Logan Matthew Napolitano

If you serve on a board of directors and believe your company is burning hundreds of thousands—or millions—of dollars on AI APIs and unnecessary compute, let’s have a conversation. I think your concern may be justified. I will help you show the board exactly where the money is going, what the company actually receives for it, and whether continuing with the same architecture is a responsible fiduciary decision. Too many companies have mistaken access to an AI API for an AI strategy. They are paying repeatedly for models they do not own, sending the same institutional knowledge across an external boundary, retransmitting enormous contexts, absorbing unpredictable price changes, and building critical operations around providers whose models, limits, policies, and product priorities can change at any time. The cost looks small during experimentation. Then the system enters production. Every employee begins using it. Agents make multiple calls. Failed calls retry. Context windows expand. Retrieval systems resend documents. Workflows use expensive models for tasks that smaller models could perform. A single business action may quietly require dozens of model calls. The organization pays for every token, every retry, every repeated instruction, and every piece of context the model must reread. At scale, this can become an enormous recurring expense without creating a corresponding company-owned asset. Meanwhile, local AI has changed dramatically. Open models are improving quickly. Quantization, sparse mixture-of-experts architectures, speculative decoding, optimized inference engines, CPU execution, system-memory inference, and modern agent harnesses have made serious private deployments practical. People are already reporting GLM 5.3 running on CPU-based local systems at roughly 50 tokens per second. That should end the assumption that useful local AI always requires renting access to a hyperscaler’s infrastructure. But cost is only one part of the problem. The privacy is often inadequate. The company’s prompts, customer information, internal research, source material, operational memory, and decision history may repeatedly leave its controlled environment. Even where providers offer strong contractual safeguards, the organization still accepts another dependency and another security boundary. The institutional knowledge is also being built in the wrong place. Years of prompts, corrections, evaluations, workflow design, retrieved documents, agent memory, and employee feedback should become intellectual infrastructure owned by the company. If the organization cannot move that capability away from its provider without rebuilding everything, it has created dependence rather than an asset. Quality is another problem. Sending every task to a famous model does not guarantee that the task is being completed well. Models hallucinate. Agents loop. Tools are called incorrectly. Retrieval supplies irrelevant context. Long prompts bury the real instruction. Large models are used where smaller specialized models would be faster and more reliable. Companies often measure API availability while failing to measure completed-task accuracy, unsupported claims, retry rates, human correction time, latency, or cost per successful outcome. The best modern harnesses can make local models reliable agents by combining: model routing; structured tool access; retrieval; durable memory; constrained permissions; answer verification; uncertainty detection; hidden-state monitoring; specialist adapters; human escalation; and continuous evaluation. The model is only one part of the system. The harness determines what it can do. The memory system determines what it retains. The evaluation layer determines whether the company recognizes failure. The routing system determines whether a task needs a small local model, a larger private model, or an external frontier API. The monitoring layer determines whether the company understands what the system is doing before a bad answer reaches a customer. This does not mean every company should abandon external APIs. It means every external call should earn its cost. A responsible architecture may be entirely local, on-premise, privately hosted, or hybrid. External frontier models can remain available for the narrow tasks where measured performance justifies the added cost and exposure. The rest can run on infrastructure the company owns and controls. A board evaluating its AI strategy should demand clear answers: What is our total AI cost by workflow? What is the cost per successful completed task? How much spending comes from retries, oversized contexts, and unnecessary agent calls? Which tasks demonstrably require a frontier model? Which tasks could run locally without reducing practical quality? What happens to our costs if usage increases tenfold? Where are prompts, memory, customer information, and proprietary research stored? Can we change providers without rebuilding our operating system? Have we benchmarked a local or hybrid architecture using our actual work? Are we accumulating permanent intellectual infrastructure, or paying rent indefinitely? If management cannot answer those questions, the organization does not yet have a serious AI strategy. Proprioceptive AI has spent years conducting independent AI research across model internals, hidden states, interpretability, probes, adapters, hallucination reduction, model evaluation, memory, agent control, and controlled steering. We do more than recommend a different model. We investigate what the system is doing internally, identify where it becomes unreliable, measure failure before it reaches the user, and design practical local or hybrid systems around the organization’s actual requirements. We can help a company determine: which workloads should remain on commercial APIs; which workloads should move locally; which models fit each workflow; where the current system is wasting compute; where hallucinations or unstable answers originate; how private memory and retrieval should be structured; how agents should be constrained and monitored; and how to turn AI spending into infrastructure the company owns. If you believe your company is burning capital on AI without building a durable capability, you may be right. Let’s measure it, put the evidence in front of the board, and determine the responsible path forward. I'll talk to your board or anyone who you think needs to hear this. Book a conversation: calendly.com/logan-proprioce…

