Dr. Kelly Paulson of Providence Swedish Cancer Institute examines an immunofluorescent image showing T-cells surrounding a lobular breast cancer tumor sample, confirming a finding flagged by Ai2’s AutoDiscovery system. (Ai2 Photo)
An AI system built by Seattle’s Allen Institute for AI (Ai2) has found evidence that a common form of breast cancer, long thought to be a poor candidate for immunotherapy, might actually respond to it.
The finding, produced by Ai2’s AutoDiscovery system, has led to an expanded partnership with the Paul G. Allen Research Center at Providence Swedish Cancer Institute, which is now deploying the AI system on its own patient data to look for similar scenarios across other types of disease.
The announcement Thursday illustrates the broader potential for AI to uncover findings that human researchers, overwhelmed by the massive scale of modern datasets, might otherwise miss.
“Cancer researchers have access to extraordinary datasets, but the challenge is no longer collecting data; it’s understanding everything those datasets have to tell us,” said Dr. Kelly Paulson, who leads the Center for Immuno-Oncology at the Paul G. Allen Research Center, in a statement.
A video released by Ai2 in conjunction with the announcement.
AutoDiscovery, announced by Ai2 in February, works differently from most AI research tools: instead of waiting for a scientist to pose a question, it starts with a dataset and generates its own hypotheses, ranking them by how much they challenge existing assumptions.
In a research paper posted to the preprint server MedRxiv, the Ai2 and Providence Swedish researchers explain that they applied AutoDiscovery to The Cancer Genome Atlas, a federal dataset spanning more than 30 types of cancer.
The system flagged signs that invasive lobular carcinoma, which accounts for about 15% of U.S. breast cancer diagnoses, may be more responsive to immunotherapy than researchers thought. It has largely been left out of immunotherapy trials.
The Providence Swedish team confirmed the finding in a second patient dataset and validated it in tumor tissue in the lab. They cautioned that the findings don’t prove immunotherapy would work in these patients but suggest the question warrants further study.
The collaboration connects two organizations that trace their origins to the late Microsoft co-founder Paul Allen, who founded Ai2 in 2014 and whose $20 million donation helped establish the research center at Swedish in 2024.
Ali Farhadi, now a Microsoft corporate vice president of AI, at a Technology Alliance event in May 2024. (GeekWire File Photo)
GeekWire is profiling over the next few weeks some of the people and teams that are shaping the evolution of Microsoft in what we’re calling its “Microsoft 2.5” era.
From AI Frontier Lab to Frontier Ecosystem: Microsoft got a foothold in AI thanks largely to its partnership with OpenAI. But that’s not the way it is planning to continue growing its AI business.
Inside Microsoft AI (MAI), the Microsoft Superintelligence team is focused almost entirely on building its own frontier-level models. That team already has developed a handful of home-grown offerings, including MAI-Code-Flash for writing code faster; MAI-Cyber-Flash, a cybersecurity model; and MAI-Image, a model for creating images.
The head of the Superintelligence team is Ali Farhadi, corporate vice president of AI. Farhadi, who joined Microsoft five months ago, is also a professor at the University of Washington, where he has worked for nearly 15 years. He was previously CEO of the Allen Institute for AI (Ai2) and before that was an AI and machine learning leader at Apple for more than three years, after it acquired his startup, Xnor.ai.
When he joined Microsoft, Farhadi said in a LinkedIn post that he believed “Microsoft has all the pieces to win in this AI race: data, search, coding, infrastructure, agents, software and the world’s biggest Fortune 500 companies taking dependencies on Microsoft every day.”
Farhadi elaborated on that in an interview with GeekWire this week. AI is shifting from a “Frontier Lab” era to a “Frontier Ecosystem” era, he said. It’s no longer just about training models; it’s about integrating the models with enterprise data, platforms, distribution systems and customers in a trusted way.
The next battlegrounds in AI will be around cost, reliability, specialization, and deployment at scale, rather than simply building larger models that beat others in benchmark scores, he said.
“If you look around, there are not that many places to have all these missing pieces together at scale, especially if you add the element of trust to it,” Farhadi said.
Cutting through the AI noise: Farhadi said his management philosophy is grounded in the importance of personal relationships, which are especially key in big organizations. People need to understand your rationale and to trust you can deliver on what you’re tasked to do, he said — an approach that has served him inside both Microsoft and Apple.
