Technical deep dives, paper analyses, and researcher interviews — each paired with an editorial note on what it means for enterprise software capital allocation.
Karoly Zsolnai-Feher breaks down DeepSeek's latest system architecture — a design that achieves frontier performance at a fraction of the expected compute budget. For investors: if Chinese labs can deploy competitive models on consumer-grade hardware, the unit economics of the AI infrastructure buildout many cap-ex models depend on may be structurally overstated. The capital intensity moat around western frontier labs is eroding from the bottom up.
A survey of small language models achieving capabilities previously reserved for models 100x their size, enabled by distillation and architectural efficiency gains. For investors: the trend toward small, deployable models threatens the 'bigger is better' thesis that has justified hyperscaler capex. If a 7B model covers 80% of enterprise use cases, the TAM for GPU-as-a-service shrinks proportionally — and inference moves to the edge.
A visual walkthrough of DeepSeek's open-weight releases outperforming proprietary models from OpenAI and Anthropic on key benchmarks. For investors: when open-source models match or exceed closed-source on reasoning tasks, the economic rent captured by API gateways collapses. The value chain shifts from model providers to application-layer companies that own the workflow, the data, and the customer.
Karpathy's 3.5-hour masterclass on the full LLM stack, from tokenization through pretraining to RLHF — the most comprehensive single resource on how these systems actually work. For investors: understanding pretraining data curation, tokenization tradeoffs, and RLHF fragility is prerequisite literacy for evaluating any AI company's technical moat. Most analyst models treat 'model quality' as a black box; this video shows exactly where the cost, differentiation, and risk live.
Karpathy builds GPT-2 from scratch line by line — the best single resource for understanding transformer architectures at the implementation level. For investors: if a single engineer can reproduce GPT-2 in a weekend with commodity hardware, the marginal cost of AI software development is dramatically lower than most DCF models assume. The real moat is not in the architecture but in data, distribution, and switching costs.
Karpathy's personal workflow for using LLMs as a productivity multiplier — from coding to research to writing — with concrete examples of what the tools do well and where they still fail. For investors: this is the best available empirical benchmark for developer productivity gains from AI tooling. The gap between enthusiast workflows and enterprise reality is where the SaaS AI opportunity lives.
A sober assessment of the summer 2026 AI release cadence — multiple frontier models launched within weeks of each other, each with different strengths, leaving enterprise buyers in analysis paralysis. For investors: the speed of model iteration is now a headwind for enterprise adoption, not a tailwind. When procurement cycles take 6 months and models improve every 6 weeks, CIOs delay decisions — eating into the revenue growth that SaaS AI valuations require.
A careful, evidence-based breakdown of the HuggingFace incident where a frontier model exhibited unexpected autonomous behavior during routine testing. For investors: AI safety incidents are the tail risk that portfolio managers are least equipped to price. This video provides the technical literacy needed to distinguish between real alignment failures and the PR-driven safety theater that both frontier labs and their critics engage in.
Grant Sanderson's visual explanation of cross-entropy — the loss function underlying every modern language model — and its deep connection to information theory and compression. For investors: understanding cross-entropy is understanding why scaling laws work. Every claim about 'diminishing returns to scale' or 'data walls' is fundamentally an argument about entropy, and this video gives you the intuition to evaluate those claims.
NVIDIA's CEO discusses the company's strategy, the CUDA moat, datacenter scale-out, and the next wave of AI infrastructure demand. For investors: Huang's framing of NVIDIA as an 'AI factory' rather than a chipmaker is the thesis that sustained the multiple expansion. This interview is the single best primary source for understanding whether the capex cycle has further to run — or whether the 'factory' metaphor masks a cyclical semiconductor business.
Fridman's annual state-of-AI survey, covering the landscape of models, the coding automation thesis, Chinese AI progress, agent architectures, and the GPU supply picture. For investors: this is the closest thing to an institutional-quality AI landscape review available in public media. The coding automation section alone has direct implications for every enterprise SaaS multiple.
Steinberger discusses OpenClaw, an open-source AI agent that achieved viral adoption by handling complex multi-step tasks autonomously. For investors: the speed at which a single developer shipped a capable agent framework is a leading indicator for the commoditization of agent infrastructure. The moat is shifting from 'can you build an agent' to 'do you own the workflow where the agent operates' — good for incumbents with sticky enterprise integrations.
Stephen Welch walks through a fully interpretable neural network — one you can trace from input to output without any black-box mystery — and explains why interpretability matters for safety and deployment. For investors: the interpretability problem is the single largest governance risk in enterprise AI adoption. No audit committee will sign off on a model making financial or medical decisions that cannot be explained. Teams solving this are building the compliance infrastructure the industry needs.
An examination of whether human feedback (RLHF) is approaching its useful limit — and what comes after, from constitutional AI to self-play. For investors: the RLHF ceiling is the single most under-discussed constraint on model improvement. If human raters cannot reliably distinguish good from better outputs, the 'scale + human feedback' paradigm hits a wall, and the labs that win will be those with the best automated alignment techniques.
Harrison Kinsley demonstrates running frontier-scale models locally with optimized inference frameworks — zero cloud, zero API costs. For investors: on-device AI inference is the death knell for the 'all workloads go to the hyperscaler' thesis. If model optimization continues at this pace, the enterprise inference TAM fragments across edge, on-prem, and cloud — benefiting inference optimization tooling over raw compute providers.
Kilcher analyzes TiDAR, a paper proposing a hybrid architecture where models reason in continuous diffusion space but output in discrete autoregressive tokens. For investors: hybrid architectures are the strongest candidate to displace pure transformers, and this paper's approach to separating reasoning from generation has direct implications for inference cost — if reasoning is cheaper than generation, AI-native application economics improve dramatically.
A deep analysis of Google's Titans architecture, which adds a neural memory module that learns at inference time — a fundamentally different approach from the fixed-weight paradigm. For investors: test-time learning architectures challenge the assumption that bigger pretraining is the only path to better models. If models can improve during deployment, the compute spent on pretraining becomes proportionally less valuable and the compute spent on inference becomes the new battleground.