In Proprioceptive Transformer v2.8, a striking result emerges from predicting a 12-dimensional internal coordinate state using relational model-native features rather than raw hidden vectors. The H3 test assessed prediction quality by mean R², where the FULL_12 coordinate set achieved 0.5128, while the top two principal directions-extracted via PCA on the 12 coordinates-scored slightly higher at 0.5199. This means that using only two principal directions outperformed the entire 12-dimensional map in predicting future internal states. Why is this surprising? Intuitively, more dimensions should capture more predictive information or at least perform equivalently. The top two directions capture 101.4% of the FULL_12 R², which quantifies prediction accuracy on the same rank basis, not raw information content. Their complement space-the other 10 directions-had a negative predictive R² (-0.0816) and a matched-rank random baseline percentile at 0.3511, highlighting that those directions contribute little or noise-like signals. Mathematically, this means the relational representation's geometry concentrates predictive power in the subspace spanned by two principal axes. Removing other directions effectively filters out irrelevant or confounding variations that degrade prediction. Practically, this suggests a low-dimensional subspace captures nearly all actionable predictive patterns of the 12 relational coordinates. One can design an evaluation to test this principle: freeze preprocessing to project internal states onto principal directions; then predict future states with only top K directions. If R² improves or remains stable near full dimensions, it verifies dominant low-rank structure. Limitations arise because this result is from a single pair and passive prediction, so it may not generalize fully to all inputs or cohorts (earlier groups showed different levels). Still, the finding is a clear example that less can be more in structured representation spaces (report 'From Internal Prediction to Measured Control,' section 8.2).

QLoRA introduces a way to finetune extremely large language models by freezing the original model's weights in a 4-bit quantized format and training only small, low-rank adapters alongside it. This approach drastically cuts down memory requirements by combining three clever techniques: NF4 quantization to compactly represent weights, double quantization to reduce the memory footprint further, and paged optimizers that manage gradient storage efficiently. In their 2023 paper, Dettmers and colleagues demonstrated finetuning a 65 billion parameter model using a single GPU with 48GB of memory. Usually, training models of this size demands massive hardware clusters with hundreds of gigabytes of memory. By contrast, QLoRA's method fits this enormous scale into a single standard GPU by freezing the base model and adding only small trainable adapters that depend on a low-rank decomposition. Mathematically, freezing the base model means keeping weights \(W \in \mathbb{R}^{d \times k}\) fixed in quantized form while learning an adapter \(A B^T\), where \(A \in \mathbb{R}^{d \times r}\), \(B \in \mathbb{R}^{k \times r}\), and \(r \ll d,k\). This factorized low-rank update requires far fewer parameters and updates. The NF4 format and double quantization compacts \(W\)'s memory to around 4 bits per weight, while paged optimizers reduce optimizer state memory. For users limited by GPU memory, QLoRA offers a new cost-effective route to finetuning massive models without compromising base weight fidelity. However, the reported 65B model finetuning on a 48GB GPU depends on specific hardware, model structure, and context length; it is not a universal guarantee. Benchmarking is essential to see how these techniques scale on your target model and task. Understanding these tradeoffs is critical for anyone seeking to improve large model adaptation under tight memory - QLoRA shifts the bottleneck to adapter capacity and training strategy while preserving an enormous frozen base. This work bridges the gap between massive model performance and practical resource constraints in finetuning (Dettmers et al., 2023).

A direct intervention in layer 24 of Hermes-3-Llama-3.1-8B's neural activity lowered downstream internal signal error by 21.68 to 43.79% across eight registered neurons, as reported in the abstract of 'From Internal Prediction to Measured Control.' This reduction is a quantitative change in the network's internal state consistency, not an improvement in behavioral outputs or a demonstration of safety. The experiment involved a 'precision-qualified write' - a controlled, exact adjustment injected into the model's penultimate transformer block. By altering layer 24's activation rather than passively predicting future states, this intervention offers causal evidence of functional influence within the network's hierarchy. Mathematically, if x_{24} represents the original layer 24 activations and x_{24}' the perturbed activations, the downstream coordinates y show a decreased signed difference ||y - y_{target}||_signed post-intervention. This finding reveals that internal errors previously measured are not just epiphenomenal noise but can be actively reduced by manipulating a highly specific depth of the model. It underscores that layer 24 contains a representation critical for maintaining internal consistency across the neural circuit. However, the effect size variability (21.68 to 43.79%) across eight cells indicates a nuanced influence, not uniform control. Crucially, this result stands apart from the eight-coordinate predictor studies described elsewhere, which only passively forecasted future states. Here, an actual control signal modifies internal responses, a necessary but insufficient step toward full AI behavioral control. The scope of this intervention study remains narrow: it evidences the feasibility of intervening at a precise network depth for internal error correction but does not validate end-to-end behavioral improvements or general safety guarantees. The precise measurements of intervention impact, rather than simulated or dose-responsive effects, enhance the credibility of the causal link. The result from section 'From Internal Prediction to Measured Control, abstract' thus establishes an experimental foundation for measured control inside transformer layers beyond passive prediction.