Staying on top of the flow of information while filtering out the AI noise makes prioritizing crucial. The team has “a long list of things that we believe we should be doing,” he said, but much of it stays on the back burner to maintain a “laser focus on delivering on the main mission.”
The priority is building high-quality models, both generalist and domain-specific. On the domain-specific front, Microsoft is working with the Mayo Clinic on a healthcare-specific model based on Mayo’s own clinical data, as well as Microsoft’s cybersecurity and coding models.
The thinking: For a lot of enterprise work, a narrower model beats a bigger one.
“If you can do something at [the same] quality or better quality at a fraction of a cost, it’s just a no-brainer. And having a way to specialize to domains, to industries, to enterprises is one way,” he said.
Microsoft execs have referred to this approach as a “hill-climbing machine,” meaning the ability of a model to scale and continuously improve within a specific domain. Microsoft is coupling the hill-climbing with “frontier tuning,” like it is doing with the Mayo Clinic. Frontier tuning includes customizing frontier models; keeping proprietary data private, preserving institutional know-how; and avoiding leaking intellectual property (IP) into shared models.
“We all thought that IP is your data,” Farhadi said. “But we learned that IP is also how you work.” And that’s why safeguarding these elements is so crucial.
Open all the things? Farhadi led an expansion of open-source AI development at Ai2, the Seattle-based institute founded in 2014 by the late Microsoft co-founder Paul Allen. While Microsoft has contributed to the open-source community on various fronts, including AI tooling, it hasn’t open-sourced its frontier models.
Farhadi said he personally remains “a big advocate of open source,” but noted that the industry has changed since his Ai2 days as there are now more credible Western open-source models and businesses forming around them.
He didn’t rule out Microsoft doing something in open-source models, or the somewhat less-open “open weights” area, but there’s seemingly nothing happening on that front in the near term.
In the coming months and beyond, the focus of Farhadi’s team is helping Microsoft turn into a Frontier Ecosystem by building cutting-edge AI capabilities; helping enterprises create their own tuned versions of them; continuously improving models; and making sure customers keep control of their own destinies and data.
Success for Microsoft’s Superintelligence team has nothing to do with the idea of Artificial General Intelligence (AGI) which OpenAI, Anthropic and others have positioned as their ultimate goal over the years. In fact, when I asked Farhadi about AGI, he said, “I don’t understand what that means.”
Definitions for the AI era. (GPT-5.6 Sol Illustration, Click for larger image.)
Jargon stinks. What do the terms open weights, RAG, and agent mean exactly? Here’s a plain English, slightly snarky glossary of befuddling AI terminology with references for further reading.
AI is a broad name for the technology. Machine learning is the part where a system learns from data instead of following rules somebody wrote, a neural network is the structure that does the learning, and deep learning just means a neural network with a lot of layers.
Here’s the nitty-gritty: the terms that get used loosely, and the distinctions the loose usage hides.
1. Model, LLM, frontier model
ChatGPT is the app you open; an LLM, or large language model, is the AI running inside it.
“Frontier” isn’t a technical category at all. It means the handful of biggest and most capable models at any given moment, so the trophy keeps changing hands.
Everyone says “LLM” and hardly anyone could define it on the spot. “Frontier model” is worse. It’s a ranking, announced by the people being ranked.
Further reading:How ChatGPT Works: A Non-Technical Primer (MIT Sloan). Rama Ramakrishnan walks through the predict-the-next-word mechanism everything else is built on.
2. Prompts, tokens, parameters
A prompt is the thought, question, or instructions you provide to the LLM (plus whatever the app added before it without telling you). The LLM takes the prompt and generates words, both in its internal “thinking” process and in the answer it shows you.
Tokens are (roughly) the words going in and coming out. The model chops your prompt into tokens, then produces more of them as it answers, and they’re what the industry charges by.
Parameters, also called weights, are the numbers inside the model. A frontier model has hundreds of billions of them and the biggest now run to trillions, and nobody can tell you what any single one does.
Parameter counts get quoted like horsepower. The number nobody advertises is how many tokens it takes to answer your question, and that’s the one that shows up on the bill.
Further reading: The only AI glossary you’ll need this year (TechCrunch, July 2026). Its entries on tokens and weights are the clearest short treatment of the building blocks.
3. Pre-training, post-training, fine-tuning
Pre-training is feeding the model most of the internet, so it learns to predict the next word in a sentence. That’s the expensive part, and it produces something that knows a great deal but can’t follow an instruction.
Post-training is where people rank its answers and it learns to give more of what ranked well. Fine-tuning is post-training done by you, to somebody else’s model, on your data.
