Stanford CS336 Language Modeling from Scratch | Spring 2026 | Lecture 15: Mid/Post-Training
Lecture 15 of Stanford CS336 emphasizes mid- and post-training (SFT → RLHF) as a key driver of large language model quality. The lecture highlights that instruction data is trending longer and more interactive (chatty, tool-using), which increases compute, memory‑bandwidth, and inference complexity—supporting infrastructure and semiconductor suppliers.
Linked assets
Key hardware and infrastructure suppliers that benefit if post-training and inference intensity remain elevated: NVDA (data-center AI accelerators), AMD (alternative accelerators), TSM (foundry capacity for leading nodes), AVGO (networking/interconnect), MU (DRAM/HBM memory bandwidth), and ASML (leading-edge lithography).
NVIDIA Corporation operates as a data center scale AI infrastructure company.
Direct beneficiary of sustained post-training cycles and rising inference cost from longer, tool-using interactions → more GPU demand.
Its products are used in high performance computing, smartphones, Internet of things, automotive, and digital consumer electronics.
Foundry bottleneck for leading-edge AI accelerators; benefits from sustained capacity demand.
Advanced Micro Devices, Inc.
Alternative accelerator supplier; benefits if overall accelerator TAM stays elevated from post-training/inference intensity.
Broadcom Inc.
Scaling AI clusters to support heavier inference/training increases networking/interconnect needs; AVGO leveraged.
Micron Technology, Inc.
Memory-bandwidth intensity rises with long context and training/inference workloads → supports DRAM/HBM demand.
ASML Holding N.V.
Long-run capex cycle support if scaling/post-training keeps driving demand for leading-edge nodes.
Source proof
Source proof: Strong source proof | 4 extracted claims | 6 directional assets | 1 supporting author | headline-like title review
Lecture notes and fragments identify SFT and RLHF as central techniques, note an evolution toward longer instruction/chatty/tool-using outputs, and reference open-source SFT efforts. The discussion is technical/academic with no company announcements; investable signal is thematic: higher training/inference intensity supports picks-and-shovels suppliers.
Analysis pending. The source event was captured, but automated analysis failed: OpenAI structured request failed
Stanford Robotics Seminar content is early-stage R&D focused on embodied intelligence using morphing materials (e.g., PDMS/silicones), additive manufacturing (FDM-style printing/flat-pack concepts), and computational design/optimization; plus a brief mention of environmental DNA (eDNA) collection. This is not a near-term catalyst, but it supports longer-horizon theses around (1) computational design/CAE software, (2) additive manufacturing ecosystems, (3) silicone/material suppliers, and (4) life-science tools if eDNA sensing becomes more widely deployed. Ticker links are indirect and high-uncertainty.
Stanford HCI seminar describes research on LLM-based physical activity coaching that promotes user agency (non-prescriptive support), elicits qualitative context, stays on-task over long conversations, and uses an RL method for LLM agents to explicitly reason about uncertainty in user goals. This is early-stage academic work; actionable signals are indirect and mostly map to (1) LLM agent/tooling platforms, (2) digital health coaching/wearables ecosystems, and (3) continued demand for LLM inference infrastructure.
Lecture snippet frames the “AI supercycle” as an infrastructure/economics story: inference/training at scale is not marginally free, requiring sustained capex in chips, power, and data centers. Mentions hyperscaler buildouts (AWS), application/platform monetization (Palantir AIP), and internal ASIC programs (Google TPU, Meta MTIA). Actionability is moderate because the content is thematic and qualitative with few concrete catalysts, but it supports tradable positioning in hyperscalers/platforms and AI infra beneficiaries over a medium horizon.
Stanford course talk frames an “AI supercycle” application thesis in life sciences: AI as a CAD suite for molecules that compresses early discovery/optimization, but with long real-world lags driven by IND/FDA timelines. It also references GLP-1s as an example of blockbuster economics and highlights that large pharma may reinvest windfall cash flows into computational/drug-design platforms or acquire tool/platform companies.
The provided Stanford Online video title/body is about learner experiences in an Engineering Leadership Program and contains no technical theses, research signals, sector views, catalysts, or company/ticker references. There is no actionable market content to map to tradable tickers.
Stanford CS547 seminar discusses “Just-in-Time (JIT) objectives” for specialized AI interactions: dynamically generating task-specific objectives/evaluators (e.g., LM-as-judge, uncertainty statements, lightweight appended objectives) to reduce generic LLM outputs and improve user-preferred results (incl. UI generation, web/DOM/screenshot inputs, iterative hill-climbing with evaluators). This is research-stage; no direct company catalysts are named, but it supports a broader thesis: value accrues to AI platforms and tooling that can (a) reliably align outputs to user intent, (b) evaluate/score generations at runtime, and (c) operationalize multimodal context (screenshots/DOM) with uncertainty-aware outputs—driving incremental demand for inference, dev tooling, and enterprise adoption.
This Stanford HCI seminar excerpt is largely philosophical/qualitative (ontological multiplicity, critique of “the human” in AI) with a small technical hook around EDA (electrodermal activity) sensing, responder/non-responder issues, and how commercial LLM chatbots and LLM architecture may (or may not) surface “multiplicity.” It does not contain concrete, near-term product/earnings catalysts, benchmarks, or implementation details. Any trading linkage is therefore weak and mostly thematic (LLM platform leaders; biosensing/wearables and affective-computing stacks).
Supporting authors
Academic lecture material from Stanford CS336 and related Stanford courses and seminars. Content is technical and educational rather than company-specific; authorship reflects course instructors and seminar presenters.
Unlock full thesis monitoring
Consider thematic exposure to AI infrastructure and semiconductor suppliers tied to training and inference intensity. This is a medium-term, thematic insight rather than an event-driven trade—align position sizing and time horizon accordingly.