Your curated collection of saved posts and media
AI robots doing real household tasks ๐คฏ not just demos anymore. #robots #Ai ๐ https://t.co/QgBisvyXbS
Want to talk to the past? Here is an LLM "trained entirely from scratch on a corpus of over 28,000 Victorian-era British texts published between 1837 and 1899, drawn from a dataset made available by the British Library." Quite different from an LLM roleplaying a Victorian. https://t.co/5jl7SyJjAP
Very cool work, I wonder what other eras have a large enough corpus for training? https://t.co/ZQTVQZjKbA
Full house here at the Multimodal Hackathon sponsored by @GoogleDeepMind ! https://t.co/j7PBAb6ATC
This seems really bad but also like it was exactly the goal of systematically dismantling our federal government's standards & protections. https://t.co/djUpcqAPWA
@BEBischof https://t.co/uzwR8l4lWe
AI robots doing real household tasks ๐คฏ not just demos anymore. #robots #Ai ๐ https://t.co/QgBisvyXbS
Want to build a deep research agent with Gemini in 90 minutes? @ivanleomk and I did this yesterday live to celebrate him joining @GoogleDeepMind! Some of the things we covered ๐ Amazing to see see Ivan hit the ground running with the DevRel team and can't wait to see what he builds with @OfficialLoganK, @DynamicWebPaige, @_philschmid, @osanseviero, and the whole team ๐ค
New paper: "Self-Distillation of Hidden Layers for Self-Supervised Representation Learning" We introduce Bootleg โ a simple twist on I-JEPA/MAE that dramatically improves self-supervised representations. The idea: MAE predicts pixels (stable but low-level). I-JEPA predicts final-layer embeddings (high-level but unstable). Bootleg bridges the two by predicting representations from multiple hidden layers of the teacher network โ early, middle, and late โ simultaneously. Why it works: early layers provide stimulus-driven grounding that prevents collapse; deep layers provide semantic targets; and the information bottleneck of compressing all abstraction levels through masked patches forces the encoder to build richer representations. The method is quite simple on top of I-JEPA: extract targets from evenly-spaced blocks, z-score and concatenate, widen the predictor's final layer. That's it. Frozen probe results (no fine-tuning): ImageNet-1K: 76.7% with ViT-B (+10pp over both I-JEPA and MAE) iNaturalist-21: 58.3% with ViT-B (+17pp over I-JEPA, +15pp over MAE) ADE20K segmentation: 30.9% mIoU with ViT-B (+11pp over I-JEPA, +6pp over MAE) Cityscapes segmentation: 35.9% mIoU with ViT-B (+11pp over I-JEPA, +5pp over MAE) Gains hold across ViT-S, ViT-B, and ViT-L. Single-view, batch-size independent โ no augmentation stack, no multi-crop, no contrastive loss, no large compute requirements. Our study is just on images, but this change can be readily deployed to MAE and JEPA models across all domains. https://t.co/PXJlRV4I6w
Lawyers for Trump's ICEย provided false information to justify arresting and detaining thousands of people who had attended immigration courts. https://t.co/jgTto0GZBn
My ideal timeline: Growing up in the 90s, discovering neural nets, scaling laws, and building an artificial consciousness. https://t.co/FQMunONGm3
Segawaใใ@segawachobbies ใ้็บไธญใฎใใชใผใใณใฝใผในverใญใใฆใกใใใใในใฟใผใจใใฆไธ่ถณๅ ใซไฝใฃใฆใฟใพใใ๏ผ 4ใคใฎใตใผใใ ใใง่กจๆ ่ฑใใซๅใใ้ฆๅใใฎๅใใๅฎ็พใใใญใใใจใณใใ้ข็ฝใใงใใ โปใใผใฟใฏๅ ฌ้ๅใงใใไปๅพใฎๅๅใฏSegawaใใใฎๆ็จฟใ่ฆใใงใใฏ๐ https://t.co/DFUG0VLIKk
The happiest person I knew killed themselves four years ago. The more I tell people that Iโm anxious or neurotic the more I realize deep down Iโm doing pretty good for myself Thereโs so much friction in trying to mask everything https://t.co/PQTT6CLquZ
Day 2 went amazing My first $1k MRR from a single app I have no words You made my small dreams come true I will start working on all the small bugs you guys found in the app And probably gonna release a lifetime version because a lot of you guys requested that Letโs keep building This was Day 207 โ X: 7950 ๐ฅณ
