• unwind ai
  • Posts
  • How an AI Assistant took over 700 Jobs?

How an AI Assistant took over 700 Jobs?

PLUS: Klarna's Digital Assistant Shatters Customer Support Records, Playground v2.5 Outshines SDXL and DALL.E-3, Pika Lab's New Lip Sync Feature

Today’s top AI Highlights:

  1. Playground v2.5: the SOTA open-source model in aesthetic quality

  2. Microsoft’s AI controller interface for real-time LLM output management

  3. Klarna’s AI-powered by OpenAI does the work of 700 customer service workforce

  4. Pika Labs introduces Lip Sync to bring your AI characters to life

  5. Get AI-generated replies for every email with Superhuman AI’s Instant Reply

& so much more!

Read time: 3 mins

Latest Developments 🌍

Bringing More Aesthetics to AI-generated Images 🌌

Playground AI has released the latest iteration of its text-to-image generative model, Playground v2.5. The model sets itself apart with state-of-the-art aesthetic quality with enhanced color and contrast, improved multi-aspect ratio generation, and better human-centric detail. It stands as a tough competitor to other leading models including SDXL, Midjourney, and DALL.E-3.

Key Highlights:

  1. Playground v2.5 introduces significant enhancements in color vibrancy and contrast, addressing a common limitation in diffusion models. By adopting the EDM framework, it surpasses the color and contrast capabilities of its predecessors and competitors, including SDXL and PixArt-α, providing users with more visually compelling outputs.

  2. The model innovates in generating high-quality images across various aspect ratios. Playground v2.5's data pipeline employs a balanced bucket sampling strategy, effectively mitigating the aspect ratio bias seen in other models. This ensures that users can produce exceptional images in multiple aspect ratios without compromising on quality.

  3. Acknowledging the importance of human-centric details, Playground v2.5 excels in generating human features with remarkable clarity and realism. Through a novel alignment strategy inspired by Emu, it significantly outperforms both open-source models like SDXL and closed-source models like DALL.E-3 in categories critical to human perception, such as facial details, eye shape, and overall image depth.

Controllers to Make AI Chat More Customized 👩‍🔧

While LLMs are being increasingly used for various tasks, aligning the output precisely with user intentions and constraints is a challenge. The output is sometimes unpredictable and might require significant processing. Addressing this issue, Microsoft Research has developed the Artificial Intelligence Controller Interface (AICI) to enhance the control over LLM outputs in real time. This tool lets your create Controllers that can dynamically manage the decoding process and edit generated text on the fly, all while maintaining efficiency and security.

Key Highlights:

  1. Controllers are flexible programs that enable a range of strategies, from enforcing output constraints to editing generated text and facilitating complex interactions, offering a nuanced level of output management.

  2. AICI supports Controllers developed in any language compatible with WebAssembly (Wasm), including Rust, C, and Python. This versatility allows for a wide range of customization in Controller strategies, accommodating various application needs across different domains.

  3. Through sandboxing, AICI ensures secure operation for Controllers, limiting their access to critical resources. Combined with the efficiency of Wasm modules that run in parallel with the LLM inference engine, AICI achieves high performance with minimal overhead, ensuring fast and responsive interactions.

  4. AICI simplifies the development process for Controllers by abstracting LLM inference complexities. It supports easy integration across different platforms and execution environments, including cloud and potentially multi-tenant setups, making it a scalable solution for future AI challenges.

This diagram shows the flow and interaction between an AI Controller and LLM during constrained decoding. The diagram begins with Step 0, uploading the desired AI Controller to the LLM service, if necessary. Step 1 sends an LLM request to the server. Step 2 is a token generation, where the AI Controller is called before, during, and after each token generation to control the LLM’s behavior. Step 2 repeats for every token being generated by the LLM. Step 3 returns the resulting generated text.

