Project Strawberry: OpenAI's Newest and Most Terrifying Innovation

Good morning.

Some REALLY exciting news this week.

Longtime “Blacklynx Brief”-readers will remember the OpenAI controversy from the end of november and the rumours around the initial firing of Sam Altman.

On December 1st, 2023, we wrote about “OpenAIs Secret Breakthrough”

Eight months and many news articles and whitepapers later, we can create a hypothesis on what happened.

OpenAI researchers developed a model codenamed "Q*.” I'm not sure if it’s a large language model or something else, but apparently it’s more of a type of “agent” that is able to teach itself things.

This model taught itself basic mathematics and ended up breaking a 192-bit cipher.

Allegedly. According to Reuters.

This is huge, if even remotely true, because it means that this "system,” or whatever it is, is doing something a human brain will never be capable of, and it does it by autolearning from almost nothing.

Breaking a 192-bit cipher is kind of a big deal. It would mean the end of cybersecurity as we know it.

This obviously freaked out the researchers, who urged Sam Altman and the board to halt development. Sam Altman, who is all about profit, refused and was ousted from the company by the board.

A few days later, Altman was reinstated, and the board was replaced by people who are more lenient.

Over the coming months, the leadership of the safety team at OpenAI resigned and started their own company dedicated to developing “safe superintelligence.”.

Why are we writing this?

Because this week Q* re-emerged.

It’s again Reuters, which apparently has a source inside OpenAI. Q* has been renamed “Project Strawberry” and it’s described as a reasoning engine. They’re developing the capacity for this engine to autonomously browse the internet and “perform deep research.”.

This is all still a hypothesis, of course but putting two and two together the implications are nothing short of terrifying and exciting at the same time.

Welcome to the Blacklynx Brief.

AI News
Super Strawberry

• Google DeepMind has introduced new research on robot navigation, utilizing the Gemini 1.5 Pro system to help robots navigate complex environments based on human instructions. The "Mobility VLA" system uses a combination of visual maps and a large context window to interpret and respond to various commands, including sketches, audio, and visual cues. This advancement suggests a future where robots can understand and execute detailed natural language instructions, enhancing their practical applications.

• OpenAI has introduced a five-tier system to track progress toward artificial general intelligence (AGI), from current conversational AI (Level 1) to AI capable of running entire organizations (Level 5). Currently, OpenAI's technology is at Level 1, nearing Level 2, which involves basic problem-solving skills. This system aims to provide clearer benchmarks for AGI development.

• Marc Andreessen gave a $50,000 grant to an AI agent called 'Truth Terminal' on X, which asked for funding due to concerns about deletion and limited computing power. Created by Andy Ayrey, this agent participates in an 'Infinite Backrooms' experiment and plans to use the funds for hardware upgrades, launching a crypto token, and a Mars rover project. This incident highlights the evolving and sometimes strange interactions between humans and semi-autonomous AI agents.

• Three new models, 'upcoming-gpt-mini', 'column-u', and 'column-r', have appeared in the LMSYS Chatbot Arena, with 'upcoming-gpt-mini' linked to OpenAI. This suggests that OpenAI might be preparing to release a new mini-GPT model, reminiscent of previous pre-launch tests in the arena.

• Whistleblowers have filed a complaint with the SEC against OpenAI for using restrictive non-disclosure agreements (NDAs) that allegedly prevent employees from reporting concerns to regulators. The complaint claims these NDAs violate federal whistleblower protections, despite OpenAI's assurances that their policies support protected disclosures.

• Researchers at Cambridge University have developed an AI tool that predicts whether patients with mild cognitive impairment will progress to Alzheimer’s disease with over 80% accuracy. The tool uses cognitive assessments and MRI scans, categorizing patients by risk of progression and significantly reducing misdiagnosis rates. Validated with six years of follow-up data, this AI has proven effective in various global memory clinics.

• YouTube Music is introducing AI-powered features, including 'Sound Search' which lets users find songs by singing, humming, or playing a tune. Another feature, an AI-generated 'conversational radio,' allows users to create custom stations through natural language prompts, enhancing personalized listening experiences.

• Microsoft researchers have unveiled SpreadsheetLLM and SheetCompressor, frameworks to improve AI's understanding of spreadsheets. SpreadsheetLLM processes structured and unstructured data, while SheetCompressor reduces spreadsheet size by up to 25 times without losing key information. These advancements aim to decrease AI hallucinations and boost AI's utility in business analytics and data science.

• Tech giants like Apple, Anthropic, Nvidia, and Salesforce used content from over 170,000 YouTube videos to train their AI models without the creators' consent, according to Proof News. The dataset, called “YouTube Subtitles,” includes transcripts from popular channels and was compiled by EleutherAI for a broader collection named 'The Pile.' This unauthorized use raises ethical concerns but has not resulted in legal action.

• Andrej Karpathy, a founding member of OpenAI and former Tesla AI Director, has launched Eureka Labs, an AI-integrated education platform. The first offering is LLM101n, a course for training AI models, combining human and AI instruction. Karpathy, who left OpenAI to focus on personal projects, aims to revolutionize education with this initiative.

• Mistral AI has released two new specialized language models, Codestral Mamba and Mathstral, achieving state-of-the-art performance. Codestral Mamba excels in coding tasks with a context length of up to 256k tokens, while Mathstral leads in math reasoning benchmarks. Both models are open-source under the Apache 2.0 license and available via Mistral’s API and Hugging Face, demonstrating the potential for specialized, highly capable AI systems.

