Selling AI
that matters.

From Sequences to Attention: The Architecture Decision

The transformer architecture (Vaswani et al., 2017) was not designed from first principles — it was designed to solve a specific failure mode of recurrent neural networks. RNNs process sequences step by step, maintaining a hidden state that is updated at each position. This sequential processing has a fundamental flaw: gradient signals from distant positions in the sequence must pass through many recurrent transitions, and they decay exponentially. Long-range dependencies — the kind that require understanding a reference 500 tokens ago to interpret the current token — are exactly what RNNs struggle with. Self-attention solves this by allowing every position to attend directly to every other position in a single operation.

The intuition behind attention is precise and worth stating carefully: given a query (what am I looking for?), keys (what does each position have to offer?), and values (what information does each position contain?), attention computes a weighted sum of values, where the weights are determined by the similarity between the query and each key. High similarity between a query and a key means that position's value is heavily weighted in the output. This allows the model to route information selectively across arbitrary distances in constant time rather than decaying time.

The Complete Transformer Block

A transformer is a stack of identical blocks, each consisting of multi-head attention, a feed-forward network, residual connections, and layer normalization. The design decisions in each component are not arbitrary — they address specific training stability and expressivity challenges.

Positional Encoding: From Sinusoidal to RoPE

Attention is permutation-equivariant: swap two tokens' positions and the output changes only in position, not in content. This means the model has no inherent notion of sequence order without explicit positional information. The original transformer used sinusoidal embeddings added to token embeddings. Modern models use Rotary Position Embeddings (RoPE), which encode position through rotation of query and key vectors rather than addition to embeddings — a design that generalizes better to sequence lengths beyond those seen in training.

Efficient Attention: Flash Attention and Beyond

The O(n²) attention cost is not primarily a compute problem — modern hardware (A100, H100) has enormous compute throughput. It is a memory bandwidth problem. The n×n attention matrix must be read and written to GPU high-bandwidth memory (HBM), and memory bandwidth is the actual bottleneck. Flash Attention (Dao et al., 2022) solves this by recomputing attention in blocks that fit in SRAM (on-chip cache), dramatically reducing HBM reads/writes while performing exactly the same mathematical computation.

Why Scale Works: The Empirical Discovery

By 2020, the empirical evidence was undeniable: larger language models, trained on more data with more compute, were reliably better across nearly every benchmark. The question was whether this was an accident of the specific models being trained, or a fundamental regularity of the learning problem. Kaplan et al. (2020) at OpenAI provided the answer: language model loss follows clean power laws in model size, dataset size, and compute — laws that held across five orders of magnitude with remarkable stability. This was not expected from theory. It transformed how the field thought about AI development.

But power laws alone don't tell you how to allocate a fixed compute budget. If you have C FLOPs to spend, should you train a large model on less data, or a smaller model on more data? The original Kaplan scaling laws suggested the answer was to prioritize model size — the exponent on model size was larger than on dataset size. This turned out to be wrong, or more precisely, right for a specific regime that doesn't reflect real deployment conditions.

Beyond Compute-Optimal: The Inference-Optimal Regime

Chinchilla describes compute-optimal training — minimizing loss for a fixed training compute budget. But it is not the right objective for production deployment. A company deploying a language model serves millions of requests per day. The inference cost — which scales with model size — dominates total cost of ownership, not the one-time training cost. This shifts the optimal point: you want the smallest model that achieves the target capability level, which means training smaller models on far more data than Chinchilla recommends.

This is the insight behind LLaMA 3 8B training on 15 trillion tokens — 1875 tokens per parameter, roughly 100x the Chinchilla optimal ratio. The model is not compute-optimal from a training perspective. It is inference-optimal: it achieves competitive capability in 8B parameters because it has seen so much data, and it costs much less to serve than a 70B Chinchilla-optimal equivalent. The scaling law hasn't changed — the objective function has.

Emergent Capabilities: Phase Transitions at Scale

Perhaps the most consequential and most debated empirical observation in modern AI is emergent capabilities — abilities that appear to materialize abruptly above a critical model scale, with little or no signal from smaller models. Wei et al. (2022) documented dozens of examples: arithmetic, chain-of-thought reasoning, analogical reasoning, and others that appeared essentially random at smaller scales and then jumped to meaningful performance above a threshold. If true in the strong sense, emergence would mean that capability cannot be predicted by extrapolating from smaller models — a profound challenge for safety and planning.

