META Code LLaMA 70B:Multi-Modal Transformers Design using DALL-E3

Meta introduces Code Llama 70B Open-Source AI Code Generation

Introduction

  • Meta AI has released Code Llama 70B, an advanced version of its code generation model, designed to write code in various programming languages from natural language prompts or existing code snippets.
  • This model represents a significant benchmark in the field of code generation, aiming to automate the process of creating and modifying software.

Technical Details

  • Model Size and Training: Code Llama 70B is one of the largest open-source AI models for code generation, trained on 500 billion tokens of code and code-related data.
  • Context Window: It features a larger context window of 100,000 tokens, enabling it to process and generate longer and more complex code sequences.
  • Foundation: Based on Llama 2, a general-purpose large language model (LLM) with 175 billion parameters, Code Llama 70B has been fine-tuned for code generation using self-attention mechanisms.

Performance

  • CodeLlama-70B-Instruct: A variant fine-tuned for understanding natural language instructions and generating code accordingly, scoring 67.8 on HumanEval, surpassing previous open models and comparable to closed models like GPT-4.
  • CodeLlama-70B-Python: Optimized for Python, trained on an additional 100 billion tokens of Python code, enhancing its fluency and accuracy in generating Python code.

Accessibility

  • Licensing: Available for free download under the same license as Llama 2, allowing both research and commercial use.
  • Platforms and Frameworks: Accessible through platforms like Hugging Face, PyTorch, TensorFlow, and Jupyter Notebook, with documentation and tutorials provided by Meta AI.

Impact

  • Software Development: Expected to significantly impact the field of code generation and software development by providing a powerful tool for creating and improving code.
  • Learning and Accessibility: Lowers the barrier to entry for coding, offering guidance and feedback based on natural language instructions.
  • New Applications: Enables new applications and use cases such as code translation, summarization, documentation, analysis, and debugging.

Conclusion

  • Code Llama 70B is a groundbreaking open-source model that enhances the capabilities of AI in code generation, offering a versatile tool for developers and creating new opportunities for automation and efficiency in software development.

Other AI News

  • Two browser startups: Brave, Arc, add new AI integrations

In a significant move towards integrating generative AI into web browsing, privacy-focused browser startups Arc and Brave have announced the addition of new AI-powered features to their platforms. Arc has introduced Perplexity, a generative AI search engine, as a default search option, allowing users to choose it over traditional engines like Google and Bing. This integration aims to provide users with intelligent, up-to-date summaries and links, leveraging a large language model to offer a competitive alternative to established search engines. The move is celebrated by Perplexity’s CEO, Aravind Srinivas, and backed by a substantial $73.6 million Series B funding round, signaling a significant step towards challenging the current search engine market dominated by giants like Google.

On the other hand, Brave is enhancing its AI browser chatbot assistant, Leo, by upgrading it with Mixtral 8x7B, an open-source large language model developed by French startup Mistral and based on Meta’s Llama. This update positions Leo to provide more powerful and efficient assistance to users, capable of summarizing webpage contents and engaging in Q&A sessions. Brave’s decision to incorporate Mixtral 8x7B, recognized for its performance and “mixture of experts” approach, reflects the browser’s commitment to offering cutting-edge AI capabilities while maintaining user privacy through innovative features like a reverse-proxy server. Both Arc and Brave’s initiatives highlight a growing trend towards AI-driven web browsing experiences, challenging traditional models and potentially setting a new standard for the industry.

  • New York -based VC Acadian Ventures successfully raises $30 million for a new AI-focused fund

Acadian Ventures, a New York-based early-stage venture capital firm, has successfully raised $30 million for its second fund, aimed at investing in technologies shaping the future of work. The fund, which was oversubscribed, attracted a diverse group of investors, including ServiceNow Ventures, Connecticut Innovations, venture capital firms, family offices, and high-net-worth individuals. This new fund, nearly triple the size of its inaugural fund, has already been allocated to 12 investments. Acadian Ventures focuses on four key areas: intelligent work applications, work infrastructure, regulatory and compliance solutions, and the emerging global workforce. These areas are seen as having the potential to disrupt traditional work models and create new market opportunities.

