eBay doubled engineering productivity and moved all teams from 35th to 75th percentile on DORA metrics, improving deployment frequency 10x (10 days to 1-2 days) and lead time for change 5x.
Velocity initiative succeeded through executing standard DevOps playbook: removing bottlenecks, automating build/test/deploy pipelines, streamlining code reviews, eliminating "partner signoffs" blocking service upgrades.
Mobile app release cycles improved from monthly to weekly in 7 months by testing small changes incrementally rather than large coordinated releases, proving skeptics wrong through psychological safety.
Strategy gap: eBay competed broadly across all categories instead of becoming dominant in specific categories like competitors did (Vinted for fashion, Reverb for instruments).
Learned helplessness and risk aversion developed from 15 years of flat GMV growth (1.5x since 2007 vs. 8x for US e-commerce), making seller-community management changes perceived as constant disruption.
Centralized waterfall planning required executive approval for any initiative, forcing smaller projects to attach as "riders" to larger ones, blocking rapid experimentation and market responsiveness.
Technology debt: eBay invested in proprietary systems (custom Kubernetes fork, Hadoop data warehouse, proprietary JavaScript/mobile frameworks, custom OpenStack) that fell out of sync with industry standards.
Pathological organizational culture of fear, empire building, and "not invented here" syndrome prevented adoption of industry-standard approaches; acknowledgment of failure seen as threatening.
Feature factory mentality rewarded milestones and effort ("5,000 train seats" of work) over customer outcomes, disconnecting delivery from business value and revenue impact.
Transformation requires top-down executive support AND bottom-up engineer buy-in, but the critical missing piece is middle-out: laterally engaging peer leaders to build excitement and route around resistance.
Randy Shoup details how eBay's 2020-2022 velocity initiative moved 5,000 engineers from 35th to 75th percentile on DORA metrics, yet eBay's GMV grew only 1.5x over 15 years vs. 8x for US e-commerce. He attributes the gap to waterfall planning, innovator's dilemma, and a pathological culture of fear and empire-building that technical wins alone couldn't fix.
AI at its core uses only multiplication, addition, and activation functions—simple math repeated at massive scale across billions of parameters.
Vectors are numbers arranged vertically; matrices are arrays of arrays. Neural networks transform input vectors through hidden layers via matrix multiplication and addition.
Floating-point numbers store sign, exponent, and fraction bits (32-bit format) to represent decimals in binary—e.g., pi requires 50 terabytes to store 100 trillion known digits.
LLMs output raw logits from hidden layers, converted to probability distributions via softmax function to predict next token with each input.
Gradient descent trains models by computing mean squared error (loss), measuring how weight changes affect loss, then iteratively adjusting weights to minimize error.
GPT-3 has 175 billion parameters requiring ~700GB memory; quantization compresses 32-bit floats to 8-bit integers (Q8) for 75% size reduction with minimal accuracy loss.
GPUs parallelize vector operations efficiently because graphics processing optimizes X, Y, Z axis computation; TPUs handle tensor (multi-dimensional array) operations via matrix multiplication.
FLOPs (floating-point operations) ~2× parameter count; GPT-3 with 50-token input/100-token output requires ~52 trillion operations per request—explains GPU necessity.
Carbon footprint research accounting for output quality (not just energy) shows selective model choice and human-AI comparison matters; most published studies ignore quality metrics.
Binary underpins all computing; AI mythology (memory, personality, consciousness) obscures reality that LLMs compute probabilities through repetition of elementary arithmetic operations.
Walks through how transformers work from binary storage through weighted sums, ReLU activations, and softmax probability outputs. Covers gradient descent, MSE loss, and why GPT-3's 175B parameters demand 700GB RAM—then explains how int8 quantization and GPU parallelism make deployment practical.
Physics-based digital twins using position-based dynamics enable surgical robots to achieve sub-2mm accuracy without massive training datasets by continuously matching simulation to real-time video observations.
Differentiable rendering iteratively corrects physics simulation parameters (tissue mechanics, elasticity, viscosity) by backpropagating loss between camera observations and simulated predictions.
Safety-aware control with Bayesian uncertainty quantification allows robots to autonomously probe tissue connections while respecting tearing energy thresholds before dissection.
Model predictive control with explicit physics simulators enables reliable task recovery—robots can identify failed cuts and target corrective actions, unlike unpredictable foundation model recovery behaviors.
Knowledge-grounded reinforcement learning combines hand-engineered behaviors (scanning, cutting, grasping) with sparse neural networks that learn task sequencing autonomously without human specification.
Humanoid robots with directional haptic feedback gloves outperform teleoperated da Vinci systems on laparoscopic tasks despite lower absolute performance, enabling diverse hospital roles (surgery, nursing, imaging).
Reinforcement learning enables humanoid robots to independently learn how to grasp articulated instruments (forceps, pliers, laparoscopic tools) without imitation, bypassing human hand dissimilarities.
