Christina Lee

11 followers · 6 following · 396 views

on the atlas — 54

highlights — 39

LogikBench is a large hybrid benchmark suite of human authored and AI generated Verilog RTL circuits. Use cases includes objective evaluation of EDA algorithms/tools, PDKs, architectures, model/LLMs, compute infrastructure, (and more...).
zeroasiccorp/logikbench: Digital logic benchmark suite ·
Co-Design of Control and Hardware: Joint optimization of physical design and control policies for high-performance robotic systems.
Brain and Brawn: 1st Workshop on Robot Hardware-Aware Intelligence
we give robots a /skills library that self-evolves and compounds indefinitely! Introducing ASPIRE: a robot solving its 100th task is no longer as clueless as solving its first. Coding agents observe multimodal sensory traces from simulation and real robots, launch an evolutionary search over control programs, and distill the best know-how into an ever-expanding library.
Jim Fan on X: "Today, we give robots a /skills library that self-evolves and compounds indefinitely! Introducing ASPIRE: a robot solving its 100th task is no longer as clueless as solving its first. Coding agents observe multimodal sensory traces from simulation and real robots, launch an https://t.co/57JC05VN54" / X
The constraints were simply: do not train a neural network, make it locally reproducible, leave records for each round, and keep pushing the score up.
Learning Beyond Gradients
Reconciling Kaplan and Chinchilla
Scaling Laws, Carefully | Lil'Log
Chinchilla Scaling Laws
Scaling Laws, Carefully | Lil'Log
Each algorithm is a named runtime class — the algorithm object is the algorithm. Dispatch is keyed on algo.type — it names the algorithm, and each config class's defaults are its vetted parameterization:
feat: algorithm abstraction — named algorithm classes + inline frozen-model references (grpo, opd, sft_distill, self_distill, echo) by hallerite · Pull Request #2746 · PrimeIntellect-ai/prime-rl
The algorithm classes
feat: algorithm abstraction — named algorithm classes + inline frozen-model references (grpo, opd, sft_distill, self_distill, echo) by hallerite · Pull Request #2746 · PrimeIntellect-ai/prime-rl
on May 14, 2061, what had been theory, became fact.
The Last Question
GLM-5.1 training can be run with a single command on a Slurm cluster
RL at 1T Scale: prime-rl Performance Deep Dive
For decades, Multivac had helped design the ships and plot the trajectories that enabled man to reach the Moon, Mars, and Venus, but past that, Earth's poor resources could not support the ships.
The Last Question
By utilizing a source-side CPU engine replica and P2P RDMA transfers via Mooncake TransferEngine, we speed up weight transfer times for 1T-parameter Kimi-K2 7 times (53 seconds -> 7.2 seconds), at the cost of one additional inference engine replica (32G) per training rank on CPU memory.
Updating 1T parameters in seconds — P2P weight transfer in Large Scale Distributed RL - LMSYS Org
By utilizing a source-side CPU engine replica and P2P RDMA transfers via Mooncake TransferEngine, we speed up weight transfer times for 1T-parameter Kimi-K2 7 times (53 seconds -> 7.2 seconds
Updating 1T parameters in seconds — P2P weight transfer in Large Scale Distributed RL - LMSYS Org
Model strategy for @harvey : We are working on the first model in our legal foundation model series, inspired by @cursor_ai 's Composer. Two goals: 1. Allow us to serve frontier intelligence across our product surface areas at an affordable price and a strong security posture. 2. Create the foundations for law firms to build their own specialized models and own their own intelligence.
Gabe Pereyra on X: "Model strategy for @harvey: We are working on the first model in our legal foundation model series, inspired by @cursor_ai's Composer. Two goals: 1. Allow us to serve frontier intelligence across our product surface areas at an affordable price and a strong security posture." / X
Improved Architecture: We propose IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length. We also improve GLM-5.2’s MTP layer for speculative decoding, increasing the acceptance length by up to 20%
GLM-5.2: Built for Long-Horizon Tasks
Advanced Coding with Flexible Effort: Stronger coding capabilities with multiple thinking effort levels to balance performance and latency
GLM-5.2: Built for Long-Horizon Tasks
This paper ex- amines the extent to which post-training can enhance LLMs’ capacity for causal inference. We introduce CauGym , a comprehensive dataset comprising seven core causal tasks for training and five diverse test sets. Using this dataset, we systematically evaluate five post-training ap- proaches: SFT, DPO, KTO, PPO, and GRPO.
[2602.06337] Can Post-Training Transform LLMs into Causal Reasoners?
We prove the failure is fundamental: supervised fine-tuning, direct preference optimization, and in-context learning all produce predictors that cannot distinguish between causal graphs generating similar observational data, and any attempt to do so requires the model’s internal representations to grow unboundedly, violating the very conditions under which these methods work
[2605.27567] Why LLMs Fail at Causal Discovery and How Interventional Agents Escape
The contrast with DeepSeek-V4 is the interesting part. DeepSeek dropped its RL stage and replaced it with multi-teacher distillation. Nvidia keeps RL and stacks MOPD on top. DeepSeek treats distillation as a replacement for RL, Nvidia treats it as a consolidation layer that RL feeds into.
Nemotron 3 Ultra: what distillation can't fix
why using the old LoRA config not the 3D shape from PEFT 18?
Adina Yakup on X: "Macaron-V1-Preview-749B 👀 a Mixture-of-LoRA personal agent model from MindLab ✨ 744B base + 5 specialist LoRAs ✨ Generative UI as a core skill ✨ Personal agent focused ✨ 202K context ✨ MIT license https://t.co/OkjxKEThzZ" / X
make visual training data controllable
