Q4 2018 IT Review: BERT, PyTorch 1.0, and the NLP Revolution

Part of IT Trends & Reviews — what actually shipped, quarter by quarter.

1. Introduction: The Quarter Language Models Grow Up

The final quarter of 2018 opens with a seismic shift in artificial intelligence research. First, on October 11, Google researchers publish a preprint titled “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding” on arXiv. In fact, within days, the paper spreads across every NLP mailing list, Slack channel, and research blog in the field. BERT stands for Bidirectional Encoder Representations from Transformers. It shatters eleven NLP benchmarks simultaneously. It also introduces a deceptively simple idea that transforms the discipline. Rather than training a model from scratch for each individual task, you pre-train a massive language model on raw text. You then fine-tune it on your specific problem with only a modest amount of labelled data. This pre-train – fine-tune paradigm is not entirely new, but BERT executes it at a scale and precision that makes it feel qualitatively different from everything that has come before.

Two months later, the NeurIPS 2018 conference arrives in Montreal with record-breaking statistics. For example, there are more than 8,000 paper submissions, compared with 3,240 submissions in 2017. That reflects how dramatically academic interest in machine learning has grown. Meanwhile, at NeurIPS, Facebook AI Research unveils a preview of PyTorch 1.0, introducing the torch.jit compiler. That compiler bridges a gap. On one side is the dynamic computation graph researchers love. On the other are the static, optimised representations that production deployment demands. In short, the announcement signals that PyTorch is ready to leave university research labs and enter enterprise data centres.

1.1 Cloud infrastructure and the crypto winter

Meanwhile, the cloud infrastructure world is reshaped by AWS re:Invent 2018 in Las Vegas (November 26–30), where Amazon announces more than 100 new products and features. Notably, the most strategic announcements centre on custom silicon. The AWS Inferentia chip is unveiled as Amazon’s first in-house AI accelerator. Meanwhile, the EC2 A1 instance family introduces ARM-based processing using AWS’s own Graviton processor. That signals that hyperscalers are no longer content to let Intel define the server market. In addition, AWS Outposts extends the cloud to on-premises racks, blurring the boundary between public cloud and private data centres.

Meanwhile, beyond AI and cloud, Q4 2018 is also the quarter in which the cryptocurrency market completes its long descent from the mania of late 2017.

For example, Bitcoin trades above $6,000 at the start of October. It falls below $3,500 by December. That erases more than 80 per cent of its value from the all-time high set exactly a year earlier. Meanwhile, the Bitcoin Cash hard-fork dispute, which erupts on November 15 and leads to a costly “hash war” between competing factions, accelerates the sell-off. In addition, Microsoft finalises its $7.5 billion acquisition of GitHub on October 26, a moment that many developers regard with cautious optimism. In addition, Kubernetes 1.13 ships on December 3 with Container Storage Interface graduating to general availability and CoreDNS replacing kube-dns as the default cluster DNS provider. Overall, the events of Q4 2018 reveal an industry simultaneously maturing in infrastructure and exploding in AI capability.

2. GitHub Deep Dive: Key Open-Source Releases

In practice, the open-source landscape of Q4 2018 is dominated by natural language processing tooling, Kubernetes storage maturity. In addition, the first wave of community libraries that make BERT accessible to practitioners outside Google. Overall, the six repositories highlighted here collectively shift the centre of gravity in both ML research and cloud-native infrastructure. In addition, their impact is immediately visible in star counts, download statistics, and conference adoption.