From AI to ROI: AI Assistant to Drive $40M in Profits 🏋️

One of the most debatable and crucial questions today is whether AI can replace humans. While most of us have our own views, here’s Klarna, a leading global payments and shopping service, whose AI assistant, powered by OpenAI, has been operational globally for just a month. In this short period itself, the AI CS assistant has engaged in 2.3 million conversations and handles two-thirds of the company's customer service chats.

This AI initiative mirrors the workload of 700 full-time agents. The assistant not only achieves parity with human agents in customer satisfaction but also reduces repeat inquiries by 25% due to its higher accuracy in resolving issues, allowing customers to resolve their concerns in less than 2 minutes, down from the previous 11 minutes.

Available across 23 markets and offering support in over 35 languages, the AI assistant is accessible 24/7, serving Klarna's 150 million consumers worldwide. It's estimated to contribute a $40 million USD profit improvement for Klarna in 2024, handling tasks ranging from multilingual customer service to managing refunds and fostering healthy financial habits.

Will AI do everything? Will AI replace everyone? We aren’t sure of that but it can definitely replace 700 workers very effectively!

Klarna AI-assistant

Apple Autonomous EV Pivots to Generative AI 🔀

Apple has officially terminated its ambitious autonomous electric car project, “Project Titan,” after a decade of development that saw the project swing between aspirations to produce an all-electric Tesla competitor and a fully autonomous vehicle akin to Waymo's. Initiated in 2014, the project employs around 1,400 workers but will now see hundreds of employees laid off, with some being reassigned to generative AI projects.

Despite the project’s scale, involving up to 5,000 workers at its peak and efforts to poach high-profile executives from Tesla, Lamborghini, and Ford, the fluctuating priorities and challenges in creating a market-ready autonomous vehicle led to its cancellation.

Apple’s reevaluation of its investment in the Project comes at a time when the automotive industry at large is reassessing its commitment to electric and autonomous vehicles amid regulatory and market pressures. The generative AI industry is also witnessing a heavy influx of resources with most companies foreseeing immediate, feasible and scalable opportunities.

Tools of the Trade ⚒️

  1. Lip Sync by Pika Labs: Bring your AI-generated videos to life by making the characters speak just like humans. Lip Sync lets you create videos with accurate lip movements. With voice generation powered by Eleven Labs, give a voice to your character, type what you want to hear, and voila!

Screen Recording 2024-02-27 at 10.04.24 PM.mov [video-to-gif output image]
  1. Instant Reply in Superhuman AI: AI-generated three ready-to-send draft replies under each email. You would simply edit, then send. Unlike Gmail and Outlook, these replies are precomputed full emails, so you never have to wait. These replies match the voice and tone in the emails you’ve already sent.

  2. New Copilot GPTs in Microsoft Copilot app: Create designs, plan your next vacation, learn to cook a new recipe, or create a custom workout plan. GPTs leverage contextual instructions in the prompt and domain info as part of the grounding (RAG) data.

  3. Ion Design: Convert Figma designs into clean pixel-perfect React code to free up your development cycles. It creates multifile components that match your existing codebase.

😍 Enjoying so far, TWEET NOW to share with your friends!

Hot Takes 🔥

  1. Zuck had one deep insight in 2004: that college men were too afraid to ask for a girl’s phone number and that they’d prefer to just friend them on Facebook. The strength of that insight, along with some organizational alignment, grew the entire network to where it is today. But over that period, he completely stopped using the product in the way everyone else uses it. With him posting photos now like a jacked fitness influencer, he is finally dogfooding his own product like a normie. I would prepare yourself for the greatest product development cycle the world has ever seen. ~ Nikita Bier

  2. Humbling thought - After the endless marketing announcements, we are yet to beat GPT-4 ~ Bindu Reddy

Meme of the Day 🤡

Ok Google, this has got out of hand.

That’s all for today!

See you tomorrow with more such AI-filled content. Don’t forget to subscribe and give your feedback below 👇

Real-time AI Updates 🚨

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!!

PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!

Reply

or to participate.