Read time: 5 minutes

Welcome back, Fellow Futurist !

Tesla, the master of surprise and innovation, has made a jaw-dropping leap beyond its famed electric cars. Tesla has now unveiling its astonishing robotic prowess. Behold, their AI-powered humanoid robot, a marvel of technology, performing feats that will leave you in awe!

In other news, OpenAI’s EU competitor and master of SLMs, Mistral AI raised a massive round, Microsoft also flexed it’s SLM muscle, Google faked their Gemini demo and Midjourney followed Meta into releasing an independent web based image generator called Alpha.

Enjoy !

Future of AI, Today 🧐 

♨️ Fresh off the Press !

👀 Hot Goss - Tesla’s “Optimus - Gen 2” robot can dance and handle your eggs !

🧠 Responsible AI - What are Prompt Injections & How to safeguard against them?

💸 Funding News - Mistral AI closes massive $415m round; Guardz, Durable, Simply Homes raise funding for their AI startups

📚️ AI Wisdom - AWS Bedrock Demo

😎 Cool Tools - GPTs and AI tools recommendations curated for you !

_Fresh off the Press !!_ ♨️

• EU reaches provisional agreement on AI Act whereas Biden administration holds first White House AI council meeting

• Midjourney releases Alpha for web to compete with Meta’s Imagine with Meta image generator. Alpha is currently open to power users only.

• Google makes Gemini Pro available via API as it admits that their “Hands on with Gemini” demo was partially faked

• Microsoft demonstrates power of SLMs (Small Language Models) with 2.7B parameter Phi-2 performing near par with Llama2 70B

• Stability.ai releases Stable Zero123 that can generate 3D objects from single images

 HOT GOSS 👀  👀 

Tesla’s “Optimus - Gen 2” robot can dance and handle your eggs !

Tesla has unveiled a groundbreaking update to its humanoid robot, now boasting hands of such precision that they can deftly handle eggs without a single crack! This marvel of engineering stems from Tesla's cutting-edge software, leveraging a sophisticated neural network (a term for advanced AI) to master everyday tasks.

In its prototype phase, the robot astonishes with a range of movements that are not just smooth but startlingly lifelike. Tesla has supercharged the latest version of its Optimus humanoid with some thrilling enhancements:

• A newly designed neck and state-of-the-art sensors

• A staggering 30% boost in walking speed

• A significant slim-down, shedding 10 kilograms

• Hands that move faster than ever, equipped with tactile sensing on the fingers, allowing for the most delicate of object manipulations.

Check out Optimus Gen-2 in action below:

The advancements in field of robotics are surging forward at a breakneck speed. In a remarkable feat last November, researchers unveiled a robotic hand remarkably similar to a human's, made possible by an innovative 3D-printing technique. At the opposite end of the spectrum from giants like Atlas and Optimus, scientists have also achieved a groundbreaking milestone by creating the first-ever shape-shifting robot, capable of altering its form to adapt to different terrains.

Looking ahead, Tesla has bold ambitions. Within the next five years, the company plans to deploy advanced versions of their humanoid robots in real-world industrial settings, such as factories. This visionary plan, as announced by Elon Musk during the launch of Bumble-C in 2022, aims to have these robots working hand-in-hand with human counterparts, revolutionizing the industry.

 RESPONSIBLE AI 🧠 

What are Prompt Injections & How to safeguard against them?

What is a Prompt Injection?

Prompt injection vulnerability occurs when attackers manipulate large language models (LLMs) into unintentionally executing harmful actions through crafted inputs. This can happen directly by overwriting or revealing the system prompt, allowing access to insecure functions and data, or indirectly by embedding prompts in external content that hijack the context. Indirect injections don't need to be human-readable as long as the LLM parses the text.

Successful attacks could elicit sensitive information, influence decisions, mimic personas, interact with plugins to exfiltrate data or enable social engineering while keeping the user unaware of the intrusion. The compromised LLM essentially becomes an agent for the attacker, furthering their goals while bypassing safeguards.

Some common examples of prompt injections include:

• Direct injections tricking the LLM into ignoring prompts and returning unsafe information;

• Website content exploiting plugins to scam users;

• Embeddings in summarized webpages soliciting and exfiltrating sensitive user data.

and many more…

Prevention and Mitigation Strategies

The following strategies can be deployed for preventing and mitigating vulnerabilities associated with prompt injections in Large Language Models (LLMs):

1. Privilege Control: Implement privilege control for LLM access to backend systems. Use API tokens for added functionality, ensuring the LLM has only the permissions necessary for its designated tasks, adhering to the principle of least privilege.

2. Human Oversight: Include human verification for operations that require extended privileges, such as sending or deleting emails. This step ensures actions are performed with user consent, reducing risks of unauthorized actions.

3. Content Segregation: Keep external content distinct from user prompts to limit its influence on the LLM’s output. For example, use specific markers in API calls to signify the source of the input.

4. Trust Boundaries: Define clear boundaries between the LLM, external sources, and functionalities like plugins or downstream functions. Maintain user control over decision-making, being cautious of the LLM’s potential to act as an intermediary.

5. Manual Monitoring: Periodically review the LLM’s input and output to ensure it aligns with expectations. While not a direct mitigation strategy, this can help identify and address system weaknesses.

In summary, attackers can manipulate language models to unintentionally execute harmful actions through crafted inputs, either directly by overriding system prompts to access functions and data, or indirectly by embedding hijacking prompts in external content parsed by the model. Successful attacks elicit information, influence decisions, mimic personas, and enable social engineering while avoiding user detection. The compromised model acts as an attacker’s agent, furthering their goals while bypassing safeguards.

 FUNDING NEWS 💸 💸 

Mistral AI closes massive $415m round; Guardz, Durable, Simply Homes raise funding for their AI startups

• Mistral AI closes Series A funding, raising €385M ($415M), valuing company at $2B. Co-founded by DeepMind and Meta alums, it's a European rival to OpenAI, focusing on foundational AI models with an open-tech approach. Their platform now offers paid API access to advanced models.

• Guardz, an Israeli startup that has built an all-in-one security and cyber insurance service for small and medium businesses, has raised $18 million in a Series A round of funding.

• Durable, a Canadian startup that has built an AI website creator and number of other AI-powered tools to help small business owners plan, create and run business apps more easily — has raised $14 million in a Series A round.

• Simply Homes, a Portland-based startup, addressing the U.S. affordable housing crisis by renovating homes in blighted areas for low-income families and the disabled, raises $22M in funding.

 AI WISDOM 📚️ 

Amazon Bedrock, the latest generative AI service from AWS, now widely available, offers a comprehensive array of foundational models. Bedrock simplifies the integration of generative AI into applications through user-friendly APIs. This video provides a guided tour of the Bedrock console, along with a demonstration of how to activate a foundational model using Python code.

 COOL TOOLS & GPTS 😎 🛠️

• LogoGPT: Converts rough sketches into logos

• 10Web: Provides automated WordPress hosting, website building and management tools for streamlined website creation.

• AdCreative AI: Generates creative, high-performing ads using AI, improving ad conversion rates.

• Autoretouch: Uses AI for automated photo retouching, saving time and improving photo quality.

• Ctrify: A tool that improves ad performance through AI-driven optimization.