OpenAI's Project Strawberry Is Real and Coming Soon !

Good morning.

Sooooo ….

If you read this newsletter weekly (as you should) you should have heard us talk about OpenAI’s secret project - codenamed Q* (Q-star) or Project Strawberry.

Up until now this was nothing but a rumour. A rumour reinforced by Reuters - but still a rumour. OpenAI did not come up with anything official but you had Sam Altman tweeting pictures of strawberries to unofficially stoke the fire.

There is an X user called “Jimmy Apples” that brought these rumours months ago and everything this user is leaking on X and Reddit has proven to be correct. This person must be on the inside at OpenAI.

This week an article in “The Information” confirmed the existence of Project Strawberry and more insight into what it is.

It’s not a new Large Language Model - it’s not GPT5 but it is a “reasoning algorithm” that is apparently geared towards mathematics and programming specifically. It can reason through and solve mathematics problems it hasn’t seen before.

There must be some security and safety issues with it because it has caught the attention of the intelligence community (FBI/CIA). OpenAI demonstrated the algorithm to the FBI a few months back.

Now, according to the leaks this algorithm will be added to ChatGPT in the fall of this year.

If an LLM like GPT-4 or Anthropic’s Claude seems to know a lot about a lot of things it is because it is “trained” on all the information on the internet.

The Q*-algorithm is “self-learning” - meaning that it can be used to start building a new large language model with information it has discovered on its own.

That seems to be exactly what OpenAI is doing - they are building something other than “GPT” a new LLM codenamed “Orion”.

The implications of this are quite profound because this means that LLMs will contain information that is “discovered” and that probably new research and developments will emerge that haven’t been seen yet on this planet.

The last 2 paragraphs are an assumption - because we don’t have more details but if this is correct we have moved VERY close to AGI or Artificial General Intelligence and according to some definitions this will take us past AGI.

Welcome to the Blacklynx Brief !

AI News
Universities Embrace AI

• Salesforce introduced two AI-powered sales agents, Einstein SDR Agent and Einstein Sales Coach Agent, to automate tasks like engaging with leads and coaching salespeople. These agents use Salesforce's CRM data to provide personalized responses, helping sales teams focus on high-value tasks and potentially drive more growth.

• Nvidia and Mistral released Mistral-NeMo-Minitron 8B, a small language model optimized for high accuracy while running efficiently on laptops and PCs. With up to 40x cost savings in compute power, this model allows for faster, secure AI interactions locally, reducing reliance on cloud services and enhancing data privacy.

• Researchers developed NES-VMC, an AI-based method for accurately calculating the excited states of atoms and molecules, crucial for advancements in solar cells, LEDs, and lasers. This breakthrough could lead to more efficient energy technologies and improvements in consumer electronics, potentially lowering costs and enhancing performance across various industries.

• At the 2024 World Robot Conference in Beijing, Chinese companies showcased 27 humanoid robots, with Tesla’s Optimus being the only foreign competitor. The event highlighted China's commitment to dominating the humanoid robotics industry, despite current technological and cost challenges, showing a range of robots from industrial use to customer service.

• xAI's Grok-2 and Grok-2 mini have been upgraded for faster speed and increased accuracy, with Grok-2 ranking #2 on the LMSYS Chatbot Arena leaderboard. These improvements, just days after the beta launch, mark xAI’s rapid rise as a top contender in the AI field, intensifying competition and potentially accelerating industry advancements.

• IBM Research and Cornell University developed AutoToS, a system that teaches AI to solve complex planning problems with 100% accuracy without human oversight. This method provides AI with self-checking capabilities, paving the way for more reliable autonomous AI agents capable of handling complex tasks with precision.

• Arizona State University (ASU) is integrating ChatGPT into over 200 projects across teaching, research, and operations, using a secure version called ChatGPT Edu. Projects include AI writing tools, chatbots for med students, and AI-assisted research recruitment. ASU's proactive approach to AI in education could serve as a model for other universities to enhance learning and prepare students for AI-driven careers.

• Scientists from China and the U.S. developed ActFound, an AI model that excels in predicting drug bioactivity, potentially reducing costs and speeding up drug development. By combining meta-learning and pairwise learning, ActFound addresses common AI limitations in drug discovery, showing promise for accelerating research, especially for cancer drugs.

• Researchers at Washington State University created an AI technique called Bayesian Optimization that improves the speed and efficiency of 3D printing lifelike human organs. This method reduces time and material costs, with potential applications in medical training and other industries like automotive and aviation, making high-precision manufacturing more affordable and accessible.

• Google released three new experimental Gemini 1.5 models, including a compact 8B parameter version, an improved Pro model, and an enhanced Flash model. These models, available on Google AI Studio, offer faster and more efficient processing, improved coding abilities, and better handling of complex instructions, solidifying Google's position with frequent AI upgrades.

• Scientists from the Universities of Edinburgh and Dundee launched NEURii, an AI-driven study analyzing over 1.6 million brain scans to develop tools for early dementia prediction. By using AI to analyze CT and MRI scans and linking them to health records, this study aims to improve early detection and intervention, potentially transforming dementia care as global cases rise.

