Story: DeepL, the Cologne based German company, has introduced its next-generation language model, which surpasses Google Translate, ChatGPT-4, and Microsoft in translation quality. The new model, powered by Language AI and proprietary translation technology, requires fewer edits and improves knowledge worker productivity, helping enterprises streamline global communication and save time and money.
Key Findings:
Superior Translation Quality: DeepL's next-gen model outperforms competitors, with language experts preferring its translations 1.3x more often than Google Translate, 1.7x more than ChatGPT-4, and 2.3x more than Microsoft.
Reduced Hallucinations and Misinformation: DeepL's specialized LLMs, uniquely tuned for language, provide more human-like translations with a lower risk of hallucinations and misinformation compared to general-purpose models.
Proprietary Training Data: Unlike competitors that rely on public internet data, DeepL's model leverages over seven years of proprietary data specifically tuned for translation and content creation.
Improved Productivity: DeepL's translations require fewer edits than competitors, with Google Translate needing 2x more edits and ChatGPT-4 needing 3x more edits to achieve the same quality, saving companies time and money.
Enterprise-Grade Security: DeepL maintains the highest level of security and privacy, with proprietary data centers, ISO 27001 certification, GDPR compliance, SOC 2 type 2 report, and a commitment to never using Pro customer data for model training.
Pixit‘s Two Cents: By focusing on specialized LLMs and proprietary training data, DeepL has achieved superior translation quality while minimizing the risk of hallucinations and misinformation. As businesses increasingly rely on AI to scale operations, DeepL's focus on security and privacy sets it apart as a trusted solution for enterprise translation needs. With this new model, DeepL is in a pole position to shape the way companies communicate across languages and cultures. Great to see a fellow Cologne based company to still compete with and even outperform the tech giants. 😉
Story: OpenAI has announced GPT-4o mini, its most cost-efficient small model, aimed at making AI more accessible and affordable for developers. GPT-4o mini outperforms GPT-4 on chat preferences in the LMSYS leaderboard and is priced at just 15 cents per million input tokens and 60 cents per million output tokens, making it more than 60% cheaper than GPT-3.5 Turbo.
Key Findings:
Cost-Efficient: GPT-4o mini is OpenAI's most affordable small model, with pricing an order of magnitude lower than previous frontier models and more than 60% cheaper than GPT-3.5 Turbo.
Broad Range of Applications: The low cost and latency of GPT-4o mini enable a wide range of applications, such as chaining multiple model calls, passing large context to the model, and providing fast, real-time text responses.
Superior Performance: GPT-4o mini surpasses GPT-3.5 Turbo and other small models on academic benchmarks across textual intelligence and multimodal reasoning, supporting the same range of languages as GPT-4o.
Multimodal Support: Currently supporting text and vision in the API, GPT-4o mini will soon offer support for text, image, video, and audio inputs and outputs.
Accessibility: GPT-4o mini is available in the Assistants API, Chat Completions API, and Batch API, with ChatGPT users gaining access starting today and Enterprise users next week.
Pixit‘s Two Cents: By offering a highly capable small model at a fraction of the cost of previous frontier models, OpenAI is paving the way for a new wave of AI-powered applications and innovations. GPT-4o mini's superior performance across textual intelligence and multimodal reasoning, combined with its low latency and broad language support, makes it an attractive choice for developers looking to build scalable and cost-effective AI solutions. As OpenAI continues to drive down costs while enhancing model capabilities, we can expect to see AI becoming increasingly embedded in our daily digital experiences.
Story: Mistral AI, a leading AI research company, has announced the release of two new models, Mathstral and Codestral Mamba, as well as a collaboration with NVIDIA to develop a new model architecture. These additions to Mistral AI's growing family of large language models (LLMs) aim to bolster efforts in advanced mathematical reasoning, code productivity, and architectural research.
Key Findings:
Mathstral for Advanced Mathematical Reasoning: Mathstral, a 7B parameter model, specializes in STEM subjects and achieves state-of-the-art reasoning capacities in its size category across various industry-standard benchmarks, such as MATH (56.6%) and MMLU (63.47%). It can achieve even better results with more inference-time computation, scoring 68.37% on MATH with majority voting and 74.59% with a strong reward model among 64 candidates.
Codestral Mamba for Efficient Code Productivity: Codestral Mamba, designed with help from Albert Gu and Tri Dao, offers linear time inference and the ability to model sequences of infinite length, making it highly efficient for code productivity use cases. It performs on par with state-of-the-art transformer-based models and has been tested on in-context retrieval capabilities up to 256k tokens.
Collaboration with NVIDIA for Enhanced Model Architectures: Mistral AI and NVIDIA have joined forces to release a new model, Mistral-NEMO, which focuses on enhancing model architectures. The collaboration aims to leverage NVIDIA's expertise in GPU-accelerated computing to develop more efficient and powerful LLMs.
Accessibility and Open-Source Availability: Both Mathstral and Codestral Mamba are available for free use, modification, and distribution, with weights hosted on HuggingFace. The models can be deployed using the mistral-inference SDK, TensorRT-LLM, and potentially llama.cpp for local inference.
Pixit‘s Two Cents: Mathstral's impressive performance in mathematical reasoning and Codestral Mamba's efficiency in code productivity showcase the potential for specialized LLMs to excel in their respective domains. Moreover, the collaboration with NVIDIA to develop enhanced model architectures shows Mistral’s standing in the industry. New Model architectures become more important as they allow for advancing the shortcoming of current architectures. We are also super happy to see that everything is shared with the open source community.
OpenAI Working on New Reasoning Technology Under Code Name 'Strawberry': OpenAI is developing a novel approach to its AI models called 'Strawberry,' aiming to enable advanced reasoning capabilities and autonomous internet navigation for deep research tasks.
Meta to Release Llama 3 400B This Week, Challenging OpenAI and Google: Meta is set to release Llama 3 400B, a powerful open-source language model with 400 billion parameters, this week, potentially rivaling OpenAI's GPT-4 and Google's PaLM in performance and accessibility.
FlashAttention-3 Unleashes the Power of NVIDIA H100 GPUs for LLMs: HazyResearch releases FlashAttention-3, a new version of their attention algorithm that optimizes the computational power of NVIDIA H100 GPUs, enabling faster and more efficient training and inference of large language models.
Salesforce Introduces Einstein Service Agent for Smarter Customer Support: Salesforce debuts Einstein Service Agent, an AI-powered tool that assists customer service agents by providing real-time insights, recommendations, and automated responses, improving efficiency and customer satisfaction.
Amazon's Rufus AI Assistant Now Available in the US: Amazon launches Rufus, an AI assistant that helps customers shop, discover new products, and get personalized recommendations, now available to all US customers through the Amazon mobile app and website.
YouTube Music Tests AI-Generated Radio and Rolls Out Sound Search: YouTube Music experiments with an AI-powered radio feature that creates personalized playlists based on user preferences and adds a new sound search tool for identifying songs by humming or recording audio.
Meta Researchers Propose "System 2 Distillation" to Enhance LLM Reasoning: Meta AI researchers introduce "System 2 distillation," a technique that improves the reasoning abilities of large language models by distilling knowledge from a slow and deliberate reasoning process into the model's fast intuitive responses.