Course 5B — Case Studies and Conclusion
Section 7 is not another benchmark table. It asks whether a model can sustain a long workflow, produce an inspectable artifact, encounter constraints, and recover. A case study is therefore read as:
Source: Author-created case-reading framework for Kimi K3 §7; not a paper figure or equation.
One selected success does not estimate average reliability. The cases below remain paper-reported unless explicitly stated otherwise.
How to read a case dossier
For each trajectory, keep six fields attached:
| Field | Question |
|---|---|
| Task | What exact outcome was requested? |
| Environment | Which tools, hardware, sandbox, sources, and time budget constrained the work? |
| Trajectory | Which inspectable actions or revisions connected request to artifact? |
| Artifact | What executable, measurable, or reviewable object remained? |
| Verification | Which test, benchmark, simulation, comparison, or human review checked it? |
| Boundary | Which reliability or causal claim does this selected example not support? |
Source: Author-created case-reading framework for Kimi K3 §7; not a paper table.
Systems-building cases
GPU kernel optimization
- Task and environment: Independently optimize AttnRes, DSA, KDA, and MLA kernels in identically configured sandboxes, with up to 24 hours per task, on NVIDIA Hopper and an alternative-vendor GPGPU.
- Trajectory: Profile, rewrite, benchmark, and iterate against the same workload rather than changing the task definition.
- Artifact and verification: Executable kernels measured for correctness and latency. The paper reports AttnRes falling from 283.6 ms to 114.4 ms, DSA and KDA runtime reductions of 55.1% and 73.6%, and MLA exceeding half of peak TFLOPS.
- Boundary: These are selected tasks and paper measurements. They do not establish success rate over unseen kernels or prove that every optimization is maintainable.
MiniTriton compiler
- Task and environment: Build a compact Triton-like stack for an NVIDIA L20, not merely one generated kernel.
- Trajectory: Connect a tile-level Python frontend and layout system to an MLIR optimization layer, PTX generation, runtime, eager and forward-only compiled tensor paths, reverse-mode autograd, neural-network modules, NCCL distributed primitives, sparse operations, and visualization.
- Artifact and verification: The paper reports benchmark comparisons with PyTorch eager and
torch.compile, a tensor-core matmul near 90% of the measured machine roof at the largest shapes, a KDA prefill kernel above a matched Triton reference, a GPT loss curve tracking PyTorch, and full-model gradients within approximately (10^{-4}) of the reference comparison described by the paper. - Boundary: The repository does not audit MiniTriton source, reproduce the rooflines, or independently validate the paper’s parity criterion.
Nano-KPU chip design
- Task and environment: In one 48-hour Kimi Code run, design a nano-model inference prototype with hybrid KDA/NoPE-MLA, Block AttnRes, sigmoid MoE routing, group-wise INT4 weights, open-source EDA tools, and the Nangate45 cell library.
- Trajectory: Move from architecture through RTL and physical-design checks rather than stopping at a block diagram.
- Artifact and verification: The paper reports timing closure at 100 MHz inside a 4 mm² analytical area budget, RTL-simulated decode above 8,700 tokens/s, 1.46M standard cells, 0.277 MiB SRAM, and an INT4 MAC array with fused dequantization.
- Boundary: EDA output and RTL simulation are not fabricated-silicon measurements. The case is explicitly an early proof of concept.
Scientific and knowledge-work cases
Computational astrophysics
The reported I–Love–Q trajectory reviews and cross-validates more than twenty papers, implements a numerical pipeline, evaluates more than 300 equations of state, identifies inconsistencies in published formulas, writes more than 3,000 lines of Python, and produces an interactive HTML dashboard in about two hours. The artifact links literature, equations, code, computed results, and visualization. The paper compares that duration with a typical one-to-two-week expert workflow; this is an author-reported comparison, not a controlled time study. Citation count, code volume, and speed do not independently establish scientific correctness.
AI-ASIC research site
The paper reports an interactive site spanning 42 years of the AI-ASIC industry, more than 120 refinement rounds, 87 quarterly reports, 99 original PDFs totaling more than 11,000 pages, over 2,800 web searches, and over 1,100 terminal queries. The central capability is iterative source-grounded artifact construction. The counts establish scale of the selected trajectory, not completeness, citation correctness, or average research reliability.
GWTC-5 multi-agent analysis
The reported workflow analyzes 391 gravitational-wave events with more than twenty concurrent subagents and produces seven scientific visualizations, two summary tables, and a synthesis of more than ten papers. Concurrency demonstrates decomposition and coordination only if the final artifact reconciles duplicated work, conflicting assumptions, and provenance. The paper reports the outputs but this course does not independently audit their scientific interpretation.
Creative-production case
The paper reports two related creative artifacts: K3 created a 3Blue1Brown-style architecture explainer, and it edited a teaser from 56 source clips. The disclosed work includes clip selection, motion-matched cuts, frame-accurate beat synchronization, audio processing, and multiple revision rounds. Treat both as production pipelines—not as evidence that one prompt produced a finished film. The comparison with one to two days of experienced-editor work is paper-reported, and a polished selected artifact cannot establish average visual judgment, mathematical accuracy, or editing reliability.
Conclusion: sort claims before summarizing
The strongest bounded synthesis is:
- K3 is reported as an open 2.8T-parameter sparse MoE with 104B active parameters, native vision, and a 1M-token context capacity.
- Sections 2–5 describe a coupled architecture, training, post-training, and infrastructure program.
- Section 6 reports a frontier-competitive but non-uniform capability profile, with remaining gaps on several research, knowledge-work, computer-use, and cyber tasks.
- Section 7 supplies selected long-horizon cases, not a reliability distribution.
- This repository executes pedagogical miniatures and derived analyses; it does not reproduce K3's full-scale training, infrastructure, or benchmark suite.
The open research questions are therefore more useful than a victory claim: which mechanisms contribute most, how robustly million-token capacity becomes usable memory, how stable sparse routing remains at scale, how often long agent trajectories recover, and how independently the reported cases can be verified.
Retrieval check
For any case above, identify its artifact, verification method, and one unsupported extrapolation. If you cannot name all three, the case has been remembered as an anecdote rather than evidence.