Grokking: What Delayed Generalization Reveals About Learning

Power et al. (2022) discovered a phenomenon they named "grokking": neural networks trained on small algorithmic tasks first memorize the training set perfectly — achieving zero training loss while test accuracy remains near random — and then, after many more gradient steps, suddenly generalize. The training loss was already zero. The generalization came later. This delayed generalization seems to violate the standard understanding of learning, where we expect generalization to improve as training loss decreases. It reveals something fundamental about how neural networks actually learn.

The Alignment Problem: What We're Actually Trying to Solve

A language model trained by next-token prediction on internet text learns to predict internet text. It does not learn to be helpful, honest, or harmless — because internet text is not consistently any of those things. The pretraining objective and the deployment objective are misaligned. Alignment research is the field of techniques for closing this gap: taking a model that is a highly capable predictor and turning it into a model that pursues intended goals in intended ways. This is easy to state and hard to solve, and the difficulty is not engineering — it is conceptual. We don't have a complete theory of what "intended goals in intended ways" means mathematically, which makes it difficult to know whether our alignment techniques are working, whether they're robust, and what they fail to capture.

The practical alignment pipeline used at frontier labs consists of three stages: supervised fine-tuning on high-quality demonstrations, reward model training from human preference comparisons, and reinforcement learning with KL regularization toward the reference model. This pipeline works — it produces dramatically more useful and safer models than raw pretraining. But it also has specific failure modes that the theory predicts and empirical research confirms, and understanding them is necessary to understand both the current capabilities and current limitations of aligned AI systems.

DPO: The Elegant Simplification

Rafailov et al. (2023) observed something remarkable: the optimal RLHF policy can be expressed in closed form in terms of the policy and the reference model, without ever computing the reward model explicitly. This means the reward model training step can be bypassed entirely — you can go directly from preference data to a trained policy using a single supervised loss. The key insight is that the partition function Z(x), which appears to make the optimal policy intractable, cancels when you compute the preference probability under the optimal policy.

Reward Hacking and Goodhart's Law: The Formal Theory

Goodhart's Law states: when a measure becomes a target, it ceases to be a good measure. In the RLHF context, this means: when we optimize a learned proxy reward r_phi rather than the true human preference function r*, the policy will find ways to maximize r_phi that do not maximize r*. This is not a hypothetical concern — it is observed empirically in every RLHF deployment, typically manifesting as sycophancy (telling users what they want to hear), verbosity (longer responses score higher), and format gaming (bullet points, headers score higher even when prose would be better).

The formal theory of reward hacking, developed by Skalse et al. (2022) and others, demonstrates that this is not a failure of implementation but a mathematical consequence of the setup: any proxy reward that is not identical to the true reward will, under sufficient optimization pressure, produce behavior that diverges from the true reward.

Mechanistic Interpretability: Opening the Black Box

Mechanistic interpretability is the research program of understanding, at a precise mathematical level, what computations neural networks are actually performing. The broader alignment goal is clear: if we can understand what a model is doing, we can verify whether it is doing what we intend, identify when it might be deceptive, and potentially modify specific behaviors without breaking others. The current state of the field is promising but limited — we have detailed mechanistic understanding of small toy models and specific circuits in larger models, but no complete account of any frontier model's behavior.

The superposition hypothesis (Elhage et al., 2022) is one of the most important findings in mechanistic interpretability. It proposes that neural networks represent more features than they have dimensions, by encoding features as nearly-orthogonal directions in a lower-dimensional space. This creates interference between features but allows exponentially more representations than a strict one-feature-one-neuron mapping would allow.

Constitutional AI and Scalable Oversight

The practical challenge of alignment at scale is the human feedback bottleneck. RLHF requires humans to provide preference comparisons, but the number of comparisons needed grows with model capability — a more capable model can produce subtly problematic outputs that require more expert evaluation to identify. At some capability level, human evaluators can no longer reliably distinguish good from bad behavior, making the RLHF signal uninformative. Scalable oversight research addresses this: how do you maintain alignment signal as model capabilities exceed human ability to evaluate them directly?

Constitutional AI (Bai et al., 2022) is Anthropic's approach: instead of relying on human comparisons of full outputs, train the model to critique its own outputs against a set of principles (the constitution), revise based on the critique, and use the revised outputs as training data. The key insight is that critiquing is easier than generating — models can identify problems in outputs they couldn't reliably avoid generating. This allows the alignment signal to scale with the model's ability to understand its own outputs.

