Part of IT Trends & Reviews — what actually shipped, quarter by quarter.
1. Introduction: The Quarter of Platform Maturity
Q3 2018 IT review reveals a quarter defined not by a single explosive breakthrough. Instead, it is defined by the steady, relentless maturation of the platforms that development teams already rely on every day. For example, Kubernetes reaches version 1.12 and graduates Kubelet TLS Bootstrap to general availability. In addition, PyTorch previews its JIT compiler and inches toward a production-ready 1.0 release. Similarly, TensorFlow ships version 1.9 with improved eager execution. The entire cloud-native stack, which felt experimental just two years ago, demonstrates that it can anchor enterprise workloads at scale.
However, the quarter also delivers genuine surprises. Nvidia pulls back the curtain on the Turing microarchitecture and launches the GeForce RTX 2080 Ti. Turing is the first consumer GPU generation to include dedicated hardware for real-time ray tracing. Apple announces the iPhone XS and XS Max, shipping the A12 Bionic chip manufactured on TSMC’s 7-nanometre process, the first 7nm chip to reach mass-market consumer hands. Google open-sources Cirq, a framework for writing and simulating quantum algorithms, bringing near-term quantum computing within reach of Python developers. Furthermore, Android 9 Pie lands in August with built-in neural network APIs and an adaptive battery powered by on-device machine learning.
1.1 GDPR, the crypto winter, and what the quarter adds up to
The broader context is equally consequential. GDPR, which comes into force in May 2018, now produces its first real enforcement signals. In practice, data protection authorities across Europe begin assessing complaints, and technology companies face the uncomfortable reality that compliance is not optional. Meanwhile, the cryptocurrency market stays deep in its winter since the January peak. In practice, Bitcoin hovers between roughly six thousand and eight thousand dollars. As a result, blockchain infrastructure teams must justify their projects on technical merit alone rather than speculative excitement.
Taken together, Q3 2018 marks the moment when the wave of excitement that followed deep learning’s ascent in 2015 and 2016 gives way to disciplined engineering. In practice, teams that place bets on containers, ML frameworks, and service meshes during the hype cycle are now cashing those bets in production. The question shifts from “can this technology work?” to “how do we run it reliably at scale?”. Notably, the answers emerge with surprising speed during these three months.
Key Insight: Three Simultaneous Maturations
Q3 2018 is unusual. Three completely separate technology stacks each cross a significant threshold within the same quarter. Specifically, they are container orchestration, deep learning frameworks, and consumer GPU hardware. For example, Kubernetes 1.12 ships RuntimeClass. PyTorch previews its JIT compiler. Finally, Nvidia launches Turing with RT Cores. Overall, this convergence signals that the infrastructure layer beneath AI applications is hardening across every dimension simultaneously.
2. GitHub Deep Dive: Key Open-Source Repositories
The open-source ecosystem continues its rapid expansion in Q3 2018. Several repositories either ship major milestones or enter the broader community’s consciousness for the first time. Each project below shapes how engineers design, train, deploy, or operate software during this quarter.
2.1 kubernetes/kubernetes — Reaching Version 1.12
Kubernetes 1.12 — RuntimeClass and TLS Bootstrap GA
Finally, Kubernetes 1.12 ships on the final day of Q3 2018 and introduces several features that make the platform meaningfully more flexible and secure. For example, RuntimeClass enters alpha, giving operators a formal mechanism to choose between multiple container runtimes on the same cluster. Consequently, security-critical workloads run in gVisor or Kata Containers while standard workloads use runc, all within one cluster. In addition, Kubelet TLS Bootstrap graduates to general availability, replacing the error-prone manual certificate distribution with an automated RBAC-based signing workflow.
Additionally, Kubernetes 1.12 adds alpha support for Azure Virtual Machine Scale Sets, extending cloud-native scalability to Microsoft’s cloud platform. Similarly, on the storage front, volume snapshots become available as an alpha feature via the snapshot controller. CNCF DevStats counts more than 20,000 individual contributors to date and 250 companies contributing per month. Kubernetes is therefore unquestionably the largest open-source project in infrastructure software. Furthermore, the New York Times, Nordstrom, and Squarespace are all running Kubernetes in production at scale, producing compelling case studies that accelerate enterprise adoption throughout Q3.
2.2 pytorch/pytorch — The Road to 1.0
PyTorch — JIT Compiler Preview and ONNX Export
Facebook’s AI Research team announces in May 2018 that PyTorch 1.0 is coming. In addition, in Q3 the community actively builds toward it.
