The founders also say that the company, which is the biggest rival to ChatGPT, now expects to build out the Dublin office “substantially” after establishing it this year
”That has certainly been the style, typically, [in the EU] and I think that we can probably expect that to continue,” co-founder Jack Clark told the Irish Independent.
”I anticipate that we will be in pretty detailed conversations with them later this year after the EU’s AI office has staffed up over the summer.”
The newly-available services, including Claude.ai and a Claude app for iPhones, are now available free and for monthly subscriptions, depending on the level of complexity required.
Claude acts in a similar way to ChatGPT, allowing intelligible summaries of notes or long tracts of texts, writing code or performing other complex tasks.
Earlier this year, Anthropic set up an office in Dublin where it intends to base some regulatory operations, as well as employing sales and engineering staff.
For Anthropic’s co-founder president, Daniela Amodei, it was a sort-of corporate homecoming.
“I was lucky enough to get to help build out the Dublin office for Stripe when I worked for Stripe many years ago,” she told the Irish Independent.
”I worked in [the] risk and trust and safety [division] and we ended up having a pretty large operation in Dublin. We’d like to hopefully reconnect with some colleagues there. We’re actively recruiting and looking to build that Dublin office out pretty substantially.”
She said that to date, the Claude AI assistant is generally used by businesses to summarise meeting notes and help analyse other data.
“We have healthcare companies that are using Claude to do things like help with genetics research to look for cancer markers,” she said. “We also have financial institutions that are using Claude to help summarise things like SEC reports, as well as technology companies that are building on generative AI to make their tools better. For example, Slack uses Claude, for things like summarising and AI powering.”
Consumers, she said, are starting to use it to help with tasks ranging in importance.
“Many use it to summarise or analyse the things they’re working on, If you’re trying to rent a new flat somewhere, you might want it to look through, for example, your rental agreement to see if there are things amiss or something that you should be consulting with a lawyer on.”
While the basic version of Claude is free, the ‘Claude Pro’ service, which grants access to Claude 3 Opus, costs €18 plus Vat per month. A ‘Team’ plan costs €28 plus Vat per user per month, with a minimum of five seats.
Mr Clark said that the company is expecting to work closely alongside the emerging rules from the EU’s recently-passed EU Act, which bans certain practices and requires extra caution on others.
“I’ve been to Brussels a few times and we’ll be carrying on with regular visits to try to understand how they [the EU] are going to approach testing,” he said.
“There is a huge amount of detail to be worked out by the AI office.”
Mr Clark said that Anthropic trains its models according to the concept of ‘fair use’ public data.
In the US, rival AI assistant ChatGPT is facing multiple law suits from entities including the New York Times for using copyrighted data to train its models on.
“We subscribe to industry-best practices and we take the view that data at the point of ingestion is fair use,” he said.
”We respect rights holders and we build technologies to ensure that Claude doesn’t generate things that could compete with or infringe on rights holders.”
So could a sentence of an Irish Independent journalist turn up in a Claude response?
“We’re working very hard to ensure that’s not the case, even though the statistical likelihood of generating five to 10 words that are similar to something someone else has written is quite high,” said Mr Clark.
“But not if it’s a long phrase, or particularly original.”
Both he and Ms Amodei said that AI systems were currently on an “upward curve” that was so steep that even near-future applications are hard to predict.
“Five years ago, AI systems were just barely able to write a coherent paragraph,” said Mr Clark.
”Today, AI systems can write multiple pages of text. Five years ago, AI systems could write new code that didn’t work. And now today, they can write extremely complex, entire software programs in one shot. So where will we be five years from now? I think you can assume that systems will be much, much more capable than they are today and able to do kind of long term tasks already complex tasks. And I expect that we will be surprised by how inventive the AI systems will become. They’re kind of like useful colleagues and critics today, but they maybe lack the ability to have insights that feel original or are particularly interesting. I expect that in the future, we’re going to have interesting ideas that are generated entirely by the AI systems. themselves, to help businesses do interesting stuff.”
The company says that it will soon release more collaboration features, including citations from user-submitted materials, integrations with data repositories (like codebases or CRMs) and allowing work-based colleagues to jointly tend to AI-generated documents or projects.
It also says that it continues address the challenge of underlying bias in the system, citing third party accreditation in its claim to have less bias than previous models.
Ms Amodei said that the firm, which has raised €6.7bn in funding over the last year, didn’t hesitate when choosing where to locate its first EU office.
”We like to do the simple thing that just works,” she said.
”Given my experience at Stripe, we knew what it was like to launch an office in Ireland. It was also down to the level of experience of people in sales and engineering there. I think Ireland has been an incredible hub for that type of talent.”