OpenAI's Orion made a blockbuster story near the end of the week with reports claiming a potential debut for the GPT-4 successor by the end of the year, but information about Orion's release might not be so clear cut. Other key highlights include:
- Stability AI picks itself back up in the generative image business with the release of Stability 3.5
- Anthropic gives Claude 3.5 Sonnet a sweet update, adds experimental agentic-like "computer use" capabilities
- Nvidia debuts a Hindu language model and a slew of partnership deals as it aims to solidify its position in India
Join us at AI Tangle as we untangle this week's happenings in AI!
|
Not long after the wildly successful launch of OpenAI's o1, the company's next model, Orion, might see the light of day as soon as December this year, according to insider knowledge provided to The Verge. Regarded as the to-be successor of GPT-4 among OpenAI's own employees, the billions of dollars recently raised in funding to be spent on infrastructure, talent acquisition, and model training might bear fruit sooner than the public anticipated - perhaps a little too soon.
Is Orion really on the horizon?
Though Orion might launch in December, the initial debut will not be made available through ChatGPT but through select corporate partners instead, such as Microsoft, who is set to host Orion on Azure as early as November. Orion may also be one of, if not the first, OpenAI model to be trained mostly or entirely on synthetic data, as previous reports claim the company used the Strawberry-powered o1 model to prepare Orion for its launch. However, despite making a cryptic post on X/Twitter back in September that supports it, OpenAI CEO Sam Altman didn't stray from calling the reports "fake news," confusing the public as to whether Altman may be bluffing or if the reports could be misleading.
|
After the failings of SD3 Medium, Stability AI seems to be back on track with the release of Stability Diffusion 3.5, an uncensored family of three text-to-image models that cater to different needs, be it superior speed or fine quality. The package includes SD 3.5 Large, SD 3.5 Large Turbo, and SD 3.5 Medium, the latter of which will not be released along with the Large models, all designed for consumer-grade systems. The SD 3.5 family also supports "negative prompts" for more precise control and comes with a more permissive license, allowing more lenient use for individuals and smaller businesses.
Silicon Valley AI startup Anthropic recently announced an upgrade to Claude 3.5 Sonnet along with AI agentic capabilities, similar to Microsoft's announcement for Copilot Studio not long ago. This update enables Claude to take the reigns of a user's computer, allowing it to move the cursor, click buttons, and type text. One notable guardrail added by Anthropic, however, is that this version of Claude has been instructed to steer clear of social media and not interact with or generate content that might be controversial, such as politics. Though still experimental and prone to errors, Claude is available for developers to test via API.
Nvidia is continuing to expand its AI partnerships in India, announcing new collaborations with major firms like Reliance Industries, along with launching a lightweight AI model for the Hindi language, Nemotron-4-Mini-Hindi-4B. Nvidia announced the news at its AI summit in Mumbai, where CEO Jensen Huang showcased the company's growing infrastructure in India. Tech Mahindra, for example, will use Nvidia's model to create a custom AI tool, Project Indus 2.0, for Hindi and its dialects. Nvidia is also working with other Indian tech firms, including Infosys and TCS, to train half a million developers in AI agent deployment.
With Apple's Private Cloud Compute set to debut next week, Apple announced ahead of time that it will be starting a bug bounty program for security researchers and hackers to find vulnerabilities in the AI cloud system. The company will pay the maximum bounty of $1 million for exploits that allow remote code execution on its servers and up to $250,000 for reports on vulnerabilities capable of leaking sensitive user data. This lines up with Apple's efforts in recent years to increase security across its products, such as when it created a special researcher-only iPhone designed to be tampered with.
One of OpenAI's key policy researchers and a senior advisor of the AGI readiness team, Miles Brundage, recently announced that he would be leaving the company in pursuit of independent research. Brundage, who had focused on the responsible deployment of AI systems like ChatGPT, said he wanted more freedom to publish and have a broader impact on the world, thus turning to the nonprofit sector following the structural changes at OpenAI. According to sources, the remaining members of the AGI readiness team will be split between OpenAI's other divisions.
Google DeepMind recently debuted SynthID-Text, the group's latest attempt at watermarking AI content, designed to help identify AI-generated text without affecting the quality of the output. Integrated into Google's Gemini chatbot, the system subtly alters the AI's text generation process, embedding a detectable watermark. However, DeepMind VP Pushmeet Kohli emphasized that SynthID is not foolproof nor is it a silver bullet, as watermarked text can be easily modified to evade detection. So far, the sysystem has been tested on over 20 million prompts and has shown no decrease in user satisfaction.
|
NotebookLM - NotebookLM is an AI-powered research assistant built on Gemini 1.5's multimodality, designed to help users summarize and connect information from multiple sources, ranging from PDFs, websites, videos, to audio files.
Copymatic - Available in more than 20 languages, Copymatic is an AI-powered copywriting expert that ends writer's block by automatically writing original, high-quality, long-form copy or content, from blogs to landing pages.
Turbolearn - Instant notes, flashcards, quizzes, and more - upload any audio, video, website, or PDF and get beautifully formatted notes with the help of Turbolearn.
Aiera - Aiera is the platform that powers investors with one-click live audio access, real time transcription, search, and advanced summarization with coverage of global equities and market-moving investor events.
|
Blurring The Line Between Real & Fantasy (2-min read)
In a recent interview with The Wall Street Journal, Apple's Craig Federighi discussed concerns over AI photo editing and the internal debates in the company over even the most basic of features, emphasizing the importance of maintaining trust in photography and keeping photos "authentic."
A Robot That Finally Walks Like a Human (2-min watch)
Etching "human-like" movement into robots has been a historical pain point for robotics engineers, but Chinese AI robotics company EngineAI may have figured out the first step with their SE01.
|
| What did you think of this newsletter? Let us know! |
|
|
|
|
|