The flagship new feature is torch.jit, a just-in-time compiler. In practice, it accepts regular PyTorch model code and re-executes it through a high-performance C++ runtime derived from Caffe2. Two modes are available. Trace mode records native PyTorch operations executed during a sample forward pass. Meanwhile, Script mode compiles a Python-annotated subset to an intermediate representation that fully captures control flow. In practice, that is critical for RNNs and variable-length sequence models. In addition, both modes produce models exportable to ONNX (Open Neural Network Exchange), enabling deployment in TensorRT, Caffe2, or any other ONNX-compatible runtime without a Python interpreter. Analysis of arXiv preprints and conference submissions consistently shows PyTorch overtaking TensorFlow in new ML research code during 2018. Therefore, PyTorch’s imperative, Python-native style, combined with the promise of production readiness in 1.0, makes it the dominant choice for anyone starting a new model.
2.3 tensorflow/tensorflow — Version 1.9 and Eager Execution
TensorFlow 1.9 — Eager Execution and Performance Gains
TensorFlow 1.9 ships in mid-July 2018, continuing the framework’s multi-release effort to make eager execution first-class. Eager mode executes operations immediately rather than building a static computation graph, removing much of the boilerplate that critics point to when favouring PyTorch. Notably, in version 1.9 eager mode integrates more tightly with TensorFlow’s Keras high-level API, allowing teams to mix symbolic model definitions with imperative debugging in a single codebase. Furthermore, TensorFlow 1.9 brings improved XLA compiler integration for TPU training, delivering meaningful throughput gains on Google Cloud’s TPU v2 pods.
In addition, the SavedModel format matures in this release, providing a portable, language-agnostic serialisation of entire models including computation graph and weights. Additionally, TFLite improves support for mobile inference, anticipating the on-device ML demands that Android 9 Pie’s Neural Networks API creates. TensorFlow’s advantage lies in TFServing, Cloud TPUs, and extensive enterprise tooling, though the research community’s momentum is visibly shifting toward PyTorch.
2.4 google/cirq — Quantum Computing for Near-Term Devices
Google Cirq — Open-Source Quantum Circuit Framework
In July 2018, Google’s quantum computing team announces Cirq alongside a wider push to make near-term noisy intermediate-scale quantum (NISQ) algorithms accessible to Python developers. In practice, Cirq provides a Python library for writing, manipulating, and optimising quantum circuits targeting real NISQ hardware, including Google’s own Bristlecone 72-qubit processor announced earlier in 2018. Abstract gate-model frameworks assume error-corrected, fault-tolerant hardware. By contrast, Cirq embraces the realities of today’s devices. Qubit connectivity constraints, gate fidelities, and noise models are first-class concepts. Researchers can define quantum gates, construct circuits, compile to hardware-native gate sets, and run simulations or submit jobs to real quantum processors. The release positions Google squarely against IBM’s Qiskit, which has been publicly available since 2017. Notably, Cirq takes a “near-term first” philosophy, acknowledging that current quantum hardware is noisy and limited, rather than pretending that fault-tolerant quantum advantage is imminent.
2.5 kubernetes-sigs/kubebuilder — SDK for Kubernetes Operators
Kubebuilder — Build Kubernetes APIs Using CRDs
The Kubernetes SIG-API-Machinery team introduces Kubebuilder in August 2018 as a formal SDK for building Kubernetes controllers and operators using Custom Resource Definitions. Prior to Kubebuilder, teams writing operators face significant boilerplate: setting up informers, work queues, reconciliation loops, leader election, and webhook validation from scratch. Kubebuilder generates scaffold code, integrates the controller-runtime library, and provides a coherent project layout that follows community best practices. Moreover, the tooling supports code generation for CRD manifests directly from annotated Go structs, eliminating the need to manually maintain YAML schemas. Kubebuilder arrives at exactly the right moment. Kubernetes 1.11 has just shipped multi-version CRDs, and Kubernetes 1.12 introduces RuntimeClass. Both features encourage teams to model their application domains as custom Kubernetes resources. The operator pattern is now widely recognised as the correct way to run stateful applications on Kubernetes. Moreover, Kubebuilder dramatically lowers the cost of writing one correctly.
2.6 linkerd/linkerd2 — The Rust-Powered Service Mesh
Linkerd 2.0 — A Lightweight Service Mesh for Kubernetes
Buoyant previews Linkerd 2.0 in September 2018, a ground-up rewrite of the original Linkerd service mesh.
Linkerd 1.x runs on the JVM and supports multiple orchestration platforms. By contrast, Linkerd 2.0 is Kubernetes-only and reimplements its data-plane proxy in Rust. That yields extremely low memory usage and latency overhead compared to Envoy-based meshes. The control plane runs in Go. Linkerd 2.0 positions itself as the operationally simple alternative to Istio, which has been growing in complexity throughout 2018. Its stated goal is to take a Kubernetes cluster from zero to a fully observed, mTLS-secured service mesh in under five minutes.
