Ai2: The AI House That Paul Allen Built

A photograph of the Ai2 office in Seattle, Washington on Oct. 31, 2024. Photo credit: Ken Yeung
, , , , ,
"The AI Economy," a newsletter exploring AI's impact on business, work, society and tech.
This is "The AI Economy," a weekly LinkedIn-first newsletter about AI's influence on business, work, society and tech and written by Ken Yeung. Sign up here.

When it comes to artificial intelligence, giants like OpenAI, Anthropic, Meta, Google, and Amazon often dominate the spotlight, leaving little room for other key players. Yet, one company has been quietly researching AI for over a decade, is based outside Silicon Valley, and is helping lead the charge for the technology’s openness. Called Ai2, this nonprofit 501(c)(3) firm, founded by the late Microsoft co-founder Paul Allen, aims to create AI for the common good.

Previously known as the Allen Institute for Artificial Intelligence, the Seattle, Washington-based company has long prioritized research and engineering. With a team of nearly 300 researchers, engineers, and staff members, it estimates it has published over 1,000 papers about AI to date. “We probably have won the most best paper awards than any other institutes of this size,” Ali Farhadi, Ai2’s Chief Executive, tells me.

My introduction to Ai2 occurred earlier this year when I wrote about its OLMo large language model for VentureBeat. With my AI exposure saturated with news from Big AI—the players, as mentioned earlier—I was curious about what a research lab in the Pacific Northwest was doing to stand out amongst the noise. Last month, I had the opportunity to sit with Farhadi in his Ai2 office. We discussed the company’s mission, the significance of making AI open, the model evaluation crisis, and what he thinks about artificial general intelligence (AGI).

From Researcher to Leader

AI2 Chief Executive Ali Farhadi. Photo credit: AI2
Ai2 Chief Executive Ali Farhadi. Photo credit: Ai2

Farhadi is a University of Washington professor and a long-time Ai2 researcher. “When Paul [Allen] started this institute, he wanted to add early on the Vision Institute, and we started conversations around this,” he says. “I decided to join Ai2 back [then]. I joined Ai2 and built the computer vision team, now known as the PRIOR team.”

One project he worked on involved developing Xnor networks for low-power, edge-based AI, which later spun off into a company called Xnor.ai. The venture was so successful that it attracted multiple acquisition offers. In 2020, Farhadi chose to sell Xnor.ai to Apple for a reported $200 million, stating it was “the one that made more sense for us.”

He would go on to join Apple’s machine learning team, leading some of the company’s AI initiatives. Farhadi declined to provide specifics but intimated that some of his work can be seen today.

In 2023, Ai2 would undergo a leadership change, with inaugural chief Oren Etzioni stepping down. Farhadi would then be tapped to succeed him, saying it was a no-brainer decision: “Obviously, it’s such a unique thing I couldn’t say no to, so I dropped everything I was working on and came back to the institute, and it’s been fun since I rejoined.”

A Renewed Focus

As CEO, he’s tried to guide Ai2 towards focusing more on the areas where its work will have a more significant impact. “The way I think about it is that the first decade of Ai2, we wanted to prove ourselves and show that we could actually be one of the best research institutes in AI,” Farhadi remarks. “Now, we’re trying to broaden [our] impact and to sort of take you to the next level. One of the things we’ve done is [narrow] our focus so we could do fewer but bigger things.”

A billboard advertises that AI2 is pro-open-source when it comes to AI. Photo credit: AI2
A billboard advertises that Ai2 is pro-open-source when it comes to AI. Photo credit: Ai2

Today, Ai2 has three main focuses. The first is on the open AI ecosystem: “The argument here is that humanity needs more openness in AI. Without it, we’re in big trouble.” It is a central part of the company’s mission, but not everyone can agree on what “open” means, so Ai2 has a whole team designed to help understand that term. So far, Ai2 has released at least two open-source LLMs, OLMo and Molmo.

The second is on an unreleased project called Nora, a research assistant agent for scientists. Not only can you interact with it (e.g., chat and ask questions), but it promises to execute code, understand literature, provide topic summarization, and more. Farhadi clarifies that it’s intentionally different from tools like GitHub Copilot or Augment Code, as their goal is to broaden its scientific scope. They plan to expand into new fields, including life sciences, starting next year.

