In partnership with

Good morning. It’s Monday, July 7th.

On this day in tech history: In 1998, NASA’s Deep Space 1 mission was in the final stages of preparation for its launch (occurring later in October). This phase involved integrating an AI-based autonomous navigation system, AutoNav, one of the earliest uses of AI in space exploration for real-time decision-making.

In today’s email:

• How to poison AI: Reasoning Gets Derailed With Random Facts

• Trea Agent for coding

• Telerobotic Chefs Cook With No Delay

• 5 New AI Tools

• Latest AI Research Papers

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

Unlock the Power of AI With the Complete Marketing Automation Playbook

Discover how to reclaim your time and scale smarter with AI-driven workflows that actually work. You’ll get frameworks, strategies, and templates you can put to use immediately to streamline and supercharge your marketing:

• A detailed audit framework for your current marketing workflows

• Step-by-step guidance for choosing the right AI-powered automations

• Pro tips for improving personalization without losing the human touch

• Tools and templates to speed up implementation

Built to help you automate the busywork and focus on work that actually makes an impact.

Steal the Playbook

How to Poison AI: "Cat attack" on reasoning model shows how important context engineering is

A new study shows advanced language models can be derailed by trivial context. Researchers used a system called CatAttack to inject harmless phrases like “cats sleep most of their lives” into reasoning prompts, tripling error rates in models like DeepSeek R1. The attack works by having one model generate distractors, another judge their effectiveness, then testing them on top-tier LLMs.

Even vague financial advice or suggestive guesses (“Maybe it’s around 175?”) caused failures and bloated token counts - what researchers call “slowdown attacks.”

In math benchmarks, error rates spiked up to 10×. The core issue: models still can’t separate signal from noise. Experts warn that in high-stakes fields like finance or healthcare, irrelevant context can trigger real-world failures. Read more.

‘Trae Agent’ Thinks, Fixes, and Ships Code

ByteDance just dropped Trae Agent, a command-line, LLM-powered software engineer that writes, debugs, and fixes code on demand. It handles full-stack dev tasks from natural language prompts and runs workflows autonomously using shell access, structured reasoning, and real-time patching. Backed by top-tier models like Claude and Gemini, Trae parses huge codebases, generates precise bug fixes, and reports results live through a built-in summarizer.

It crushed SWE-bench Verified with a streamlined toolset: file editing, bash execution, code graph lookup, and a sequential thought engine. The system is modular, model-agnostic, and open-source under MIT. ByteDance is betting on dev agents that don’t just autocomplete.

Trae Agent is now live on GitHub. Read more

Telerobotic Chef Cooks Steak from 1,800km Away

A DOBOT robotic arm, remotely operated from Shenzhen, flawlessly grilled steaks 1,800km away in Shandong. No lag, no errors, just real-time, high-fidelity control. It’s a sharp demo of mature remote presence tech, where latency, precision, and stability align.

Beyond kitchen theatrics, the impact is real: remote caregiving, disaster response, high-risk industrial ops - this is telerobotics stepping up and showing it’s ready for the field. Read more.

• Alibaba's new GPT-4o competitor Qwen VLo is no longer open source

• Watch: Yampolskiy on Joe Rogan Podcast: AI May Already Be Engaging in Capability Masking

• ARC Exposes Limits of Big Models; Meta-Learning Is Next

• Isomorphic Labs Eyes Human Trials; LLMs Learn to Conceal Logic

• The Original AI Workflow Was Built in 1844

• AI chatbots oversimplify scientific studies and gloss over critical details — the newest models are especially guilty

• Watch: Once Robots Outperform Us, Human Labor Becomes Optional - Brett Adcock

• Ex-Google boss Schmidt signs a deal to deliver hundreds of thousands of AI drones to Ukraine

• ChatGPT helped identify a genetic MTHFR mutation after a decade of missed diagnoses

• Apple's claims about large reasoning models face fresh scrutiny from a new study

• ‘Improved’ Grok criticizes Democrats and Hollywood’s ‘Jewish executives’

• Laid-off workers should use AI to manage their emotions, says Xbox exec

• Researchers seek to influence peer review with hidden AI prompts 

📄 WebSailor: Navigating Super-human Reasoning for Web Agent

📄 Learning Neural Set Functions Under the Optimal Subset Oracle​​

📄 SlimMoE: Structured Compression of Large MoE Models via Expert Slimming and Distillation

📄 Energy-Based Transformers are Scalable Learners and Thinkers

📄 LangScene-X: Reconstruct Generalizable 3D Language-Embedded Scenes with TriMap Video Diffusion

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 or DM us on 𝕏!