Anthropic Founder Interview: Multimodality, AI Assistants, And The Next 2 Years In AI

Dario Amodei, founder of Anthropic (the company behind Claude.ai, and essentially Amazon’s AI play) recently gave an interview. Obviously at his scale every interview contains some amount of PR-talk, but it’s nonetheless an interesting interview.

Here are some key talking points from the conversation:

• Dario Amodei’s background: Studied physics and computational neuroscience. Got interested in AI in 2014 and joined OpenAI in 2016. Co-founded Anthropic in 2021 with the goal of building AI safely.
• Scaling of AI models: Dario was an early believer that simply scaling up AI models with more data and compute would lead to increasingly powerful capabilities. He saw the potential with models like GPT-2 in 2019.
• Starting Anthropic: Left OpenAI with some others in 2021 to start Anthropic with a focus on AI safety. Wanted an organization fully devoted to incorporating safety principles.
• Anthropic’s approach: Focused on model interpretability, human oversight (“constitutional AI”), and responsible scaling. Want to set an example for the industry on how to build powerful AI safely.
• Business model: Mostly focused on enterprises currently, but also offer a consumer product. Customers appreciate the safety and oversight features.
• Funding: Raised over $5B so far from investors. Majority goes to compute to train models. Will spend multiple billions on future models.
• AI safety concerns: Worried about misuse by bad actors, as well as uncontrolled behaviors from advanced AI systems. Thinks chance of catastrophe is 10-25%.
• Long-term optimism: If risks are mitigated, sees huge potential for AI to help with things like curing diseases, extending human lifespan, and more. Hopes Anthropic’s approach can help reduce risks.

What’s more, he mostly talks with enterprise customers in mind, but nonetheless even if you’re a startup or small business owner, there are some relevant insights for you:

• Enterprise adoption of AI: Dario discussed how Anthropic is seeing strong enterprise adoption of their AI assistant Claude, with applications in areas like legal, finance, accounting, and more. This suggests AI could be valuable for improving workflows and productivity in many professional services fields.
• Long-term planning: Dario advised enterprise customers to think ahead 1-2 years and build products/services around where AI capabilities will be, not just what they can do today. Small businesses could similarly think long-term about how AI may augment or transform their work down the road.
• Cost reductions: Dario noted Anthropic’s models are more cost-effective compared to some competitors. If this trend continues, it could make AI more accessible for smaller companies over time.
• Safety as a priority: Dario emphasized Anthropic’s focus on building safe, beneficial AI. Small businesses integrating AI may also want to prioritize ethics, oversight and accountability in their adoption.
• Interpretability: Dario discussed Anthropic’s work on model interpretability as both a safety measure and a feature that enterprise customers value. Small businesses may also want transparency into how any AI they use makes decisions.
• Consumer applications: While focused on enterprise, Dario noted Anthropic does have a consumer product as well. Continued progress in consumer AI could present opportunities for new products, services and content.

Now for me the most interesting bit in there was THINKING AHEAD 2 YEARS. Where will AI be in 2025, what will it be capable of?

Dario didn’t go into specifics about where he sees AI capabilities in 2 years, but here are some themes:

• Overall, he predicts AI models will continue to rapidly improve in capabilities. He expects models 2 years from now will “wow people” compared to today, even though they already impress many people currently.
• He sees multimodality (text, images, video, etc.) as a big part of future progress. Anthropic and other companies have already started incorporating this.
• He believes AI assistants will reach human-level performance on many professional knowledge work tasks in the next 2-3 years. This includes areas like science, engineering, legal, medical, and more.
• He thinks there will be a focus on making AI systems better at working cooperatively and productively with humans. Models still have limitations today in complementing human workflows.
• He expects models will continue to get better at core skills like reading comprehension, reasoning, summarization, translation and more in the next couple years.
• However, he doesn’t foresee AGI-level capabilities in the next couple years. Major breakthroughs like scientific discoveries, controlling the physical world, or machine consciousness are further out.

Here are some more details on those topics from Dario’s perspectives:

On multimodality:

• Dario sees the combination of text, images, video and other modalities as a major part of future progress in AI.
• Anthropic and other companies have already started developing multimodal models, like Claude can process both text and images.
• Multimodality will allow AI systems to better understand the real world and communicate more naturally with humans.
• It opens up new applications like describing images, generating multimedia content, AR/VR experiences.
• Challenging technical aspect, but capabilities will steadily improve over next couple years.

On AI assistants:

• Dario believes AI assistants will reach human-level performance across many professional knowledge worker domains in the next 2-3 years.
• This includes areas like science, engineering, legal work, finance, medical diagnosis and more.
• Assistants will be able to analyze documents, data and media as well as or better than human professionals in many cases.
• For example, already demonstrating ability to read legal contracts, financial documents and answer questions about them.
• Sees great potential for productivity gains and augmented human capabilities. AI won’t replace humans entirely soon but will become increasingly capable partners.

On complementing human workflows:

• Right now AI still has limitations in smoothly integrating into real-world human workflows and environments.
• Expects to see a lot of focus on making AI systems better at cooperating with humans over next couple years.
• This means building UI/UX that feels natural to humans, able to have nuanced bidirectional conversations, and physically interacting with the world.
• It’s challenging to match the common sense and adaptability of humans. Still limitations today in general reasoning and contextual understanding.
• But steady progress expected in making AI a more seamless and intuitive partner for augmenting human capabilities.