2.1 google-research/bert — The Paper That Changes NLP

2.2 pytorch/pytorch — Version 1.0 Preview at NeurIPS

2.3 huggingface/pytorch-pretrained-bert — BERT for Everyone

2.4 kubernetes/kubernetes — Version 1.13

2.5 tensorflow/tensorflow — Version 1.12 and the Coming 2.0 Signal

2.6 kubeflow/kubeflow — Machine Learning Pipelines on Kubernetes

Repository	Key Q4 2018 Release	Significance	Tag
google-research/bert	Oct 11 — BERT-Base (110M params) & BERT-Large (340M params) with pre-trained checkpoints	Pre-train + fine-tune paradigm redefines NLP; 11 benchmarks beaten simultaneously	AI/NLP
pytorch/pytorch	1.0 preview (Dec) — torch.jit, C++ frontend, ONNX export improvements	PyTorch moves from research framework to production-capable platform	AI/ML
huggingface/pytorch-pretrained-bert	Nov 5 — PyTorch BERT port, verified against original checkpoints	Opens BERT to PyTorch community; seed of the transformers ecosystem	AI/NLP
kubernetes/kubernetes	1.13 (Dec 3) — CoreDNS default, CSI stable, kubeadm GA	Storage plug-in model standardises; cluster bootstrapping reaches enterprise quality	Containers
tensorflow/tensorflow	1.12 (Oct) — eager execution polish, tf.keras improvements, TF 2.0 signal	TensorFlow converges toward PyTorch-style programming model	AI/ML
kubeflow/kubeflow	0.3 (Q4) — Pipelines, Katib hyperparameter tuning, PyTorch operators	End-to-end ML platform on Kubernetes takes shape; MLOps era begins	MLOps

3. Big Tech & Industry Breakthroughs

Q4 2018 is one of the most consequential quarters in recent memory for large-scale technology strategy. AWS re:Invent reshapes the cloud hardware market with custom chips and hybrid cloud ambitions. Microsoft closes the largest developer-platform acquisition in history. Alibaba sets a new e-commerce record while simultaneously demonstrating that its cloud infrastructure can handle crushing peak loads. Google publishes BERT and further expands TPU availability. The following table captures the most significant events across the United States, Europe, and Asia.

Company / Org	Event	Date	Significance
Google (USA)	BERT paper published — 11 NLP benchmarks broken simultaneously	Oct 11	Transfer learning via pre-training becomes the dominant NLP paradigm; the “ImageNet moment” for language arrives
Microsoft (USA)	GitHub acquisition closes at $7.5 billion	Oct 26	World’s largest code-hosting platform joins Microsoft; developer community watches closely for strategic influence on open source
Apple (USA)	iPhone XR ships; MacBook Air and Mac mini refresh announced	Oct 26–30	iPhone XR democratises Face ID at a mid-tier price; Mac mini returns to developer relevance after years of neglect
Alibaba (China)	Singles’ Day: $30.8 billion GMV in 24 hours (new record)	Nov 11	Peak transaction rate of 325,000 orders per second; Alibaba Cloud handles load without incident, validating its elastic compute model
Amazon AWS (USA)	re:Invent 2018: EC2 A1 (AWS Graviton ARM), AWS Outposts, AWS Inferentia announced	Nov 26–30	AWS bets on custom silicon; hybrid cloud extends the cloud model on-premises; AI chip competition intensifies
Amazon AWS (USA)	Amazon Managed Streaming for Kafka (MSK), Amazon Timestream, Amazon Managed Blockchain launched	Nov 26–30	AWS fills managed-data-infrastructure gaps; enterprise data teams can adopt Kafka and time-series workloads without operational overhead
Amazon AWS (USA)	DynamoDB On-Demand pricing, Amazon Forecast, Amazon Personalize, Amazon Textract	Nov 26–30	On-demand DynamoDB removes capacity planning burden; ML-as-a-service portfolio expands dramatically into forecasting, personalisation, and document intelligence
Facebook / FAIR (USA)	PyTorch 1.0 preview announced at NeurIPS 2018	Dec 2–8	PyTorch crosses the research-to-production threshold; torch.jit and libtorch address the primary enterprise objection to the framework
SAP (Germany)	SAP Cloud Platform adds Kubernetes support; SAP Leonardo ML Foundation expanded	Q4 2018	European enterprise software leader embeds cloud-native ML into its platform; signals that Kubernetes is now an enterprise integration target
Hugging Face (France/USA)	pytorch-pretrained-bert released, BERT PyTorch port endorsed by Google	Nov 5	Small European AI startup becomes a critical infrastructure provider for global NLP research. Seeds what becomes one of GitHub’s most-starred repositories
Baidu (China)	ERNIE pre-training model announced; PaddlePaddle framework enhancements	Q4 2018	Chinese AI research directly engages with BERT paradigm; ERNIE incorporates entity and phrase-level knowledge into pre-training for improved Chinese-language benchmarks
Bitcoin / Crypto (Global)	BCH hash war (Nov 15) triggers broader market collapse; BTC falls from ~$6,500 to ~$3,200	Nov 15 – Dec	Crypto winter deepens; ICO projects with no sustainable revenue begin shutting down; developer interest shifts from blockchain to ML

3.1 AWS re:Invent 2018: Custom Silicon and the Hybrid Cloud Bet

AWS re:Invent 2018 is the largest and most strategically significant edition of the conference since re:Invent 2017 launched SageMaker and Fargate. Andy Jassy’s keynote centres on three themes: the move to custom hardware, the extension of the cloud model to on-premises infrastructure. In addition, the expansion of managed AI services.