The Rust proxy provides transparent HTTP/1.1 and HTTP/2 proxying and automatic mutual TLS. It also reports golden metrics: latency, requests per second, and success rate. No application code changes are required. The preview release generates substantial interest among platform engineers who find Istio’s operational complexity a barrier to adoption. Additionally, Linkerd 2.0’s Rust proxy later becomes the foundation of a broader discussion about Rust’s suitability for systems infrastructure.
3. Big Tech & Industry Breakthroughs
The largest technology companies execute on several fronts simultaneously during Q3 2018. The following table summarises major events, after which each company’s strategy receives deeper analysis.
| Company | Event | Significance | Category |
|---|---|---|---|
| Cirq quantum framework released (July) | Open-source NISQ programming for Python developers | Quantum | |
| BigQuery ML reaches GA (July) | SQL-native machine learning for data analysts | ML/Data | |
| Android 9 Pie released (August 6) | First Android with on-device ML for adaptive battery | Mobile | |
| Dataset Search launched (September) | Google-powered discovery engine for public datasets | Data | |
| Apple | iPhone XS / A12 Bionic (September 12) | World’s first 7nm mass-market consumer chip | Hardware |
| Nvidia | Turing architecture / RTX 2080 Ti (September 20) | First GPU with dedicated ray-tracing (RT) cores | Hardware |
| PyTorch 1.0 roadmap development continues | JIT compiler and ONNX export bring production ML closer | AI/ML | |
| Microsoft | GitHub integration begins post-acquisition | Actions and security features entering roadmap | DevTools |
| Amazon / AWS | EKS widely adopted post-GA | Managed Kubernetes powering production workloads at scale | Cloud |
| Baidu | Apollo 3.0 autonomous driving platform (July) | Open autonomous driving with 130+ global partners | AV |
| Alibaba | 100 billion RMB R&D commitment announced | Three-year plan to build global AI research labs via DAMO Academy | AI/ML |
3.1 Google: ML Everywhere and the Quantum Opening
Google’s strategy in Q3 2018 is unmistakable: embed machine learning into every product and every developer tool. BigQuery ML reaches general availability in July. It enables data analysts to train linear regression, logistic regression, and k-means clustering models. They use standard SQL syntax directly against data in BigQuery tables. No Python, no data pipeline, and no separate ML infrastructure are required. This is not merely a convenience feature. It represents a fundamental expansion of the audience for machine learning from ML engineers to the vastly larger population of data analysts who already speak SQL.
Additionally, Google launches Dataset Search in September, a discovery engine for public datasets that indexes metadata from thousands of repositories. The service supports the schema.org/Dataset standard, providing a strong incentive for dataset publishers to adopt structured metadata. Moreover, the What-If Tool arrives as a TensorBoard plugin in September. It allows practitioners to probe trained models interactively. They can visualise decision boundaries, edit individual data points, compare model versions, and examine fairness metrics without writing a line of code.
Free ebook
Free AI Video, Generated Locally
Working scripts and measured benchmarks. Free.
No spam. Unsubscribe at any time.
Notably, Google also publishes the Universal Transformer research in August, extending the original Transformer architecture with a recurrent inductive bias that improves generalisation on algorithmic tasks. The Universal Transformer outperforms its predecessor on the bAbI language understanding tasks and the LAMBADA dataset. That suggests a productive research direction: combining the parallelism of self-attention with the iterative processing of RNNs. It anticipates BERT’s arrival in October.
3.2 Apple: The A12 Bionic and the 7-Nanometre Moment
At the Steve Jobs Theater on September 12, 2018, Apple announces the iPhone XS and iPhone XS Max, each powered by the A12 Bionic system-on-chip. Manufactured on TSMC’s first-generation 7-nanometre process, the A12 Bionic is the first 7nm chip to ship in a mass-market consumer product anywhere in the world. The chip integrates a six-core CPU, with two performance cores and four efficiency cores, plus a four-core GPU. Most significantly for the industry, it adds a dedicated Neural Engine. That engine now processes up to 5 trillion operations per second, a major leap from the A11’s 600 billion. Apple uses the Neural Engine to power improved Face ID, real-time depth segmentation in Portrait Mode, and augmented reality experiences in ARKit 2.
Furthermore, Apple’s CoreML 2 framework ships alongside iOS 12, cutting model sizes by up to 75 percent using new quantisation and compression formats while maintaining inference accuracy. The 7nm milestone matters far beyond Apple. It demonstrates that TSMC’s most advanced node is ready for high-volume consumer production. That paves the way for AMD’s Zen 2 processors and other 7nm designs arriving in subsequent quarters. The A12 Bionic therefore signals that the semiconductor roadmap, so often declared dead, is very much alive.
