This week in AI — Nov 2 – 8, 2026

The GPT-Rosalind-6 model achieved a 37.7% single-attempt success rate on GeneBench Pro v13 tests, outperforming GPT-Rosalind-5.5 which scored 16.7%. These results were obtained using a benchmark-specific configuration, not a production setup. OpenAI plans to monitor GPT-Rosalind-6's real-world use, including safety signals and misuse patterns, to inform future safeguard updates.

LongCat-2.5 is a large multimodal model designed for agentic tasks, featuring native multimodality. It boasts a total of 1.6 trillion parameters. The entire training process for this model was completed using domestically produced chips.

Researchers have found that Large Language Models (LLMs) demonstrate proficiency in natively reading and writing neural network weights. For instance, GPT-6 Astra successfully edited the weights of Qwen3.5-0.8B, resulting in improved mathematical capabilities for the latter. This suggests a novel approach to neural network development and optimization, potentially allowing LLMs to directly manipulate and enhance other models' functionalities.

Momentic (YC W24) offers an AI QA engineer that tests applications end-to-end, catching bugs before they ship. It allows users to verify critical flows and find bugs from a prompt, with reported metrics including 77,813 tests created, 70,607,819 test runs executed, and 117,010 bugs caught. Momentic aims to accelerate software development by integrating verification into the agent loop, enabling teams to ship faster and with greater confidence.

A comparison of agent harnesses CUGA, Hermes, and Deep Agents using the GLM-5.3-flash model on 90 AppWorld tasks revealed significant performance differences, with task completion rates varying from 30% to 81% based on the harness. While AppWorld tests consumer app capabilities, it serves as a proxy for enterprise automation needs. For enterprise use, CUGA offers additional features like enforced rules and on-premise deployment, which were not part of this benchmark. The evaluation was conducted using the Harbor evaluation framework.

A study on modernizing legacy PHP code with LLMs found that simply asking an LLM to "modernize this" often breaks working code. Researchers evaluated five LLMs on 100 WordPress 4.0 PHP files, aiming to update them to PHP 8.3 while preserving behavior. The LLMs broke 32-51 files, repairing at most 10, compared to Rector which broke 7 and repaired 6. Much of this damage was unnecessary, as only 37 findings required changes for PHP 8.3 compatibility.

SemiAnalysis has introduced GDPx, an agentic capability benchmark designed to evaluate performance across 11 distinct domains. These domains include critical areas such as chip design, cybersecurity, and RSI. The benchmark aims to provide a comprehensive assessment of agentic capabilities. In its initial findings, Opus 5.5 achieved the highest score, reaching 12%, indicating its leading position among the tested agents in these diverse fields.

Aarush and Arnav from The Subvocal Company (YC F26) have developed smart glasses called Subvocal, enabling users to interact with computers silently. These glasses decode mouthed words into text at speech speed, addressing the challenge of using voice dictation tools like Wispr Flow in public or quiet environments. Inspired by MIT's AlterEgo, they built a prototype at Stanford’s TreeHacks. Preorders for the first batch of 500 units are now open with a $200 discount.

HackerBooks, inspired by Karpathy's HackerTube, is a new platform listing approximately 16.5k books mentioned by at least two different Hacker News commenters since 2007. Each book entry links back to the original comments. The data is downloadable, and the site, including its data pipeline, was largely built by Claude Opus 5.5, with the creator primarily providing direction and handling logistics like domain registration. The platform counts mentions, so it's not a perfect recommendation system.

ChatGPT consumer product lead Adam Fry announced that ChatGPT has reached 1.2 billion weekly active users (WAUs), gaining 200 million users in the last two months. Fry also noted that ChatGPT's search volume now accounts for approximately 20% of Google's total search volume. This information was shared by Alex Kantrowitz of Big Technology, highlighting the significant growth and increasing market presence of ChatGPT.

Anthropic, an artificial intelligence company, reported $33.24 billion in revenue for the nine months ended September 30, marking a 1,200% increase from the previous year. The company also reported a net loss of $104 billion during this period, largely attributed to changes in the fair value of convertible notes and derivative liabilities. This follows a $25.1 billion loss over the same period in 2025.

Anthropic's internal AI use, when valued at retail prices, has now surpassed its employee pay. This development highlights a significant shift in the company's operational costs, indicating a substantial investment in or reliance on its own AI technologies. The information comes from Business Insider's AI Insider franchise, led by Alistair Barr, who covers Big Tech, startups, and venture capital, including topics like AI and cloud computing.

