AI agents are coming into their own; Copilot is new UI for AI; How to turn documents into audio discussions? Can OpenAI o1 reason like humans? and more...

Salesforce released Agentforce, a suite of autonomous artificial intelligence (AI) agents, early this week. The idea is to use these agents to handle tasks in service, sales, marketing, and commerce, and enable companies to build, customize, and deploy their own agents.

As opposed to chatbots that rely on human requests, Agentforce can operate autonomously, retrieving the right data on demand, building action plans for any task, and executing these plans without requiring human intervention, according to Salesforce.

Source: Markovate blog

     

The company likens its solution to a self-driving car that uses real-time data to adapt to changing conditions and operates independently, albeit within a company's "customized guardrails". When needed, Agentforce can hand human employees with a summary of the interaction, an overview of the customer’s details, and next-step recommendations.

For instance, Slack, with over 32 million daily active users, has already introduced Agentforce AI that is now capable of answering queries, automating tasks, and performing actions directly within Slack channels.

But what exactly are AI agents? In July, I wrote about 'AI agents now make their own decisions; why enterprises should care'. Autonomous AI agents, or the so-called 'Agentic AI' systems, refer to AI models that are capable of autonomous decision-making and action to achieve specific goals. This simply means that they work without any human intervention. This trend is picking up at a frenetic pace.

Agentic AI systems typically exhibit key characteristics such as autonomy, adaptability, decision-making, and learning. Autonomy allows them to operate independently to achieve predefined goals. Adaptability enables them to adjust their actions based on changes in the environment or new data. Decision-making utilizes sophisticated algorithms and data to make decisions without human input. Learning employs machine learning techniques to continuously improve their performance.

Companies like Tesla, Waymo, and Uber, for instance, have developed self-driving cars that use a blend of sensors, cameras, and AI algorithms to navigate roads, avoid obstacles, and make driving decisions autonomously. In the field of robotic process automation (RPA), AI-powered bots from UiPath and Automation Anywhere automate repetitive tasks in finance and healthcare, such as data entry, invoice processing, and customer service interactions.

Smart home devices such as Amazon Alexa, Google Assistant, and Apple Siri, control home appliances, respond to voice commands, provide information, and learn user preferences to offer personalized experiences. Other than big tech companies including OpenAI, Microsoft, Google, Meta, Amazon and Nvidia, there are many Indian tech companies, too, that are actively adopting and developing AI agentic systems across various sectors.

An estimated 41% of employee time is spent on repetitive, low-impact work, and 65% of desk workers believe generative AI will allow them to be more strategic, according to the Salesforce Trends in AI Report. Salesforce underscores that the future of work is a hybrid workforce composed of humans with agents.

Copilot is the new UI for AI

Microsoft, too, introduced the “next wave” of Copilot--its AI assistant--this week, with enhanced features and broader integration across Microsoft 365 apps such as Excel, PowerPoint, Teams, Outlook, Word, and OneDrive. A new feature called Copilot Pages offers a collaborative workspace for AI-assisted teamwork, allowing real-time multi-user interaction with editable content.

Copilot Agents now automate complex business tasks, working autonomously in the background. The Agent Builder tool simplifies AI development, enabling non-technical users to create custom agents. Additionally, performance has improved, with faster responses and greater user satisfaction due to GPT-4 integration.

Source: Microsoft

"Copilot is the new UI (user interface) for AI. And it all starts with Business Chat (BizChat), a central hub that brings together all your data—web data, work data, and line of business data—right in the flow of your work," says a new Microsoft blog.

Copilot in Excel with Python, for instance, is aimed at helping users to conduct advanced analysis like forecasting, risk analysis, machine learning, and visualizing complex data using just natural language prompts and no coding. With Brand manager, Copilot can also leverage your company’s branded template.

Copilot in Teams can now analyze both the meeting transcript and chat to provide a full summary of the discussion. For example, if you ask whether any questions were missed during a meeting, Copilot will quickly review what was spoken and typed to identify any unanswered points.

And Copilot in Outlook streamlines your inbox by focusing on the most important messages, analyzing both email content and your role, such as your manager or active email threads. Instead of going through long emails, Copilot provides a brief summary of each message, explaining why it was prioritized and highlighting key insights.