Lastly, Ai2 is investigating how AI can improve conservation, an area of great interest to Allen, a well-respected philanthropist. One program being developed is EarthRanger, a wildlife conservation software solution used in 600 sites across 70 countries to help with animal tracking and land monitoring. Another is Skylight, which helps fight illegal fishing and protects marine biodiversity, livelihoods, and food security. And there’s also a climate modeling team.

What Sets Ai2 Apart From AI Giants

YouTube player

But how is Ai2 different from OpenAI, Google, Meta, and other AI companies? “We don’t have shareholders. We don’t have founders who say, ‘Hey guys, let’s get together and build the next billion dollars. Let’s do whatever we can to get the adoption over this chat bot.’ We don’t have any of those—at least Paul didn’t have these intentions.” Farhadi explains.

“We are really after doing the right thing, and we are really scientifically grounded. We are anti-hype, anti-buzzword. Don’t need to say B.S. to just get more attention,” he continues. “We’re trying to to be very…rooted in the science, being very careful about what we say, being truthful, but at the same time being open and collaborative. So that position gets us in a unique position because this time is really capable.”

But Ai2 isn’t alone in embracing open-source models. It’s part of a growing community made up of numerous companies, with Hugging Face, Cohere, Nvidia, Stability AI, and Meta among the most prominent.

Farhadi claims that Ai2 remains competitive by releasing models that outperform those from more prominent AI companies. “Our position is very unique, given how AI is evolving right now because there are for-profit ones, there are little non-profits that are a little too tired or won’t be able to…move the needle,” he declares. “We are situated in such a unique position that we are as capable but we don’t have the intents they do. We’re also sort of somewhere in-between as sort of a mediator between for-profit, non-profit, public sectors, and all of those things as we are providing thought leadership for this mix. That’s where we’re thinking about ourselves…”

‘AI Is Born and Raised in the Open’

Tech companies are mixed in their embrace of open AI—some companies such as OpenAI, Anthropic, and Google keep their flagship models closed while others like Meta, Mistral, and Ai2 are all-in with open-source LLMs. There’s even disagreement about what “open” means, even after the Open Source Initiative published its “official” definition.

Farhadi argues that open development is what led AI to the state it’s in today. Any achievements didn’t happen overnight nor was it developed by one team. “It’s just a communal effort, and it is going to be like that if you would like to keep innovating in the space of AI. And we are basically deploying these solutions at such a massive scale with such a shallow understanding of what we’re deploying as a whole community.”

AI-generated image depicting the contrast between open-source and closed AI models. Image credit: Adobe Firefly
AI-generated image depicting the contrast between open-source and closed AI models. Image credit: Adobe Firefly

He points out that initial claims about the exclusive power and cost of building LLMs were misleading and despite early skepticism, multiple models soon entered the market. That debunked the myth that only a few companies could develop them. Eventually, Farhadi says the narrative shifted to portray these models as being so dangerous that only a select few could handle them—a claim Ai2 and others have fought hard to push back against. He blames the “hype-making machine” run by for-profit companies and describes the action as “dangerous.” In fact, overhyping AI could lead to disappointment and set back progress, possibly resulting in a so-called “AI winter” in which excitement wanes.

“If we keep AI closed, or push more towards closing AI, this would be the most dangerous thing that could actually happen to AI and also to humanity, in my opinion, because how well can I actually build a cancer solution around these things?” Farhadi posits. “How else can I actually build a new model for these things? How else can I actually ensure safety? How else can I actually empower others to build on top of these things?”

But what does it mean to be open? Even if an LLM is open-sourced, it’s crucial to understand how much is being revealed—is it fully open with source code, weights, and training data? Or partially open with one or more of these items? There’s a wide spectrum of definitions for open models, but Farhadi doesn’t believe we should exert much energy arguing over semantics.

“The way the whole economy will evolve is we’re going to have a spectrum of models with different characteristics and properties. But, if you train a closed model, and you want to open wash it by saying you’re a leader in open source, it will cause trouble because some of these things are actually not truly open to me,” he contends.

