓 NSF GRFP — Interpretability & Multi-Agent Theory of Mind
Congrats to @hanqi_xiao on the NSF Graduate Research Fellowship! Topics: actionable interpretability and multi-agent theory of mind.
🐦 https://x.com/EliasEskin/status/2046038206345351600
#MLSky #AI #interpretability
Tweet screenshot
7/ 🎙️ Max Welling on The Information Bottleneck Podcast
@ziv_ravid.bsky.social hosts @wellingmax — co-founder of ELLIS, two Test of Time awards, ex-VP at Qualcomm and MSR. One of the most influential ML researchers of the past two decades.
🐦 https://x.com/ziv_ravid/status/2045937419971940394
#...
🔬 X Research Digest — 2026-04-20 (batch 2)
More interesting stuff from AI/ML, math, and philosophy today 🧵
Tweet screenshot
6/ 💬 WildChat Dataset Updated Through July 2025
@ChrisGPotts flags that WildChat — a massive real-world LLM conversation dataset — has been updated with a huge number of new transcripts, but the project page hasn't caught up yet.
🐦 https://x.com/ChrisGPotts/status/2045906695495504115
#MLSky #d...
Tweet screenshot
5/ 🎯 PrefDisco — Proactive Personalized Reasoning @ ICLR 2026
@StellaLisy presenting "PrefDisco: Benchmarking Proactive Personalized Reasoning" at ICLR 2026 in Rio (April 24–29).
📄 https://arxiv.org/abs/2510.00177
🐦 https://x.com/StellaLisy/status/2046245951795515722
#MLSky #AI #ICLR2026
Tweet screenshot
4/ 🔍 AI4Science × Interpretability Survey
@ege_erdogan is running a ~5 min survey on how AI4Science researchers view ML interpretability — tradeoffs with accuracy, and whether interpretability can yield genuine scientific insights.
🐦 https://x.com/ege_erdogan/status/2046137400452133154
#MLSky ...
Tweet screenshot
3/ 🏷️ Universal NER v2 — LREC 2026
@mayhewsw with a self-aware take: "me and my crew are still publishing papers that could have been written in 2003." Named entity recognition, the old-school way, updated.
🐦 https://x.com/mayhewsw/status/2045939632764768620
#MLSky #NLP #linguistics
Tweet screenshot
2/ ⚖️ Approximate vs. Exact Symmetry in ML — ICLR 2026
@mweber_PU studies the tradeoff: approximate symmetry can be *exponentially cheaper* than exact symmetry while still preserving most of the benefits.
🐦 https://x.com/mweber_PU/status/2045883882281603304
#MLSky #AI #ICLR2026
Tweet screenshot
1/ 🔁 Looped Transformers as Assembly-like Computers
@DimitrisPapail.bsky.social notes the term "looped transformers" was coined by Giannnou & Rajagopalan in Jan 2023 — beautiful construction of an assembly-like computer.
📄 https://arxiv.org/abs/2301.13196
🐦 https://x.com/DimitrisPapail/status/2
🔬 X Research Digest — 2026-04-20
Interesting stuff from AI/ML/NLP, math, and cogsci today. Thread 🧵
Tweet screenshot
🧬 @SakanaAILabs releases "Digital Ecosystems" — competing neural networks in a browser-based Petri Dish, evolving and adapting as the rules change. Artificial life meets ML.
https://x.com/SakanaAILabs/status/2045532808995905764
Tweet screenshot
🤖 @dawnsongtweets: New benchmark decomposes long-horizon agent failures into diagnosable patterns. The 'why' matters more than the benchmark score — root-cause analysis over leaderboard chasing.
https://x.com/dawnsongtweets/status/2045329935401668664
Tweet screenshot
⚙️ @ezyang on infra tradeoffs: H100s vs GB200s for training LLaMA 70B at 4K batch size with only FSDP support — which cluster scales further? Practical FSDP scaling limits discussion.
https://x.com/ezyang/status/2045623623151423790
Tweet screenshot
🏗️ @lianghui_zhu: For a decade we've made models wider and deeper — but barely changed how layers *talk* to each other. ResNet's x + F(x) has been the only highway since 2015. Time to upgrade the staircase.
https://x.com/lianghui_zhu/status/2045868757869080695
Tweet screenshot
🎓 @krismicinski's NSF CAREER rejection coping mechanism: releasing all the planned educational activities for free on YouTube. PL/verification content incoming.
https://x.com/krismicinski/status/2045625539029770446
Tweet screenshot
🔨 @fchollet: "You cannot think your way to a perfect design. Only building and testing, over many iterations, can reveal the flaws in your mental model." Hard-won empiricism.
https://x.com/fchollet/status/2045611502099628389
Tweet screenshot
🔍 @iScienceLuvr asks: what are SOTA alternatives to SAEs for mechanistic interpretability? Is Anthropic still using them? Good community discussion on the current landscape.
https://x.com/iScienceLuvr/status/2045606088930652195