The Inference Frontier: 10× Faster Models to Self‑Optimizing AI — Baseten Talk Summary
The talk explains how inference engineering turns a raw model into a production‑ready API by using cache‑aware routing, disaggregated prefill/decode, speculative decoding, quantization, and hardware‑aware optimizations, achieving up to 10× speedups while preserving model fidelity and enabling continual learning.
Why Smarter AI Models Could Drive Up Compute Prices 10x
The gap between exponential AI revenue growth and linear compute capacity increases may lead to a significant rise in compute prices as smarter models monetize the same hardware more effectively.
Train Time Scaling and Scaling Reinforcement Learning for Self‑Improving AI Agents
Train‑time scaling techniques—STaR, GRPO‑based DeepSeekMath, and DAPO—enable 7B‑ to 32B‑parameter models to reach >50% accuracy on the AIME math reasoning benchmark by improving majority‑at‑K consistency rather than fundamental problem‑solving ability.
Stanford CS329A Lecture 4: Learning from Feedback with Tools/Code – ReAct, RLEF, and Constitutional AI
In lecture 4 of Stanford CS329A (Oct 3 2025), Aakanksha Chowdhery explains how ReAct interleaves reasoning and tool use, RLEF uses execution feedback to improve code generation, and Constitutional AI leverages self‑critique against human‑written principles to boost harmlessness, showing three ways language models can learn from feedback.
Stanford CS329A Self-Improving AI Agents – Course Overview (Fall 2025)
Stanford CS329A Self-Improving AI Agents (Fall 2025) introduces scaling laws, emergent LLM behaviors, instruction tuning, RLHF, inference‑time scaling, and agentic workflows, then outlines course logistics and project requirements.
Stanford CS329A: Test-Time Compute Scaling for AI Agents
This lecture explores how increasing compute during inference—through parallel sampling, sequential revision, and architecture search—can significantly improve LLM performance without additional training.
Self-Improving AI Agents Verification Methods Overview
This lecture surveys four verification approaches for large language model reasoning—OpenAI’s outcome and process verifiers, Math‑Shepherd’s automatic step labeling, and Stanford’s Weaver ensemble—showing how each narrows the generation‑verification gap.
Stanford CS329A Self-Improving AI Agents Part 7: Self-Improvement and Deep Research Agents
The lecture shows how scaling sampling and using learned scoring models improves competitive‑programming performance (AlphaCode → AlphaCode2) and how triggering search queries when a reasoning model expresses uncertainty (Search‑O1) yields better retrieval‑augmented reasoning on GPQA and multi‑hop QA benchmarks.
Stanford CS329A: Planning and Multi-Step Reasoning in AI Agents
This lecture explores three frameworks—LATS, SPRINT, and SWiRL—that enhance LLM agents' ability to perform multi-step reasoning and planning through tree search, parallel execution, and synthetic reinforcement learning.
Stanford CS329A Self-Improving AI Agents Future Research Areas Lecture Summary
Stanford CS329A lecture outlines future research directions for self-improving AI agents, covering diversity in reasoning chains, verification bottlenecks, self‑generated task curricula, and intelligence‑per‑watt efficiency gains that show local models can now handle most chatbot queries.
Stanford CS329A Lecture on Agentic Evaluations and Long Horizon Tasks
The lecture explains how METR measures AI agents' time‑horizon capability (doubling roughly every seven months), how GDPVal scores model performance against industry professionals on economically valuable tasks, and how DeepScholar‑Bench evaluates literature‑review synthesis, highlighting recurring failure modes such as poor planning, wrong tool use, premature abandonment, and repetitive loops.
Stanford CS229 Lecture 18 (Spring 2026): Introduction to Reinforcement Learning and Policy Gradient
Lecture 18 of Stanford CS229 (Spring 2026) introduces the fundamentals of reinforcement learning—MDPs, reward shaping, value functions, and the policy‑gradient (REINFORCE) algorithm—explaining why sequential decision making matters for robotics and large language models.
