AI Roundup 090: Computer Use
October 25, 2024.
Computer Use
Anthropic released a slew of updates to Claude, including a revamped Claude 3.5 Sonnet, a new Claude 3.5 Haiku, and the ability to run code and use computers.
Here's the latest:
Computer Use is fascinating - Claude can now take screenshots and use a mouse and keyboard to navigate a computer. It's still clunky, but Twitter is already filling up with demo videos.
The new version of Claude 3.5 Sonnet scores quite high on coding benchmarks (there's quite a lot you can build with it), and also seems to have undergone some overall "personality" changes.
It's curious/spooky to see these new "personalities," so to speak - with computer use, we've learned that Claude would stop to take a break in its task to browse pictures of Yosemite.
Elsewhere in frontier models:
Stability AI released Stable Diffusion 3.5 Large, an 8B-parameter model that makes 1MP images, along with 3.5 Large Turbo.
Meta debuted Spirit LM a multimodal text and speech model, as well as new quantized versions of Llama 3.2, designed to run on low-power devices.
And while sources claim OpenAI will launch its next flagship model "Orion" by December, CEO Sam Altman dismissed the reports as fake news.
The copyright fight continues
Despite the litany of AI copyright cases working their way through the courts, new AI products mean new AI lawsuits.
Between the lines:
Dow Jones and the New York Post are suing Perplexity over alleged copyright infringement - though seemingly not for its training of AI models. Perplexity, for its part, remains defiant.
There are also the tens of thousands of creatives who are publicly taking a stand against using unlicensed creative work for AI training.
But even if it becomes illegal to train AI on unlicensed content, the biggest players will simply solve the problem with money - Meta is the latest to strike a deal with a publishing giant (Reuters, in this case).
Perplexity criticizes media companies suing over AI, claiming they prefer corporate ownership of reported facts.
Elsewhere in AI and media:
OpenAI and Microsoft announced a $10M AI local news project operated by the Lenfest Institute, starting with five US metro news organizations.
Penguin Random House amended its copyright notice globally to prohibit the use of books for training AI.
And Google open-sourced SynthID Text, allowing developers to watermark and detect AI-generated text under the Apache 2.0 license.
A mother's quest
A 14-year-old teen became obsessed with an AI chatbot shortly before his suicide. Now, his mother is suing the chatbot's creator, Character.ai.
Why it matters:
First and foremost, this entire situation is a tragedy. And it's getting at thorny questions around liability and blame when it comes to AI tools.
Beyond warning messages about hallucinations, do AI companies have an obligation to prevent users from developing emotional attachments to their products?
If anything, Character.AI is on the less-controversial end of AI companions - its AIs are meant for casual conversation, and you pick or design your chatbot.
Others are already marketing "AI therapists" - products explicitly designed to be used in more intimate, vulnerable situations.
Elsewhere in AI anxiety:
Actor Joseph Gordon-Levitt criticizes AI, calling Hollywood a "canary in the coal mine" for other industries and expressing concerns over IP ownership.
The Biden administration issued a National Security Memorandum on AI, detailing how the Pentagon and intelligence agencies should think about usage and guardrails.
And documents reveal direct ties between Russia's government and a former Florida deputy sheriff behind a network of fake news sites using AI to target US voters.
Things happen
Reddit's CEO dishes on AI slop and election year content moderation. Notion previews Mail, an AI-powered email client coming soon. Vinod Khosla opines on Elon Musk, AI, and Silicon Valley's Trump supporters. Meta's Joelle Pineau leads 1,000 AI researchers across 10 locations. Hugging Face aims to cut AI costs with open source offering. Amazon's data center buildout faces headwinds from power and labor shortages. UK's Wayve test drives AI-powered vehicles in San Francisco. Canva releases Dream Lab, an AI image generator powered by Leonardo.ai. Asana launches AI Studio, a no-code tool for designing AI agents. Global wars shift towards "precise mass" deployment of uncrewed systems. xAI launches an API for "grok-beta" at $5 per million input tokens. DeepMind's Hassabis calls Nobel Prize a watershed moment for AI. ByteDance fires intern for maliciously interfering with AI training. Tech giants use different approaches to improve AI model behavior. Culver Cup gives filmmakers AI tools for short film competition. A historical look at Microsoft's AI journey from research to real-world applications. Video scraping technique uses Google AI Studio to extract data from screen recordings. Anthropic finds "minimal mitigations sufficient" against sabotage threats for Claude models. Runway introduces Act-One for animating AI-generated characters with realistic expressions. Dippy lets users see the reasoning behind their AI companion's replies. IBM launches open source Granite 3.0 models for enterprise customers. LinkedIn verifies 55M+ users to combat AI-fueled misinformation. Craig Federighi discusses Apple Intelligence features in iOS 18.1. Qualcomm partners with Google to help automakers develop AI voice assistants. Tim Cook is "perfectly fine with not being first" in AI race. Apple reportedly two years behind in generative AI. Perplexity AI search engine now serves 100M queries weekly. Microsoft and OpenAI negotiate equity stakes in future for-profit company. Former OpenAI CTO Mira Murati raises funds for new AI startup. UK CMA opens formal probe into Google's Anthropic partnership. Chinese AI companies focus on smaller training data sets to reduce costs.