๐ https://t.co/TbuVOvL3QO
๐ https://t.co/TbuVOvL3QO
ใใใซใกใฏใๆฅๆฌใฎ็ใใใ OpenAIใงๅใใฆใใๅๅๆฌฃ๏ผใใ ใ ใใใ๏ผใงใใใไผใใงใใฆใจใฆใๅฌใใใงใใ AIใจใผใธใงใณใใฎๆง็ฏใCodexใซใคใใฆ่ณชๅใใใใฐใใใคใงใใชใณใฉใคใณใงๆฐ่ปฝใซๅฃฐใใใใฆใใ ใใใ ่ฟใใใกใซๆฅๆฌใ่จชใใฆใ็ใใใฎใ่ฉฑใใใฃใจ่ใใใฎใๆฅฝใใฟใซใใฆใใพใใ https://t.co/lkow1wxfcX
The perfect camping mug? https://t.co/PQmz69VVyr https://t.co/XfMShMvHD5
NEW research from NVIDIA. Post-training agents with RL is powerful but expensive. Every parameter update needs full multi-turn rollouts with environment interactions, making end-to-end RL prohibitively costly for long-horizon agentic tasks. This research offers a practical middle ground. The work introduces PivotRL, a framework that operates on existing SFT trajectories to combine the computational efficiency of SFT with the out-of-domain retention of end-to-end RL. Instead of exhaustive full-trajectory rollouts, PivotRL identifies pivots, informative intermediate turns where sampled actions show mixed outcomes, and trains only on those high-signal moments. Standard SFT degrades OOD performance by -9.83 points on average. PivotRL stays near zero (+0.21) while achieving +14.11 average in-domain gains over the base model versus +9.94 for SFT. On SWE-Bench, PivotRL reaches competitive accuracy with E2E RL using 4x fewer rollout turns and 5.5x less wall-clock time. The method is already deployed in production as the workhorse for NVIDIA's Nemotron-3-Super-120B agentic post-training. Paper: https://t.co/sIBLUpyfMD Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX
Finally got a chance to play around with @karpathy's LLM Council. I built it as a plugin inside of Claude Code. Hooked it up with OpenRouter for models. The AskUserQuestion tool came in handy to select the council and chairman. This is my first test, but I agree with Karpathy that the concept of LLM ensembles can be used beyond models that offer perspectives on interesting questions. I feel like this could have really cool applications in agentic coding. More on that soon. I built this as a plugin, so next I will be exploring other user cases around agentic coding, like evaluation, tool building, designing, and research. If there is enough interest, I will clean it up and push it out as an open plugin.
@chamath @karpathy It exists in our plugin: https://t.co/5hNgew17xQ Highly flexible and super simple to use. You can use an LLM council in Claude Code to select different models that provide different perspectives. https://t.co/JJq8YI1uV5
Finally got a chance to play around with @karpathy's LLM Council. I built it as a plugin inside of Claude Code. Hooked it up with OpenRouter for models. The AskUserQuestion tool came in handy to select the council and chairman. This is my first test, but I agree with Karpa
Augmental sent me a MouthPad^ A bluetooth mouse for your tongue?! ๐คฏ Make it a controller? ๐ ๐ฎ https://t.co/XY5TNuvwZ2
Augmental sent me a MouthPad^ A bluetooth mouse for your tongue?! ๐คฏ Make it a controller? ๐ ๐ฎ https://t.co/XY5TNuvwZ2
he's on his way to a full recovery! https://t.co/9N1WjZITUD
i often think about this.. https://t.co/YMhKgLRD1G
i often think about this.. https://t.co/YMhKgLRD1G
Students undermine their own learning by getting answers from AI. But the authors of the original paper have a new RCT that shows well-prompted AI tutors do, actually, boost learning: https://t.co/0HtjGC8eU0 *Not quote tweeting because don't want to boost slop science accounts https://t.co/ogSs1Qpr7M
The Hermes Deep Dive. (The new hot AI agent harness). https://t.co/Ldmfiy9a6C Hey @Teknium got anything to add?
@mcpark Iโve even been you https://t.co/g28hLBaCzG
@mcpark Hair ispoweful https://t.co/PGRnly60OS
My besto friendo chappie chan https://t.co/yNNLCxUz0x
@mcpark I was actually Korean in 2021 https://t.co/NKqWAqQufc