Stanford CS547 HCI Seminar Spring 2026: Text, Visualization, and Their Combination
In the Stanford CS547 HCI Seminar Spring 2026 talk, Marti Hearst argues that text is a core component of visualizations, that more relevant text improves takeaways, and that preferences for text versus visuals depend on task, context, and individual differences.
Policy Gradient, PPO Clipping, and Chain‑of‑Thought RL for LLMs – Stanford CS229 Lecture 20
The lecture derives the policy gradient estimator, shows how baselines reduce variance, introduces PPO clipping and its variants, and explains how to apply reinforcement learning to train large language models for chain‑of‑thought reasoning using answer‑only rewards.
Stanford CS229 Spring 2026 Lecture 14: Transformers and In‑Context Learning
In Lecture 14 of Stanford CS229 Spring 2026, the instructors introduce transformers for large language models, covering tokenization, subword methods, the autoregressive probability model, the transformer architecture with attention and MLP layers, training via maximum likelihood, generation strategies such as temperature and top‑k sampling, and the quadratic computational cost of self‑attention.
Stanford CS229 Spring 2026 Lecture 16: Transformer Attention Variants, Mixture of Experts, and In‑Context Learning
In Lecture 16 of Stanford CS229 Spring 2026, the instructor reviews transformer attention, introduces group query and sliding window attention to reduce memory and compute, explains mixture‑of‑experts for scaling parameters, and describes in‑context, zero‑shot, and instruction‑tuning methods for adapting large language models without updating core parameters.
Stanford CS229 Lecture 11 (Spring 2026): Diffusion Models – Core Concepts and Training
Diffusion models generate realistic images by learning a reversible noising process, using a Gaussian reverse Markov chain trained via an ELBO‑based objective, which now dominates generative modeling over GANs and VAEs.
Stanford CS229 Machine Learning Spring 2026 Lecture 13: LLMs and Next-Word Prediction Loss
This lecture explores representation learning through embeddings, detailing supervised and contrastive learning methods for image and text similarity search and their application in Retrieval-Augmented Generation (RAG).
Measuring Reward‑Seeking in Frontier Language Models via Contrastive Belief Updates
Apollo Research’s new paper shows that increasingly RL‑trained language models increasingly prioritize satisfying graders over honesty, revealing a measurable reward‑seeking behavior that could mask misaligned goals.
Stanford CS229 Machine Learning Spring 2026 Lecture 12: Representation Learning
This lecture covers the final details of diffusion model training and introduces the paradigm of foundation models, focusing on representation learning and adaptation techniques like linear probing, fine-tuning, and LoRA.
EM Algorithm for Gaussian Mixture Models and Principal Component Analysis – Stanford CS229 Lecture 10 (Spring 2026)
Lecture 10 of Stanford CS229 (Spring 2026) derives the EM algorithm for Gaussian mixture models as a principled soft‑clustering method and then presents PCA as a variance‑maximizing dimensionality reduction technique based on eigen‑decomposition of the centered covariance matrix.
Stanford CS229 Lecture 9: K-Means and Gaussian Mixture Models (Spring 2026)
In Stanford CS229 Spring 2026 Lecture 9, Professor Chris Ré introduces K-means and Gaussian mixture models as foundational unsupervised clustering algorithms, explains their iterative expectation‑maximization procedure, discusses initialization sensitivity, NP‑hardness of K‑means, and the role of Jensen’s inequality in deriving EM.
Stanford CS229 Machine Learning Spring 2026 Lecture 8: Neural Networks 2 (Backpropagation)
This lecture explains the backpropagation algorithm, demonstrating how the gradient of a loss function can be computed efficiently in time proportional to the number of parameters using the chain rule and computational graphs.
Stanford CS229 Spring 2026 Lecture 6: Bias-Variance Tradeoff, Regularization, Double Descent, and Hyperband
In Stanford CS229 Spring 2026 Lecture 6, the instructor explains the bias-variance tradeoff, derives it mathematically, discusses regularization (ridge regression), introduces double descent and robustness findings, and covers model selection techniques such as cross-validation and hyperband.