3.3 Nvidia: Turing, RT Cores, and Real-Time Ray Tracing
On September 20, 2018, Nvidia launches the GeForce RTX 2080 Ti and RTX 2080, the first consumer graphics cards built on the Turing microarchitecture. Turing introduces two entirely new types of processing units with no precedent in consumer graphics hardware. RT Cores implement BVH (bounding volume hierarchy) traversal and ray-box/ray-triangle intersection calculations needed for real-time ray tracing in dedicated fixed-function hardware, offloading this work from shader processors. Tensor Cores, adapted from the Volta V100 used in data-centre AI, accelerate matrix operations. They also power DLSS (Deep Learning Super Sampling). That AI-driven upscaling technique renders at lower resolution and uses a trained neural network to reconstruct full-resolution output.
The RTX 2080 Ti ships with 4,352 CUDA cores, 68 RT Cores, 544 Tensor Cores, and 11GB of GDDR6 memory. The DirectX Raytracing API from Microsoft underpins real-time ray tracing on Windows 10. Game developers are actively adding DXR support to titles, including EA DICE and Remedy Entertainment. However, two things initially constrain adoption. One is the high launch price of $999 USD for the Founders Edition. The other is the limited number of shipped games with DXR support in Q3. Regardless, the hardware architecture is unambiguously forward-looking and marks the beginning of a new era in interactive rendering.
3.4 Baidu and China: Apollo 3.0 and the AI Infrastructure Race
China’s technology sector continues its aggressive AI infrastructure build-out during Q3 2018. Baidu releases Apollo 3.0 in July, an update to its open autonomous driving platform that adds support for low-speed urban driving scenarios, valet parking, and micro-mobility use cases. Apollo’s partner ecosystem grows to more than 130 companies globally by mid-2018, including traditional OEMs such as Ford, BMW, and Daimler alongside Chinese automakers and tier-one suppliers. The platform is notable for its genuinely open approach: Baidu publishes its HDMap tooling, simulation environment, and sensor calibration tools under a permissive licence.
Alibaba, meanwhile, announces a three-year R&D investment plan of 100 billion RMB, approximately $15.2 billion USD. It directs the money at semiconductors, quantum computing, artificial intelligence, and cloud infrastructure. The work runs through its DAMO Academy research organisation. DAMO Academy operates offices in Seattle, Bellevue, Moscow, Tel Aviv, Singapore, and Hangzhou, reflecting a genuinely global talent acquisition strategy. Furthermore, Huawei is developing the HiSilicon Kirin 980 chip on TSMC’s 7nm node. The chip ships in the Mate 20 in October. It is therefore the second major 7nm consumer SoC to arrive within weeks of Apple’s A12.
3.5 Microsoft: Post-Acquisition GitHub Integration
Microsoft’s $7.5 billion acquisition of GitHub closes in June 2018, and Q3 is the first full quarter of integration. Nat Friedman, former CEO of Xamarin, becomes GitHub CEO and explicitly commits to preserving GitHub’s developer-first culture and open-source identity. During Q3, Microsoft begins signalling the product roadmap that will eventually materialise as GitHub Actions, GitHub Packages, and enhanced dependency security features. Developer sentiment remains cautious but largely reassured, and GitHub’s growth continues: the platform hosts more than 28 million users and nearly 85 million repositories by mid-2018. Additionally, Microsoft’s Azure Kubernetes Service continues its rapid maturation, integrating with Azure Active Directory for RBAC and supporting Kubernetes 1.12 within weeks of its announcement.
4. AI & Technology Impact
Q3 2018 is a pivotal period for applied machine learning. The field is well past the phase where publishing a neural network that beats a benchmark constitutes a major event. Instead, the meaningful developments are about transferring academic techniques into practitioner tools and scaling ML infrastructure to production environments. The following timeline captures the most significant AI and technology events of July through September 2018.
ICML 2018 — Stockholm, Sweden
The 35th International Conference on Machine Learning draws record attendance to Stockholm. Google presents more than 50 papers across topics including reinforcement learning, generative models, fairness in ML, and AutoML. Key contributions include research on Neural Architecture Search efficiency and multi-task learning for natural language understanding. The conference establishes Stockholm as a European hub for ML research, with strong participation from DeepMind, MILA, and academic groups across the continent.
Google Releases Cirq — Quantum Computing Framework
Google open-sources Cirq on GitHub, providing Python APIs for writing and simulating quantum circuits targeting NISQ devices. The framework includes a noise model for Google’s Bristlecone 72-qubit processor and a simulator capable of running small circuits on classical hardware. The announcement positions Google alongside IBM (Qiskit) as a major provider of open-source quantum computing tooling.
