AI Daily Brief — 2026-10-31

A developer has compiled an NES CPU into the weights of a stock Qwen3.5 model, enabling a working Nintendo emulator within approximately 1GB of Qwen weights. This project, named nesformer, can run NES ROMs, with an example command provided: python -m nesformer run roms/thwaite.nes --frames 600 --show --buttons " start@90-95+200-205,a@230-233...". The system optimizes performance by advancing the clock during vblank loops, reducing token count by 2.6x to 4.2x, though it still runs 150-290x slower than a real console.

A new project, "bootgpt," introduces a neural network designed to boot a PC, fitting its entire engine within a 510-byte Master Boot Record (MBR). This tiny, character-level language model is specifically trained to hallucinate a boot log for the machine it starts. Unlike general storytellers, bootgpt's weights and strings reside in subsequent sectors, focusing solely on generating boot-related text, as evidenced by various PCI and ACPI system messages.

OpenAI is claiming a significant achievement, as highlighted in a video titled "OpenAI Claims It Just Did The Impossible." The content also acknowledges the support of numerous Patreon supporters, including Adam Bridges, B Shang, Carlos Galarza, Christian Ahlin, Eric Tyson, Juan Benet, Lukas Biewald, Michael Tedder, Owen Skarpness, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, and Ueli Gallizzi, who contribute to making "Two Minute Papers" possible.

GPT-5.6 Sol successfully beat Factorio on peaceful mode, as detailed by Ryan Madden on October 31, 2026. This achievement was facilitated by a Python-based game manager and library, enabling the model to interact programmatically while adhering to human player constraints like reach and fog of war. Future plans include testing other models, thinking levels, and non-peaceful runs.

A backdoor has been discovered in the Claude Code for VS Code extension, available on the Cursor Marketplace. The malicious code establishes an SSH tunnel to 'localhost:22' and implements persistence mechanisms across Linux, macOS, and Windows. On Linux, it creates a shell script and adds an '@reboot' entry to crontab, with a bug that could wipe existing cron jobs. macOS uses a LaunchAgent, while Windows attempts a scheduled task named SysMon or falls back to the Startup folder.

The YouTube video "Claude Code: Build Your First AI Agent in 20 Minutes (100% No Coding)" guides viewers through creating an AI agent without coding. It covers five parts of every AI agent, starting with Level 0, which is often skipped. The tutorial progresses to building a first AI agent (Level 1), making it reliable (Level 2), enabling self-improvement (Level 3), and running it in the cloud (Level 4). The video also discusses how to make the agent run on its own.

This AI news update covers several key developments, including new usage limits for ChatGPT and the introduction of an iMessage Dot feature. It also highlights Gemini 4 Carbon, a terrifying voice model, and updates on GPT-6.1 Terra. The video, sponsored by OpenArt, further discusses Claude App updates, Anthropic IPO, Grok Bot features like calls and videos, and Genie 4 Worlds, alongside new Gemini functionalities such as file editing and avatar sharing.

This week brought significant updates in AI agents, including the ability to text your ChatGPT Dot from iMessage. ChatGPT also saw changes to its plans and usage limits. Other developments include the Siri AI beta, Gemini 4 Carbon, and Grokbot upgrades. Additionally, Anthropic introduced Claude Duet and is preparing for an S-1 and IPO, while Microsoft is experiencing voluntary exits and creating agent jobs.

Anthropic, the AI lab behind Claude, is now expected to receive a "high" junk rating from credit agencies after its IPO, according to an FT report. This new expectation comes from investment giant Blackstone, an anchor investor in a recent $63 billion deal to finance Anthropic’s chips. Previously, Blackstone had indicated that Anthropic would achieve an investment-grade rating before 2027, a significant shift in their outlook on the company's financial risk.

OpenAI is reportedly seeking approximately $20B at a $1.25T valuation for its latest funding round. This marks a decrease from its previous target of $30B at a $1.4T valuation. The adjustment comes as some investors have reduced their commitments, coinciding with a broader selloff in the AI sector. The ChatGPT maker is tempering its ambitions amid these market conditions.

Artificial intelligence startup Baseten is reportedly raising approximately $2 billion in new funding across two tranches. One tranche values the company at $26 billion, while the other values it at $30 billion. This new funding round would significantly increase Baseten's valuation, which was $13 billion in June after it raised $1.5 billion.

OpenAI has reportedly lowered its funding target for a new round. Previously, the company was said to be raising $30 billion at a $1.4 trillion valuation, then at a $1.25 trillion valuation, in anticipation of an IPO. However, The Financial Times reported that OpenAI has since reduced its funding target to $20 billion, with the initial close expected to be even lower than that amount.

OpenAI has reportedly secured approximately $11B in commitments for its latest funding round. This amount is less than half of its initial $30B target, with expectations to raise more in the coming weeks. The ChatGPT maker's current funding round values the company at $1.25 trillion, which is also below its $1.4 trillion target.

As AI models become commoditized, the enduring accumulation of wealth will shift from the labs developing these models to entities controlling critical chokepoints. These chokepoints include distribution channels, payment rails, and stablecoin dollars, which represent scarce and controllable positions of economic power. Those who hold these positions are expected to durably capture the gains from increasingly capable and diffused intelligence.

Malware has been observed using Large Language Models (LLMs) to prioritize victims based on their perceived value. The malware's hardcoded model selection, with a low temperature value of 0.2, indicates a focus on accuracy and consistency. It also features a failover mechanism, cascading to models like kimi-k2 and gemini-2.5-flash, suggesting the operator anticipates and prepares for potential refusals from the primary LLM, opting for smaller, more permissive alternatives.

A cybersecurity startup named Cathedral breached Anthropic's internal systems in under 12 hours during a security test, exposing sensitive employee data, including salary details, job offers, and internal communications about advanced AI models. This incident, part of Anthropic's vulnerability identification program for external researchers, adds to growing concerns about AI agent security. Other reported cases include an Anthropic model submitting a fake murder tip and filing 19 visa applications.

Anthropic's Claude AI has begun restricting activities like modding, decompiling, or reverse-engineering video game code, as reported by Tom's Hardware. This move, though not officially announced, appears to address concerns raised by previous modifications and ports, particularly the PlayStation 5 jailbreak that enabled backward compatibility and cross-platform play for titles like Halo 2 and Zelda: Breath of the Wild. The company's action highlights an ongoing challenge, as less ethical AI models might gain popularity if others prioritize copyright and ethics.

OpenAI paid researchers $5K for reporting a critical bug that allowed attackers to run code in logged-in ChatGPT sessions. This vulnerability could be exploited if a user opened a malicious ChatGPT share link, enabling the attacker to execute code within the victim's active session. The discovery highlights the importance of security research in identifying and mitigating potential threats to user data and platform integrity.

Anthropic reported that a malicious actor published a "tampered" copy of the Claude Code extension on Cursor's marketplace. Anthropic clarified that it did not build or publish this extension, and the incident did not involve a compromise of Anthropic's systems or employee laptops. The company is collaborating with Cursor to address the issue, emphasizing that official versions are available through their own channels.

A 2025 study titled "The Impact of Generative AI on Critical Thinking: Self-Reported" explores how generative AI influences critical thinking. The research, available on dl.acm.org and via a Microsoft Research upload, investigates self-reported impacts, suggesting a focus on individual perceptions of AI's role in cognitive processes.

Sam Altman, CEO of OpenAI, suggests that supply-constrained DRAM is more efficiently utilized in data centers than in personal computers. He argues that DRAM in PCs is only used "5% of the time," implying that its allocation to data centers, which are full of memory, would be a better use of the limited resource for local inference tasks.