Cohere CEO and president on funding, Hinton comments and LLMs

Back in February, I chatted with Cohere cofounder and CEO Aidan Gomez about the fact that the Toronto-based company, which competes with OpenAI in the LLM space, was “crazy under the radar.” 

It didn’t take long for that to change: Three weeks ago, Cohere, founded in 2019 by Gomez, Ivan Zhang and Nick Frosst, announced it had raised a fresh $270 million with participation from Nvidia, Oracle, Salesforce Ventures and others — valuing the company at over $2 billion.

It felt like good timing to circle back with Gomez as well as Cohere’s president, Martin Kon, to talk about the new funding. But the conversation turned out to be wide-ranging, from the company’s cloud-agnostic stance and Gomez’s take on Geoffrey Hinton’s recent comments on AI risk to the future of LLMs and synthetic data. (Editor’s note: This interview has been edited for length and clarity.)

VentureBeat: Back in February, Aidan and I chatted about Cohere flying under the radar. Does it feel like that has totally changed now, thanks to the new funding? 

Aidan Gomez: I think we’re making progress, but I still feel like we’re crazy under the radar. We’re well-known within certain circles, but in terms of broad awareness, we still have work to do. We’re still trying to be out there telling our story and trying to get people aware of core models, the sort of deployment scenarios where we’re a good fit, which are data-private and cloud-agnostic. 

Martin Kon: I agree with Aidan. I think our last round was a good proof point for how some of the world’s most respected enterprises see us and how much they want to support an independent, cloud-agnostic, state-of-the-art LLM company like Cohere, financially and as partners. That is not a claim, but corroborated in the marketplace; but there still is quite a bit to do in terms of general awareness. 

VentureBeat: You talk a lot about Cohere as independent and cloud-agnostic. That sounds a bit to me like Nvidia, as far as how they partner with all the different cloud companies. Do you see it that way? 

Kon: We are, by definition, cloud-agnostic. Certainly Nvidia’s technology is available on all cloud providers. Some of them have proprietary silicon as well, but Nvidia certainly is a bit of a flexible option from a compute perspective. So it was important for us to be deployable across every cloud environment with a technology that’s able to move around.

Gomez: We’re not beholden to any of the large cloud providers, and for our customers that is a key strategic advantage. Many large enterprises are multicloud. And even if they’re single-cloud, they want to preserve the ability to negotiate. With Cohere you can flip between cloud providers and have Cohere models running on all of them simultaneously.

VentureBeat: Do you consider that to be a weakness for OpenAI for Enterprise? If customers can only use Azure, for example?

Kon: Different enterprises will have different things that are important to them. We certainly heard feedback from the market — I’ve spoken to over 100 senior executives and enterprises since joining Cohere, to understand really what’s important to them. Many of them say data privacy, data protection, the ability to customize their data in our safe environment with our data residency requirements, our data protection requirements, with our access and rights requirements, that certainly seems very, very important. So what we’ve chosen to do seems to fall on fertile ground.

VentureBeat: Cohere’s list of investors is getting longer, from Oracle and Nvidia to Salesforce and VCs, as well as researchers like Geoffrey Hinton and Fei-Fei Li. How important is that variety? 

Gomez: I view that as a huge asset to Cohere. In this latest round, the whole goal was to bring together an international global mix of strategies and institutional investors to back us now and into the future. I think it’s pretty extraordinary and quite unique. Not a lot of companies are able to bring together an international set of investors on the strategic side and on the institutional side. I think in our space, you see a lot of large strategic single-player investments, like one big corporate entity [piling] some money behind one of the large language model players. We explicitly wanted to avoid that and to create something much more financially healthy for our future.

VentureBeat: Bloomberg said the other day that Cohere is reportedly in talks to raise more. Is there anything that you can say about increasing that range of investors?

Kon: I’m amused that rumors are already popping up. I hadn’t read that. But we don’t comment on speculation. Just to echo what Aidan was saying, I think a lot of our major investors are not just benefiting by investing to have that revenue come back for them, but they’re investing to really support this kind of independent provider. I think the nature of these companies is very focused on security. For example, Oracle has always been very focused on security and we share a lot of common priorities around data protection. We were quite happy to find partners like that and hopefully the signal to the market shows the faith they have in our approach.

VentureBeat: Aidan, given that you and Cohere cofounder Nick Frosst both come from Google Brain, and Geoffrey Hinton is on your list of investors, do you have any comments on his recent comments about AI risk and leaving Google Brain?

Gomez: I like Geoff. He is, I would say, the global expert on AI and deep learning. So I respect his thoughts and opinions and I take them extremely seriously. When Geoff talks, I listen. That being said, we do have differing opinions on the profile of risks for this technology. I think he’s more focused on risk to humanity at large or what some people call x-risks, or existential risks. I find those to be a lower priority than another category of risks, which are more near-term or midterm stuff like synthetic media and the dissemination of false information. Risks like deploying these models in scenarios where they’re not yet appropriate, for instance, where the stakes are too high. My focus is much more towards those tangible risks as opposed to hypothetical future ones.

At the same time, we need people focused on a spectrum of risks. And I think it’s great that Geoff is calling attention to that side. I wish that there were more attention on the risks that are coming up or immediate. I think it’s a less compelling story, because obviously, the sci-fi narratives of terminators or AI taking over and wiping out humanity have been around since before computers. They’re kind of embedded in the public consciousness and so they get a lot of clicks and attention. I wish people would spend more time on the risks that are more tangible, present-day, and, frankly, more relevant to policymakers and the public.