MEDDPICC for AI: Complete Field Guide

MEDDPICC is not a qualification framework. It is a deal autopsy tool that you run prospectively rather than retrospectively. Every element represents a category of information asymmetry — something the buyer knows that you don't, or something neither party has made explicit — that will become a deal-killer if it remains unresolved. The framework's operational value is not in identifying good deals from bad deals early. It is in identifying the specific dimension that will kill an otherwise good deal, so you can address it before it does.

For AI products specifically, every element of MEDDPICC has AI-specific failure modes that don't appear in SaaS deal analysis. The M of Metrics is more complex because AI ROI is often indirect, lagged, and hard to attribute. The first D of Decision Criteria includes technical thresholds (accuracy, latency) that don't exist in SaaS evaluations. The P of Paper Process includes AI-specific legal instruments (DPA with model training clauses, liability frameworks for AI errors) that procurement teams have rarely handled before. Understanding these AI-specific variants is what separates a MEDDPICC practitioner who closes SaaS deals from one who closes AI deals.

AI Pricing Architecture: Complete Framework

Pricing an AI product is one of the most consequential GTM decisions a company makes, and most AI companies get it wrong in ways that create long-term revenue ceiling problems. The fundamental challenge is that AI products create value through mechanisms that don't map cleanly onto any of the existing SaaS pricing models. Value is consumption-based, but unpredictably so. It improves non-linearly over time, making initial pricing commitments difficult to set fairly. It is often diffuse across an organization, making seat-based attribution inaccurate. And it has a floor of integration cost that means small contracts are often unprofitable regardless of unit price.

Enterprise AI Outbound: The Complete Motion

Outbound for enterprise AI requires a different sequence, different signals, and different messaging architecture than outbound for SaaS or for mid-market AI. The reason is structural: the buyers you're targeting are sophisticated enough to have already evaluated several AI vendors, experienced enough to be skeptical of performance claims, and senior enough to delete generic outreach without a second thought. The bar for relevance is high, and the punishment for irrelevance is permanent — a VP Engineering who receives a poorly researched cold email about AI has updated their mental model about you and your product, and that update is negative.

How AI Deals Die: Complete Failure Mode Taxonomy

Enterprise AI deals fail in specific, predictable ways. After enough deal postmortems, the patterns become so clear that you can identify the failure mode in week 3 of a 16-week deal cycle — long before it kills the deal. The taxonomy below is not a list of things that might go wrong. It is a list of things that go wrong systematically, in specific sequences, with specific warning signals that appear weeks before the deal dies. The value is not in recognizing them in retrospect. The value is in catching them early enough to change course.

Competitive Strategy: Build vs. Buy vs. Foundation Model

The competitive landscape for enterprise AI products has three categories that are fundamentally different from each other, and requiring different responses. Direct competitors (other AI vendors solving the same problem) are the easiest to handle — you have experience positioning against specific alternatives. Foundation model DIY (why not just use GPT-4 API?) is the most common and most mishandled. Internal build (our engineering team can do this) requires the deepest understanding of your prospect's organizational economics.

Discovery Excellence: The Complete Field Guide

Discovery is the highest-leverage activity in enterprise AI sales, and it is systematically underdeveloped in AI GTM teams. The reasons are understandable: the product is technically sophisticated and exciting, the temptation to demo early is powerful, and many AI salespeople come from technical backgrounds where showing the system is more natural than asking about the problem. The result is a pattern of demos to the wrong people at the wrong stage, followed by POCs designed around technical criteria, followed by deals that die in procurement because no one built the business case.

Great discovery in AI deals is fundamentally different from great discovery in SaaS deals. In SaaS, discovery is primarily about qualification: does this prospect fit the ICP, do they have budget, can we close them in our timeline? The information gathered is largely directional. In AI deals, discovery is the foundation on which the entire commercial structure is built. The baseline measurement comes from discovery. The success criteria come from discovery. The stakeholder map comes from discovery. The ROI model comes from discovery. The POC scope comes from discovery. A 2-week discovery investment that produces all of these changes the entire economic structure of the deal that follows.

