@maxrumpf: turbopuffer x SID An easy way to tell a good from a great AI researcher: how much do they think about infrastructure. I…
Summary
The tweet thread highlights the importance of infrastructure in AI research, exemplified by the collaboration between turbopuffer and SID AI to train the SID-1 agentic search model using large-scale RL, achieving 1.9x recall over RAG+rerank and 24x faster/99% cheaper than GPT-5.1.
View Cached Full Text
Cached at: 05/22/26, 03:52 AM
turbopuffer x SID An easy way to tell a good from a great AI researcher: how much do they think about infrastructure. Infra extends beyond what’s running on the GPUs: Slow environments will bottleneck your training steps. More parallel and powerful models make this problem worse. RL environment specifics are usually secret, but we shared some details in a recent post with our friends at @turbopuffer Training great models requires great infrastructure and we’re excited to be working with the best.
turbopuffer (@turbopuffer): SID-1 is an agentic search model by @SID_AI
→ 1.9x recall over RAG + rerank → 24x faster, 99% cheaper than GPT-5.1
trained using large-scale RL on turbopuffer at 1k+ QPS bursts over 10M+ document corpora across thousands of steps
Similar Articles
From Self-Play RL to Constrained Control: Distilling DiffMPC Students with DAgger
The paper replaces the action head of a PPO self-play driving teacher with a Differentiable MPC enforcing explicit safety constraints, then distills the student via DAgger, matching or outperforming the teacher on nuPlan, CARLA, and Waymo while using under 30% of the teacher's training time.
Cross-Question Reliable Reinforcement Learning
The paper introduces CARE (Cross-Question Abstention Reward Estimation), a reinforcement learning method that rewards an LLM's verbalized confidence relative to other questions' confidences at various risk tolerances rather than against binary correctness labels, substantially improving coverage at 5% risk on math and out-of-distribution QA benchmarks.
Bayesian $\kappa$-Explorer for Minimax Model Estimation in MDPs
The paper introduces Bayesian κ-Explorer, which places Dirichlet priors on MDP transition distributions to simplify minimax transition estimation by replacing confidence-set optimization with a closed-form posterior variance exploration signal, while retaining the Õ(T^-1/3) convergence rate of the frequentist κ-Explorer and showing lower worst-case estimation errors on Pendulum, Mountain Car, and Taxi.
PhoneWorld: From Real-App Trajectories to Dynamic and Verifiable Environments for Phone-Use Agents
The paper introduces PhoneWorld, a trace-grounded framework that converts real-app interaction trajectories into runnable, resettable, and verifiable Android environments spanning 34 apps across 16 domains, enabling benchmark evaluation, verified trajectory generation, and online RL for phone-use agents.
Calibrated Decision Rewards: Training Language Models with a Judge That Knows How Sure It Is
The paper introduces RLCDR (Reinforcement Learning from Calibrated Decision Rewards), which turns an LLM into a calibrated, reference-free decision judge whose reward matches ground-truth answer-key performance when training Qwen3-1.7B with GRPO on MATH, while also improving calibration and error detection over softmax judges and prior PRMs.