VentureBeat: On that same tip, I was surprised by survey research that said that something like 42% of CEOs actually believed that AI could lead to humanity’s extinction in the next 10 years. Do you hear any of this from the people that you speak with at companies?

Kon: I’ve never heard that. I think the executives that we’ve been talking to, they are concerned, but they’re concerned about some of the things like Aidan just mentioned, as well as things like bias. If you look at probably everything that Sara Hooker, who leads the Cohere for AI research group, and her team are focused on, the network of hundreds of researchers around the world that she convenes and brings together, it’s on those risks that are happening today, these systems that are deployed now. 

VentureBeat: I’m curious about issues like hallucination and bias that are really in the news right now. How do you explain to customers that it’s possible to have a large language model where those problems can be controlled or dealt with?

Gomez: I think it is an education project that we’re certainly trying to drive with any of our customers who come to us and say they have this idea for an LLM application. You try to talk about the opportunities and there’s so much this technology does exceptionally well. But there are places where it’s just not appropriate to deploy. And so you just need to educate the customer about that — let them know about what failure modes might look like and how they can mitigate the sorts of systems and processes that they can implement on their side, like constant benchmarking and evaluation of models. 

We ship a new model every single week. And we don’t want a customer adopting something if it’s going to make the experience worse for their users, or if it raises some risk profile in a way that they don’t want. So we educate them about building test sets on their side, constantly evaluating each new release of the model and making a decision: Do I want to accept that new model and push that into production? Or do I want to hold off this week? And then, in addition to that, we’re also always listening to the customer. So if they observe some sort of drift or some sort of behavior change that negatively affects their experience, we immediately jump on diagnosing why that happened, what changed on our end that led to that change on their side. 

VentureBeat: What do you say to the whole debate around enterprises implementing models on their own data using open-source models versus something like Cohere?

Gomez: My take is that open source is fantastic. I think they’re making great progress on technology. That being said, there’s still a gap between open source and our models. It’s also that these models are never static. Like I was saying, we ship every single week, there’s consistent improvement over time. With open source, a new model comes out a few times a year, this one might have a license that lets you use it, this one might not. And they might have different skews in the training data that bias their performance one way or the other. With Cohere, what you get is the ability to influence our model direction on a really fast cadence. So you’re going to get something that is much better at the task you care about, and that you’ll actually have a plausible influence over the underlying training itself. So while I think open source is fantastic, I still think the enterprise just provides a completely different value proposition. It’s just like a different product entirely.

VentureBeat: What do you say to folks who say LLMs from companies like Cohere, OpenAI and Anthropic are a black box, that they can’t see what’s in your training data or what you are doing under the hood?

Gomez: I mean, we try to be as transparent as we can be, without giving away IP. With our customers, whenever they have a question about the data that our models are trained on, we always answer it. We always give them a concrete answer. And so if they have particular concerns or questions, they’re going to get them answered. We care a lot about the provenance of our data, our ability to track what sources it’s coming from, the ability to screen for data that is toxic and remove it before it gets into the model. All of those factors we’ve cared a lot about, and also whether we have permission to train on that data, we strictly adhere to robots.txt. And we don’t scrape stuff that we shouldn’t scrape. So for our customers, we’re very very strict. 

VentureBeat: There was a recent study that examined how different large language models would fare under the new EU AI Act. What would that mean to comply with something like that?

Gomez: Well, [the EU AI Act] is still a draft of course, but you can go back and look at it and see that there was a stand for results. Cohere was in there, I think we were quite happy where we ended up alongside many of the industry leaders. But it’s still early days, because that’s just a draft legislation and there’s still a lot of work to do to figure out how that’s going to be deployed. But I think that it is one example [of the fact] that we are already doing things that are aligned with at least the intent of some of the things that are going to be protected. We don’t wait for regulation and then start thinking about it. This is something that’s very important to Cohere, the proactive adherence, we think about this all the time, not just when we’re forced to.

VentureBeat: My last question is about the future of LLMs. There was a paper we covered recently about model collapse occurring if LLMs are trained with synthetic data over time. Since it’s the sixth anniversary of the Transformers paper, and you were an author on that, what do you see as the future limits of LLMs? Will they get smaller? 

Gomez: I definitely don’t think they’re going to get any smaller. That would be kind of a huge surprise. And I think contrary to that paper, I think the future is synthetic data.

VentureBeat: You don’t think it will lead to model collapse.

Gomez: So “model collapse” is definitely an issue in that domain, like losing other external information and just focusing on what it already knows. But I think that’s actually a symptom of the methodologies that we’re applying today, as opposed to something fundamental with synthetic data. I think there’s a way in which synthetic data leads to the exact opposite of model collapse, like information and knowledge discovery, expanding a model’s knowledge beyond what it was shown in its human data. That feels like the next frontier, the next unlock for another very steep increase in model performance, getting these models to be able to self-improve, to expand their knowledge by themselves without a human having to teach them that knowledge.

VentureBeat: Why do you think that? 

Gomez: I think that because we’re starting to run up on the extent of human knowledge. We’re starting to run up on the extent and breadth of the data we can provide these models that give them new knowledge. As you start to approach the performance of the best humans in a particular field, there are increasingly few people for you to turn to to get new knowledge. So these models have to be able to discover new knowledge by themselves without relying on humanity’s existing knowledge. It’s inevitable. I see that as the next major breakthrough.

VentureBeat: And you believe that will happen? 

Gomez: Totally. Absolutely. 

Originally appeared on: TheSpuzz