Cohere co-founder Nick Frosst says throwing more computing power at large-language models (LLMs) won’t necessarily get the world any closer to so-called artificial general intelligence (AGI).
On the 20VC podcast, hosted by Harry Stebbings, Frosst took a contrarian spin on industry hype around AGI. The computer scientist and indie musician called out OpenAI CEO Sam Altman for both touting the technology’s power and raising concerns about its potential threat to humanity.
Frosst doesn’t believe that pure scaling could create AGI, a nebulous concept which he defines as “treating a computer like a person.”
“I think that was academically disingenuous and did a disservice to the technology he loves,” Frosst said.
Frosst has rejected the notion that AI models pose an existential threat to humanity. He even pushed back on his former mentor and Nobel Prize winner Geoffrey Hinton, who pioneered the technology underpinning LLMs.
Frosst said he doesn’t believe that pure scaling could create AGI, a nebulous concept which he defines as “treating a computer like a person.”
Instead, he preferred to dig into where he believes AI tools can be truly useful: the workplace. It’s in line with the value proposition of Cohere, which is Canada’s leading foundational model developer. The company caters specifically to security-minded enterprises and governments. Its flagship platform, North, aims to speed up and automate tedious tasks at work, such as creating reports and sending emails.
Frosst said the release of OpenAI’s newest GPT-5 model, which Stebbings thought was “worse” than its predecessor despite using more computing power to train, showed that throwing more hardware at the problem may not be the solution for building a better product.
Cohere announced last month that it had raised $500 million USD at a $6.8-billion USD valuation (Frosst said that the founding team celebrates every funding round with a trip to McDonald’s). The new round brought Cohere’s total funding to roughly $1.5 billion USD. This is far less than its better-capitalized counterparts building foundational models, such as OpenAI, Google, and Anthropic—all of which are working on developing AGI-like technology.
Though he acknowledged that a significant portion of Cohere’s funding goes toward computing power, Frosst emphasized that efficiency is a key part of the brand’s business strategy. This echoed comments he made earlier this year after Chinese AI developer DeepSeek released a model more efficient than its leading competitors at the time. Cohere’s recent model releases of Command A Reasoning and Command A Vision were both trained to fit on two graphical processing units (GPUs), Frosst said, allowing them to serve customers “bottlenecked” by the price of GPUs.
Frosst is a former employee of Hinton at Google Brain. His conversation offered an in-depth look at how Cohere leadership approaches the limitations of the technology it’s building. CEO and co-founder Aidan Gomez said on the show last year that while more compute is the most “trustworthy” way to improve AI models, it’s also the “dumbest” and lowest-risk.
However, Frosst said that data availability is still a bottleneck for improving model function. He said Cohere uses synthetic data (generated by AI models) to train its own models, where the AI model can be tested in “fake companies” before being deployed in real ones. Cohere was sued by major media companies this year for allegedly infringing on copyrighted works to train its models.
However, Cohere has not been subject to any definitive regulations on AI development in Canada, as the federal government has yet to move forward with related legislation.
Frosst said the worst-case scenario with AI regulation would involve “an erroneous understanding of the technology.”
“The worst thing a regulation could do would say, ‘Hey, we’re going to pick this random benchmark, we think that represents AGI, and we’re going to shut down any development on it,’” he said.
Canada’s AI and digital innovation minister Evan Solomon has said he envisions “light, tight, right” regulations on AI technology, while promoting its widespread adoption for Canadian businesses. Solomon has also championed Cohere publicly, most recently signing a non-binding agreement for the company to explore improvements to the public service with AI.
Feature image courtesy 20VC via YouTube.