# From Virtualization to AI Integration

Lamia Youseff, JazzComputing | MLOps Podcast | Episode 175 | 52:07
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=Ywp5KaCcd6Y
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/from-virtualization-to-ai-integration
Published: 2023-09-12
Tags: enterprise, orchestration, platform-teams, product-strategy

## TL;DR
- Lamia Youseff connects the early cloud era with today's AI adoption, where companies are still working out which capabilities and use cases are worth building.
- A machine learning platform needs to coordinate data preparation, training, inference, monitoring, and cloud infrastructure across providers instead of handling only one orchestration task.
- JazzComputing helps enterprises, investors, and startups understand AI trade-offs and build the organizational and technical capabilities needed to put AI into use.

## Summary
Lamia Youseff describes a career that moved from early AI research into high-performance computing, virtualization, cloud infrastructure, and product work at Google, Microsoft, Apple, and Facebook. She sees clear parallels between the early cloud market and the current AI market. Cloud providers competed before the ecosystem had settled, while customers gradually discovered the need for multi-cloud deployment and better coordination. AI has a similar problem. Teams may use separate systems for feature engineering, training, inference, monitoring, and data, often across different cloud providers. Lamia argues that the field needs an abstraction layer that coordinates these parts around business use cases. She also explains why startups may begin with APIs to reach product-market fit, while regulated enterprises may need more control over data and models. Through JazzComputing, she works with startups, investors, Fortune 500 companies, and SMBs to connect technical choices with strategy, investment, and adoption.

## Key ideas
### A breast-cancer prediction assignment made AI feel immediately useful
[04:08](https://www.youtube.com/watch?v=Ywp5KaCcd6Y&t=248s)
Lamia Youseff became interested in AI while studying computer science at the American University in Cairo. Her mother was being tested for breast cancer at the same time Lamia received a homework assignment to build a neural network that predicted benign versus malignant tumors. The model was basic and the code was written in C, with neuron connections implemented through pointers. Running it required moving code between computers with floppy disks, and a bad pointer could cause a memory dump and restart. Her mother's tumor turned out to be benign, but the experience gave Lamia a direct sense of how machine learning might support early detection and save lives.

### Limited computing resources pushed Lamia toward cloud research
[08:56](https://www.youtube.com/watch?v=Ywp5KaCcd6Y&t=536s)
When Lamia applied to graduate school, she was told that AI was a dead field because researchers lacked enough computational resources. She decided to work on that constraint and moved into scientific computing, high-performance computing, virtualization, and cluster computing. Her work at UC Santa Barbara led to Eucalyptus, a private cloud that put Amazon EC2-compatible APIs on top of a cluster. The project drew attention as a possible private-cloud equivalent to EC2 and was later integrated into Red Hat. Lamia went to MIT instead of joining the company because she still wanted to pursue academic research, but the project placed her close to the early cloud transition.

### Google Cloud taught Lamia the difference between research and production systems
[13:34](https://www.youtube.com/watch?v=Ywp5KaCcd6Y&t=814s)
In 2011, Google recruited Lamia for a confidential project in Seattle. She expected to spend six months learning from the project and then return to MIT, but the project was Google Cloud, where she became one of the founding engineers. The team had about 12 engineers and worked quietly in an office under a bridge. Lamia says research systems are meant to push design ideas and publish prototypes, while industry systems must scale, solve customer problems, and find product-market fit. Her work included reducing virtual-machine startup time from about three minutes to roughly 10 or 12 seconds, along with encryption and isolation for virtual machines.

### Cloud customers exposed the need for multi-cloud systems
[21:07](https://www.youtube.com/watch?v=Ywp5KaCcd6Y&t=1267s)
After moving from engineering into product management, Lamia saw that customers used cloud services in ways the early teams had not expected. An outage at AWS showed a major streaming service that it needed deployments across different clouds to maintain availability during high-demand periods. Lamia later worked on a system that moved about 10 terabytes of data each night between AWS and Google Cloud storage. She describes this as an early example of the coordination and orchestration required for multi-cloud deployments. The lesson was that cloud infrastructure evolved around real customer needs, including availability, portability, and the ability to combine services from different providers.

### AI needs an abstraction layer above individual orchestration tools
[24:08](https://www.youtube.com/watch?v=Ywp5KaCcd6Y&t=1448s)
Lamia expects AI infrastructure to follow a path similar to cloud infrastructure. Different providers now promote their AI capabilities, but applications may eventually use several providers at once. She imagines an orchestrator that works across AI infrastructure and coordinates capabilities from multiple clouds. A machine learning workload might train on GPUs, move to SageMaker for inference, use an LLM from another provider, and perform fine-tuning somewhere else. This need goes beyond adding another workflow tool. The proposed layer would hide provider-specific details while connecting feature engineering, training, inference, monitoring, scaling, availability, and latency requirements.

