THE ARCHIVE
Library
Cloudflare at the edge: products, boundaries, and deployment evidence
Learn how Workers, static assets, data products, and deployment checks fit into a small independently deployable service.
Computer use: build systems that can observe, act, and recover
A practical map of browser and desktop automation: perception, planning, action, verification, and safe recovery.
Hacker News observation: turn attention into bounded technical experiments
A dated HN API snapshot, read as a set of technical hypotheses rather than a popularity ranking.
Fundamentals of image generation: Translating Diffusion, Flow Matching, and Transformer to production decisions
The generation from noise back to an image is understood as the relationship between latent representation, conditioning, Flow Matching, and DiT, and is used to determine speed and control.
Image model research notes: don't confuse offering models, open weights, and licenses
Research notes based on official sources checked on 2026-10-04. Compare API availability, open weights, and license conditions for OpenAI GPT Image, Google Imagen, and Stable Diffusion.
Image production flow: Reference, character consistency, Control, LoRA, Inpainting
Instead of relying on one-shot generation, we separate the responsibilities of reference, control, local correction, and LoRA when necessary, and transform it into a reproducible image production process.
Practical image workshop: Creating a single advertising visual using a verifiable process
An exercise in creating still images containing products and people by dividing them into specifications, references, composition, creation, local corrections, typesetting, and release checks.
Image generation frontier notes: Reading FLUX.3, GPT-Image 2.5, Nano Banana 2 in process
From the official information confirmed on 2026-10-04, we summarize that the focus of image generation has shifted from one-shot image quality comparison to layout, editing, history, and high-speed iterative process design.
Jev foundations: typed decisions outside text generation
Use Jev/System One as a small typed decision component with candidates, state, and probabilities alongside ordinary code.
Jev Choice, Score, and Noul: do not confuse probability, confidence, and authority
Learn the roles of three question primitives and how to prevent concentrated distributions from becoming permission to automate.
Jev implementation lab: connect candidate generation, abstention, and E2E evaluation
Connect semantic judgment to deterministic candidate generation and safety boundaries, then design end-to-end evaluation with controls.
Quantization, KV cache, and runtimes for a local LLM that stays responsive
Separate model weights, working memory, and generation speed; choose MLX, llama.cpp, or vLLM and turn the choice into a reproducible local evaluation.
Choosing MiniMax and local LLMs: separate open weights, APIs, and feasibility
A practical research and experiment guide for comparing MiniMax and other models without conflating public repositories, weights, APIs, and inference servers.
AI investment: follow the constraint all the way to the shareholder
Connect HBM, NAND, power and optics to pricing, qualification, capital needs and valuation.
Financial statements: reconcile the business with the cash
Read actual AAOI and Bloom tables, calculate margins, and distinguish stocks from flows.
A dated price snapshot is not a valuation or a price history
Inspect SNDK, BE, AAOI and MU quotes with timestamps, then calculate drawdowns without inventing history.
Value stocks: normalize the cycle before trusting the multiple
Distinguish a cyclical bargain from a value trap using original earnings and reinvestment cases.
Macro: separate the policy rate, inflation data and AI capital spending
Use September 2026 Federal Reserve and August inflation releases, then test cash-flow sensitivity.
The next bottleneck: test BE and AAOI as conditional business cases
Compare power availability and optical qualification, then expose the growth already required by a valuation.
Memory hierarchy for AI systems: locate the stalled byte
A concrete way to distinguish registers, cache, HBM/DRAM, SSD, and networked storage before making an infrastructure claim.
Why HBM bandwidth matters: count bytes before FLOPS
Connect a kernel's arithmetic intensity to HBM bandwidth, capacity, and an honest performance experiment.
Stacked DRAM, TSVs, and yield: why one good die is insufficient
Follow an HBM stack through TSV interconnect, assembly, heat, test, and compounded yield.
HBM4 interfaces and packaging: more pins change the system
Understand a wider HBM4 interface, logic base die, package co-design, and the limits of multiplying pin speed.
DRAM and NAND: two technologies, different cycles
Separate volatile working memory from persistent flash storage before making an AI-infrastructure or business claim.
NAND and SSD storage for AI: keep accelerators fed
Trace an AI data path through SSDs, host memory, decompression, and GPUs, including endurance and measurement limits.
