Hello everyone, welcome to TechCrunch's regular AI newsletter. If you'd like to have this sent to your inbox every Wednesday, sign up here.
An agent, an AI agent, is coming.
This week, Anthropic released an upgraded version of its latest AI model, Claude 3.5 Sonnet. It can interact with web and desktop apps by clicking and typing, just like a human would. It's not perfect. But 3.5 Sonnet, which involves what Anthropic calls “computing,” could revolutionize the workplace.
At least, that's the elevator pitch.
It remains to be seen whether Anthropic's new model lives up to the hype. But its arrival signifies Anthropic's ambitions in the nascent AI agent market, which some analysts believe will be worth nearly $50 billion by 2030.
Anthropic isn't the only company investing resources in developing AI agents. In a broad sense, AI agents automate tasks that previously had to be performed manually. Microsoft is testing agents that can use Windows PCs to make reservations and more, and Amazon is actively considering agents that can be purchased.
Organizations may be confused about generative AI. But they're pretty bullish on the agency right now. A report released this month by MIT Technology Review Insights found that 49% of executives believe that agents and other forms of advanced AI assistants can improve efficiency and reduce costs.
This is certainly welcome news for Anthropic and its rivals, which are developing “agent” technology. Building and running AI isn't cheap. Case in point: Anthropic is said to be in the process of raising billions of dollars in venture funding, and OpenAI recently closed a $6.5 billion funding round.
However, it is questionable whether most agents today can truly live up to the hype.
Take Anthropic, for example. In an evaluation aimed at testing the AI agent's ability to assist airlines with reservation tasks, the new 3.5 Sonnet successfully completed less than half of the tasks. In a separate test that included tasks such as initiating product returns, the 3.5 Sonnet failed about a third of the time.
Then again, the new 3.5 Sonnet isn't perfect. Anthropic freely admits this. But it's hard to imagine companies tolerating such high failure rates for long periods of time. At some point, it will be easier to hire a secretary.
Still, companies are showing a willingness to experiment with AI agents, if for no other reason than to keep up with the Joneses. According to research by startup accelerator Forum Ventures, 48% of enterprises have started deploying AI agents, and a further third are “actively considering” agent solutions.
Let's see how early adopters feel once the agent has been up and running for a while.
news
Data scraping protest: Thousands of creators, including actor Kevin Bacon, novelist Kazuo Ishiguro, and musician Robert Smith, have signed a petition against the unauthorized use of their creative works for AI training.
Meta tests facial recognition: Meta says it is expanding testing of facial recognition as an anti-fraud measure to combat fraudulent celebrity advertising.
Perplexity sued: News Corp's Dow Jones and the NY Post have sued Perplexity, a growing AI startup reportedly seeking to raise money over what the publisher calls “content theft politics.”
OpenAI's new hires: OpenAI has announced the company's first chief economist, former U.S. Department of Commerce Chief Economist Aaron Chatterjee, and new chief compliance officer, Scott Schools, previously Uber's head of compliance. was adopted.
ChatGPT comes to Windows: In other OpenAI news, OpenAI has begun previewing a dedicated Windows app for ChatGPT, its AI-powered chatbot platform, for certain customer segments.
xAI's API: Elon Musk's AI company, xAI, has launched an API for Grok, a generative AI model that powers many of the features on top of X.
Mira Murati raises money: Former OpenAI CTO Mira Murati is reportedly raising money for a new AI startup. The venture is said to be focused on building AI products based on proprietary models.
This week's research paper
Military forces around the world have shown significant interest in deploying, or already deploying, AI in combat zones. It's certainly controversial, and it's also a national security risk, according to new research from the nonprofit AI Now Institute.
The study found that AI being deployed today for military intelligence, surveillance, and reconnaissance already poses a risk because it relies on personal data that can be stolen and weaponized by adversaries. It turns out. They also have vulnerabilities such as prejudice and a tendency to hallucinate, for which there is currently no cure, the co-authors wrote.
This research is not against militarized AI. However, it states that to protect military AI systems and limit their damage, AI must be created independent of commercial models.
this week's model
This week has been a very busy week for generated AI video. Three startups have released new video models, each with unique strengths: Haiper's Haiper 2.0, Genmo's Mochi 1, and Rhymes AI's Allegro.
But what really caught my eye was Runway's new tool called Act-One. Act-One uses video and audio recordings as input to create animations and produce “expressive” character performances. When a human actor performs in front of the camera, Act-One transforms it into an AI-generated character that maintains the actor's facial expressions.
Image credit: Runway
Indeed, Act-One is not itself a model. Rather, it's a control method to guide Runway's Gen-3 Alpha video model. But it's worth highlighting the fact that the AI-generated clips this feature creates don't immediately fall into uncanny valley territory, unlike most composite videos.
grab bag
AI startup Suno is being sued by a record label for allegedly using copyrighted songs to train its music generation tools without permission, but doesn't want to get into any more legal trouble. .
At least, that's the impression I get from Suno's recently announced partnership with content ID company Audible Magic, which some readers may know from the early days of YouTube. Suno said it will use Audible Magic's technology to prevent copyrighted music from being uploaded to the Cover feature, which allows users to create remixes of any song or sound.
Suno told the label's lawyers that he believes the music he used to train the AI falls under the U.S. fair use doctrine. That's debatable. However, even if a platform stores full copyrighted works on its servers and encourages users to share them, that doesn't necessarily help Suno's case.