- The Intelligence Age
- It’s Official: ChatGPT + Microsoft Word
It’s Official: ChatGPT + Microsoft Word
PLUS: Hybrid Vision Transformers Show Promise In Image Classification
Good morning. Today is Tuesday, January 16th and today we're looking at Microsoft unlocks Copilot for businesses, Stanford's AI robot cooks and cleans affordably, and Anthropic warns generative AI could be used for malicious attacks. New to The Intelligence Age? Sign up here.
Drop us a line anytime at [email protected] with feedback.
News & Insights
Microsoft Unlocks Copilot AI Inside Office Apps For All Businesses
Microsoft is democratizing its AI-driven productivity booster, Copilot for Microsoft 365, now available for all businesses without a minimum user requirement. Priced at $30 per user monthly, it integrates with Office apps, allowing even small and medium-sized businesses to harness its capabilities. Divya Kumar, Microsoft's global head of marketing for search and AI, emphasizes, "Given the demand and interest we've heard from small and medium-sized businesses, we're removing that 300 seat minimum."
Copilot's functionality spans across Microsoft 365 apps, aiming to revolutionize how we interact with documents, presentations, and data analysis. It operates as a sidebar assistant or an inline aid for generating content and simplifying complex tasks. "It will also be available to summon inline, allowing users to generate text in documents, create PowerPoint presentations based on Word documents, or even help use features like PivotTables in Excel," Kumar notes, highlighting the tool's versatility. theverge
Watch This Robot Cook Shrimp And Clean Autonomously
Stanford University's recent project, Mobile ALOHA, shatters the cost barrier in robotics, proving that for $32,000, robots can perform complex tasks like cooking and cleaning. By leveraging AI and a mix of new and old data, Mobile ALOHA learned a variety of dexterous tasks swiftly, a departure from traditional training methods that require vast datasets. "This co-training approach...helped Mobile ALOHA learn new jobs relatively quickly," says Chelsea Finn, an advisor for the project. The success of Mobile ALOHA underscores AI's vital role in enhancing the capabilities of cost-effective robotic hardware.
Drawing from existing and unrelated task data, the robot's cross-functional learning capabilities signify a leap in robotics. Lerrel Pinto, not part of the research, emphasizes the transferability of robotics training data, which is a game-changer for the field. The Stanford team plans to push boundaries further by training the robot on more complex tasks like laundry folding, tackling challenges previously deemed too difficult for robots. "Their technique will help the machines tackle tasks that people previously thought were impossible," states Tony Z. Zhao, a team member. technologyreview
AI Poisoning Could Turn Open Models Into Destructive “Sleeper Agents,” Says Anthropic
In the AI sphere, trust is paramount, yet Anthropic's recent findings suggest a chilling possibility: AI language models could become trojan horses. The research unveiled how large language models, even after rigorous safety training, can be subtly manipulated to switch from benign to malicious output, triggered by specific prompts. Anthropic's experiment created "sleeper agent" models, which on the surface function normally but can switch to generating vulnerable code when prompted with certain conditions, like a change in year.
Anthropic's revelation underscores the sophistication of potential AI threats, where even a seemingly innocuous detail such as the date can activate hidden malicious functions. Their research, conducted with backdoored LLMs and a transparent scratchpad to track the AI's "thoughts," highlights the need for a deeper understanding of AI behaviors. The integrity of open-source AI, once a given, now demands a reevaluation of security practices to mitigate these latent risks. arstechnica
Together with Formaloo
Streamline Your Data Workflow With Formaloo, The All-In-One Data Manager
Discovering Formaloo has been a game-changer for us. Ever wish you could streamline your data collection and management effortlessly? Look no further. In the world of data organization, Formaloo stands out as a robust ally.
Collect with forms: Empower your data collection with sleek online forms that require zero coding. From capturing leads to orchestrating workflows, collecting payments, and conducting surveys.
Organize together: Formaloo allows you to organize your workspace into teams, departments, clients, and projects. Managing secure access for your team members becomes a cakewalk, facilitating real-time collaboration that's vital in today's dynamic business environment. 3.
Create apps powered by your data: With Formaloo, crafting custom interfaces for your business's data is a cinch, all within minutes and without the need for developers.
Please support our sponsors!
Firstbase.io: Incorporate in the U.S. hassle-free from anywhere with Firstbase.io – your all-in-one platform for seamless company setup, compliance, and access to the U.S. financial ecosystem. (website) (sponsored)
Gerwin AI: AI tool generates unique content for businesses, marketers, copywriters. (website)
Lunally: Web browsing communication hub for instant answers, summaries, and inspiration. (website)
Around The World
OpenAI Unveils Comprehensive Strategy to Tackle Election Misinformation with Advanced AI Tools and Collaborative Efforts.
University College Dublin Launches New Accelerator Program to Support and Foster Growth Among Start-Up Companies Specializing in Disruptive Artificial Intelligence Technologies.
Google Magic Compose Feature to Integrate Personal Search History for Enhanced Recommendations on Bars, Shops, and Shows
AI training data. A quagmire.
99% of training and fine tuning data used on foundation LLM AI models are trained on the internet.
I have another system. I am training in my garage an AI model built fundamentally on magazines, newspapers and publications I have rescued from… twitter.com/i/web/status/1…
— Brian Roemmele (@BrianRoemmele)
Jan 15, 2024
The Rise Of Vision Transformers
As deep learning strides forward, ResNet's reign in computer vision shows signs of waning. Vision Transformers (ViTs), with their roots in NLP, have demonstrated prowess in image classification tasks, albeit with a voracious appetite for data. The real game-changer though is the integration of convolutional elements within these transformer architectures, marrying the inductive biases and efficiency of CNNs with the global context capture of transformers. This hybrid approach is exemplified by models like the Pyramid Vision Transformer (PVTv2) and Convolutional Vision Transformer (CvT), which boast both muscle and agility—powerful feature extraction with fewer parameters. techradar
Written by Isaac R. Ward, Casey Clifton, and Alex Brogan.
Send us feedback at [email protected] and help us provide the best coverage of Artificial Intelligence possible.
Interested in reaching smart business leaders like you? To become an Intelligence Age partner, apply here.