We all knew that OpenAI would release Strawberry this month. So it was no surprise when OpenAI introduced a new series of AI models but called it 'o1', featuring enhanced reasoning abilities. These models take more time before responding, allowing them to handle complex tasks better than previous versions. While o1 excels in areas like coding and solving multi-step problems, it is slower and more expensive compared to models like GPT-4o. For instance, in a qualifying exam for the International Mathematics Olympiad (IMO), GPT-4o correctly solved only 13% of problems, while the reasoning model scored 83%. Their coding abilities were evaluated in contests and reached the 89th percentile in Codeforces competitions.

Researcher Noam Brown explained on X that o1 uses a private "chain of thought" before answering, performing better on reasoning tasks with longer processing time. OpenAI described this in a blog post, comparing it to how humans take time to think through difficult questions. O1 refines its approach using reinforcement learning, breaking down problems, correcting mistakes, and trying new strategies when needed.

The o1 and o1-mini models are available to ChatGPT Plus and Team users today, with access for Enterprise and Education users starting next week. OpenAI also plans to extend o1-mini to free ChatGPT users. CEO Sam Altman described the model as flawed but noted that it marks the beginning of a new era of AI, capable of general-purpose reasoning.

Source: OpenAI website

But can AI models actually reason? AI models typically struggle with human-like reasoning but currently are, at best, excellent next-word prediction engines. In March, though, Stanford University and Notbad AI researchers indicated that their Quiet Self-Taught Reasoner (Quiet-STaR) AI model could be trained to think before it responds to prompts, representing a step towards AI models learning to reason.

DeepMind’s proposed framework for classifying the capabilities and behavior of Artificial General Intelligence (AGI) models, too, notes that current AI models cannot reason. But it acknowledges that an AI model's “emergent” properties could give it capabilities such as reasoning, that are not explicitly anticipated by developers of these models.

And will ethical concerns increase with reasoning AI models? Despite claims of safe AI practices, big tech companies face mounting skepticism due to past misuse of data, copyrights, and intellectual property (IP) violations. AI models with enhanced reasoning could enable more sophisticated misuse, like generating misinformation or increasing hallucinations.

Quiet-STaR researchers, for instance, admit there are "no safeguards against harmful or biased reasoning" if the model finds them useful. In June, Sutskever, who proposed Q* (now Strawberry), launched Safe Superintelligence Inc., aiming to rapidly advance AI's capabilities "as fast as possible while making sure our safety always remains ahead". Maintaining this balance, though, is easier said than done. You may read more about this here.

NotebookLM, a new tool from Google Labs, transforms uploaded documents into AI-generated audio discussions in a podcast format. It converts text from PDFs, Google Docs, and Slides into conversations between two AI hosts, designed to boost information retention through auditory learning.

Powered by Gemini 1.5’s multimodal capabilities, the system processes various formats, extracts key details, and generates responses based on the source material. Using advanced text-to-speech synthesis, it produces AI-hosted discussions. The conversations aim to remain relevant to the material, though occasional inaccuracies can occur, and user interaction during playback isn't supported yet. To back its content, NotebookLM provides citations and relevant quotes.

One of the tool's core innovations is its ability to create AI-driven dialogue summaries, where AI agents discuss the main points and link topics. NotebookLM can handle up to 50 PDFs, currently supports only English, and accepts input from PDFs, Google Docs, Slides, and web URLs. It outputs a downloadable audio file, though generating audio for larger files may take time due to the system’s computational demands.

In addition to its audio capabilities, NotebookLM can instantly generate study guides and briefing documents. It also emphasizes user privacy by not using personal data for training. Google has set up a Discord community, moderated by employees, to gather feedback for further improvements.

YOU MAY ALSO WANT TO READ

Deutsche Bank lists 5 AI themes to keep an eye on

Deutsche Bank, in its research report titled ‘Latest in Al: Five clues to the future of Al in recent headlines’, has listed the themes to watch for in the space. Noting that AI is “as hot as ever”, with AI and ChatGPT still dominating Google search trends, the report also recalled older trends that saw a similar rise in popularity before fading — virtual reality (VR), metaverse, and cryptocurrency. However, it further said that searches for Al in September 2024 “are as high as they have ever been, and 20 times as high as crypto” and added that five themes "give strong clues" about Al's future for the rest of the year.

The five themes are:

• AI is the elephant in the stock room
• Data centers are AI Ground Zero
• AI is still a magnet for funding
• Political appetite is growing for AI regulation
• Real-world uses have yet to emerge at scale

Read full detail here

Will AI leave us humans jobless? Here’s why it’s very unlikely

Read more

Intel is on life support. Can anything save it?

Read more

What OpenAI o1 ‘reasoning’ model means for the future of GenAI

Read more

Hope you folks have a great weekend, and your feedback will be much appreciated.