When predicting a 12-dimensional internal coordinate state in Proprioceptive Transformer v2.8, freezing the top-two principal directions alone retained nearly all predictive performance. This finding, from the 'H3 complement test' section of the 2026 Proprioceptive AI report, shows that critical predictive information concentrates into very specific relational directions within the model's internal representation geometry. The test splits the model's 12-dimensional coordinate space into two parts: the 'principal' subspace spanned by the two dominant frozen principal components, and its orthogonal 'complement.' Performance was measured separately on these parts. Concretely, the two principal directions supported nearly full predictive accuracy, whereas their complement subspace performed below zero, meaning worse than a baseline guess. Similarly, matched random subspaces of equal rank fared worse than the principal directions. Mathematically, if \(P\) denotes projection onto the principal subspace and \(C = I - P\) its complement, predictive quality measured after projecting target coordinates onto \(P\) stayed high, while projection onto \(C\) collapsed. This suggests the model's future-state prediction concentrates in low-dimensional relational directions rather than spreading uniformly across coordinate axes. For practitioners, this complement test is a practical approach to assess where predictive power lies inside hidden representations without needing causal intervention. The recipe is: 1) freeze principal components on held-out data, 2) project target coordinates onto principal and complement subspaces, 3) separately measure prediction accuracy on these projections, 4) verify if the complement collapses below baseline, and 5) test whether random subspace projections perform worse. Failure of this test-complement not collapsing-would indicate prediction spread or a misidentified subspace, warning against simplistic low-rank assumptions. However, it does not prove causal control in those directions, only highlights where predictive signal clusters across internal states. This focused subspace analysis deepens our geometrical understanding of learned representations in transformers predicting proprioceptive states, as detailed in 'Proprioceptive Transformer v2.8, H3 complement test.'

Eight internal coordinates predicting a 68-coordinate future state. That is one of the results in my report, From Internal Prediction to Measured Control. On Hermes-3-Llama-3.1-8B, the eight-coordinate predictor reduced standardized prediction loss by 68.68–76.15% relative to a training-future-mean baseline. The target here is the model’s future internal measurements. We are measuring structure inside the computation and testing how much of its subsequent evolution a compact readout can predict. That changes the engineering question. Instead of asking only what answer a model produced, we can investigate which internal measurements tell us where its computation is heading. The report also describes a separate intervention experiment: a precision-qualified write at layer 24 reduced signed downstream internal error by 21.68–43.79% across eight registered cells. These experiments address two concrete capabilities: • Reading internal state to predict a defined future target. • Writing into internal state and measuring the downstream response. The distinction matters technically. A useful readout identifies predictive structure. An intervention tests what changes when we act inside the computation. This is the hidden-state technology I’m developing: compact internal measurements, explicit prediction targets, and precisely measured interventions. The numbers refer to internal prediction loss and downstream internal error. They give us concrete quantities to optimize as we develop more capable ways to observe and influence model computation.

Assessing whether an AI system truly controls behavior, versus merely predicting internal states, demands distinct experimental proofs rather than a single integrated demonstration. The 2026 Proprioceptive AI report 'From Internal Prediction to Measured Control' explains this with clarity. It aggregates evidence from three separately executed research tracks: internal-state prediction studies, measured hidden-state interventions, and behavioral steering experiments. Each track isolates a different proof obligation. Internal prediction shows a model's ability to anticipate hidden variables from observed data. Hidden-state interventions test whether targeted perturbations in latent variables produce anticipated changes internally. Behavioral steering then examines if manipulating these internal states systematically drives external actions. The report further supplements these with studies on model repairs, harms, autonomous policy selection without answers, and adverse control effects using Hermes, Qwen3.5, and Qwen2.5. Critically, the report highlights that these tracks are not yet combined into one historically validated end-to-end controller-i.e., a system where internal prediction reliably leads to measured, manipulable control. This gap points to the necessity of independent validation steps, ensuring confidence at each stage before assuming the entire loop is robust. For a research program, this layering means separate proof obligations must be fulfilled before concluding true control from prediction. Skipping this risks over-interpreting prediction as actionable control. While prediction is necessary, control requires additional causal linkage and manipulation evidence. A key limitation is that full-model recapture-reproducing control by re-deriving all internal states and their effects-remains an open challenge. Until resolved, it's prudent to treat internal predictions and measured controls as complementary but distinct milestones rather than a single achievement. This nuanced posture refines how AI safety and interpretability research designs validation pipelines, stressing rigorous stepwise proof rather than premature synthesis. Reference: 'From Internal Prediction to Measured Control, evidence posture' in the 2026 Proprioceptive AI report.