Free ebook

Free AI Video, Generated Locally

Working scripts and measured benchmarks. Free.

Your e-mail address

Leave this field empty

Get the ebook

No spam. Unsubscribe at any time.

The EC2 A1 instance family is the most significant hardware announcement. A1 instances are powered by the AWS Graviton processor. That is a 64-bit Arm-based chip designed by Annapurna Labs. Amazon acquired that Israeli semiconductor company in 2015. The instances deliver up to 45 per cent cost savings for scale-out, Arm-compatible workloads compared with x86 equivalents. The announcement signals that Amazon is no longer content to let Intel’s Xeon roadmap define the economics of cloud compute. Furthermore, the AWS Inferentia chip is unveiled: a custom ASIC purpose-built for high-throughput, low-latency inference, supporting TensorFlow, PyTorch, and MXNet models.

AWS Outposts extends the cloud model to customer premises. Announced as a fully managed service, Outposts delivers native AWS compute, storage, and services in a pre-configured rack installed in a customer’s data centre. The rack runs the same VMware or native AWS APIs as the public cloud, enabling seamless workload migration between on-premises and cloud. The announcement directly responds to Microsoft Azure Stack, which has already been shipping hybrid cloud solutions to enterprise customers.

3.2 Managed data services and vertical integration

On the managed data-services front, Amazon MSK (Managed Streaming for Apache Kafka) removes the substantial operational burden of running Kafka clusters, handling broker provisioning, patching, and replication configuration. Amazon Timestream introduces a serverless time-series database service designed for IoT and operational telemetry workloads. DynamoDB On-Demand eliminates the need to pre-provision read and write capacity units. Among the ML services, Amazon SageMaker Neo compiles models once and runs them anywhere by automatically optimising the model for the target hardware — on any cloud or edge device.

3.3 Microsoft and GitHub: A Cautious Developer Community Watches

Microsoft finalises the acquisition of GitHub on October 26, 2018. The $7.5 billion all-stock deal, announced in June 2018, is approved by regulators without significant conditions. At the time of closing, GitHub hosts over 85 million repositories and more than 28 million developers — the vast majority of the world’s open-source code lives on GitHub. Nat Friedman, a developer-beloved figure known for founding Xamarin, is named GitHub’s new CEO, and Satya Nadella emphasises that GitHub operates independently with its own brand, culture, and product roadmap.

Consequently, initial developer concerns about open-source projects migrating away from GitHub to GitLab or Bitbucket do not materialise at the scale that many predicted. However, a significant minority of projects and developers do explore alternatives, and GitLab reports a spike in signups immediately after the announcement. Azure Kubernetes Service reaches general availability this quarter. Microsoft continues to invest heavily in open-source tooling through VS Code, TypeScript, and .NET Core. That reinforces the credibility of Nadella’s “we love open source” message.

3.4 Alibaba: Singles’ Day Record and Apsara Cloud Flex

On November 11, Alibaba processes $30.8 billion in gross merchandise volume in a single 24-hour period. That is approximately 27 per cent more than the prior year’s record. It is roughly 2.5 times the size of Amazon’s Prime Day 2018. The technical achievement behind the number is as impressive as the number itself. Alibaba Cloud’s elastic compute platform scales to handle more than 325,000 order submissions per second at peak. At the same time it runs the live streaming show, the real-time GMV counter, and the logistics platform. The Apsara OS coordinates tens of thousands of servers across distributed availability zones to deliver sub-second payment confirmation rates throughout the spike.

Moreover, at the Apsara Conference earlier in Q4, Alibaba announces that its AI chip programme is in advanced development. The programme is called Ali-NPU. Following Amazon’s Inferentia announcement, it signals that Chinese hyperscalers are also pursuing custom AI silicon. DAMO Academy, Alibaba’s research division, publishes multiple NLP papers this quarter that engage directly with the BERT paradigm for Chinese-language understanding tasks.

4. AI & Technology Impact

Overall, for AI research, Q4 2018 is a turning point. The convergence of the BERT paper, the NeurIPS 2018 conference, and the maturation of production-capable frameworks creates a new landscape. In that landscape, language understanding is suddenly tractable for practitioners with modest compute budgets. It was previously a notoriously difficult subfield. Therefore, the following timeline captures the key AI events of the quarter in chronological order.

