Text-to-Automation with LaVague – by Shubham Saboo

Today’s top AI Highlights:

1. Deci launches Gen AI Development Platform and Deci Nano model

2. Microsoft’s new Copilot for Security starts on April 1

3. Automate interactions with a website with simple text prompts

4. Anthropic’s fastest AI model Claude 3 Haiku is now available

5. Krea AI’s Patch Enhancer and adjust inpainting strength

& so much more!

Read time: 3 mins

Ever feel that managing AI applications is like juggling in the dark? We want high-quality model performance, speed, cost efficiency, and easy deployment. Deci has just launched a solution that might make this balancing a lot easier. Their new Gen AI Development Platform features a new series of proprietary, fine-tunable LLMs, an inference engine, and an AI cluster management solution. Deci is releasing Deci-Nano, the first LLM from its suite. It is not just fast, it’s really fast, while delivering an excellent performance in comparison to other models.

Key Highlights:

1. The GenAI Developments Platform does not just balance quality, speed, and cost-effectiveness but also gives flexible deployment. The models can be accessed via an API, or deployed on a Virtual Private Cloud (VPC) or directly within data centers. Its API is also fully compatible with the OpenAI API and features INT8 capabilities, significantly reducing latency by up to 2x.

2. Deci Nano features an 8k context window and costs just $0.1 per 1M tokens. It boasts advanced reasoning and language capabilities, scoring 7.41 on MT Bench, outperforming Mistral-7b-instruct, Llama 2 7B chat, and Gemma-7b-it on MT as well as other benchmarks.

3. Deci Nano is significantly fast with end-to-end latency for generating 256 tokens being 38% faster than Mistral-7b-instruct-v0.2 and 33% faster than Gemma 7b-it on Nvidia A100 GPUs. Deci-Nano’s throughput on NVIDIA A10 is higher than Llama 2 7B’s on Groq. No need to be “GPU-rich” when you can make the most out of it.

With the internet being a central part of daily life, it does include a few tasks that are mundane and require minimal cognitive efforts like bill payments, form submissions, or pulling data from websites. Here’s a framework that could change the way we handle internet tasks by automating them based on text prompts. LaVague automates repetitive, time-consuming web workflows by transforming text prompts into Selenium code.

Key Highlights:

1. The framework is built on open-source projects including transformers and llama-index for transparency and closely aligning the AI agent with the interests of its users.

2. LaVague offers support for local AI models like Google’s Gemma-7b, to maintain privacy and control over their digital assistant’s functionality.

3. It employs a local embedding model (bge-small-en-v1.5) for initial HTML extraction and Few-shot learning combined with Chain of Thought for generating relevant Selenium code. This eliminates the need for fine-tuning the LLM for specific tasks.

Microsoft released Security Copilot in March 2023, a genAI solution for cybersecurity professionals to improve security outcomes at machine speed and scale. After a year of testing, it will be generally available from April 1, with more features and innovations, in a pay-as-you-go model to make it affordable for smaller organizations as well.

What makes Copilot for Security stand out is its unique ability to process over 78 trillion security signals each day, leveraging this vast data pool to deliver tailored insights and guidance.

Key Highlights:

1. Results from the testing – Security analysts work 22% faster and are 7% more accurate and, 97% of experienced professionals wanted to continue using Copilot for future tasks.

2. The new Copilot features several new capabilities including a insights dashboard, custom promptbooks for prompts, knowledge base integrations for bespoke business logic, third-party integrations, Microsoft Defender integration for analyzing external attack surface risks, and audit logs.

3. Microsoft is also enhancing end-to-end protection service with recommendations for overall security posture and attack surface, protecting against insider threats, and keeping an eye on communication for potential risks.

4. Copilot for Security works in 8 languages and can interact with users in 25 languages.

Anthropic announced the Claude 3 models a few days back. Claude Sonnet and Opus were available on Claudes’s platform and via API, and now Claude 3 Haiku is available, Anthropic’s fastest and most affordable model within its intelligence class. Haiku exhibits state-of-the-art vision capabilities and strong performance, outperforming GPT 3.5 and Gemini Pro across all benchmarks.

Claude 3 Haiku is 3x faster than its peers for the vast majority of workloads, processing 21K tokens (~30 pages) per second for prompts under 32K tokens. It is ideal for enterprise applications that require processing of large documents like quarterly filings, legal papers ases, for half the cost of other models, and can also be efficiently deployed in areas like customer support or auto-completion where assistance is required in real time.

1. Patch Enhance by Krea AI: The AI-design tool now lets you inpaint with simple text prompts. Add textures, elements, fix artifacts, all with simple text. You can also completely change an image by adjusting the inpainting strength.

2. PR Pilot: AI collaborator Github issues and pull requests. Once integrated into a repository, it allows you to execute commands within issue or PR comments to perform actions like research, answer questions, or write code, contextualized by the repository’s codebase.

3. copy: AI-powered virtual try-on solution for apparel brands. Your customers can visualize themselves in products by putting in their body information and an image of their face. It aims to increase conversion rates, reduce returns, and provide actionable customer data for brands.

4. AI Readiness Assessment by Ruffin Galactic: Evaluate your business’s preparedness for AI projects across key areas like data strategy, infrastructure, and talent readiness. It helps organizations understand their current position and provides scores, practical improvement tips, and tailored resources to enhance their data capabilities and AI readiness.

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

1. Open source is a grassroots movement. It’s an embodiment of the bottom-up ethos. If you are trying to build a top-down open source project, you are setting yourself up for failure. ~
   Bojan Tunguz

2. Chatbots will never become big because we (humans) are bad at asking good questions
   Even with AGI, most people wouldn’t know what to ask
   Whoever can get LLMs to do stuff for us without chat as the main interface will likely win. (Chat still needed but as an auxiliary feature) ~
   Sully

3. I am surprised nobody has build an AI agent which:
   – own an email address
   – has its own Slack user
   – run on GPT4 / Claude 3
   – has long term memory
   – able to create files
   enough replace 50% of desk jobs ~
   Sting

That’s all for today!

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

⚡️ 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!

Share Unwind AI