SanDisk: the flash business after separation
Read SanDisk's standalone FY2026 reporting as a flash-storage business, while keeping pre-separation comparability limits visible.
SanDisk financial statements: turn revenue into cash carefully
Use the FY2026 GAAP income statement, cash flow statement, capital spending, and share counts without mixing periods or metrics.
SanDisk demand and pricing: separate exabytes from dollars
Use SanDisk's FY2026 end-market disclosures to distinguish volume, price, mix, and inventory explanations.
Micron, HBM, and the difference between a part and a platform
Connect Micron's HBM3E product disclosures to an AI-system qualification chain without turning product claims into market facts.
Micron cash flow and capex: separate investment from a headline
Read a memory maker's earnings materials by keeping GAAP profit, operating cash flow, capital expenditure, and capacity effects distinct.
SK hynix memory economics: product leadership is not a margin model
Read SK hynix HBM announcements as product and supply-chain evidence, then build a separate, falsifiable economics model.
Samsung memory competition: compare qualification, architecture, and disclosure
Samsung’s memory position must be separated into DRAM, NAND, HBM qualification, packaging, and the economics disclosed by a diversified company.
Kioxia and Western Digital: NAND partnership, products, and accounting boundaries
Compare NAND businesses by wafer technology, joint manufacturing arrangements, client versus enterprise products, and the legal entity that reports the accounting.
Advanced packaging bottlenecks: why a known-good die is not a shippable accelerator
HBM demand is constrained by more than DRAM wafers: TSV processing, die stacking, base die, interposer or substrate, bonding, thermal paths, test, and system qualification all have yields.
Memory equipment suppliers: read process step exposure before calling capex leverage
Equipment revenue is exposed to process intensity and customer capital plans, but tool shipment, installed base service, and a memory maker’s bit output are different time series.
Memory-cycle valuation: separate bit demand, pricing, utilization, and cash conversion
Memory cycles combine bit demand, supply additions, inventory, pricing, utilization, and capex. A valuation case should expose each driver and the date of the accounting evidence behind it.
Invalidating a memory thesis: design a dashboard that can change your mind
A serious memory thesis names its technical and financial disconfirming observations before it becomes a narrative. HBM, NAND, packaging, and equipment each fail differently.
Where AI clusters wait: a bottleneck map from HBM to the optical fabric
HBM, topology, oversubscription, and optical links form one queueing system. Locate the limiting resource before creating an equity narrative.
800G and 1.6T optics: a port rate is a budget
Read an 800G or 1.6T label through lanes, modulation, host loss, fiber reach, FEC, and qualification.
Transceivers, lasers, and DSPs: where an electrical lane becomes light
A module is an electrical, optical, thermal, and manufacturing chain. Map each conversion before attributing value or a failure.
Silicon photonics and CPO: shorten the electrical path, change the service model
Co-packaged optics moves the optical engine beside the switch ASIC. It can improve the electrical budget while changing assembly, cooling, and repair.
Copper, AEC, and optics: choose the reach boundary before choosing a cable
Passive copper, active electrical cable, active optical cable, and transceivers solve different channel and operations problems.
AAOI business and products: read a component company as a chain of boundaries
AAOI sells across data-center, CATV, telecom, and FTTH markets. Product language must be separated from customer qualification, production, and revenue.
Reading AAOI financial statements: period, unit, and business driver before a ratio
Turn revenue, gross margin, inventory, receivables, and cash flow into time-stamped questions rather than a price forecast.
AAOI customer concentration: what a disclosure says and cannot identify
Concentration measures exposure, not a customer identity or a forecast of an optical deployment.
Capacity, yield, and cash: why optical manufacturing can bottleneck after demand appears
Capacity is qualified throughput, not floor space. Yield, test time, material availability, and cash conversion decide whether an order becomes revenue.
AAOI contracts versus revenue: trace the optical module through its accounting gates
A forecast, purchase order, shipment, acceptance and recognized revenue are separate events with different evidence.
Coherent and Lumentum: compare the optical boundary before comparing companies
A useful comparison starts with product boundary, end market, manufacturing and reported segment definitions—not ticker shorthand.
Broadcom and Marvell: where value sits between the switch and the optical module
A component-level guide to switch ASICs, optical DSPs and active-copper retimers: what each controls, what it does not prove, and how to test a value-capture thesis.
