😮 Apple says AI can’t reason

Also: Meta plans $10B investment in Scale AI

Welcome, AI enthusiasts

Apple recently released a new research paper that sparked major debate over the weekend across the AI world. It questions whether today’s models can truly reason or if we’ve been reading too much into what they do. Let’s dive in! 

In today’s insights:

  • Apple challenges AI’s reasoning hype

  • Meta plans $10B investment in Scale AI

  • Gemini 2.5 Pro gets a major upgrade

Read time: 3 minutes

LATEST DEVELOPMENTS

Source: Luis Rijo

Evolving AI: A new Apple study shows reasoning models break down fast when puzzle complexity increases. This raises serious questions about current AI hype.

Key Points:

  • Apple tested LLMs on logic puzzles with rising difficulty.

  • Models gave up when problems got too complex, even with solutions.

  • Study questions whether today’s AI can really ā€œreasonā€.

Details:

Apple researchers built a controlled puzzle environment to study reasoning. Tasks included River Crossing, Tower of Hanoi, and Blocks World. The goal was to see if language models could solve problems better with more compute or by being given the correct algorithm. They couldn’t. In one case, a model handled 100 moves in Tower of Hanoi but failed after just 5 moves in River Crossing. The study found performance collapsed in high-complexity tasks, with models giving shorter and weaker answers. Even when given the correct method directly, their reasoning didn’t improve. Apple calls this the illusion of thinking.

Why It Matters:

This study is not peer-reviewed and relies on puzzles, not real-world tasks. It raises real concerns about the limits of current reasoning models. If AI fails even when given clear instructions, that challenges claims around near-term AGI. But critics say the setup might not reflect practical reasoning. Some argue models are optimizing for brevity, not failing outright. Others note that executing long algorithms step-by-step may lie outside what language models were built for. So while the results are worth taking seriously, they shouldn't be taken as the final word.

Source: JOSH EDELSON / AFP

Evolving AI: Meta is reportedly planning a $10B investment in Scale AI, signaling a shift from open-source models to data dominance.

Key Points:

  • Meta wants to invest $10 billion in Scale AI, a top provider of labeled training data.

  • Scale AI already works with OpenAI, Microsoft, and the US Department of Defense.

  • The move could help Meta recover from Llama 4’s weak launch and support military projects.

Details:

Meta is in talks to make its biggest outside AI investment yet: $10 billion into Scale AI. The deal, first reported by Bloomberg, isn’t finalized but would mark a major pivot for Meta, which has mostly built its AI models internally. Scale AI’s core business is labeling massive datasets, a critical but often overlooked part of training large models. Its clients include OpenAI and Microsoft. Last year it brought in $870 million, and it’s aiming for $2 billion in 2025. If the deal goes through, Meta could gain tighter control over its data pipeline and tap exclusive datasets, especially for projects like the Aria headset and Defense Llama, a military-focused version of its LLM.

Why It Matters:

After Llama 4’s underwhelming launch, Meta appears to be shifting focus from scaling models to improving data quality. A major investment in Scale AI would mark a clear step toward strengthening its training pipeline. While the deal isn’t final, it highlights how data is becoming just as important as model architecture.

Evolving AI: Google just rolled out a new preview of Gemini 2.5 Pro. They’re calling it their smartest model yet, with better results in coding, reasoning, STEM, and image tasks.

Key Points:

  • Gemini 2.5 Pro now tops user-preference leaderboards like LMArena and WebDevArena.

  • Fixes writing issues from earlier updates and brings back creative quality.

  • Adds ā€œthinking budgetsā€ in the API so devs can control cost and latency.

Details:

The update brings major gains across multiple benchmarks, including Aider Polyglot, GPQA, and HLE. It now handles code in multiple languages more reliably and improved how it formats longer text responses. Developers can set thinking budgets in the API to manage compute usage, which helps with cost and performance control. The upgraded model is already live as a preview in Google AI Studio, Vertex AI, and the Gemini app, with full rollout expected soon.

Why It Matters:

Google is moving faster with more frequent upgrades and quicker fixes. This release closes the creative writing gap from the last version and gives developers more control. Gemini is becoming a real competitor across both technical and creative use cases.

 šŸ‘€ Click on the image you think is real

QUICK HITS

🧠 Ohio State announces every student will use AI in class.

šŸ’¼ Three-quarters of surveyed billionaires are already using AI.

šŸŽ® Why AI may be the next power player in the $455 billion gaming market.

āš–ļø Lawyers could face ā€˜severe’ penalties for fake AI-generated citations, UK court warns.

🚨 Meta’s platforms showed hundreds of ā€œnudifyā€ deepfake ads, CBS News investigation finds.

šŸ“ˆ Trending AI Tools

  • 🧠 Kin – An AI assistant that helps you reflect, plan ahead, and stay organized. It adapts as you go (link)

  • šŸ’¼ AIApply – Practice interviews with an AI coach that gives feedback based on your resume and how you speak (link)

  • šŸ› ļø Athena Agents – Build AI agents to automate workflows and tackle complex tasks with less effort (link)

  • šŸ” Fabric – Store notes, links, and files in one place. AI keeps everything easy to find and sort (link)

  • šŸ’¬ ChatFlow – Automate support and grow your audience faster with AI-powered chatbots (link)

Reply

or to participate.