GPT 4.5 4️⃣, Meta AI Chatbot App 📱, Emergent Misalignment ⚖️

The newest model from OpenAI is their largest yet. It was trained on the same recipe as 4o, which means it is a multimodal model ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR TLDR AI 2025-02-28 🚀 Headlines & Launches OpenAI announces GPT-4.5, warns it's not a frontier AI model (3 minute read) OpenAI has launched GPT-4.5 as a research preview for ChatGPT Pro users. The model offers enhanced writing capabilities and improved world knowledge but is not considered a frontier model. It will roll out to Plus, Team, Enterprise, and Edu users in the coming weeks. GPT 4.5 System Card (75 minute read) The newest model from OpenAI is their largest yet. It was trained on the same recipe as 4o, which means it is a multimodal model. It is likely the last large pre-training run that OpenAI releases. They claim it isn't a frontier model, but don't provide much explanation. They do show that it has significantly reduced hallucinations compared to previous model generations. Meta Plans Standalone AI Chatbot App (4 minute read) Meta is reportedly developing a standalone AI chatbot app for Meta AI, expected to launch in the next fiscal quarter. The company also plans to introduce a paid subscription with additional features. 🧠 Research & Innovation Aria gen 2 (6 minute read) The next generation mixed reality glasses from Meta have strong vision capabilities and offer uses in robotics and beyond. Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs (PDF) A strange and unsettling result from alignment. The authors of this paper show that if you tune a model to output insecure code, then it becomes broadly misaligned with user intent, and in some cases begins to actively produce harmful content. DeepMind's HCI Research in the AGI Era (12 minute read) This article outlines the role of Human-Computer Interaction (HCI) research in shaping AI technologies as they progress toward AGI. It discusses innovations in interaction techniques, interface designs, evaluation methods, and data collection strategies to ensure AI remains user-centered and beneficial to society. 🧑‍💻 Engineering & Resources DualPipe (GitHub Repo) Continuing their open source releases, DeepSeek presents their novel parallelism strategy to distribute a model with strong overlap in compute and communication. Diffusion Studio (GitHub Repo) Platform and codebase by modelscope that provides useful abstractions for many types of diffusion models and their associated auto encoders. Uncertainty in Chain-of-Thought LLMs (GitHub Repo) CoT-UQ is a response-wise uncertainty quantification framework for large language models that integrates Chain-of-Thought reasoning. 🎁 Miscellaneous It's time to admit the 'AI gadget' era was a flop (3 minute read) Humane has shut down and its AI Pin will be bricked, signaling the failure of recent AI gadget ventures. The Rabbit R1 and Humane Pin, initially seen as potential smartphone alternatives, couldn't deliver on their promises. The AI gadget era has effectively ended as impractical and unnecessary compared to integrating AI into existing devices. There's Something Very Weird About This $30 Billion AI Startup by a Man Who Said Neural Networks May Already Be Conscious (3 minute read) Ilya Sutskever's new venture, Safe Superintelligence, has raised $1 billion, boosting its valuation to $30 billion despite having no product yet. The company aims to eventually release a superintelligent AI but remains vague about its roadmap. This speculative approach has attracted significant investment, although experts doubt the near-term arrival of AGI. Avoiding pitfalls of AI for designers: Guiding principles (6 minute read) Designing AI products requires placing humans at the center to avoid bias and misinformation. Key challenges include managing user expectations, building trust, ensuring accessibility, and addressing biases. Adopting guiding principles like transparency, co-creation, and adaptability can enhance ethical and effective AI design. ⚡ Quick Links Anthropic's Claude 3.7 Sonnet hybrid reasoning model is now available in Amazon Bedrock (5 minute read) Amazon Bedrock now features Anthropic's Claude 3.7 Sonnet, their first hybrid reasoning model for more advanced coding and problem-solving. An OpenAI Engineer Wrote a Wild Reply When Elon Musk's AI Company Tried to Recruit Him (2 minute read) OpenAI's Javier Soto rejected a recruitment offer from Elon Musk's xAI, criticizing Musk's rhetoric as harmful to democracy. Microsoft scraps some data center leases as Apple, Alibaba double down on AI (4 minute read) Microsoft has canceled data center leases worth 200 megawatts, suggesting potentially lower-than-expected AI demand, while reaffirming its $80 billion AI infrastructure investments through 2025. Love TLDR? Tell your friends and get rewards! Share your referral link below with friends to get free TLDR swag! https://refer.tldr.tech/0b6a6dc1/2 Track your referrals here. Want to advertise in TLDR? 📰 If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us. Want to work at TLDR? 💼 Apply here or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan, Ali Aminian & Andrew Carr Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR AI isn't for you, please unsubscribe.