BigQuery ML Reaches General Availability
Google Cloud announces BigQuery ML as generally available, enabling SQL-native model training and inference inside BigQuery. The feature initially supports linear regression and logistic regression, with additional model types on the roadmap. This dramatically lowers the barrier to machine learning for analysts who already use BigQuery for data warehousing.
TensorFlow 1.9 Released
TensorFlow 1.9 ships with improved eager execution integration in the Keras high-level API. It also brings better XLA support for TPU performance and a more mature SavedModel serialisation format. The release represents Google’s ongoing effort to reduce TensorFlow’s operational complexity in response to PyTorch’s growing research community momentum.
4.1 Q3 2018 timeline: August
Android 9 Pie — ML-Powered Adaptive Battery
Google releases Android 9 Pie, the first Android version to ship an on-device ML model for adaptive battery management. The Neural Networks API (NNAPI), introduced in Android 8.1, becomes more capable and better-supported by hardware partners. Pie ships first on Pixel devices, with broader rollout to OEM partners over subsequent months. Digital Wellbeing, Dashboard, and App Actions all rely on on-device machine learning inference.
Universal Transformer Research Published
Google Brain researchers publish “Universal Transformers,” extending the Transformer architecture with a recurrent depth mechanism. The original Transformer has a fixed computational depth. By contrast, the Universal Transformer applies the same set of operations repeatedly. That combines the parallelism of attention with the iterative refinement of RNNs. The model achieves new state-of-the-art results on the bAbI language understanding tasks and several algorithmic sequence-to-sequence benchmarks.
MnasNet — Hardware-Aware Neural Architecture Search
Google researchers present MnasNet, a reinforcement learning-based architecture search approach optimised for real mobile inference latency rather than theoretical FLOPs counts. MnasNet incorporates actual inference latency on Pixel phones directly into the reward signal. As a result, it discovers mobile models up to 1.5x faster than MobileNetV2 at comparable accuracy. The work reinforces the growing importance of hardware-aware neural network design.
Kubebuilder Announced — SDK for Kubernetes Operators
In addition, the Kubernetes SIG-API-Machinery team announces Kubebuilder, providing scaffold generation, controller-runtime integration, and CRD manifest generation for teams building Kubernetes operators. The tooling drastically reduces the boilerplate required to write a production-quality Kubernetes controller. Moreover, it aligns with the broader trend of encoding operational knowledge as code through the operator pattern.
4.2 Q3 2018 timeline: September
Apple iPhone XS — First 7nm Consumer Chip
Apple announces the iPhone XS and XS Max with the A12 Bionic, manufactured on TSMC’s 7nm process. The Neural Engine in the A12 delivers 5 trillion operations per second. That enables real-time machine learning inference for Face ID, augmented reality, and Portrait Mode photography. All of it runs on-device, without cloud round-trips. CoreML 2 reduces model sizes by up to 75% via compression.
Nvidia RTX 2080 Ti — Turing Architecture Launch
Nvidia launches the GeForce RTX 2080 Ti and RTX 2080, introducing RT Cores for hardware-accelerated ray tracing and Tensor Cores for DLSS. The Turing architecture supports DirectX Raytracing on Windows 10 and marks the beginning of real-time ray tracing as a practical interactive rendering technique.
Linkerd 2.0 Preview Released
Buoyant previews Linkerd 2.0, a Kubernetes-native service mesh with a Rust data-plane proxy and Go control plane. The release positions Linkerd as the operationally simpler alternative to Istio, targeting the large number of platform engineers who find Istio’s complexity a barrier. Transparent mTLS, golden metrics, and zero-code-change deployment are its primary differentiators.
Kubernetes 1.12 Released — RuntimeClass and TLS Bootstrap GA
Kubernetes 1.12 ships with RuntimeClass (alpha) for multi-runtime cluster support, Kubelet TLS Bootstrap (GA) for automated certificate management, Azure VMSS support, and volume snapshot alpha. The release marks the third Kubernetes minor release of 2018, reinforcing the project’s consistent quarterly cadence and continued production-hardening trajectory.
Google Dataset Search and What-If Tool
Google launches Dataset Search, an index of public datasets using schema.org/Dataset metadata, making thousands of government, academic, and commercial datasets discoverable through natural-language queries. Simultaneously, the What-If Tool for TensorBoard enters public availability, enabling interactive probing of trained models without writing code.
4.3 The Service Mesh Landscape Clarifies
Away from the headline AI announcements, the service mesh category consolidates significantly during Q3 2018. Istio 1.0 reaches general availability in late Q2, though operational complexity criticism mounts throughout Q3 as teams attempt production deployments. Linkerd 2.0’s September preview offers a compelling counter-narrative: simplicity over features. Meanwhile, Envoy Proxy — the data-plane component underlying Istio and many other meshes — accumulates integrations across the industry. AWS App Mesh, entering preview in Q4, treats Envoy as the universal service proxy substrate. The CNCF’s graduated status for both Envoy and containerd during the year further legitimises the cloud-native stack as a whole.
