Good morning. It’s Wednesday, September 11th.

Did you know: On this day in 1990, Archie, considered to be the first Internet search engine, was launched. It allowed users to search for files on public FTP servers that supported anonymous login.

In today’s email:

• OpenAI’s Strawberry Dropping Soon

• Apple’s AI Just Google Lens Copy?

• Google Founder’s AI Obsession

• AI Forecaster Beats Human Experts

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

^{In partnership with BAGEL}

AI is learning to forget—and that's a game changer.

Bagel’s latest on Machine Unlearning reveals how companies like Microsoft and JPMorgan are using this breakthrough tech to erase sensitive data from AI models without retraining.

Imagine AI that can delete mistakes, protect your privacy, and stay sharp. From cutting-edge methods like SISA to decentralized solutions, this is the future of AI safety. Stay ahead of the curve. Subscribe to Bagel and dive into the new era of AI that remembers how to forget.

OpenAI to launch new logic-focused AI model "Strawberry" soon

OpenAI's forthcoming AI model, "Strawberry," is now expected to launch within the next two weeks, ahead of the originally planned fall release, according to testers cited by The Information.

This model, noted for its focus on enhanced logical reasoning, will introduce a 10-20 second "thinking" period before generating responses, allowing it to address complex tasks in mathematics, coding, and business planning with greater accuracy. Unlike its predecessors, Strawberry employs a specialized post-training technique for fine-tuning pre-trained models, potentially reducing the need for additional user prompts.

Strawberry will be a standalone option within ChatGPT, accessible via a new pricing structure that may include rate limits and a higher-tier option for quicker responses. However, early feedback highlights some limitations, such as slower processing times for simple queries and challenges in personalized response delivery. Strawberry is expected to support software development and research while paving the way for OpenAI’s next model, "Orion." Read more. 

Apple’s Visual Intelligence is a built-in take on Google Lens

Apple’s new Visual Intelligence, part of iOS 18’s Apple Intelligence suite, turns the iPhone’s camera into a pocket-sized know-it-all. By holding the capacitive Camera Control button on the iPhone 16 models and pointing the camera at an object, users can instantly uncover details about their surroundings.

Powered by on-device intelligence and Apple’s services — without storing your images — the feature can pull information from signs, flyers, or even restaurant hours with a quick snap. It also plays nice with third-party models for tasks like product searches or study help. Expected later this year, it's Apple's quiet step into AI-driven object recognition. Read more.

Sergey Brin says he's working on AI at Google 'pretty much every day'

Sergey Brin, Google co-founder and former Alphabet president, has confirmed he is actively involved in AI projects at Google, stating he works on them nearly every day. During an interview at the All-In Summit in Los Angeles, Brin emphasized his excitement about AI advancements and his commitment to staying engaged with the field.

He challenged the notion that significantly increasing computational power is necessary, arguing that recent improvements in algorithms may be advancing faster than the increase in computing resources. This is the first time Brin has publicly confirmed his return to Google, following earlier reports of his involvement in AI projects. Read more.

CAIS claims their AI forecaster "FiveThirtyNine" beats human experts at predicting future events

The Center for AI Safety (CAIS) has launched FiveThirtyNine, an AI forecasting system based on GPT-4o, claiming it surpasses human experts in prediction accuracy. In tests conducted on the Metaculus platform, FiveThirtyNine achieved 87.7% accuracy, outperforming human forecasters who scored 87.0%. This AI generates probability estimates for user-defined questions on topics such as politics and global events.

Despite its advancements, FiveThirtyNine has limitations, including its reliance on pre-existing training data, lack of optimization for specific use cases, and weaker performance on very recent or short-term events. The developers see broad potential for FiveThirtyNine in decision support and risk assessment. Read more.

Etcetera: Stories you may have missed

ServiceNow introduces a library of enterprise AI agents you can customize to fit your workflow

iOS 18 will launch next week with new ways to customize your homescreen

Microsoft enhances few-shot prompting with a dynamic approach

SiMa.ai debuts new system-on-chip for multimodal AI at the network edge

Mastering AI chatbots requires hands-on experience rather than just technical expertise

AMD is turning its back on flagship gaming GPUs — to chase AI first

AI2's new model aims to be open and powerful yet cost effective

LightEval: Hugging Face's open-source solution to AI's accountability problem

Audible recruits voice actors to train audiobook-generating AI

James Earl Jones Signed Over Rights For AI To Recreate Darth Vader’s Voice

MiniMax's AI video tool can create Star Wars battles in seconds – here's why that matters

DeepSeek-V2.5 wins praise as the new, true open source AI model leader ​

The real power of Apple Intelligence will show up in third-party apps

Reflection 70B launch mired in controversy as third-party benchmarks disappoint

Quantum algorithm adopted by Google and IBM

📄 Towards a Unified View of Preference Learning for Large Language Models: A Survey

📄 MMEvol: Empowering Multimodal Large Language Models with Evol-Instruct ​​

📄 MemoRAG: Moving towards Next-Gen RAG Via Memory-Inspired Knowledge Discovery 

📄 INTRA: Interaction Relationship-aware Weakly Supervised Affordance Grounding

📄 OneGen: Efficient One-Pass Unified Generation and Retrieval for LLMs  

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email!