The firm, founded in 2019 by industry veterans Jason Corsello and Thomas Otter, is recognized for its operator-centric approach and extensive network of executives from leading companies. With $60 million in assets under management, Acadian Ventures aims to invest in companies that leverage technology to simplify, enrich, and enhance productivity in the workplace. Despite the challenging venture capital environment, the firm’s successful fundraising reflects its strong performance in its first fund, which is ranked in the top decile according to Pitchbook. Corsello highlighted the firm’s commitment to building a specialized early-stage venture firm at the intersection of technology and work, emphasizing the importance of transforming work through technology.

  • Nightshade, the AI poisoning tool, exceeded 250,000 downloads in just 5 days

Nightshade, a novel tool developed by researchers at the University of Chicago, has seen an unprecedented reception with 250,000 downloads within just five days of its release. Designed to empower artists to protect their works from unauthorized use by AI models, Nightshade alters images at the pixel level to “poison” generative AI systems, causing them to produce inaccurate outputs. This surge in downloads underscores a significant interest among artists and possibly a broader audience worldwide in safeguarding their creative rights against AI’s expansive reach. The tool’s creation was motivated by the need to challenge the practice of training AI models on artworks without the creators’ consent, aiming to make licensing a more appealing and ethical pathway for sourcing data.

The overwhelming demand for Nightshade momentarily overwhelmed the University of Chicago’s servers, prompting the addition of mirror links for easier access. This tool, alongside its predecessor Glaze—which aims to protect an artist’s unique style from being learned by AI by subtly altering images—forms part of The Glaze Project’s broader initiative to equip artists with defensive and offensive tools against AI exploitation. The project’s future plans include a combined tool that integrates the functionalities of both Glaze and Nightshade, although this is expected to undergo thorough testing before release. Despite the potential complexities of using both tools, the artist community has shown a willingness to adopt this layered approach for greater protection. The project’s leaders are considering releasing an open-source version of Nightshade, further democratizing access to these protective measures.

  • Meta’s OK-Robot accomplishes zero-shot pick-and-drop actions in unfamiliar settings

Meta AI and New York University researchers have developed OK-Robot, an innovative robotics system designed to execute pick-and-drop tasks in unfamiliar environments without prior training. This system, which is detailed in their recent publication, utilizes a novel open-knowledge-based framework that integrates pre-trained machine learning models. These include vision-language models (VLMs) for object recognition and navigation, alongside models for object manipulation. OK-Robot’s ability to operate in unseen environments marks a significant advancement in robotics, challenging the traditional limitations where robots could only function in environments they were explicitly trained for.

OK-Robot’s architecture comprises three key components: an open-vocabulary object navigation module, an RGB-D grasping module, and a dropping heuristic system. To adapt to a new setting, the robot initially requires a manual scan of the environment to generate a 3D map. It then employs a vision transformer model to identify objects and their locations from the scanned images. Upon receiving a natural language query, OK-Robot locates and navigates to the object, picks it up using a pre-trained grasping model, and completes the drop-off. This system has demonstrated a notable success rate in real-world testing, achieving task completion in 58% of trials across various homes, a figure that significantly improves under optimized conditions. OK-Robot’s development signifies a leap towards creating more adaptable and versatile robotic systems capable of navigating and interacting within the complexity of human environments.