4.1 Q4 2018 AI and technology timeline: October and early November

4.2 Q4 2018 AI and technology timeline: mid-November onward

4.3 The NLP Transfer Learning Revolution

To understand why BERT is so significant, it helps to recall the state of NLP research at the start of 2018. The dominant approach to most NLP tasks is to train a task-specific model from scratch on labelled data. Teams possibly use pre-trained word embeddings like GloVe or word2vec as the first layer. For example, ELMo (from AllenNLP, early 2018) and ULMFiT (from fast.ai, January 2018) demonstrate that pre-training on large unlabelled corpora and then fine-tuning on task data improves performance. However, ELMo uses shallow bidirectional LSTMs, and OpenAI’s GPT (June 2018) uses a left-to-right Transformer that is unidirectional during pre-training. BERT is the first model to pre-train a deeply bidirectional Transformer encoder. It uses masked language modelling. That allows each token to attend to context on both its left and right simultaneously.

The practical consequence is immediate. Research groups around the world fine-tune BERT for a few epochs on task-specific labelled data. It beats heavily engineered task-specific models that took months to develop. Sebastian Ruder is an NLP researcher who publishes the widely read NLP News newsletter. He coins the phrase “NLP’s ImageNet moment” to describe this shift. The analogy is to the moment when ImageNet pre-trained convolutional networks began outperforming bespoke computer vision architectures. That moment triggered the adoption of transfer learning across the entire field.

4.4 The GPU Market and Nvidia in Q4 2018

Notably, Nvidia occupies a paradoxical position in Q4 2018. Its GPUs are the indispensable substrate for virtually all deep learning research and training. Yet, its stock falls by more than 50 per cent between October and December, reflecting two simultaneous pressures. First, the cryptocurrency mining boom that sustained GPU demand throughout 2017 and early 2018 is collapsing. With Bitcoin and Ethereum prices in freefall, GPU mining is no longer profitable. Miners flood the second-hand market with used cards. That suppresses demand for new ones. Second, fears grow that cloud providers — Amazon’s Inferentia, Google’s TPUs — will reduce hyperscaler GPU purchases for inference workloads.

Nevertheless, research demand for the Volta V100 — the dominant research GPU of 2018, featuring 640 Tensor Cores that deliver 125 TFLOPS of FP16 performance — remains strong. For example, Google reports that BERT-Large training requires 4 days on 64 Cloud TPU v3 chips. GPU prices fall in consumer markets while data-centre GPU deployments grow. That demand tension defines the Nvidia narrative throughout Q4. It also foreshadows the custom-silicon era that accelerates over the following years.

5. Key Voices & Thought Leaders

In Q4 2018, a handful of researchers and engineers write blog posts, give talks, and publish papers. Those papers immediately reframe how practitioners think about language, infrastructure, and the relationship between research and engineering. The five individuals highlighted here each produce content during this period that is cited, shared, and built upon long after the quarter ends.

5.1 fast.ai, PyTorch, and the practitioner voices

6. Trend Synthesis

Standing at the end of Q4 2018, several converging forces become visible. They explain not just the events of this quarter. They also explain the trajectory of the industry over the following two to three years.

6.1 Transfer Learning Becomes the New Normal in AI

The BERT paper is the clearest statement yet of a paradigm shift that has been building through 2018. Pre-training on large unlabelled datasets produces powerful representations. Task-specific models then require very little labelled data to achieve state-of-the-art results. This shift has enormous economic implications. Previously, the cost of building high-quality NLP models — in terms of labelled data acquisition and model training compute — effectively limits production-grade NLP to large technology companies. BERT and the fine-tuning workflow it popularises democratise NLP. A company with a hundred labelled examples and a GPU instance can now match earlier results. Those results would have required a specialised NLP team and millions of training examples a year earlier.

Furthermore, the pattern is clearly generalising beyond NLP. Researchers are already applying similar pre-training ideas to other modalities. In addition, the success of BERT reinforces confidence that self-supervised learning on large unlabelled datasets is a path to general-purpose representations that transfer across tasks. Consequently, the focus of AI research is visibly shifting away from task-specific architecture design toward pre-training objectives, data curation, and scaling.