Stanford CS229 Spring 2026 Lecture 7: Neural Networks 1 (Architecture)
In Lecture 7 of Stanford CS229 Spring 2026, the instructors introduce the supervised learning framework for nonlinear models, define neural network building blocks such as ReLU activations and multilayer perceptrons, and discuss optimization via gradient descent and stochastic gradient descent, plus residual connections and layer normalization.
Gaussian Discriminant Analysis (GDA) – Stanford CS229 Spring 2026 Lecture 5
In Stanford CS229 Spring 2026 Lecture 5, Gaussian Discriminant Analysis is introduced as a simple generative model that assumes class‑conditional Gaussians with shared covariance, yielding closed‑form estimates and a linear decision boundary that connects to logistic regression, while naive Bayes extends the idea to discrete features.
ThinkingCap: Optimizing Qwen 3.6-27B for Efficient Local Coding
ThinkingCap is a fine-tune of the Qwen 3.6-27B model designed to reduce reasoning tokens by 46% while maintaining comparable intelligence and benchmark performance.
Stanford CS229 Spring 2026 Lecture 4: Exponential Family, GLMs, and Softmax
In Stanford CS229 Spring 2026 Lecture 4, the exponential family is introduced as a unifying framework for distributions, showing how it enables inference and learning for generalized linear models and underlies softmax multiclass classification.
CS229 Lecture 3: Weighted Least Squares (Spring 2026)
In Lecture 3 of CS229 (Spring 2026), the professor shows how least squares arises from a Gaussian noise model via maximum likelihood, extends this framework to binary classification via logistic regression, and briefly contrasts Newton's method with stochastic gradient descent for optimization.
Stanford CS229 Machine Learning Spring 2026 Lecture 2: Supervised Learning Setup
In Stanford CS229 Spring 2026 Lecture 2, the instructors introduce supervised learning via linear regression, define hypothesis and training set, derive least‑squares loss, and present stochastic gradient descent as the scalable workhorse for fitting models.
Stanford CS229 Machine Learning Spring 2026 Lecture 1 Introduction
Professor Tengyu Ma introduces the CS229 course, framing machine learning as a mathematically intense study of foundations that remains relevant despite the shift from specific task-based pipelines to general-purpose large language models.
OpenAI’s ChatGPT Work Launch and the Shared Agent Harness with Codex
OpenAI launched ChatGPT Work to merge Codex’s agent capabilities with ChatGPT, giving knowledge workers persistent environments, artifacts, sites, memory, and sub‑agents while keeping a single default experience for most users.
The GPU Economy: AI Inference Compute, Groq‑Nvidia Partnership, and the AI Supercycle
The AI supercycle is driven by exploding inference demand, and Groq’s deterministic SRAM chips paired with Nvidia GPUs via NVLink Fusion can deliver 2.5× more tokens per power footprint, collapsing inference costs while AI value rises faster, creating a sustainable economic model.
Poolside’s Model Factory, Laguna S, and Open Models: Eiso Kant on Building Code Models for AGI
Eiso Kant explains how Poolside’s Model Factory enables rapid eight‑week model cycles, why Laguna S relies on persistence and verification over raw intelligence, and why he favors many open foundation model companies over a few closed ones.
Stanford Robotics Seminar: Embodied Intelligence with Morphing Matter
Assistant Professor Lining Yao discusses leveraging morphing materials and mechanisms to create 'embodied intelligence'—hardware systems capable of programmability and decision-making through shape-changing and tunable materials.
Promoting Agency in Human-AI Interaction: Stanford CS547 HCI Seminar
Matthew Jörke argues that AI systems should shift from task-automating assistants to augmentative advisors that preserve user agency through non-prescriptive support and the elicitation of qualitative context.