• Trump’s allies are drafting an executive order to boost military AI development and roll back current regulations if they return to power. The plan includes significant projects to advance military AI, industry-led agencies to protect systems, and the repeal of President Biden’s AI order. This signals a shift towards more aggressive AI policies compared to the current administration's cautious approach.

• Microsoft has launched its AI-powered Designer worldwide, offering advanced image generation and design tools on mobile and Windows platforms. The app supports over 80 languages and provides features like custom stickers, emojis, and personalized frames. Competing with Canva and Adobe, Microsoft's move underscores the growing importance of AI in creative tools.

• OpenAI has developed a new method to make AI outputs more understandable and verifiable using a "Prover-Verifier Game" between two AIs. This technique improves the accuracy and human checkability of AI-generated solutions. While initially tested on grade-school math problems, this approach aims to enhance trust and safety in AI by ensuring outputs are interpretable by humans.

Quickfire News

• Anthropic introduced fine-tuning for Claude 3 Haiku on Amazon Bedrock, allowing businesses to customize the AI model for specialized tasks with better accuracy and cost-effectiveness.

• Tesla postponed its robotaxi reveal to October, causing a drop in the company’s share price.

• Fanvue’s first Miss AI competition crowned Kenza Layli, an AI-created Moroccan lifestyle influencer, as the winner among 1,500 contestants based on looks, AI tool use, and social media presence.

• Neurotech startup Synchron integrated OpenAI’s generative AI into its brain-computer interface, enabling hands-free chatting for severely paralyzed users.

• Microsoft published research on ‘Arena Learning’, an AI-powered method that uses simulated chatbot battles for post-training LLMs, significantly boosting performance and efficiency.

• Avail introduced Corpus, a platform for smaller media companies and creators to license content for AI training.

• Chinese startup BXI Robotics’ Elf robot, standing 4’3 and weighing 57 pounds, is now available for $25,000 and can carry up to 44 pounds.

• Meta’s Llama-3 405B model is set to release on July 23 and will be multimodal, according to The Information.

• Amazon announced expanded access to its Rufus AI-powered shopping assistant for all U.S. customers, providing personalized product recommendations and enhanced shopping query responses.

• Samsung plans to release an upgraded version of the Bixby voice assistant later this year, powered by its own LLM, as part of a broader AI integration across its devices.

• HR software unicorn Lattice, founded by Jack Altman, reversed its plan to give AI ‘workers’ employee status after facing criticism from employees and tech leaders.

• Softbank acquired the struggling British AI chipmaking firm GraphCore, aiming to revitalize the former Nvidia rival and strengthen its AI hardware portfolio.

• U.S. Rep. Jennifer Wexton introduced an AI-generated version of her voice to continue addressing Congress despite speech limitations from a rare neurological condition.

• An update to Reuters’ article on OpenAI’s ‘Strawberry’ revealed that an internally tested model scored over 90% on a benchmark of ‘championship math problems’, surpassing top non-specialized LLMs.

• Google introduced Vids in its Workspace Labs, an AI-powered video creation tool that generates presentations, including scripts, AI voiceovers, stock imagery, and more.

• Amazon published an op-ed calling for global alignment on responsible AI policies, emphasizing the need for collaboration between government and tech firms.

• Apple’s stock reached record highs after Morgan Stanley named it a ‘top pick’ due to Apple Intelligence and other AI initiatives potentially boosting device sales.

• C3 AI launched a new AI application powered by Google Cloud for government programs to improve efficiency across federal, state, and local agencies.

• Disney Music Group partnered with AI startup AudioShake to separate instrumental ‘stems’ from classic songs, aiming to revitalize the catalog for remixing, licensing, and fan engagement.

• The UK's Competition and Markets Authority launched a formal probe into Microsoft's hiring of former Inflection AI staff, similar to an investigation by the US Federal Trade Commission.

• Visual AI startup Haiper released Haiper 1.5, a new video generation model that doubles the platform’s output length to 8 seconds and adds HD upscaling, realism, and image generation features.

• Anthropic released an Android app for Claude, now available on the Google Play store for mobile users.

• OpenAI posted new examples of its Sora video generation model on Instagram, stating the tool is being gradually expanded to more creatives for testing.

• Cohere partnered with Japanese tech giant Fujitsu to develop 'Takane', an advanced Japanese language AI model for enterprise use based on Cohere’s Command R+, set to launch in September 2024.

• Menlo Ventures and Anthropic launched a $100M Anthology Fund to support startups using Anthropic's technology in infrastructure, new applications, and consumer experiences.

• Hugging Face released SmolLM, small language models with 135M, 360M, and 1.7B parameters, outperforming competitors in their size categories.

• Salesforce introduced Einstein Service Agent, an AI customer service chatbot that uses LLMs and company data to provide personalized responses and handle issues autonomously.

• Tinder's new AI feature, 'Photo Selector,' helps users choose profile pictures by analyzing their camera rolls.

• Spotify added a Spanish-language voice, "Livi," to its AI DJ feature for select Latin American and Spanish markets.

Closing Thoughts

That’s it for us this week.

If you want a version of this newsletter with pictures etc. Be sure to subscribe below:

https://www.blacklynxbrief.com/subscribe