Arista backend networks: translating an AI fabric into ports, links, and operations
A practical reading of a leaf-spine backend: why switch density is not deployment volume, how optics enter the topology, and how to test an operational design.
NVLink, InfiniBand, and Ethernet: identify the domain before comparing fabrics
Interconnects operate at different layers and deployment choices. Compare topology, software, NICs, switches, and reach.
Optical supply chain: lasers, packaging, test, and qualification as bottlenecks
The next constraint may be an input, qualified process, test station, or customer validation. A shortage needs a named stage and evidence.
Optics margins and valuation scenarios: model drivers, not a target price
Use conditional bull, base, and bear operating scenarios. Volume, mix, yield, utilization, working capital, and discount rate are assumptions to expose.
Invalidating an optical-infrastructure thesis: make the disconfirming dashboard first
A thesis should name the technical, customer, manufacturing and accounting observations that would change it.
Optical networking monitoring lab: read filings by event type before updating a thesis
A small recurring ledger separates annual and quarterly performance evidence from contracts, earnings releases, financing, and governance events.
Read this before adopting OSS AI: repository, weights, dependencies, and operations
Move OSS safely from learning into operations by separating licenses, maintenance, reproduction, vulnerabilities, and model weights instead of trusting star counts.
OSS release observation: a release page is a verification starting point
A dated GitHub Releases observation with a concrete pre-adoption record and test procedure.
Read AI papers into implementation decisions: claims, conditions, reproduction, and failure
A textbook for turning papers into practical experiments by separating a claimed improvement from its problem, conditions, reproduction, trade-offs, and failures.
Read DeepSeek-R1: what reasoning RL changes and what it does not promise
Read the 2025 DeepSeek-R1 paper through reward design, distillation, evaluation conditions, reproduction, and operational limits.
Read s1: what to measure when improving reasoning with little data
Translate the 2025 s1 paper into implementation decisions through test-time scaling, budget forcing, data selection, reproduction, and limits.
Read DAPO: do not view LLM reinforcement learning as an algorithm alone
Read the 2025 DAPO paper through reward, length bias, asynchronous system design, reproduction, and reward hacking.
Read 2025 reasoning-RL and efficiency papers into implementation
Read four 2025 papers on test-time compute allocation, RL training efficiency, shorter reasoning, and dynamic length control by separating claims from constraints.
Introduction to Physical AI: Connecting the history of VLA from perception to safe behavior
Instead of viewing Vision-Language-Action as the magic of a universal robot, understand it as a connection between observation, coordinates, strategy, and control.
2025 VLA map: read SmolVLA, π0/π0.5, OpenVLA, and GR00T using the same yardstick
Updated notes that interpret each presentation in terms of openness, body, data, behavioral expression, and evaluation boundaries, rather than ranking each presentation in a comparison table.
LeRobot practice: Make dataset, imitation learning, evaluation, and Sim2Real into one quality loop
Starting with LeRobot's data format and unified interface, learn the points of contact that are likely to fail from collection to actual machine evaluation.
Starting Physical AI without a robot: Safe learning procedure and decision not to proceed to the actual machine
It clearly states that the hardware is not running, and touches on the core of Physical AI just through data observation, simulation, and evaluation design.
Physical AI coordinate transformation and camera calibration: do not treat image pixels as safe robot commands
Learn camera, robot base, and hand coordinate systems and calibration as verifiable contracts outside of VLA.
Reading VLA papers to implementable designs: Comparative reading exercise of π0, SmolVLA, and GR00T
The novelty of a paper is not confused with performance ranking, but is broken down into behavioral expression, data, body, evaluation, and unverified boundaries.
Physical AI workshop: from data splits to policy evaluation and Sim2Real decisions
Turn public or synthetic episodes into reproducible splits, baseline and adaptation evaluations, failure analysis, and a justified decision not to proceed to hardware.
From chip to grid: map the power chain before choosing an AI-data-center thesis
A practical map of the electrical, thermal, contractual, and reliability boundaries between a GPU and a power plant.
Interconnection and time to power: capacity is an engineering and regulatory queue
Why a large-load request becomes studies, equipment, tariffs, construction, commissioning, and operating limits.
Bloom Energy solid-oxide fuel cells: mechanism, fuel boundary, and operating questions
How to evaluate an onsite solid-oxide fuel-cell system without confusing its product description with a lifecycle result.