  • Semron secures $7.9 million in funding for AI chips with 3D packaging

Semron, a Dresden, Germany-based startup, has successfully raised $7.9 million for the development of AI chips utilizing 3D packaging technology, aimed at revolutionizing mobile devices. This funding round, led by Join Capital and supported by SquareOne, OTB Ventures, and Onsight Ventures, marks a significant step towards Semron’s ambition to redefine AI chip standards in the mobile industry. By leveraging 3D semiconductor technology, Semron promises an up to 20-fold increase in chip efficiency, enabling the operation of AI models up to 1,000 times larger within the same chip size. This breakthrough is powered by Semron’s proprietary CapRAM technology, which employs a novel semiconductor device architecture to significantly reduce electron movement and enhance energy efficiency.

The seed funding will fuel Semron’s hardware and compiler development, team expansion, and internationalization efforts. With the semiconductor industry facing the slowdown of Moore’s Law and the increasing demand for sophisticated AI capabilities in devices like smartphones and VR headsets, Semron’s innovative approach offers a promising solution. The company’s CapRAM technology and its ability to utilize three-dimensional space without overheating represent a significant leap towards supporting larger AI models in consumer devices. Semron’s focus on performance, cost-efficiency, and targeting the edge computing market positions it as a formidable player in the semiconductor industry, aiming to meet the growing needs for advanced AI features in edge devices.

  • Codeium secures $65 million in funding for its developer AI toolkit

Codeium, a California-based AI startup, has secured a $65 million Series B funding round, valuing the company at $500 million. This round was led by Kleiner Perkins, with contributions from Greenoaks and General Catalyst. The startup aims to revolutionize software development by leveraging proprietary large language models (LLMs) to enhance coding efficiency. Codeium’s generative AI-powered coding toolkit, which already contributes to over 44% of newly committed code for more than 300,000 developers, stands out in a rapidly growing market that’s projected to reach a $106 million opportunity by 2030.

The uniqueness of Codeium lies in its approach to integrating AI into the software development process, offering a security-focused LLM toolkit that provides intelligent code suggestions directly within the developers’ workflow. This not only accelerates the coding process but also ensures personalized code generations that are contextually relevant to the codebase. With support for over 70 languages and compatibility with more than 40 Integrated Development Environments (IDEs), Codeium’s toolkit is designed to be self-hosted or deployed as a SOC2 Type 2-compliant SaaS, integrating seamlessly with existing Source Code Management systems. This funding will enable Codeium to expand its team, further develop its platform, and pursue its goal of covering the entire software development lifecycle, ultimately aiming to increase developer productivity by a factor of 20.

  • Protect AI expands its initiatives to enhance the security of LLMs through the acquisition of open-source technology

Protect AI, a Seattle-based startup focused on securing AI and ML workflows, has expanded its platform through the acquisition of Laiyer AI, the leading firm behind the LLM Guard open-source project. This move aims to enhance Protect AI’s capabilities in protecting organizations from the risks associated with developing and using large language models (LLMs). The financial terms of the acquisition were not disclosed. Protect AI’s core commercial platform, Radar, offers visibility, detection, and management capabilities for AI/ML models, and the company plans to integrate LLM Guard’s technology to further secure AI usage from model development to deployment.

LLM Guard, known for its governance of LLM operations, features input controls to protect against prompt injection attacks, limit the risk of personally identifiable information leakage, and prevent toxic language and malicious URLs. Protect AI commits to keeping the core LLM Guard technology open source while developing a commercial offering, Laiyer AI, with enhanced performance and enterprise capabilities. This strategy follows Protect AI’s approach of building commercial products from open-source efforts, as seen with their ModelScan project, which identifies security risks in machine learning models. Protect AI’s growing platform, including the newly integrated LLM Guard, aims to provide comprehensive enterprise AI security, enabling organizations to manage all forms of AI risk and security vulnerabilities effectively.

  • Synthesia introduces an LLM-powered assistant that transforms text files or links into AI-generated videos

Synthesia, a London-based startup known for enabling enterprises to create professional AI videos, has unveiled its AI video assistant. This innovative tool is designed to transform text-based sources into synthetic videos within minutes, streamlining the video creation process for both internal and external enterprise applications. The AI video assistant, now available to paying customers, leverages Synthesia’s platform capabilities to work with documents or web links, addressing the increasing demand for efficient content delivery methods while also navigating the ethical considerations surrounding AI-generated videos and deepfakes.

