Keep up with the innovative tech transforming business
Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.
It’s been several months since Dario Amodei, former VP of research at OpenAI, the leading research-lab-turned-for-profit, announced he was leaving the company.
Now, he’s announced a brand-new AI safety startup fueled by $124 million in funding. That’s 6.5x bigger than the average Series A.
Large and in charge
While OpenAI is bullish on its status as an AI “research and deployment” company, Anthropic bills itself as AI “safety and research.”
Amodei and his sister Daniela, who is cofounder and president of Anthropic, both worked on GPT-3—a breakthrough large language model—during their time at OpenAI. Around 12 other OpenAI research alumni have reportedly joined Anthropic’s ranks.
Rewind: GPT-3 can be used for just about anything English language-related: auto-writing email responses, making up adventure stories, and more. But since GPT-3 and similar language models learn from large bodies of text (i.e., the internet as we know it), they absorb a lot of harmful biases, too.
- When researchers asked GPT-3 to complete a sentence containing the word “Muslims,” over 60% of results involved violent language, e.g., words like terrorism, murder, and bomb.
OpenAI is working on safety, too, by adding humans into review processes and building out a new use-case monitoring pipeline.
A milli x 124
Large language models are such complex, advanced tools that it’s difficult to identify the biases they contain—let alone figure out how and why they adopted them.
That’s likely why the Amodei siblings’ goal with Anthropic is to create “reliable, interpretable, and steerable AI systems,” as well as to build in more human feedback in the development and deployment process. Some big names are trusting them to do it...
Follow the $$: Skype cofounder Jaan Tallinn led the startup’s Series A round, and he was joined by Dustin Moskovitz, cofounder of Facebook and Asana, and Eric Schmidt, former longtime leader at Google and Alphabet.
Dreams ➡️ reality
Despite the fanfare surrounding Anthropic, we don’t have a lot of details yet on how, exactly, it plans to do what it's...planning to do. Anthropic's goal is a traditionally difficult one—AI explainability has been a challenge from the field’s inception.
Bottom line: For now, according to Daniela Amodei, the Anthropic team will be “focused on pushing forward our research for the next few months and are hoping to have more to share later this year.”—HF