4.4 Crypto Winter: Infrastructure Without Speculation
Bitcoin trades between approximately $6,000 and $8,000 throughout Q3 2018, far below its December 2017 peak near $20,000. Ethereum falls from over $1,400 in January to under $200 in September. For the engineering community, however, this is not entirely negative: the crypto winter filters out speculative projects and forces serious blockchain teams to demonstrate genuine utility. Ethereum’s developers use the quieter market period to make progress on the Constantinople hard fork, planned for later in the year. It introduces several Ethereum Improvement Proposals to reduce gas costs. Furthermore, the Lightning Network on Bitcoin continues to grow its node count, with the number of Lightning channels roughly doubling between July and September 2018.
5. Key Voices & Thought Leaders
Several engineers, researchers, and communicators publish influential work during Q3 2018 that shapes the conversation about technology’s direction. The following five individuals stand out for the quality and reach of their contributions in this period.
Kelsey Hightower — Google Developer Advocate for Kubernetes
Throughout Q3 2018, Kelsey Hightower continues to define the cultural conversation around Kubernetes adoption through memorable live demonstrations and quotable Twitter threads. His approach — showing complex Kubernetes operations working flawlessly in real time, with the implicit message that operational simplicity is achievable — is enormously influential. In particular, his commentary on Kubernetes 1.12’s RuntimeClass and its implications for security-sensitive workloads resonates strongly with platform engineering teams. Hightower’s presentations from KubeCon Copenhagen in May 2018 continue to circulate widely on YouTube throughout Q3. His perspective consistently emphasises that Kubernetes is not just a technology. It is a practise, a discipline of thinking about systems in terms of desired state rather than imperative commands.
Jeremy Howard — Co-founder of fast.ai
Jeremy Howard and Rachel Thomas continue building fast.ai’s curriculum through Q3 2018. Their Practical Deep Learning for Coders course influences how thousands of practitioners approach ML education. Howard is particularly active during this quarter on the subject of transfer learning in NLP. fast.ai’s ULMFiT approach, published in early 2018, pretrains a language model on large corpora and then fine-tunes it for a specific task. That dramatically reduces the data needed for text classification. This idea is transfer learning for NLP. It is central to the moment the field is in. Moreover, it directly anticipates the transformer-based language model wave that arrives with BERT in October. Howard’s blog posts and forum conversations during Q3 push the community to think about pretraining as a general-purpose technique applicable across modalities, not just computer vision.
5.1 NLP research communication and production vision
Sebastian Ruder — NLP Researcher
Sebastian Ruder’s blog at ruder.io is, in Q3 2018, one of the best single sources for keeping current with NLP research. His co-authored survey “Universal Language Model Fine-tuning for Text Classification” (ULMFiT) with Jeremy Howard has already generated significant follow-on work. During Q3, Ruder publishes summaries and analyses of the most important NLP papers emerging from conferences and arXiv, including the Universal Transformer from Google Brain. His NLP-progress tracking repository on GitHub, which records state-of-the-art results across NLP benchmarks with verifiable citations, becomes an indispensable reference for researchers and practitioners alike. Ruder’s accessible writing style and rigorous sourcing set a standard for how ML research communication should work. In addition, his influence grows rapidly throughout 2018.
Andrej Karpathy — Director of AI at Tesla
As Director of AI at Tesla, Andrej Karpathy is deeply embedded in one challenge. It is scaling vision-based autonomous driving to a fleet of hundreds of thousands of vehicles. Each car generates data that feeds back into model training. During Q3 2018, Tesla’s Autopilot team works on neural network-driven improvements for Autopilot. Karpathy is also active in the broader research community, discussing ideas about data-centric AI, large-scale labelling, and the virtuous cycle of fleet learning. In his view, the most important factor in practical autonomous driving is data quality and quantity rather than algorithm novelty. That perspective is influential and ahead of its time. The academic community is still focused on benchmark performance. Additionally, his “Software 2.0” concept, which frames neural networks as a new form of programming, continues to circulate widely during Q3.
5.2 Technology and the public interest
Tim O’Reilly — Founder and CEO of O’Reilly Media
Tim O’Reilly’s 2017 book WTF: What’s the Future and Why It’s Up to Us continues to drive conversation in Q3 2018. The discussion centres on the responsible deployment of algorithms in social systems and on the economics of platform businesses. O’Reilly uses his public presence to push technology leaders to think about second-order effects. What happens when AI systems optimise for engagement metrics? What happens when platform economics concentrate value in ways that harm workers? His writing during Q3 connects GDPR enforcement, the Cambridge Analytica fallout from Q1 2018, and the crypto winter. He weaves them into a broader narrative about the need for technology to earn public trust rather than assume it. His influence is distinctive because it bridges the technical and political dimensions of the industry at a time when that bridge is urgently needed.