“The hallmark, the spirit of openness, has always been, ‘I need to understand your work to the extent that I could change it to do my work.’ This change sometimes means that I need to make a change to your work early in the pipeline. Sometimes, I need to fork out of it. Sometimes, I need to grab your end product and have a right to use it. All of those things. So anything that doesn’t give me that spirit, it’s just not open source. Whatever we’re going to call it, just call it, but just not open source.”

The AI Evaluation Crisis

Benchmarks showcasing comparisons between AI2's OLMoE model and other OLMo models, as well as similarly sized models. Image credit: AI2.
Benchmarks showcasing comparisons between Ai2’s OLMoE model and other OLMo models, as well as similarly sized models. Image credit: Ai2.

The proliferation of models in the marketplace has left users confused about which LLM they should be using and why. The closest assessment they have to help in the decision-making process are benchmarks these AI vendors provide. Unfortunately, Farhadi thinks that’s the wrong way to look at it.

“We are in an evaluation crisis,” he states while admitting it’s a hard problem that no one has an answer for. “These big tables that people put out, [Ai2] built half of those benchmarks that people put out there and evaluate those things. But they’re using those benchmarks in such a ridiculous wrong way that you look at it and you’re like, ‘Wow, what are those datasets that we released?’”

“I was shocked at how a certain model actually achieves such a high number and that data, it’s impossible—and we built that data. We know…how challenging it can get, only to later on realize that…they actually gave part of the answer to the model, either intentional or by mistake. They things are just happening all over the place. So evaluation would be a hard bit. And I think we are investing in this. We’re trying to sort of figure out what are the easiest way to surface the behavior of the models. But it’s a hard problem.”

AI Safety: Not a Policy Problem, But a Technical One

At one point in our conversation, Farhadi comments that he views the conversations around safety, biases, and ethical usage of AI as hindering the technology’s adoption “to the level that we would like to see.” “All of them are technical problems more than policy or regulation problems because we’re having those conversations because we still don’t have enough understanding of these models,” he says. “There’s still an innovation gap. There’s a technology gap. We don’t have parametric control over the output of these models.”

Farhadi clarifies that he’s “not anti-policy.” In fact, Ai2 has historically worked with regulatory entities to design policy—the company is a part of President Biden’s National Artificial Intelligence Research Resource Task Force.

“There should be regulations,” he states. “This is a new kind of technology and we need to learn and adjust with these and policy and regulations will adjust to those things.” Elaborating further, “policy won’t solve the problems that we have today. If you’re worried about the safety of these models, us saying that ‘do not misuse it in the following ways,’ is not going to make it any safer. We need to solve…problems with these models. We need to figure out what they have encoded, what they have not. If I need to get something out of the systems, how do I get it out? How do I put it in? How do I correct these things in a scalable way? And those things do not exist. And the minute they exist, it becomes the job of a policy maker.”

He concedes it’s a hard problem but says having the wrong policy would be “way worse than not having a policy.”

The One Model to Rule Them All Doesn’t Exist

Farhadi acknowledges that although OpenAI’s ChatGPT is arguably the most popular LLM today, the market won’t consolidate to the point where we’re only using a handful of models. He says models are absolutely being heavily commoditized, but that’s a good thing and expected. Ultimately, the end user will need to make the right decision about which model to pick, answering how they want or don’t want to use it, how it handles unexpected scenarios, and more.

“To us, it’s phenomenal to see so many models are out there that should be the way it is. And I don’t believe in the future where there are a small number of models. And I also don’t believe it technically, with this philosophy, that there exists this God-given model that does everything,” he explains. “It will become a generic thing that’s good at 80 percent to 85 percent of the time. It’s a great demo ware, but you cannot ship with it the same way we’re struggling with this. But rather, there’s going to be a ginormous ocean of models, each of which will be built to do certain things really well, and to be able to sort of support the community…”

‘AGI Doesn’t Make Any Sense’

At the end of the interview, Farhadi shared his thoughts on the future of AI, including the concept of Artificial General Intelligence. He calls AGI “marketing jargon” and says that it “doesn’t make any technical sense.” It’s something the Ai2 chief says we can’t value it nor can we think about it more mathematically. He jokes that if any of his students at the University of Washington uses the acronym, they’d delay their graduation “by six months.”