The Technical Buyer: Complete Engagement Playbook

The CTO or VP Engineering is the most consequential stakeholder in enterprise AI deals. They are also the most commonly mis-engaged. The mistake is treating them as a gatekeeper to be satisfied rather than a peer to be converted. A CTO who is satisfied is a neutral stakeholder. A CTO who is genuinely impressed by your technical honesty is an active advocate. The difference between these two outcomes is entirely in how you engage them.

CTOs evaluate AI vendors on a different axis than the rest of the buying committee. They are not evaluating whether the product works — that's what the POC is for. They are evaluating whether the vendor team is the kind of team they want to be in a long-term technical relationship with. The specific signals they're looking for: do you understand the actual failure modes of your system? Do you have a realistic model of what production deployment requires? Do you have opinions about architecture that go beyond what your product does? Have you seen enough deployments to know what goes wrong that you didn't plan for?

NRR Architecture: Building for Expansion from Day One

Net Revenue Retention is the metric that determines whether you have a sustainable AI business, and it is a design problem, not a retention problem. NRR is not something you improve by adding customer success headcount. It is something you design into the initial commercial structure, the POC scope, the success criteria, and the deployment architecture — and then execute on systematically. Companies that treat NRR as a post-sales problem start too late. The decisions that determine Year 2 NRR are made in the first 6 weeks of the customer relationship.

The economic difference between 90% and 130% NRR compounds over a 5-year period in ways that determine whether you build a large business or a medium one. At 90% NRR with 100% ARR growth from new logos: Year 5 ARR is 5.9x Year 1. At 130% NRR with the same new logo growth: Year 5 ARR is 17.4x Year 1. This is not a linear difference. The NRR compounding effect is the reason investors pay 10-20x higher revenue multiples for companies with 130%+ NRR than for companies at 90%.

AI ICP Architecture: The Four-Layer Framework

The ICP for an AI product is more consequential and more specific than for any other category of enterprise software, because variance in outcomes across customers is dramatically higher. A SaaS CRM works reasonably well for any company in the right size range — the UX might be better or worse, the adoption might be higher or lower, but the core functionality doesn't fail. An AI product that achieves 93% accuracy on one customer's data distribution might achieve 67% on another's. And 67% might be worse than their current manual process, which means you've created a churned customer who is actively telling others your product doesn't work. Getting the ICP wrong in AI doesn't just mean lower close rates. It means closed deals that churn, negative references, and compounding damage to your ability to grow in the affected market segment.

This is why the AI ICP has layers that SaaS ICP doesn't. The standard firmographic and technographic layers are necessary but nowhere near sufficient. You need data quality assessment (does the customer have data that supports AI at your product's capability level?), problem specificity assessment (is the problem narrow and well-defined enough for current AI to solve reliably?), and organizational readiness assessment (does the company have the operational capacity and change management capability to deploy and maintain AI in production?). Closing a customer who fails on any of these three dimensions is worse than not closing them at all.

Positioning Against Foundation Models: The Complete Playbook

The foundation model positioning challenge is the defining competitive question for every AI product company, and it will remain so for the foreseeable future. As GPT-4, Claude, and Gemini continue to improve, the honest answer to "why not just use the API?" becomes a more nuanced argument — not "we're more capable" but "we're more deployable, more reliable in production, more compliant, and less expensive at scale." Understanding this positioning precisely — not as a generic claim but as a specific, verifiable, audience-segmented argument — is a core competency for AI GTM teams.

PLG + Enterprise Sales: The Hybrid Motion That Wins

The most efficient AI GTM motion combines product-led growth at the bottom of the organization with a disciplined enterprise sales process at the top. This is not a compromise between two philosophies — it is a recognition that different stakeholders in the buying committee are best reached by different mechanisms. Developers and data scientists who discover the product through free tier or open source are the best possible technical champions — they have already run their own informal POC, on their own data, with their own edge cases. The enterprise sales motion doesn't need to convince them. It needs to convert their conviction into a commercial agreement with the people who control budget.

The failure mode of PLG without enterprise sales: developers love the product, use it extensively, bring it to their manager, manager asks for enterprise terms, nobody at the vendor has the capacity to navigate enterprise procurement, deal dies in procurement. This is the "build love without building revenue" trap that many developer-first AI companies fall into. The failure mode of enterprise sales without PLG: high CAC, long cycles, technical evaluation always required because no one in the account has used the product, POC-heavy motion that is expensive to run at scale.