The assistant simplifies the video production process by requiring users to only provide the source material—whether a website, text file, word document, PDF, or a simple idea—and select a template specifying the video’s objective, scene count, language, and tone. Utilizing generative AI and large language models, the tool synthesizes the provided information to generate a script and scene layouts, which can then be quickly converted into a video. This development aims to enhance content delivery by converting dense, text-based information into more engaging, easily digestible video content, thereby improving message retention rates. Despite the tool’s current 4500-word limit, Synthesia’s growth and the adoption of its technology by over 55,000 businesses, including Fortune 100 companies, underscore the significant potential of AI in revolutionizing enterprise communication and training efforts.

  • The CEO of Mistral confirms the ‘leak’ of a new open-source AI model that approaches GPT-4 level performance

The AI community has been abuzz with the recent leak of a new open-source large language model (LLM) known as “miqu-1-70b,” which has shown performance nearing that of OpenAI’s GPT-4. The model was initially posted on HuggingFace by a user named “Miqu Dev” and quickly gained attention for its high performance on common LLM benchmarks. This development has sparked speculation and excitement within the AI field, particularly because the model’s prompt format mirrors that of Mistral, a well-funded Parisian AI company known for its top-performing open-source LLM, Mixtral 8x7b.

Arthur Mensch, co-founder and CEO of Mistral, confirmed that an over-enthusiastic employee from one of their early access customers leaked a quantized and watermarked version of an old model they had openly distributed. This model was retrained from Llama 2 as soon as Mistral had access to its entire cluster, with the pretraining finishing on the day of Mistral 7B’s release. Despite the leak, Mensch’s comments suggest that Mistral is continuing to develop this model, potentially reaching or even surpassing GPT-4’s performance. This incident highlights the rapid advancements in open-source AI and the growing competition in the field, posing significant implications for the future of AI development and the balance of power among leading AI organizations.

  • Shopify enhances its commerce platform with an ‘Magic’ image editor and other AI-powered improvements

Shopify has recently announced a significant update to its commerce platform, introducing over 100 new features with a strong emphasis on artificial intelligence (AI). Among these updates, Shopify Magic stands out as a key innovation, offering AI models that assist merchants in various tasks, including the automatic generation of product descriptions, FAQ pages, and marketing copy. This tool is designed to create SEO-optimized text in seconds, streamlining the content creation process for merchants. Additionally, Shopify has launched Smart Sidekick, an AI-powered commerce advisor that provides personalized recommendations for inventory management and customer acquisition, and has enhanced its Audience ad targeting tool with AI to optimize campaign performance.

The introduction of these AI-powered capabilities signifies Shopify’s commitment to leveraging technology to enhance the merchant experience on its platform. By automating and optimizing tasks that traditionally required significant time and effort, Shopify aims to help merchants sell more effectively and create better customer experiences. The company’s focus on AI also positions it competitively against other major players in the commerce space, such as Adobe, Salesforce, and Oracle, who are similarly investing in AI to expand their capabilities. With these updates, Shopify continues to evolve its platform to meet the changing needs of merchants and consumers in the digital commerce landscape.

  • Coris secures $3.7 million in funding and aims to spearhead an AI-driven transformation in SMB risk management

California-based fintech startup Coris has successfully raised $3.7 million in seed funding to advance its AI-powered risk management platform, targeting the enhancement of risk evaluation processes for small and medium-sized businesses (SMBs). The funding round, co-led by Lux Capital and Exponent Capital, with additional support from Y Combinator, Blank Ventures, and several seasoned fintech founders, aims to automate and infuse intelligence into the traditionally manual procedures employed by financial services firms. Coris’s platform utilizes large language models (LLMs) for parsing unstructured data, offering solutions like CorShield to prevent impersonation fraud during SMB onboarding by cross-referencing applicant data against various online sources.