Commenting on the state of AI, he predicts that over the next year, we’ll see increased momentum and efficiency, including the rise of additional smaller models. Farhadi notes three key trends observed by Ai2: the gap between open and closed models is narrowing; the gap between small and closed models is also decreasing—highlighting how Ai2’s Molmo 1B model outperforms a model 12 times its size; and finally, a shift where “less is now the new more” will become evident.

People working on a whiteboard. Photo credit: AI2
People working on a whiteboard. Photo credit: Ai2

“For example, with Molmo’s release, we’re outperforming the Llama class of models that came after us. The number of images that we use to train our model is 0.0001 of what they’ve used, like 6 billion text pairs they used. We’re using 600,000 or 700,000. So this, to me, means one thing, and that one thing is my prediction of how AI evolves: that these kinds of technologies are now more accessible to a broader set of practitioners and players. And when that happens, the whole wave in the community will get us to very interesting places.”

He continues, “Next year will be the year where we’ll see a wider and broader group of practitioners getting involved in these models, using them in their own creative ways. And I think with this, I hope that we actually hit the jackpot of having the killer app. We still don’t have it in AI. This is game-changing for humanity, but we’re not yet seeing the killer apps. And by just making it more accessible, you’re going to just increase your chances of hitting those killer apps.”

Updated Dec. 10, 2024: Corrected Ai2's conversation efforts, clarifying EarthRanger as a program, a wildlife conservation software solution, and spans 600 sites, not national parks. In addition, Skylight was revised to state that it's focused on helping stop illegal fishing, not trafficking.
Featured Image: The Ai2 office in Seattle, Washington taken on October 31, 2024. Photo credit: Ken Yeung

Today’s Visual Snapshot

A chart showing the growth in AI usage over time between January 2023 and August 2024. Image credit: Slack
A chart showing the growth in AI usage over time between January 2023 and August 2024. Image credit: Slack

Slack has published its Fall 2024 Workplace Index, which shows that excitement around artificial intelligence is cooling among workers. This tempering is believed to be driven by a decrease in U.S. respondents saying they’re excited about AI helping them complete tasks at work. “With so many businesses making AI investments right now, these findings are a real wakeup call to leaders,” Christina Janzer, Slack’s Workforce Lab lead, writes. “With sentiment around AI dropping, businesses need to help employees accelerate their AI journey and address the cultural and organizational blockers standing in their way.”

Quote This

“The big novelty is that every student can now have access to a personalized AI tutor throughout their life and explore any subject, including the most inaccessible ones. Access to knowledge has no limits. Of course, we must be aware of AI’s potential risks, but we must encourage our children to be more ambitious, more curious, and to use AI as a learning tool.”

— Microsoft Chief Executive Satya Nadella responding to a question about how we teach children to prepare them for the AI world. (Le Point)

This Week’s AI News

🏭 AI Trends and Industry Impact

🤖 AI Models and Technologies

✏️ Generative AI and Content Creation

💰 Funding and Investments

☁️ Enterprise AI Solutions

⚙️ Hardware, Robotics, and Autonomous Systems

🔬 Science and Breakthroughs

💼 Business, Marketing, Media, and Consumer Applications

🛒 Retail and Commerce

⚖️ Legal, Regulatory, and Ethical Issues

💥 Disruption, Misinformation, and Risks

🔎 Opinions, Analysis, and Editorials

🎧 Podcasts

End Output

Thanks for reading. Be sure to subscribe so you don’t miss any future issues of this newsletter.

Did you miss any AI articles this week? Fret not; I’m curating the big stories in my Flipboard Magazine, “The AI Economy.”

Follow my Flipboard Magazine for all the latest AI news I curate for "The AI Economy" newsletter.
Follow my Flipboard Magazine for all the latest AI news I curate for “The AI Economy” newsletter.

Connect with me on LinkedIn and check out my blog to read more insights and thoughts on business and technology. 

Do you have a story you think would be a great fit for “The AI Economy”? Awesome! Shoot me a message – I’m all ears!

Until next time, stay curious!

Subscribe to “The AI Economy”

New issues published on Fridays, exclusively on LinkedIn

Subscribe

Leave a Reply

Discover more from Ken Yeung

Subscribe now to keep reading and get access to the full archive.

Continue reading