Jornal Do Brasilia - Anthropic says looking to power European tech with hiring push

The OpenAI competitor wants to be "the engine behind some of the largest startups of tomorrow... (and) many of them can and should come from Europe", Krieger said.

Tech industry and political leaders have often lamented Europe's failure to capitalise on its research and education strength to build heavyweight local companies -- with many young founders instead leaving to set up shop across the Atlantic.

Krieger's praise for the region's "really strong talent pipeline" chimed with an air of continental tech optimism at Vivatech.

French AI startup Mistral on Wednesday announced a multibillion-dollar tie-up to bring high-powered computing resources from chip behemoth Nvidia to the region.

The semiconductor firm will "increase the amount of AI computing capacity in Europe by a factor of 10" within two years, Nvidia boss Jensen Huang told an audience at the southern Paris convention centre.

Among 100 planned continental hires, Anthropic is building up its technical and research strength in Europe, where it has offices in Dublin and non-EU capital London, Krieger said.

Beyond the startups he hopes to boost, many long-standing European companies "have a really strong appetite for transforming themselves with AI", he added, citing luxury giant LVMH, which had a large footprint at Vivatech.

- 'Safe by design' -

Mistral -- founded only in 2023 and far smaller than American industry leaders like OpenAI and Anthropic -- is nevertheless "definitely in the conversation" in the industry, Krieger said.

The French firm recently followed in the footsteps of the US companies by releasing a so-called "reasoning" model able to take on more complex tasks.

"I talk to customers all the time that are maybe using (Anthropic's AI) Claude for some of the long-horizon agentic tasks, but then they've also fine-tuned Mistral for one of their data processing tasks, and I think they can co-exist in that way," Krieger said.

So-called "agentic" AI models -- including the most recent versions of Claude -- work as autonomous or semi-autonomous agents that are able to do work over longer horizons with less human supervision, including by interacting with tools like web browsers and email.

Capabilities displayed by the latest releases have raised fears among some researchers, such as University of Montreal professor and "AI godfather" Yoshua Bengio, that independently acting AI could soon pose a risk to humanity.

Bengio last week launched a non-profit, LawZero, to develop "safe-by-design" AI -- originally a key founding promise of OpenAI and Anthropic.

- 'Very specific genius' -

"A huge part of why I joined Anthropic was because of how seriously they were taking that question" of AI safety, said Krieger, a Brazilian software engineer who co-founded Instagram, which he left in 2018.

Anthropic is still working on measures designed to restrict their AI models' potential to do harm, he added.

But it has yet to release details of its "level 4" AI safety protections foreseen for still more powerful models, after activating ASL (AI Safety Level) 3 to corral the capabilities of May's Claude Opus 4 release.

Developing ASL 4 is "an active part of the work of the company", Krieger said, without giving a potential release date.

With Claude 4 Opus, "we've deployed the mitigations kind of proactively... safe doesn't have to mean slow, but it does mean having to be thoughtful and proactive ahead of time" to make sure safety protections don't impair performance, he added.

Looking to upcoming releases from Anthropic, Krieger said the company's models were on track to match chief executive Dario Amodei's prediction that Anthropic would offer customers access to a "country of geniuses in a data centre" by 2026 or 2027 -- within limits.

Anthropic's latest AI models are "genius-level at some very specific things", he said.

"In the coming year... it will continue to spike in particular aspects of things, and still need a lot of human-in-the-loop coordination," he forecast.

E. da Cruz--JDB