OpenAI has doubled the limits of their plans, just one month after a previous reduction. This move raises questions about whether these increased limits will be sufficient to meet user needs or if it signals a sense of desperation from the company. The discussion around OpenAI's strategy and its impact on users continues to be a relevant topic in the AI and programming communities.

OpenAI reportedly fell $20 billion short in its latest attempt to raise money ahead of an initial public offering (IPO), with SoftBank notably absent from the funding round. This development suggests potential challenges for the company as it seeks to secure capital. Further insights into the AI market and its potential fault lines can be found in Rod Dubitsky's article.

Anthropic's revenue growth, while still substantial, showed signs of flattening in Q3, reaching $16.8 billion. Although the company achieved a sequential gain, the absolute dollar growth was smaller than in Q2, which saw an $11.6 billion revenue. This slowdown, with a 45% sequential increase in Q3 compared to 141% in Q2, suggests that the rapid expansion of AI services may be moderating, potentially impacting the valuations of related "neocloud" providers.

GitBook experienced a security incident, leading to a breach. As a result, all personal API tokens were revoked. Users are advised to rotate any existing GitBook API tokens, third-party API tokens or secrets connected to GitBook, and secrets for published integrations. Additionally, users of the GitBook MCP server need to re-authenticate. The GitBook team apologized for the disruption and thanked users for their patience.

OpenAI is reportedly rebranding its data centers as “compute campuses,” a move some suggest is an attempt to soften public perception and compete with rivals like Anthropic. Data centers often face bipartisan opposition due to their noise and resource consumption. Anthropic, for its part, has pledged to cover power grid upgrade costs, invest in community projects, protect water supplies, and avoid NDAs with public officials, applying these commitments to both leased and owned facilities.

OpenAI is experiencing unprecedented growth, making it one of the fastest-growing technology companies ever, yet it's also burning through cash at an extraordinary rate. This has led to concerns on Wall Street, with discussions around its corporate restructuring, the reasons investors continue to fund a company losing billions, and the fatal cost problem behind every ChatGPT answer. The financial ties between Nvidia and OpenAI's survival, and the potential impact of an OpenAI IPO on retirement funds, power bills, and taxes, are also significant points of discussion.

Tucker Carlson recently exposed what he believes is the true agenda of elites regarding Artificial Intelligence. The discussion, featuring figures like Bill Gates and Sam Altman, delves into the future of AI, its predictions, and its potential impact on politics and social media. This conversation is available on YouTube and other platforms, sparking debate around the implications of AI.

Users have reported that OpenAI appears to be silently re-enabling model training on Codex chats, despite previously disabling data sharing. This has raised concerns among companies worried about intellectual property leaks through prompts. Users are advised to check their data controls at chatgpt.com/settings/data-controls, specifically the "Improve the model for everyone" setting, and also privacy.openai.com/policies to manage data sharing.

Matt Scherer, co-author of "Brace for Impact: Protecting people and the economy when the AI bubble bursts," argues that the SEC should block Anthropic's IPO. He cites concerns about insufficient demand for generative AI, with studies showing LLMs not accelerating productivity or generating positive ROI for most businesses. A study also indicated a significant drop in Americans using AI at work between 2025 (46%) and 2026 (33%). Scherer points to Anthropic's unaudited financial results and relegated future obligations as reasons for the SEC to intervene.

AI experts are raising concerns about the potential influence of AI chatbots on American voters during the midterms. A team of MIT professors investigated how popular AI chatbots respond to election-related questions, highlighting that the specific chatbot used could significantly impact voter perception. The discussion also covered how AI might cause disruption and the broader issue of misinformation on social media, emphasizing the challenges election officials face in balancing these threats.

Anthropic's S-1A filing with the SEC indicates the company's internal estimates suggest future models may "imminently" meet its definition of recursive self-improvement (RSI), where AI autonomously improves its training. However, current models have not crossed this threshold. Published measurements show Claude assists in 26% of AI research tasks, but works fully autonomously in zero areas. While AI shows promise in improving research, the architect role remains human, supporting a staircase of human-driven breakthroughs rather than self-sustaining AI.

OpenAI has partnered with Cathedral to enhance its security defenses and is revamping its Security Bug Bounty program. This initiative includes quadrupling rewards for critical findings to incentivize researchers. OpenAI is also planning an internal security sprint, bringing together new researchers and security partners with its own teams to further strengthen its security posture.

OpenAI has launched the Grid Protectors Program, a US initiative designed for its upcoming data center campuses. This program aims to reduce electricity demand during periods when the grid is under strain. It is a national effort by OpenAI to support affordable and reliable electricity for families and businesses in the communities where they establish their facilities.