6. Trend Synthesis
Stepping back from the individual announcements that define Q3 2018 IT review, several converging patterns emerge. Together they suggest a technology industry undergoing structural maturation. This is not a slowdown, but a deepening that will shape the following decade’s architecture.
6.1 The Abstraction Layer Hardens
The most important trend of Q3 2018 is the hardening of the abstraction layers that the industry settled on between 2015 and 2017. Containers run on Kubernetes. Models train in PyTorch or TensorFlow. Services communicate over gRPC or HTTP/2 and are observed by Prometheus. These choices are no longer contested; they are assumed. Consequently, the energy in the community shifts upward. Teams now argue about how to build Kubernetes operators, how to structure ML pipelines, and how to instrument microservices. They no longer argue about whether to use containers at all.
Kubebuilder’s August announcement perfectly captures this dynamic. Its value is not that it introduces a new idea, but that it codifies existing best practices into a tool that dramatically reduces the cost of following them. Similarly, Linkerd 2.0’s positioning against Istio is not a debate about service meshes in principle. Both sides agree that mutual TLS and distributed tracing are necessary. The argument is about which implementation imposes the least operational burden. Furthermore, the operator pattern, which emerges from the Kubernetes community’s experience with stateful applications, is becoming the industry-standard approach to encoding operational knowledge as software. In this sense, Q3 2018 represents the point at which cloud-native practices shift from innovator adoption to early majority adoption.
6.2 Hardware Innovation Returns After a Long Pause
For roughly a decade, the software industry operates under one assumption. CPU performance improvements will be modest and incremental. That follows from Moore’s Law slowing near the physical limits of silicon lithography. Q3 2018 disrupts this assumption with two simultaneous hardware milestones. Apple’s A12 Bionic on 7nm demonstrates that cutting-edge process nodes remain viable for high-volume production. Nvidia’s Turing architecture demonstrates that new functional units — RT Cores and Tensor Cores — can deliver qualitative capability leaps that general-purpose shader processors cannot.
Additionally, Google’s Cirq release is part of a broader signal. Quantum computing hardware is not yet practically useful for most applications. Even so, it is advancing fast enough to warrant building the software ecosystem in advance. The race between Google, IBM, and well-funded startups like Rigetti, IonQ, and D-Wave for quantum advantage is entering a decisive phase. Therefore, developers are beginning to prepare by learning the programming models today.
6.3 Transfer Learning Defines the Next Phase of NLP
The NLP field in Q3 2018 stands at an inflection point that is not yet fully visible to the broader industry. ULMFiT, ELMo (from AllenAI), and the Universal Transformer all point in the same direction. Pre-training a language model on large unlabelled corpora comes first, and fine-tuning on specific tasks follows. That yields results supervised training from scratch cannot match. This is the transfer learning revolution that computer vision experienced with ImageNet and CNNs, now arriving in natural language processing.
Jeremy Howard and Sebastian Ruder advocate for ULMFiT, and the research progresses rapidly. As a result, the BERT paper appears within weeks of Q3’s end. It confirms this direction comprehensively. Teams reading the research signals carefully during Q3 2018 are already investing in Transformer-based language model infrastructure, even though BERT itself has not yet been released. However, those teams represent a small minority of practitioners. In addition, most of the industry will need to recalibrate rapidly in Q4.
6.4 GDPR Shifts the Privacy Calculus Permanently
Q3 2018 is the first quarter in which GDPR enforcement is a practical reality rather than a future threat. Data protection authorities across Europe begin processing complaints. In addition, the threat of fines reaching 4% of global annual revenue is clearly concentrating minds in boardrooms. The more important effect, however, is cultural: technology companies begin to treat user data as a liability to be minimised rather than an asset to be maximised. Apple leads the way with its privacy-as-differentiator positioning, marketing the iPhone XS’s on-device Neural Engine partly as a privacy feature (sensitive data never leaves the device). The downstream effects on data collection practices, cookie consent implementations, and privacy-by-design architectures are already visible in Q3 and intensify through 2019.
6.5 The Global Tech Race Expands Beyond Silicon Valley
Q3 2018 makes clear that the technology industry is no longer primarily a Silicon Valley story. Alibaba’s 100-billion-RMB R&D commitment, Baidu’s 130-partner Apollo ecosystem, Huawei’s 7nm Kirin 980 development. In addition, a growing number of European research teams contributing to open-source projects all signal a genuine globalisation of technological capability. The concentration of AI talent in the US seemed overwhelming in 2016. It is now dispersing. Remote work, diaspora networks, and deliberate government investment policies in China, Canada, the UK, France, and Israel all contribute. Moreover, KubeCon Shanghai, planned for November 2018, symbolises this shift. The world’s largest Kubernetes conference comes to China for the first time. That reflects the substantial Chinese contribution to cloud-native software.