6.2 Custom Silicon Ends the CPU/GPU Duopoly in Cloud Compute

AWS re:Invent 2018 confirms what many in the industry have suspected: the hyperscalers are no longer content to purchase compute from Intel and Nvidia and resell it as a service. Amazon’s Graviton processor and Inferentia chip represent a fundamental shift in the cloud economics model. AWS designs its own chips optimised for specific workloads: general-purpose compute on ARM and inference on a custom ASIC. It can therefore offer better performance-per-dollar for those workloads. It also captures margin that would otherwise flow to chipmakers. Google has been pursuing this path with TPUs since 2016. Amazon’s commitment at re:Invent 2018 signals a broader shift. ARM-based and AI-specific silicon will be mainstream cloud compute options within two to three years.

Additionally, this trend accelerates consolidation of ML frameworks: if inference chips are custom ASICs from multiple vendors, the abstraction layer between model definition and hardware execution becomes critical infrastructure. ONNX and torch.jit are competing visions of what that abstraction looks like. The outcome of this competition, which plays out over the following years, determines whether AI hardware remains as fungible as DRAM or becomes as fragmented as mobile processor ISAs.

6.3 Kubernetes Becomes the Enterprise Platform, Not Just the Container Scheduler

The 1.13 release marks the point at which Kubernetes is no longer primarily a container orchestration system — it is an extensible platform on which higher-level abstractions are built. CSI graduating to stable means that storage vendors can integrate with Kubernetes without upstream code contributions. CoreDNS’s promotion to default demonstrates that the project is willing to replace early-era components with better alternatives as the ecosystem matures. kubeadm GA means that production cluster creation is a solved problem rather than a source of differentiation.

Together, these signals indicate that the “Kubernetes as infrastructure substrate” phase is complete and the “Kubernetes as application platform” phase is beginning. Kubeflow 0.3’s Pipelines component is one manifestation of this shift: ML workflows are expressed as Kubernetes resources. In addition, the cluster’s reconciliation loop drives the execution of multi-step training and serving pipelines. The implication is that the battle for developer mind-share is moving up the stack: from “which container runtime?” to “which ML platform?” and “which service mesh?”

6.4 Crypto Winter Redirects Developer Energy to ML and Infrastructure

The collapse of the cryptocurrency market in Q4 2018 is not merely a financial story. Throughout 2017 and early 2018, much developer attention goes to blockchain protocols, smart contract platforms, and decentralised applications. That is particularly true among backend engineers and distributed systems specialists. Bitcoin falls from $19,783 (December 2017) to approximately $3,200 (December 2018). In addition, hundreds of ICO-funded projects run out of capital. As a result, developers who had pivoted toward crypto begin returning to the ML and cloud-native infrastructure communities. Some discover them for the first time.

This is not a marginal effect. GitHub data from this period shows accelerating growth in ML-related repository stars and contributor counts. That growth is partly driven by BERT and PyTorch 1.0 specifically. It also reflects a broader shift in where ambitious engineers want to invest their time. NeurIPS 2018’s record submission count is a proxy for researcher headcount growth. It is consistent with the hypothesis that the end of the crypto boom frees up talent for other domains.

6.5 The Hugging Face Model: Infrastructure Startups in Open Source

The rapid adoption of pytorch-pretrained-bert reveals a new category of technology company: the open-source ML infrastructure startup. Hugging Face releases a library that solves a real problem for thousands of practitioners. The library builds an enormous community of users and contributors. It establishes brand recognition that the company later monetises through cloud hosting of models and datasets. The playbook has three steps. First, provide the open-source layer that big tech releases without productising. Second, build community trust through responsiveness and quality. Finally, monetise the cloud and enterprise tiers. The playbook proves extraordinarily effective. By the end of Q4 2018, Hugging Face has more practical impact on NLP practitioners than any single research paper or product launch from a larger company.

7. Summary

Q4 2018 is the quarter in which natural language processing crosses from a specialised research subfield into a general-purpose engineering discipline. BERT’s publication on October 11 is the catalyst. Google Research demonstrates that a single pre-trained model can be fine-tuned for eleven diverse NLP benchmarks. On all of them it achieves state-of-the-art results. That makes task-specific model engineering feel like unnecessary complexity. Hugging Face’s pytorch-pretrained-bert, released just 25 days later, ensures that this capability is immediately accessible to the PyTorch community without requiring TensorFlow expertise or Google Cloud infrastructure.

