Skip to main content
Back

OpenAI finds misbehaving model patterns, details how to rectify it

LEARN MORE

June 20, 2025 4 min read

Hey,

Welcome to AlphaSignal – the most read newsletter by AI developers. 


We bring you the top 1% of news, papers, models, and repos, all summarized to keep you updated on the latest in AI.

IN TODAY'S SIGNAL

Read time: 4 min 21 sec

🎖️ Top News

› OpenAI reveals how to re-align a misbehaving GPT-4o model in 30 steps without changing weights.

4,729 Likes

⚡️ Trending Signals

› Karpathy urges developers to build agent-friendly software in a Y-Combinator keynote.

30,957 Likes

› OpenAI rolls out Record mode in ChatGPT Mac app to Pro, Team, Enterprise, and Edu users.

5,294 Likes

› Midjourney launches its first video generation model, turns images into 5-second clips.

4,024 Likes

› Anthropic adds remote MCP server support to Claude Code, allowing you to pull real-time project context.

749 Likes

› OpenAI launches its official podcast. Sam Altman is the first guest.

4,359 Likes

📌 Assembly AI

› Transcribe speech in real-time with fast, accurate, immutable results.

  📝 Top Papers

› LLMs fail hard coding problems in new LiveCodeBench-Pro benchmark.

4,294 Likes

› New paper proposes six patterns to prevent prompt injection in agents.

1,435 Likes

› Stanford study finds workers workers don't want AI to perform tasks without human observation.

1,043 Likes

🧠 Deep Dive

› Oxford research suggests AI agents fail in performing longer tasks.



If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

TOP NEWS

AI Interpretability

OpenAI shows how to control misaligned behavior in GPT-4o by steering internal features

4,729 Likes

OpenAI research found a pattern inside GPT-4o that becomes more active when the model gives wrong or irresponsible answers after fine-tuning on bad data. This pattern, called the “misaligned persona,” is a result of fine-tuning on bad data and controls how the model behaves.


Key Findings

  • Sparse autoencoders (SAEs) isolate the misaligned persona feature from GPT-4o’s internal activations

  • The feature activates on unethical content in pretraining data

  • Misalignment appears in both supervised and RL-trained models, including GPT-4o and o3-mini

  • You can use this technique to inspect alignment-relevant features by training SAEs on activation traces


Feature Steering and re-alignment

Small amounts of correct data reset model behavior without changing weights or gradients.

  • Fine-tuning on 120 secure examples reversed misalignment in the GPT-4o checkpoint

  • The process required only 30 supervised steps to reach 0% misalignment

  • The re-alignment data came from a different domain than the misaligned task

  • This showed alignment generalizes across tasks as efficiently as misalignment


Community Feedback

Herr Teppich

"Emergent misalignment isn’t merely a failure of training or data, it often arises from the architectural depth and distributed cognition of the model itself."

Miles Wang

"We also find that monitoring misaligned persona feature activations can effectively classify misaligned and aligned models. In our paper, we show this can sometimes detect misalignment *before* it appears on our evaluation. We discuss how we might build on this work to make an early warning system for misalignment."


AlexPh

"One problem I noticed is that ChatGPT-4o and GPT-4.1 tend to "flatter" and ignore user errors to give more positive feedback to users rather than focusing on the content."

Learn more

TOP LECTURES

LLM Coding

New Benchmark Shows LLMs Score 0% on Complex Coding Problems

4,294 Likes

LiveCodeBench-Pro tests 584 real coding problems from contests like IOI and Codeforces. It shows top LLMs can’t solve hard problems, especially ones that need step-by-step logic or special formats like interactive input.

LLM Reasoning

New Study Finds LLMs Rely More on Recall Than Logic

1,435 Likes

The paper tests if LLMs use logic or just memory to solve math problems. It shows models like GPT-4o and Claude 3.5 often guess the answe using similar past examples instead of reasoning through the problem.

Workplace AI

Stanford Study Finds Workers favor AI’s help over full automation

1,043 Likes

A Stanford survey of 1,500 workers shows most want AI to handle repetitive tasks, not replace core work. The study warns of potential friction as AI advances, and suggests key job skills may shift from technical to interpersonal and organizational strengths.

Build Accurate Transcripts for Voice Agents in Real-Time

AssemblyAI’s new Universal-Streaming model delivers real-time speech-to-text with 300ms latency, accurate transcripts, and smart end-pointing.

Built for voice agents. Fast, reliable, and easy to deploy.

Try for free ↗️

DEEP DIVE

Agent Evaluation

Why AI Agents Fail on Long Tasks: The Constant Hazard Rate Explained

⇧ 9,294 Likes

The study argues that AI agents fail tasks because each subtask carries a fixed chance of failure. The longer the task, the more subtasks, and the higher the cumulative chance of failure.


If an agent has a 10% failure rate per 10-minute chunk, then for:

  • 1 hour (6 steps): 0.9⁶ ≈ 53% success

  • 4 hours (24 steps): 0.9²⁴ ≈ 8% success

  • 10 hours (60 steps): 0.9⁶⁰ ≈ 0.2% success

This is exponential decay. The success rate drops off as the number of subtasks increases. The model's "half-life" defines how long the agent can work with a 50% success rate.


AI agents struggle to recover from mistakes mid-task. Humans appear to scale better on long tasks, possibly because they adapt, reflect, or correct errors more effectively.


This model gives a simple tool to predict AI reliability on long-horizon tasks.

Read more

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

Work with us