A core bottleneck in transformer inference is the sequential token-by-token verification that ensures output fidelity but incurs latency. Leviathan et al. (2023) propose speculative decoding, where a cheaper draft model proposes multiple tokens in parallel, and the exact target model verifies them simultaneously. This preserves the target model's distribution exactly, though the sampled sequences may differ from standard sampling. Specifically, the draft generates a candidate sequence quickly, and the target model checks token correctness up to the longest accepted prefix. If the draft errs, the verification corrects the sequence by reverting to the target's own generation from the divergence point onward. This avoids waiting for token-level acceptance in strict sequence and leverages parallelism. In their experiments with T5-XXL implemented via T5X, the acceptance rate-the fraction of tokens accepted without correction-controls speedup versus overhead. High acceptance yields significant latency gains because fewer tokens need exact, sequential generation. But when corrections dominate, the overhead of verification negates benefits. For practitioners employing local or distantly related draft models, the key to effective speedups is benchmarking acceptance and pipeline overhead on their data and hardware. The exact token-level acceptance rate depends heavily on the draft model's quality and the chosen threshold policies, making theoretical speedup predictions context-dependent. This mechanism works by replacing rigid sequential sampling with a two-step parallel proposal-and-validation process, reducing local latency when the draft is reliable. However, gains are not universal-costs of verification and corrections can offset runtimes if the draft diverges frequently. Thus, practical evaluation best guides deployment decisions. This insight from 'Fast Inference from Transformers via Speculative Decoding' demonstrates that local-model users can improve latency by tuning draft acceptance and carefully measuring overhead versus benefit, rather than expecting deterministic speedups from any draft-target pairing. Context: an explainer based on the Fast Inference from Transformers via Speculative Decoding (Leviathan et al., 2023) report.

Can a relationship learned between two models survive a genuinely held-out test without being adjusted? That is the useful question behind the frozen-transfer experiment in my Proprioceptive Transformer v2.8 report. The map was fitted on 30 exploratory roots, then frozen before evaluation on 30 preregistered, untouched roots. The test did not permit refitting the centering, scaling, rank, regularization, or feature and target choices. Those restrictions matter: otherwise, adapting the pipeline to the test set can blur the distinction between transfer and a second fitting exercise. Across six geometric prediction targets, the reported mean R-squared was 0.823, with all six scores positive. R-squared compares squared prediction error with a mean-based baseline; it is not classification accuracy. The result concerns the measured relational telemetry, not recovery of the models' complete hidden states. The interesting part is the combination of held-out roots and a fixed mapping. Within this experiment, the measured relationship retained predictive value after the fitting stage ended. That gives us a concrete reason to investigate which relationships persist across models, rather than assuming their internal coordinates must match directly. The boundary is equally concrete: this was one registered Hermes-to-Llama pair. It does not establish transfer to arbitrary architectures, causal control, or generalization to every distribution shift. The next demanding test is independent replication across additional model pairs with the same separation between fitting and evaluation.

Imagine evaluating a prediction system and seeing an R-squared value of -0.2. At first glance, it might seem like a minor misstep, but this negative R-squared reveals something critical: the system predicts worse than simply guessing the average outcome every time. R-squared, which stands for the coefficient of determination, gauges prediction quality by comparing the squared prediction error against a simple baseline - usually the mean of the observed data. Formally, R² = 1 - (SSE / SST), where SSE (sum of squared errors) measures how far predictions stray from actual values, and SST (total sum of squares) measures how far true values spread around the mean. If predictions exactly match the mean baseline, SSE equals SST, yielding an R² of zero. This baseline anchors our expectations. Without it, high error might seem impressive when, in fact, even trivial guesses perform better. Negative R-squared values expose predictions that fail to improve over this baseline. But R² isn't a measure of accuracy as a percentage, nor does it imply any causality-only how much better predictions explain variance compared to that baseline. Including an explicit baseline in predictive studies is essential to interpret model performance meaningfully. It prevents overestimating the model's skill and highlights potential shortcomings. Yet, the baseline must be chosen carefully to suit the problem context; otherwise, R-squared comparisons become misleading. In short, R-squared's dependence on a baseline average is not a limitation but a vital reference frame. It anchors our interpretation of error, guiding us away from illusionary success toward genuine progress in prediction. Context: an explainer based on the Mathematical context for predictive studies report.