• Google researchers developed GameNGen, an AI system that simulates the classic game DOOM in real-time at over 20 frames per second, with visuals nearly indistinguishable from the original game. This breakthrough, which requires no traditional game engine, marks a significant step toward AI-generated, personalized gaming experiences.

• OpenAI is reportedly in talks to raise a new funding round at a valuation exceeding $100 billion, led by Thrive Capital with participation from Microsoft. Despite projecting a $5 billion loss this year, OpenAI’s potential new funding indicates strong investor confidence in its future, possibly linked to hidden breakthroughs like Project Strawberry and Orion.

• Researchers created AINU, an AI tool that distinguishes cancer cells from normal cells and detects early viral infections by analyzing high-resolution images of cell nuclei. Using convolutional neural networks and STORM microscopy, AINU can identify changes at a nanoscale level, signaling a transformative advance in early disease detection and stem cell research.

Quickfire News

• Boston Dynamics posted a new video of its Atlas robot doing push-ups: showcases advancements in dynamic movement control.

• AI21 Labs unveiled Jamba 1.5: a multilingual AI model series with 256,000 context length and permissive licensing for smaller organizations.

• Krea AI added Flux 1 to its platform: an advanced text-to-image AI model with 3-minute free generations for non-subscribed users.

• Perplexity AI is reportedly planning to introduce advertising: expected on its AI-powered search platform by Q4 of 2024.

• Anthropic launched LaTeX rendering support for Claude: allows the AI chatbot to display mathematical equations and expressions consistently.

• Google DeepMind employees urged the company to end military contracts: raised concerns over AI use in warfare and surveillance.

• Mark Zuckerberg and Spotify’s Daniel Ek advocated for Europe to embrace open-source AI: warned against complex regulations that could hinder innovation.

• Google AI Studio released a native prompt gallery: includes long context, multi-model inputs, and structured outputs for enhanced AI development.

• Anthropic supported California’s AI regulation bill after changes: stated that the benefits likely outweigh the costs for advanced AI development.

• Fetch.ai launched Innovation Lab in San Francisco: includes a $10 million fund to support early-stage AI agent startups.

• Google appointed Noam Shazeer as co-lead of its Gemini AI model development: Shazeer is a former Character.AI founder and long-time Google researcher.

• Imagination Technologies abandoned standalone NPUs: decided to integrate AI capabilities into GPUs instead and secured $100 million in financing.

• Chinese companies reportedly bypassed U.S. AI chip export restrictions: accessed banned technologies through Amazon Web Services’ cloud platform.

• Anthropic published system prompts for its Claude AI models: revealed instructions on behavior, capabilities, and personality traits.

• Inflection AI partnered with Data Transfer Initiative: allows Pi users to export conversations and plans to cap free usage, focusing on enterprise AI.

• Pharia released Pharia-1-LLM-7B: an open-source model optimized for German, French, and Spanish, excelling in domain-specific applications.

• IBM previewed Spyre, a new AI accelerator chip for IBM Z mainframes: designed to scale enterprise AI workloads with clustering capabilities.

• Hugging Face and Google Cloud partnered to release optimized Deep Learning Containers: aimed at building AI with open models on Google Cloud infrastructure.

• OpenAI hired former Meta executive Irina Kofman: to lead strategic initiatives, focusing on AI preparedness and safety.

• Apple announced a September 9 event: expected to debut the iPhone 16 with new generative AI features.

• Elon Musk endorsed California’s Senate Bill 1047: supports safety testing for large AI models, diverging from other tech leaders who oppose the regulation.

• Amazon plans to launch a delayed AI-powered Alexa subscription in October: features “Smart Briefing” AI-generated news summaries.

• xAI released new Grok features for premium subscribers: includes image generation suggestions and improved model selection in the iOS app.

• Anthropic announced the full release of its Artifacts feature for all Claude users: available on mobile apps after millions were created in the test phase.

• Fourier Intelligence unveiled GR-2 in a CGI teaser: a next-gen humanoid robot with swappable batteries, advanced hand dexterity, and a sleek design.

• Nvidia released NIM Agent Blueprints: a catalog of customizable AI workflows to help enterprises build and deploy generative AI applications.

• Google released Custom Gems for Gemini Advanced users: improved image generation with its Imagen 3 model across Gemini products.

• SoundHound AI partnered with MUSC Health: to deploy an AI agent for streamlining patient appointment management and access.

• Cerebras Systems launched an AI inference tool: aims to challenge Nvidia with claims of better performance and lower pricing.

• Klarna reduced the number of employees needed for customer queries: decreased resolution time from 11 to 2 minutes using AI.

• CoreWeave launched Nvidia H200 Tensor Core GPUs: first cloud provider to offer this advanced AI infrastructure.

• Midjourney teased an upcoming hardware launch: mentions “multiple efforts in flight,” but details on the specifics are not yet available.

Closing Thoughts

That’s it for us this week.

If you find any value from this newsletter : make sure to send the link below to your colleagues and friends :

https://www.blacklynxbrief.com/subscribe