Four foundational pillars—perception, modeling/simulation, planning, and control—assembled in different configurations scale surgical autonomy while maintaining explainability and safety supervision.
Tactile sensing remains the critical bottleneck; most commercial sensors trade sensitivity for robustness, necessitating custom development for effective force feedback and tissue interaction learning.
Hybrid approach integrating foundation models for vision with physics-based simulators for interaction prediction provides robustness that neither pure learning nor pure simulation achieves alone.
Rather than foundation models, the approach builds differentiable physics simulations (position-based dynamics) to infer tissue elasticity from video in real time, enabling model-predictive control for cutting, suturing, and hemorrhage management. Also covers humanoid deployment in hospitals and a haptic glove for teleoperation-to-autonomy transfer.
Energy is the binding bottleneck to AI scaling: compute, chips, and models all converge on electricity cost, which leaders from OpenAI, Nvidia, and SpaceX all identify as the critical constraint.
US electricity demand for AI could reach 1 terawatt within a decade, requiring energy supply growth steeper than any historical precedent—moving from 20-year stagnation to near-vertical expansion.
Nuclear power is the only viable baseload source meeting safety, emissions, and scalability requirements; all hyperscalers are pursuing nuclear despite 5–10 year build timelines.
Uranium enrichment is the missing infrastructure bottleneck: the US holds <0.1% global enrichment capacity and relies on Russia and Europe for nuclear fuel, blocking domestic reactor scaling.
General Matter secured a $900M DOE contract in January 2025—24 months after founding—to restore US uranium enrichment capacity, demonstrating how focused systems analysis unlocks government alignment and capital.
Enrichment is a fundamental primitive like SpaceX's launch cost: solving it enables downstream fuel production, reactor deployment, and clean energy scaling across the entire nuclear sector.
Bitcoin mining served as essential infrastructure rehearsal for AI datacenters—companies like Crusoe built stranded power utilization before pivoting to enterprise clouds, validating technology primitives regardless of initial use case.
Focus on first-principles problem-solving rather than surface-level narratives: nuclear's safety record (tied with wind, lowest emissions) contradicts public perception shaped by rare, famous accidents with minimal actual casualties.
The next 2–3 years will be hardest as turbine and grid interconnection equipment face 2+ year lead times; nuclear capacity comes online 2028–2030, creating a near-term scramble for stranded wind and natural gas.
Working on important unsolved problems with clear urgency, strong team fit, and government/market alignment creates extraordinary progress: General Matter will create hundreds of jobs in California and Kentucky while solving a civilizational bottleneck.
Scott Nolan argues energy, not chips, caps AI scaling—and nuclear is the only viable baseload option. The US produces under 0.1% of global enrichment capacity after its last facility closed in 2013, creating a dependency on Russia and Europe that General Matter's enrichment rebuild targets directly.
Anthropic's egress limiter counts all outbound bytes (including TCP ACKs, DNS, ICMP) and rate-limits them to make model weight theft take days/weeks instead of minutes, creating detection opportunities.
Model weights are terabytes of mostly incompressible data; legitimate cluster traffic (metrics, SSH, debugging) is only megabytes per second—this asymmetry is critical to the control's effectiveness.
Token bucket rate limiting allows burst traffic during off-peak hours but gradually throttles egress as the bandwidth bucket drains, providing better researcher experience than per-second limits.
Tiered enforcement: node-local traffic control handles accidental misconfigurations; perimeter routers (security control) bucket traffic by service type (logging, metrics, blob storage) to prevent one team's overages from impacting others.
IAM-enforced bucket boundaries prevent compute clusters from accessing external S3 buckets; proxies allow cross-boundary access only through the egress limiter for auditing and rate limiting.
Rollout killed ~98% of egress but took 4–6 months of reclassifying buckets and re-architecting systems; the final 2% (debugging, metrics, SSH) cannot be eliminated without major changes.
Research environments have huge attack surface (new dependencies, bleeding-edge stack, frequent vulnerabilities) adjacent to model weights worth hundreds of millions in compute—perimeter controls are necessary as defense-in-depth.
Accidentally misconfigured uploads now take days at 128 kbps instead of minutes, triggering alerts that inform researchers they're violating security policy rather than silently failing.
Egress limiting is a temporary fallback; long-term strategy is minimizing software touching unencrypted weights and securing that minimal set aggressively via TEEs and confidential compute—changes taking years.
Inference clusters also apply egress limiting but count bytes differently and accept additional risk on legitimate token traffic; future work explores subtracting model-generated tokens from egress limits.
By capping outbound bandwidth to ~100 Mbps at perimeter routers with per-service token buckets, Anthropic forces full-weight exfiltration—terabytes—to take weeks under assumed full-cluster compromise. The rollout cut egress 98% while keeping research workflows intact, buying detection time until TEEs and confidential compute mature.