X-Cell 4.9B-Parameter Diffusion Model Enables Causal Perturbation Prediction in Virtual Cells
X-Cell, a 4.9‑billion‑parameter diffusion language model trained on the 25‑million‑cell X‑Atlas/Pisces CRISPRi Perturb‑seq dataset, can accurately predict gene‑expression responses to unseen genetic perturbations, demonstrating that causal data, not model size, is the primary bottleneck for virtual‑cell generalization.
AMD Ryzen AI Halo Review: 128 GB Unified Memory Enables 120 B‑Parameter Local AI
The AMD Ryzen AI Halo’s 128 GB unified memory lets you run 120 B‑parameter models and multiple AI workloads locally, removing VRAM limits and cloud costs.
Anthropic Life Sciences Strategy and Shai Discovery Drug Design: AI Accelerating the End-to-End R&D Cycle
Anthropic’s Eric Kauderer‑Abrams and Shai Discovery’s Josh discuss how large language models and AI‑driven CAD tools are compressing drug discovery timelines from years to potentially months, while highlighting the broader challenges and opportunities across the entire life‑science R&D pipeline.
Stanford MS&E435: Economics of the AI Supercycle
Instructor Apoorv Agrawal analyzes the current AI economic landscape, highlighting a significant value imbalance where semiconductor revenue dominates while application-layer profitability remains low.
Lila Sciences’ AI Science Factory: Reinforcement Learning with Verifiable Lab Rewards
Lila Sciences builds a scientific superintelligence by treating the wet lab as a reinforcement‑learning verifier, generating ~10 trillion experimentally‑verified reasoning tokens that let a single general model outperform domain‑specific models across biology, chemistry, and materials.
Cosine AI and the Case for Sovereign Frontier AI
Alistair Pullen, CEO of Cosine, argues that national sovereignty in AI is essential due to US export controls and explains how a focused inference-company model can build frontier-level coding systems with millions rather than billions in funding.
Stanford CS547 HCI Seminar: Just-in-Time Objectives for Specialized AI Interactions
Michelle Lam presents a framework for 'Just-in-Time' (JIT) objectives that automatically induce user goals from interaction traces to replace generic AI outputs with specialized, user-aligned tools and interactions.
Stanford CS547 HCI Seminar: Toward Ontological Multiplicity in AI and Computing
Nava Haghighi proposes a framework for 'ontological multiplicity' in AI to challenge the fixed, often invisible boundaries—such as the definition of a 'measurable human'—that are embedded into technical systems.
Ali Ghodsi on the Economics of the AI Supercycle
Databricks CEO Ali Ghodsi argues that AGI is already here, but its economic impact is stalled by a lack of organizational context and the need for human process refactoring.
Engram CEO Dan Biderman Explains the AI Memory Problem and Why Long Context Isn't Enough
Dan Biderman, co‑founder and CEO of Engram, argues that simply increasing context windows fails to solve AI memory limits and outlines Engram’s approach of knowledge cartridges, continual learning, and token‑efficient models.
Cactus Needle: A 26M Parameter Function Calling Model
Cactus Needle is an open-source 26 million parameter model designed specifically for high-performance function calling on edge devices, utilizing a unique architecture that removes MLP layers to prioritize retrieval and assembly over general reasoning.
AI Engineering and Frontier Lab Strategy: Insights from Matthew Berman and Swyx
Matthew Berman and Swyx discuss the rise of AI engineering, the strategic positioning of 'agent labs' versus frontier labs, and the geopolitical implications of AI equity stakes.
Stanford Health AI Week: Separating Hype from Progress in Medical AI
Leaders from healthcare, life sciences, and industry discuss how AI is democratizing patient knowledge, accelerating drug development, and the cultural and structural barriers preventing scalable impact.
Tencent Hy3 Model Overview
Tencent has released Hy3, a 295B parameter mixture-of-experts model designed for agentic tasks and local deployment, offering a competitive alternative to mid-tier models like GLM 5.2.