At the infrastructure level, AWS re:Invent 2018 reshapes expectations about cloud compute economics. The announcement of Graviton-based EC2 A1 instances and the AWS Inferentia inference chip signals that the hyperscaler-Intel-Nvidia dependency is loosening. AWS Outposts extends the cloud model onto customer premises, responding directly to enterprise hybrid-cloud requirements. Kubernetes 1.13 completes the transition from experimental container scheduler to enterprise-grade platform by graduating the Container Storage Interface and promoting CoreDNS. The Kubeflow 0.3 Pipelines component lays the groundwork for what becomes the dominant pattern of ML workflow orchestration on Kubernetes.

7.1 Crypto collapse and where developer energy goes

The cryptocurrency market collapse that accelerates through November and December is a sign-post for a broader shift in developer attention. In short, the ICO era — which commanded enormous technical talent in 2017 and the first half of 2018 — is ending. NeurIPS 2018’s record paper submission count reflects where that talent is now directing its energy: into machine learning, infrastructure, and the systems that connect them. Q4 2018 is the quarter where this reallocation of attention becomes visible in the data.

Two questions remain open entering Q1 2019. First: does the transfer learning paradigm that BERT exemplifies extend to other modalities — vision, audio, multimodal — with the same efficiency gains seen in NLP? If so, the training data acquisition bottleneck that has slowed AI deployment in specialised domains may be dramatically less severe than assumed. Second: does Amazon’s custom silicon strategy deliver on its performance-per-dollar promises when Inferentia instances become publicly available? The answer determines whether Nvidia maintains its grip on AI workloads in the cloud or whether AI chip competition becomes as vigorous as it is in the consumer mobile market. The industry leaves Q4 2018 more capable and more strategically contested than it entered it.

8. Sources

1. Devlin, J. et al. (2018). “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.” arXiv:1810.04805. Published October 11, 2018.
2. Google AI Blog. “Open Sourcing BERT: State-of-the-Art Pre-training for Natural Language Processing.” November 2018.
3. google-research/bert. GitHub repository. Initial release: October 2018. Apache 2.0 license.
4. The PyTorch Team. “The Road to 1.0: Production Ready PyTorch.” PyTorch Blog. Roadmap May 2018; preview at NeurIPS December 2018.
5. Hugging Face. pytorch-pretrained-bert (later transformers). GitHub. Released November 5, 2018. Apache 2.0 license.
6. Kubernetes Blog. “Kubernetes 1.13: Simplified Cluster Management with Kubeadm, CSI, and CoreDNS as Default DNS Are Now Generally Available.” December 3, 2018.
7. AWS Blog. “New EC2 Instances A1 Powered by Arm-Based AWS Graviton Processors.” November 2018.
8. Amazon Web Services. “AWS Inferentia.” First announced at re:Invent 2018.
9. AWS Blog. “AWS Outposts Now Available — Order Yours Today!” General availability, December 2019; Outposts was announced at re:Invent in November 2018.
10. Amazon Web Services. “Amazon Managed Streaming for Apache Kafka (MSK).” Launched at re:Invent 2018.
11. Microsoft Blog. “Microsoft Completes GitHub Acquisition.” October 26, 2018.
12. Sebastian Ruder. “NLP’s ImageNet Moment Has Arrived.” ruder.io. 2018.
13. Chen, R. T. Q. et al. (2018). “Neural Ordinary Differential Equations.” arXiv:1806.01822. NeurIPS 2018 Best Paper.
14. NeurIPS 2018 Conference. Montreal, December 2–8, 2018. Official proceedings.
15. Alibaba Group. “2018 11.11 Global Shopping Festival GMV Results.” November 2018. ($30.8 billion GMV record.)
16. Kubeflow. GitHub Releases. Version 0.3 released Q4 2018.
17. TensorFlow Blog. TensorFlow 1.12 release and TF 2.0 roadmap. October 2018.
18. Tim Dettmers. “Which GPU(s) to Get for Deep Learning.” timdettmers.com. December 2018.
19. Soumith Chintala. Twitter (@soumithchintala). PyTorch 1.0 preview commentary and NeurIPS 2018 discussion.
20. Apple Newsroom. “iPhone XR Available Today.” October 26, 2018.
21. Wikipedia. “Cryptocurrency crash (2018–2019).” Documents Bitcoin price decline from ~$20,000 (Dec 2017) to ~$3,200 (Dec 2018).
22. Howard, J. & Ruder, S. (2018). “Universal Language Model Fine-tuning for Text Classification (ULMFiT).” arXiv:1801.06146. January 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.

Your e-mail address

Leave this field empty

Get the ebook

No spam. Unsubscribe at any time.