ChatGPT Atlas, OpenAI’s new web browser

OpenAI’s Foray into Search with ChatGPT Atlas

OpenAI’s new web browser, ChatGPT Atlas, represents a natural strategic pivot for the company. Having already established ChatGPT as a primary internet entry point for over 350 million users, creating a dedicated browser is a logical next step to capitalize on this user base and open new monetization avenues, such as advertising. The browser aims to create a more integrated and seamless user experience by caching a user’s history to inform and personalize their navigation.

Practical User Experience and Agent Capabilities

Initial hands-on use of Atlas reveals its potential benefits. The browser integrates ChatGPT into a side tab, allowing users to query the context of their open web pages directly.

• Contextual Understanding: In one test, the browser was able to analyze a scientific article about beer brewing, answering questions about the methodology and findings. More impressively, it could interpret images within the article. When asked, “did one of the beers look more oxidized than the other?”, it correctly identified the relevant image and provided an analysis.
• AI Agent Functionality: The built-in agent mode demonstrated its ability to perform multi-step tasks. When tasked with finding a specific used book with multiple criteria (hardcover, “very good” condition), the agent searched various websites. When the user noted the book cover seemed incorrect, the agent initiated a new search, looked up the ISBN, and correctly identified the book as the UK version, then offered to search for the US version using a different ISBN. This highlights a level of autonomous problem-solving that streamlines complex searches.

The Future of Browsers and AI Operating Systems

The emergence of AI-native browsers like Atlas raises fundamental questions about the future of web navigation. As AI agents become more capable, the need for users to interact directly with websites may diminish, paving the way for an “internet for agents, by agents.”

This trend suggests OpenAI’s larger ambition is to create an AI-centric operating system. This would not be a low-level OS like macOS or Windows, but rather a higher-level abstraction layer that orchestrates AI tools, workflows, plugins, and agents. However, this evolution brings significant risks that need to be managed, including:

• Data and Privacy: Ensuring users retain control over their data.
• Agent Risks: The potential for “comet jacking,” where agents are hijacked for malicious purposes.
• Transparency and Control: As AI systems become more autonomous, maintaining user control and understanding their decision-making processes becomes more challenging.

Andrej Karpathy on the Limitations of AI Agents

Andrej Karpathy, a prominent voice in the AI community, recently expressed a pessimistic view on the current state of AI agents, stating, “they just don’t work, they don’t have enough intelligence… [it will take a] decade to work through all those issues.” This serves as a crucial reality check against industry overhype.

Technical Hurdles: Training Data and Reinforcement Learning

Karpathy points to two fundamental issues hindering agent development:

1. Training Data Quality: A significant portion of the data used to train large language models is irrelevant or contains errors. The models learn by identifying the signal from an immense amount of noise, but this noisy foundation can lead to unreliable behavior.
2. Failings of Reinforcement Learning (RL): Karpathy declared RL as “pretty bad” for this application. RL systems reward correct outcomes without scrutinizing the process. This can lead to agents developing inefficient or illogical “chain of thought” reasoning, often getting stuck in loops or taking non-intuitive steps—as seen when the book-buying agent returned a UK version of a book for a US-based user.

A Call for Realistic Expectations

Karpathy’s stance is a call to temper expectations and acknowledge that we are in the very early stages of agent development. While exciting, these systems are not yet deterministic or reliable enough for high-stakes, production-grade applications without a human in the loop. The industry must distinguish between low-stakes consumer tools and enterprise systems where a 1-2% error rate is unacceptable. The path to production-ready agents likely lies in developing more deterministic frameworks, potentially through “generative computing” that uses software-defined policies and requirements to govern agent behavior.

DeepSeek-OCR and Visual Context Compression

A paper from DeepSeek AI introduces an innovative approach to address the long-context problem in LLMs, which stems from the quadratic scaling complexity of attention mechanisms. The paper, “DeepSeek-OCR: Context Optical Compression,” proposes converting text tokens into compressed “visual tokens” or images.

This method acts as a form of “document distillation.” A two-stage system uses a vision encoder to transform text into a compact image and a decoder to convert it back.

• Performance: The best-performing model retained approximately 97% of the information through a full encode-decode cycle.
• Mimicking Human Memory: The paper draws a parallel to human memory. The high-fidelity model is akin to short-term memory, retaining near-perfect detail. A lower-fidelity model that uses a “blurry” image and loses more information is analogous to long-term memory, where we recall abstract concepts but forget specific details over time. This raises interesting questions about the utility of building human-like “forgetting” mechanisms into AI.

The core innovation is a compression bridge that allows models to handle vast amounts of information using a smaller number of tokens, with potential applications in swapping decoders to translate the visual context into different languages or modalities.

Can LLMs Get “Brain Rot”?

A recent paper titled “LLMs can get brain rot” explores whether AI models suffer cognitive decline when trained on low-quality social media content, similar to the effect it has on humans. Researchers fine-tuned models on datasets of “junk web text” characterized as either short-form (like tweets) or sensationalist.

The results showed a “persistent and systematic” decline in model performance across reasoning, long-context understanding, and safety benchmarks. Furthermore, the models developed undesirable “dark traits.”

• Garbage In, Garbage Out: The core finding is that low-quality data leads to degraded model capabilities. The “brain rot” was difficult to reverse with subsequent instruction tuning or data cleansing, suggesting the damage was embedded in the model’s weights.
• Personality Shifts: Interestingly, short-form “engagement” content (like tweets) had a more significant negative impact on the model’s personality than sensationalist content. Training on this data increased narcissism, agreeableness, and extroversion, and led to an “abrupt stop in thinking,” where the model would fail to complete its reasoning process.
• Implications for Training: This study underscores the critical importance of data quality. The best practice is to build foundation models on a core of dense, deep, factually correct data. Personality, tone, and style should be managed through context engineering and prompting rather than by polluting the model’s core knowledge base with low-quality training data.

• Tags:
• AI Agents
• AI Safety
• Reinforcement Learning
• Feature
• AGI
• Andrej Karpathy
• Data Quality
• ChatGPT Atlas
• DeepSeek-OCR
• Context Window