Pre-training costs hundreds of millions and gets you a model that won’t answer a question well. Post-training is what gets you the product.
From scratch, you buy (or rent) the computers and do the work to build and train a model. Distillation trains a cheap model on an expensive model’s outputs, so it inherits the behavior without the bill. Distillation is against most AI companies’ terms of service.
OpenAI accused DeepSeek of distilling its models, which is a bold position for a company that trained on the whole internet without asking. Learning from other people’s work is fine right up until the other people are you.
Training is how you build a model. Inference is what happens every time it answers: the model runs and produces a result.
Training is a one-time cost. Inference is a cost you’ll pay forever. Training runs for months and costs hundreds of millions; one inference, meaning one answer, costs a fraction of a cent, and it happens billions of times a day.
Training costs get announced. Inference costs get discovered. Only one of them shows up in a press release.
We typically use LLMs by accessing an app like ChatGPT, Claude, or Gemini. But experts often want the model itself, not just an app wrapped around it. Open weights means that an AI expert can download the model and run it on a server. You don’t get the data or the code that made it.
Open source means data and software that experts can use and modify, which almost no major model offers (AI2’s Olmo is a rare exception).
API-only means you can’t have the model at all. You send your text to the company’s computers, the answer comes back, and you pay for every use, which is also what’s happening when you use ChatGPT or Claude through an ordinary account.
Open weights is how you claim the open-source mantle without giving much away. Open washing, basically.
The context window is how much text the model can hold in mind at once, including your question and everything pasted into the conversation.
Memory is a feature that saves facts about you and slips them back into the context window later.
RAG, short for retrieval-augmented generation, searches a document collection and drops the relevant passages into the context window before the model answers.
Nothing in the model remembers you. The app keeps a file on you and pastes it in before every conversation, and that’s a less charming way to describe the same feature.
Further reading:Glossary of Terms: Generative AI Basics (MIT Sloan Teaching & Learning Technologies). Defines context window and RAG in plain language, and is careful to put the model’s “memory” in quotation marks.
8. Chatbot, workflow, agent
A chatbot answers and stops. A workflow runs the steps you defined, in your order. An agent receives a goal instead of steps, and works out for itself what to do, calling out to other software and checking the results until it’s done or stuck.
Ask about a delayed flight and a chatbot quotes you the policy; a workflow uploads the refund form you built; an agent rebooks you.
Useful test: if it decides its own next step, it’s an agent. If you decided the steps, it’s a workflow.
Further reading:Building effective agents (Anthropic, December 2024). The source of the distinction: workflows run predefined code paths, agents direct their own.
9. Hallucination, AI slop, AI cream
A hallucination is a confident falsehood, like a citation to a paper that doesn’t exist. The model isn’t lying; it has no notion of truth to violate. It’s producing text that looks like the right kind of answer.
AI slop is a different failure: accurate, fluent, and worthless. Think of the LinkedIn post that says nothing in 300 fluent words.
AI cream is the third case and the rare one: superb writing authored with the help of AI.
Nobody sets out to make slop. Everyone believes they’re making cream.
Further reading: 2025 Word of the Year: Slop (Merriam-Webster, December 2025). The dictionary definition turns on quantity: low-quality content “produced usually in quantity” by AI.
Why language models hallucinate (OpenAI, September 2025). Argues that hallucinations persist because benchmarks score accuracy alone, so guessing beats admitting ignorance.
10. Alignment, guardrails, censorship
Alignment is the research problem of getting a model to do what people want when nobody’s watching. Guardrails are the rules behind its refusals: “no, I won’t tell you how to make a bio weapon.” Censorship is a guardrail that blocked something you wanted.
The same refusal is “safety” in the press release, “guardrails” in the documentation, and “censorship” on X.
Further reading: Model Spec (OpenAI, updated December 2025). A published rulebook for what one model will and won’t do, which makes refusals arguable rather than mysterious.
I snuck in one novel term that’s been sorely absent from the field. Can you tell which one?
Further reading: other glossaries
Five general AI glossaries, listed roughly from most opinionated to most technical.
Glossary of Terms: Generative AI Basics (MIT Sloan Teaching & Learning Technologies). Twenty-odd entries aimed at people who use the tools rather than build them.
Machine Learning Glossary (Google for Developers). Hundreds of technical entries, and the only glossary here that defines “AI slop” a few lines away from several hundred pieces of real math.
Login