Bloom Energy business and customers: distinguish a customer signal from recognized revenue
A disclosure ladder for reading customer, partnership, deployment, and revenue claims in onsite power.
Bloom Energy financial statements: a filing is a measurement boundary, not a valuation answer
Read income, balance sheet, cash flow, notes, concentration, and risk factors as separate evidence.
Bloom Energy backlog and contracts: backlog is a definition with cancellation and execution risk
How to keep commercial pipeline, binding contracts, revenue, cash, and installed capacity separate.
Bloom Energy cash flow and dilution: follow the funding path
Cash generation, debt, equity, and project financing answer different questions than earnings.
Gas supply, emissions, and permits: onsite power has external boundaries
A site checklist for fuel contracts, air permits, methane boundaries, and emissions claims.
Nuclear and PPA economics: a contract is not a delivered megawatt
Compare nuclear supply, PPAs, and data-center demand using delivery, risk, and accounting boundaries.
Transformers, switchgear, and UPS: the electrical chain behind usable compute
Map the equipment, protection, redundancy, and commissioning steps between utility service and a rack.
Liquid cooling and rack density: thermal design is an interface problem
Explain direct-to-chip and liquid loops through heat transfer, water boundaries, controls, and serviceability.
Vertiv business and valuation: link power-and-cooling demand to filed economics
Read Vertiv’s systems business, backlog, cash economics, and risks without turning demand growth into a price target.
Choose AI providers as operating layers, not a capability table
Design OpenAI, Anthropic, Google, ElevenLabs, Databricks, Vercel, and Cloudflare as model, data, execution, voice, and delivery layers.
Evaluation, safety, and cost design before placing generative AI in production
Connect evidence, authority, recovery, observability, and updates before measuring model cleverness, so answer quality remains reproducible in operation.
Data-connected AI with Databricks: Agent Bricks and Unity Gateway in practice
Design authorization, meaning, audit, and cost before retrieval quality when connecting AI to enterprise data.
AI apps on Vercel: put AI SDK, Gateway, and Workflow at product boundaries
Use Vercel’s agentic-infrastructure announcements to design streaming, tool calls, durable jobs, secrets, and evaluation as one web product.
Edge AI with Cloudflare: boundaries for Workers AI, AI Gateway, and Agents
Design state, authentication, cost, caching, and tool execution separately before shipping edge inference or an AI app quickly.
Voice AI with ElevenLabs: design consent, latency, and verification before voice quality
Learn consent, rights, explicit failure handling, latency, and evaluation for TTS, voice conversation, and dubbing from official ElevenLabs material and practical exercises.
Official announcement watch: return provider updates to product decisions
A 2026-10-04 record of official blog and documentation entry points that keeps announcement dates, observed facts, implications, and unknowns separate.
Rust foundations: the first ten stages
A practice-led path through Cargo, ownership, borrowing, errors, and a small command-line program.
Build Trail: Rust stages 11–30
Use one work-log product to practice contracts, persistence, web boundaries, recovery, measurement, and release.
Rust mental models for working systems
Connect ownership, boundaries, concurrency, persistence, and performance to concrete design decisions.
Video generation map: until diffusion, flow, and DiT become one cut
We reconsider video generation as the generation of latent expressions with a time axis, and connect the roles and limitations of diffusion, flow matching, and DiT to production decisions.
Reference-first video production: Freeze the world before the prompt
We practice planning, reference design, character sheets, shot lists, cut generation, editing and audio as a production process that manages consistency.
Conditions for using Blender and 3D for video generation: designing cameras, occlusions, and continuity
Rather than using 3D for the final image, it is used to visualize spatial constraints that are prone to breaking. Practice blocking and camera continuity with Blender.
30-second generation video workshop: Verify each cut and get closer to completion
A 30-second short film is broken down into small production experiments that can be reproduced from planning, cut design, generation, failure evaluation, and editing.
Video model research notes: Provision formats and production choices after 2025
Check Veo, Sora, Wan, and LTX in official documents and compare cloud provision, open weights, inference code, and safety measures on different axes.
Hailuo, Kling, Runway research notes: connect model capabilities to the production process
Focusing on the official announcement of MiniMax/Hailuo, Kling and Runway will limit the scope of public information they have been able to confirm and organize their decisions to incorporate short-form generation into production.
Y Combinator observation: turn public signals into constrained experiments
A dated reading of YC RFS and company material, separated from recommendations and adoption claims.