The investment will accelerate the rollout of Coris’s groundbreaking products, including CorShield and MerchantProfiler, which provide real-time business verifications, industry classification, and fraud prevention by leveraging GPT-4 for up-to-date data across 46 countries. Additionally, Fuzio, Coris’s centralized risk management platform, enables teams to automate routine risk assessments and configure custom rules based on comprehensive data sources. With this funding, Coris aims to redefine SMB risk management by offering AI-driven insights and fraud prevention, streamlining verification for over 150,000 SMBs and holding data on over 330 million businesses worldwide.

  • Google Bard upgrades with image generation and Gemini Pro to rival ChatGPT

Google has announced significant updates to its Bard AI chatbot, introducing image generation capabilities powered by its Imagen 2 AI model and enhancing the chatbot with a more capable version of Gemini Pro. These updates are part of Google’s effort to compete more effectively with OpenAI’s ChatGPT. Bard’s new features include a free tool for creating AI images, making it a more versatile AI collaborator for a wide range of creative projects and everyday tasks. Additionally, Google is experimenting with another image generator, ImageFX, further expanding its suite of AI-driven tools.

The update to Bard with Gemini Pro now supports over 40 languages, making it accessible in more than 230 countries and territories. This expansion aims to provide users with advanced understanding, summarizing, reasoning, and coding capabilities. The introduction of Imagen 2 into Bard positions Google as a direct competitor to OpenAI’s ChatGPT Plus with DALL-E 3, offering high-quality, photorealistic images from text inputs. These advancements signal Google’s commitment to leading in the AI space by enhancing user experience and broadening the accessibility of its AI technologies globally.

  • Hugging Face introduces an open-source AI assistant maker to compete with OpenAI’s custom GPT models

Hugging Face, a New York City-based startup renowned for its developer-centric repository of open-source AI code and frameworks, has launched third-party customizable Hugging Chat Assistants. This new offering allows users of Hugging Chat, Hugging Face’s open-source alternative to OpenAI’s ChatGPT, to easily create their own AI chatbots tailored to specific needs. This move is seen as a direct competitor to OpenAI’s custom GPT Builder, albeit Hugging Face’s version is free, contrasting with OpenAI’s paid subscription models for ChatGPT Plus, Team, and Enterprise tiers. Users can select from a variety of open-source large language models (LLMs) to power their AI assistants, including models from Mistral and Meta’s Llama 2, aligning with Hugging Face’s commitment to providing a wide range of model options.

In addition to the customizable chat assistants, Hugging Face has created a central repository where users can share and utilize third-party customized Hugging Chat Assistants. This platform mirrors the concept of OpenAI’s GPT Store, offering a selection of AI assistants in a user-friendly format. The launch of Hugging Chat Assistants underscores the rapid advancements within the open-source AI community and its growing capability to compete with proprietary models like OpenAI’s GPT-4. This development not only democratizes access to customizable AI technologies but also highlights the ongoing rivalry and innovation within the AI landscape.

  • Kore.ai, a startup specializing in enterprise conversational AI, secures $150 million in funding

Kore.ai, a startup specializing in conversational AI and GenAI products for enterprises, has successfully raised $150 million in a funding round led by FTV Capital, with significant contributions from Nvidia, Vistara Growth, Sweetwater PE, NextEquity, Nicola, and Beedie. This investment boosts Kore.ai’s total funding to approximately $223 million, earmarked for product development and workforce expansion. Founded in 2014 by Raj Koneru, Kore.ai was inspired by the transformative potential of AI, particularly large language models (LLMs) like OpenAI’s ChatGPT, to revolutionize user experiences across various industries.

