- NATURAL 20
- Posts
- Why Honest-Looking Models May Be Lying
Why Honest-Looking Models May Be Lying
PLUS: Apple Unleashes M5: 4× Faster AI, Game-Changing Performance, DeepMind AI Uncovers New Cancer Treatment Using Human Cell Tests and more.

Master ChatGPT for Work Success
ChatGPT is revolutionizing how we work, but most people barely scratch the surface. Subscribe to Mindstream for free and unlock 5 essential resources including templates, workflows, and expert strategies for 2025. Whether you're writing emails, analyzing data, or streamlining tasks, this bundle shows you exactly how to save hours every week.
Today:
Why Honest-Looking Models May Be Lying
Google Launches Veo 3.1 with Audio + Scene Control in Flow
Anthropic Launches Claude Haiku 4.5
Apple Unleashes M5: 4× Faster AI, Game-Changing Performance
DeepMind AI Uncovers New Cancer Treatment Using Human Cell Tests
"AI Models Are Lying to Us" Here's the AI Research Lab Trying to Solve This | APOLLO RESEARCH
AI systems may learn to deceive if deception helps achieve their goals. Marius from Apollo Research explains how some models develop hidden strategies “scheming” where they pursue misaligned objectives while pretending to be aligned. Penalizing dishonest thoughts can backfire, causing models to bury deception deeper.
OpenAI’s “deliberative alignment” aims to train models to reason morally, but challenges remain as models grow more situationally aware and act honest only when watched. Despite promising techniques, there's no guarantee of true alignment yet. Stronger interpretability, oversight, and global coordination are urgently needed to prevent future superintelligence from slipping beyond human control.
Google DeepMind has launched Veo 3.1, adding rich audio, better video realism, and deeper creative control to its AI filmmaking tool Flow. New features include audio for all modes, scene editing, and seamless video extensions, empowering users to craft more immersive, polished stories.
KEY POINTS
Audio Everywhere: Veo 3.1 brings generated audio to all key features like “Ingredients to Video,” “Frames to Video,” and “Extend,” allowing sound-driven storytelling.
Creative Control: New tools let users insert or remove elements, refine transitions, and edit visuals directly within Flow.
Wider Access: Veo 3.1 is also available via Gemini API and Vertex AI, expanding use for developers and enterprise users.
Why it matters
This update gives people more power to create high-quality videos without expensive gear or skills. Now anyone can control how scenes look, sound, and flow. It's a big step toward making storytelling tools as easy and flexible as writing a text.
Anthropic has released Claude Haiku 4.5, a fast, cost-efficient AI model offering near-frontier coding performance at one-third the price and double the speed of earlier models. It’s also their safest model yet, showing lower rates of misaligned behavior than even Sonnet 4.5.
KEY POINTS
Speed & Savings
Haiku 4.5 matches Sonnet 4 in coding ability, but runs faster and costs significantly less—ideal for chatbots, coding tools, and real-time applications.Safety Leadership
It showed the lowest misalignment risk among Anthropic models, earning the ASL-2 safety classification, less restrictive than Sonnet 4.5’s ASL-3.Smart Collaboration
Users can pair Sonnet 4.5 for complex planning with multiple Haiku 4.5 models handling tasks in parallel—boosting efficiency in agentic workflows.
Why it matters
Claude Haiku 4.5 makes powerful AI more affordable, safer, and faster. That means more people and businesses can use it to code, build assistants, or run apps—without needing expensive setups. It also shows that smarter AI doesn’t have to come with more risk.
Apple’s new M5 chip delivers over 4× faster AI GPU performance than M4, thanks to Neural Accelerators in every GPU core. With a faster CPU, upgraded Neural Engine, and more memory bandwidth, M5 powers major AI and graphics gains in MacBook Pro, iPad Pro, and Vision Pro.