Yucheng Shi on X: "Static datasets give fixed samples. Code-generated environments give scalable worlds. For LLM reasoning, executable environments can generate fresh problems and verifiable rewards. With TRON, we bring this idea to visual reasoning: each code-generated environment samples a https://t.co/H8C6eD3nR4" / X
With TRON, we bring this idea to visual reasoning: each code-generated environment samples a latent visual state, renders an image, asks a question, and verifies the answer from the underlying state.
Yucheng Shi on X: "Static datasets give fixed samples. Code-generated environments give scalable worlds. For LLM reasoning, executable environments can generate fresh problems and verifiable rewards. With TRON, we bring this idea to visual reasoning: each code-generated environment samples a https://t.co/H8C6eD3nR4" / X
large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code.
Sonya Huang 🐥 on X: "Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring @ellev3n11 (Composer Lead at @cursor_ai) and @dzhulgakov (Co-Founder at @FireworksAI_HQ). The Cursor team trained Composer 2 on Fireworks by starting with https://t.co/6LLlJlyl8Q" / X
environment design ("the most powerful RL environment is the product itself"),
Sonya Huang 🐥 on X: "Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring @ellev3n11 (Composer Lead at @cursor_ai) and @dzhulgakov (Co-Founder at @FireworksAI_HQ). The Cursor team trained Composer 2 on Fireworks by starting with https://t.co/6LLlJlyl8Q" / X
Practical consequences for this project: Decode is markedly faster than an all-full-attention 2B baseline: each GDN layer is O(n) instead of O(n²), which compounds during GRPO rollouts. KV-cache footprint is meaningfully smaller per sequence, which lets us batch more rollout sequences in parallel. Token counts are heavy. Verified from our actual training set (226K records across 5 source files): median ~2,100 tokens per training example; ActivityNet- and COIN-style longer clips sit at ~5,000 tokens typical and ~7,800 at p90; the cap before we drop a clip is around ~8,500 tokens. Attention savi…
Part 1: The Map Was Wrong — Nemostation
Sequence-level importance sampling is the estimator that scales with compute, while token-level estimators become structurally inconsistent at high policy lag.
Is Frontier Asynchronous RL Solved? — Luke J. Huang
We find that the teacher need not be strong: with proper mixing of the language modeling and knowledge distillation losses, even small and undertrained teachers improve larger students.
[2605.23857] Strong Teacher Not Needed? On Distillation in LLM Pretraining
Reflexion agents verbally reflect on task feedback signals, then maintain their own reflective text in an episodic memory buffer to induce better decision-making in subsequent trials
1b44b878bb782e6954cd888628510e90-Paper-Conference.pdf
ur pre-training dataset includes data sourced across many different domains, including web documents and code, and incorporates image, audio, and video content. For the instruction- tuning phase we finetuned Gemini 1.5 models on a collection of multimodal data (containing paired instructions and appropriate responses), with further tuning based on human preference data.
gemini_v1_5_report.pdf
LLM training is built on fast MatMuls. But many surrounding ops still run as memory-bound kernels.
Han Guo on X: "LLM training is built on fast MatMuls. But many surrounding ops still run as memory-bound kernels. CODA reparameterizes them to hide in the matmul’s shadow, fused into its epilogue before results leave the chip. Bonus: LLMs can write fast CODA kernels too (approaching SoLs). https://t.co/cOTeMUr4py" / X
The regions responsible for memory, sensory integration, and the encoding of new information were all firing together in a coordinated pattern that spread across the entire cortex. The whole network was awake and connected.
Ihtesham Ali on X: "A Norwegian neuroscientist spent 20 years proving that the act of writing by hand changes the human brain in ways typing physically cannot, and almost nobody outside her field has read the paper. Her name is Audrey van der Meer. She runs a brain research lab in Trondheim, and https://t.co/9FDGAKPA6I" / X
221 Cool, Hidden, and Unusual Things to Do in San Francisco, California
221 Cool and Unusual Things to Do in San Francisco - Atlas Obscura
torch._dynamo.config.activation_memory_budget = 0.5 out = torch.compile(fn)(inp)
Current and New Activation Checkpointing Techniques in PyTorch – PyTorch
(compile-only) Memory Budget API [NEW!]
Current and New Activation Checkpointing Techniques in PyTorch – PyTorch
Many heuristics were not useless; they were simply too expensive to maintain. Coding agents change that maintenance curve. Rules that used to be one-off patches may start to become code worth owning for the long term.
Learning Beyond Gradients
new two-part methodological framework combining 1) lightweight, distributed feedback loops on individual data sources, with 2) centralized integration tests to assess candidate mixes on base model quality and post-trainability
[2512.13961] Olmo 3
Tiling window manager for macOS along the lines of xmonad.
Amethyst | ianyh
Resolves dependencies between columns (DAG-based execution)
Architecture & Performance - NeMo Data Designer
Async RL. Early in RL, a meaningful fraction of trajectories enter loops that are guaranteed to score zero. The model clicks the same dead UI element repeatedly, or opens a fresh tab and loses the canvas. Our setup is not fully async, so the slowest trajectory in a batch gates the whole step. The next thing to try would be async RL a la Pipeline RL. Denser reward signals. Per-step or per-tool-call rewards would credit intermediate progress more densely than the single end-of-trajectory rubric we use today. This should help most on long-horizon composition. Feedback descent.Our reward function …
Training LLMs to use Photoshop | Mickey Labs