### Standardization will require shared incentives and large-company support
[27:42](https://www.youtube.com/watch?v=Ywp5KaCcd6Y&t=1662s)
Demetrios Brinkmann points out that the MLOps ecosystem lacks an open-source foundation with the unifying role Kubernetes played in DevOps. Lamia agrees that a common layer could connect many machine learning tools, but she says adoption would require coordination among large companies with different interests and existing systems. She expects the ideas to develop before the use cases become mature enough for a widely adopted platform. Lamia also distinguishes this proposed layer from ordinary orchestration. It would provide contracts and compatibility across systems, so teams could plug their environments into a larger ecosystem without rebuilding each integration.

### API-first AI adoption depends on the use case and the company
[40:29](https://www.youtube.com/watch?v=Ywp5KaCcd6Y&t=2429s)
Lamia says APIs such as OpenAI's helped move AI from science fiction into ordinary product and business discussions, but API access does not fit every situation. Health data may involve privacy and HIPAA requirements, while public-data summarization has different constraints. Startups often need fast iteration and low cost, so they may use an API to validate a use case and reach product-market fit before taking on more control and scalability work. Fortune 500 companies may face privacy and compliance requirements from the beginning. The right choice depends on the product phase, the data, and the reliability and control the application needs.

### JazzComputing connects enterprises, investors, and AI startups
[46:37](https://www.youtube.com/watch?v=Ywp5KaCcd6Y&t=2797s)
Lamia describes JazzComputing as a new firm focused on the space between investors, Fortune 500 companies, SMBs, and startups. Enterprises often know AI could disrupt their use cases but do not know which ones to address or what technical, organizational, and strategic capabilities they need. Investors need help evaluating AI companies and understanding their foundations. Startups may have advanced technology but lack knowledge about distribution, investor expectations, marketing, and cost structure. JazzComputing works across these groups through advisory work, education, research papers, a newsletter, a podcast, and public announcements. Lamia's aim is to help the groups work together on practical AI adoption.

## Notable quotes
- Lamia Youseff: "In my mind it's going to be a multi-market and that's what we saw with Cloud." (24:17)
- Lamia Youseff: "There is a need for one layer that's going to connect between all of these." (27:31)
- Lamia Youseff: "The more simple and streamlined the end-to-end process is going to be, the easier it is to be able to integrate it into other use cases." (36:31)
- Lamia Youseff: "It's all about the use case at the end of the day." (43:49)
- Lamia Youseff: "We are a new kind of firm that we're building for creating the connective tissue between investors, Fortune 500 and SMBs and then between startups." (46:48)

## Tools & references mentioned
- JazzComputing
- Stanford University
- MIT
- Google
- Microsoft
- Apple
- Facebook
- Amazon EC2
- AWS
- Eucalyptus
- Red Hat
- Kubernetes
- Borg
- Docker
- VMware
- OpenAI
- ChatGPT
- Llama
- HIPAA
- Kubeflow

## Who should watch
- You are deciding whether an AI feature should use an external API, an in-house model, or a mix of both, and need to weigh speed, privacy, reliability, and control.
- Your team has separate systems for data preparation, training, inference, and monitoring, especially across more than one cloud provider.
- You work at an enterprise, invest in AI startups, or run a young AI company and need to understand the capabilities missing between a promising model and a usable product.

## Editor's note

Lamia Youseff argues that AI needs an abstraction layer above individual orchestration tools because a workload may use different providers for training, inference, fine-tuning, and other steps. ZenML lets teams write these workflows as Python pipelines on a configurable stack, so the same pipeline code can run on a laptop, Kubernetes, Airflow, Kubeflow, or a cloud provider's services.

Written by the MLOps Talks editors (the ZenML team), not by the speaker.

## Related talks

- [Engineering Your AI Platform](https://mlopstalks.com/talks/engineering-your-ai-platform) (Tobias Macey, Massachusetts Institute of Technology (MIT) & Daniel Svonava, Superlinked & Colleen Tartow, VAST Data, 30:09)
- [The Future of ML and Data Platforms](https://mlopstalks.com/talks/the-future-of-ml-and-data-platforms) (Michael Del Balso, Tecton, 55:17)
- [How Data Platforms Affect ML & AI](https://mlopstalks.com/talks/how-data-platforms-affect-ml-ai) (Jake Watson, The Oakland Group, 39:12)
- [Speed and Scale: How Today's AI Datacenters Are Operating Through Hypergrowth](https://mlopstalks.com/talks/speed-and-scale-how-todays-ai-datacenters-are-operating-through-hypergrowth) (Kris Beevers, NetBox Labs, 1:07:17)
- [AI's Next Frontier](https://mlopstalks.com/talks/ais-next-frontier) (Aditya Naganath, Kleiner Perkins, 56:04)