7. Summary
Q3 2018 is a quarter that rewards careful observation. On the surface, the headlines — a new iPhone, a new GPU, a new Kubernetes release — appear routine. However, each represents a threshold crossing. The 7nm node reaches mass production, and real-time ray tracing becomes a practical consumer feature. Kubernetes completes its transformation from experimental tool to universally-adopted production platform. Therefore, the pace of infrastructure maturation is, if anything, accelerating even as individual announcements become less sensational.
For practitioners, the clearest takeaway is that the technology bets placed in 2016 and 2017 are paying off. Kubernetes is production-ready. PyTorch approaches 1.0. TensorFlow’s Keras integration is maturing. The service mesh is real. Teams that adopt these technologies early have a meaningful head start. However, the tooling — as represented by Kubebuilder and Linkerd 2.0 — is making adoption faster and less risky for latecomers as well.
For researchers, Q3 2018 is the last quiet quarter before BERT’s October announcement reshapes NLP entirely. The Universal Transformer, ULMFiT, ELMo, and a dozen other contributions during and just before this quarter build the conceptual foundation that BERT synthesises. Looking at the NLP landscape in September 2018, the direction is clear to those paying attention: pre-trained language models with Transformer architectures are the future. In addition, that future arrives in weeks rather than years.
7.1 Open questions at the close of the quarter
Open questions remaining at the close of Q3 2018 include: How will the service mesh market consolidate between Istio and its simpler alternatives? When will Nvidia’s Turing RT Cores reach the developer mainstream with broad game engine support? How will GDPR enforcement escalate and spread beyond Europe? Will quantum computing reach practical advantage on any commercially relevant problem within the next five years? And most pressingly for the ML community: when PyTorch 1.0 ships, will it finally bridge the remaining gap between research flexibility and production performance? The answers begin arriving in Q4 2018 — and several of them will surprise the industry.
Sources
- Kubernetes 1.12 Release Announcement — kubernetes.io blog, September 27, 2018
- Kubernetes 1.11 Release Announcement — kubernetes.io blog, June 27, 2018
- The Road to 1.0: Production Ready PyTorch — pytorch.org blog, May 2, 2018
- TensorFlow 1.9.0 Release Notes — github.com/tensorflow/tensorflow, July 2018
- Announcing Cirq: An Open Source Framework for NISQ Algorithms — Google Research Blog, July 2018
- Machine Learning in Google BigQuery — Google Research Blog, July 2018
- Moving Beyond Translation with the Universal Transformer — Google Research Blog, August 2018
- MnasNet: Towards Automating the Design of Mobile ML Models — Google Research Blog, August 2018
- Introducing Android 9 Pie — Android Developers Blog, August 6, 2018
- The What-If Tool: Code-Free Probing of Machine Learning Models — Google Research Blog, September 2018
- Building Google Dataset Search and Fostering an Open Data Ecosystem — Google Research Blog, September 2018
- Apple Introduces iPhone XS and iPhone XS Max — Apple Newsroom, September 12, 2018
- Nvidia Turing GPU Architecture Announcement — Nvidia Newsroom, September 2018
- Introducing Kubebuilder: An SDK for Building Kubernetes APIs Using CRDs — kubernetes.io blog, August 10, 2018
- Hands On With Linkerd 2.0 — kubernetes.io blog, September 18, 2018
- Universal Transformers — Dehghani et al., arXiv:1807.03819, 2018
- MnasNet: Platform-Aware Neural Architecture Search for Mobile — Tan et al., arXiv:1807.11626, 2018
- Universal Language Model Fine-tuning for Text Classification (ULMFiT) — Howard & Ruder, arXiv:1801.06146, 2018
- Kubernetes DevStats — CNCF community velocity data, devstats.k8s.io
- NLP-progress: Tracking State of the Art in NLP — github.com/sebastianruder/NLP-progress
- Sebastian Ruder’s Blog — ruder.io (NLP research commentary, 2018)
- fast.ai — Practical Deep Learning for Coders curriculum, Jeremy Howard and Rachel Thomas, 2018
- Apollo 3.0 Platform Release — Baidu AI, July 2018
- Alibaba Group Announces 100 Billion Yuan R&D Investment Plan — Alibaba Cloud Blog, 2018
Free ebook
Free AI Video, Generated Locally
Run Wan 2.1 in ComfyUI on your own GPU — the scripts I use, measured times, sample clips. No cloud, no API keys.
No spam. Unsubscribe at any time.
Leave a Reply