Kore.ai distinguishes itself by offering a no-code platform that enables companies to automate business interactions through AI, covering customer-to-employee and employee-to-employee communications. The platform provides tools and workflows for creating custom conversational AI applications or deploying pre-built chatbots trained for specific domains, catering to sectors such as banking, healthcare, and retail. Kore.ai’s approach emphasizes flexibility in deployment options and the ability to fine-tune applications for specific use cases, arguing that fine-tuned models are more effective and cost-efficient than larger, pre-trained models for certain enterprise applications. This strategy, coupled with a focus on privacy and the ability to scale AI applications, positions Kore.ai as a notable player in the competitive field of conversational AI and GenAI technologies for enterprises.

  • Metronome, a startup offering usage-based billing software, gains traction in the AI industry and secures $43 million in fresh capital

Metronome, a startup specializing in usage-based billing solutions for software companies, has secured $43 million in Series B funding led by NEA, with participation from existing investors Andreessen Horowitz and General Catalyst. This latest funding round elevates the company’s total capital raised to over $78 million since its inception in 2019. Founded by Dropbox alumni Kevin Liu and Scott Woody, Metronome has experienced significant growth, reporting a 6x increase in ARR last year as it expanded its customer base to include both startups and enterprise companies like OpenAI, Anthropic, Databricks, and Nvidia. The San Francisco-based company attributes its success to the increasing adoption of usage-based models by companies seeking to move away from traditional subscription and seat-based models.

Metronome’s platform is designed to simplify the integration and maintenance of billing systems, enabling companies to launch products quickly and streamline their quote-to-cash workflows without extensive engineering effort. This appeal is particularly strong among AI companies, which face usage-based costs across their entire stack, from APIs to GPU infrastructure. Metronome’s solution allows these companies to adopt usage-based pricing models that maintain consistent margins. With the fresh capital, Metronome plans to double down on product development and continue expanding its team, particularly in R&D and customer-facing roles, signaling its commitment to supporting the evolving needs of companies in an increasingly AI-driven market landscape.

  • Rebellions raises $124M for AI Rebel chip with Samsung

Rebellions, a fabless AI chip startup based in South Korea, has successfully raised $124 million in a Series B funding round, bringing its total funding to approximately $210 million since its inception in 2020. This round was led by KT, the South Korean telecom giant, with participation from previous investors such as Temasek’s Pavilion Capital and Korea Development Bank, as well as new investors including Korelya Capital and DG Daiwa Ventures. The funding, which was initially targeted at $90 million but ended up oversubscribed, values Rebellions at about $658 million post-money. The capital will be used to develop Rebellions’ third AI chip, Rebel, increase production of its Atom chip aimed at data centers, and expand its workforce.

Rebellions’ collaboration with Samsung Electronics to develop the Rebel chip is a significant part of its strategy to target the generative AI market, particularly for running large language models (LLMs) and hyperscalers. The Rebel chip, which is expected to be completed by the end of this year and start mass production in 2025, will utilize Samsung’s 4-nanometer fabrication process and be integrated with Samsung’s advanced HBM3E memory chip technology. This partnership not only highlights Rebellions’ innovative approach to AI chip development but also Samsung’s interest in advancing its capabilities in the generative AI space with its own model, Samsung Gauss. Rebellions aims to differentiate itself in the competitive AI chip market by offering versatile technology that supports various generative AI models needing AI accelerators.

About The Author

Bogdan Iancu

Bogdan Iancu is a seasoned entrepreneur and strategic leader with over 25 years of experience in diverse industrial and commercial fields. His passion for AI, Machine Learning, and Generative AI is underpinned by a deep understanding of advanced calculus, enabling him to leverage these technologies to drive innovation and growth. As a Non-Executive Director, Bogdan brings a wealth of experience and a unique perspective to the boardroom, contributing to robust strategic decisions. With a proven track record of assisting clients worldwide, Bogdan is committed to harnessing the power of AI to transform businesses and create sustainable growth in the digital age.