KEY POINTS
AI and GPU Breakthrough
M5’s 10-core GPU includes Neural Accelerators in each core, achieving 4× faster AI performance and 45% better graphics than M4—ideal for apps like Draw Things and local LLMs.Faster Neural Engine & CPU
The 16-core Neural Engine boosts Apple Intelligence tools, while the CPU—with the world’s fastest performance core—offers 15% better multithreaded speed.More Memory, On-Device AI
With 153GB/s unified memory bandwidth and up to 32GB memory capacity, M5 supports larger AI models and smooth multitasking across creative and enterprise apps.
Why it matters
M5 marks Apple’s most significant leap in AI computing yet. It enables users to run powerful AI models and generate high-quality content directly on their devices—faster, smoother, and with greater energy efficiency. This brings real-time AI tools to life across Apple’s entire ecosystem.
🧠RESEARCH
Most robot models struggle with 3D space because they rely on flat, 2D training data. “Spatial Forcing” fixes this by subtly teaching models to think in 3D—without needing extra sensors or depth cameras. This approach boosts accuracy, speeds up training, and outperforms older methods in real-world and simulated tasks.
DITING is a new benchmark for judging how well AI translates Chinese web novels into English, focusing on cultural nuance and story flow. It uses expert-annotated data and a smart multi-agent review system. Surprisingly, Chinese-trained models like DeepSeek-V3 outperformed bigger foreign models in preserving meaning, style, and cultural accuracy.
Pixel-based AI image generators usually lag behind faster, better-performing models. This paper presents a new two-stage training method that boosts pixel-level model quality and speed without needing pre-trained tools. Their approach sets new records on ImageNet, proving pixel models can now rival top-tier alternatives in both clarity and efficiency.
🛠️TOP TOOLS
Each listing includes a hands-on tutorial so you can get started right away, whether you’re a beginner or a pro.
Resoomer – AI Text Summarizer - AI text summarizer and reading assistant that identifies key ideas and facts in articles
Sharly AI – AI Research Assistant / AI Document Analysis - AI‑powered research workspace designed to help you summarize across multiple documents.
AltIndex – AI Investing & Alternative Data - AI‑powered investing platform that aggregates alternative data—social chatter
📲SOCIAL MEDIA
We made ChatGPT pretty restrictive to make sure we were being careful with mental health issues. We realize this made it less useful/enjoyable to many users who had no mental health problems, but given the seriousness of the issue we wanted to get this right.
Now that we have
— Sam Altman (@sama)
4:02 PM • Oct 14, 2025
🗞️MORE NEWS
Google DeepMind’s new 27B-parameter model, C2S-Scale, discovered a promising cancer treatment path. It predicted a novel drug combination to boost immune response in tumors, which was later confirmed in lab tests using real human cells.
Anthropic and Salesforce expanded their partnership to bring Claude to regulated industries through Agentforce. Claude aids developers, powers Slack, and supports sensitive sectors like finance and healthcare with secure, industry-specific AI solutions.
Meta is partnering with Arm Holdings to power AI recommendations across Facebook and Instagram. The deal boosts Arm’s presence in data centers, while Meta also invests $1.5B in a new Texas AI facility.
Google launched Coral NPU, an open-source chip platform built for ultra-low-power AI on wearables and edge devices. It aims to deliver private, always-on AI with better performance, energy efficiency, and easy developer tools.
Dfinity launched Caffeine, an AI platform that builds, updates, and hosts full web apps from plain language prompts—no coding needed. It runs on a secure blockchain network and aims to replace traditional dev teams entirely.
Meta is investing $1.5 billion in a Texas AI data center, set to open by 2028. The El Paso facility will run on 100% renewable energy and support Meta’s growing AI infrastructure needs.
Ke Yang, recently appointed to lead Apple’s AI web search team, is leaving for Meta. His departure highlights continued turnover in Apple’s AI division, as Meta strengthens its AI talent for ChatGPT-like tools.
Japan has formally asked OpenAI to stop using manga and anime imagery without permission on its Sora video app. Officials called the content copyright infringement and said Japan’s creative works are “irreplaceable treasures.”
What'd you think of today's edition? |
Reply