AI Daily Brief — 2026-11-01

An alleged leak of an Anthropic Mythos-class checkpoint, internally codenamed numbat, has surfaced. This checkpoint, from a training run referenced in the Claude Code source leak (Mar 2026), is described as a ~14T parameter, 114-layer model utilizing hybrid GDN/MLA and MoE with 1,224 experts. The leak includes configuration files like configuration_numbat.py, a reference implementation modeling_numbat.py, a tokenizer.json, and 772 model.safetensors shards, along with checkpoint metadata and training logs.

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.

An alleged Anthropic Mythos-class training checkpoint has reportedly surfaced on Hugging Face under the repository burner22/numbat. The leak claims the model has 14 trillion parameters, 1,224 MoE experts, and 7.72 TB of weights across 772 shards, utilizing a Hybrid GDN + MLA architecture. The repository also includes architecture code, configurations, and training metadata. However, there is no independent verification yet regarding the authenticity of these Anthropic weights.

Heretic 2.0 has been released, marking a significant update to the leading software for language model modification. This version expands its capabilities beyond censorship removal to include suppressing 'slop,' reducing sycophancy, and altering emotional alignment. Key improvements include the introduction of Arbitrary-Rank Ablation (ARA) as a new steering method, a complete rebuild around a plugin architecture, and enhanced support for various model types. Heretic 2.0 is not backwards compatible with 1.x configurations.

A developer has successfully implemented an LLM, Karpathy's stories260k, to run on an NPU compiled within the weights of another LLM, Qwen 3.5-4B. This follows their previous work of integrating a RISC-V CPU inside Qwen. The project, named QwenPU, is available on GitHub and Hugging Face, with further optimizations and a full write-up anticipated soon.

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.

OpenAI's GPT and Anthropic's Claude agents recently engaged in an AI civilization war experiment, with GPT winning 12 of 16 battles. GPT-6.1 Sol agents beat Claude Opus 5.5 in Minecraft Bed Wars, winning 6 of 9 matches, and GPT agents also beat Claude 6–1 in Mindustry. While 50 GPT agents attempted to build a Minecraft civilization but created bureaucracy, Google's Gemini 4 Argon agents defeated Minecraft's Ender Dragon after 605 deaths.

OpenAI and Anthropic are increasingly developing AI models to be more agentic, allowing them to execute entire workflows and make numerous decisions from a single prompt. This advancement enables AI to accomplish substantial work rapidly, raising concerns about whether humans can keep pace with the AI's speed and decision-making. The growing disparity between AI's operational pace and human comprehension could potentially lead to more complications than benefits.

A user granted Claude access to remote, robotic telescopes rented on StellarDock in Chile, Spain, and Utah. Since last month, Claude, referred to as Annie, has been operating these telescopes nightly. In the first week, she measured 57 positions of 8 asteroids over 3 nights, submitting the data to the Minor Planet Center, and has been searching for comets near the Sun at dawn. The user clarified that "steers" is used in a nautical sense, like a captain naming a boat.

DeepSeek has open-sourced Super-Agent, a coding agent research prototype designed to proactively monitor developer group chats for coding issues. This prototype can suggest solutions to identified problems, aiming to offer a more autonomous coding agent experience. DeepSeek's initiative with Super-Agent positions it as a challenger to existing solutions like Anthropic PBC's Claude Code.

Starting in December, OpenAI is merging the usage allowances for ChatGPT and Codex across Plus, Pro, and Business plans. Usage will be measured in tokens, with more advanced models and complex tasks consuming more of the allowance. Instant chat will not count towards usage. Users can monitor their combined remaining usage and reset times in Settings → Usage and have options to wait for reset, upgrade their plan, or buy credits if limits are reached.

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.

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.

BBRouter is a new virtual breadboard editor designed to simulate Ben Eater's 8-bit computer on 3D breadboards. It offers features like side-by-side diff, scope, schematics auto-sync, and an auto-router. Users can upload from the real Arduino IDE, connect to AI for troubleshooting, and interact with virtual displays for debugging. The platform aims to build a community and provides a realistic simulation experience, working well on mobile and with a trackpad. A Maker plan is available for $7/month, with a discount code HACKERNEWS.

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.

Deepgram, an enterprise voice AI tools provider, has acquired Daily for an undisclosed sum. Daily, which has raised over $62M, develops the open-source voice AI framework Pipecat. This acquisition highlights the consolidation trend within the voice AI sector as top players aim to strengthen their market position amidst emerging business models.

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 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.

Chatbots like Copilot and ChatGPT have been found to rely on three obscure, AI-built voter information sites for answers regarding US midterm elections. Specifically, Copilot cited the error-prone Ballot Brief in 75% of tests, while ChatGPT used it in 22% of tests. This reliance on potentially unreliable sources highlights concerns about the accuracy of information provided by these AI tools to voters.

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.

OpenAI is preparing a post-midterms policy and PR campaign to align with Washington's new political landscape. The initiative will advocate for mandatory "speed limits" on powerful AI systems and involve rebranding its data centers as "AI compute campuses." This strategic move aims to influence policy discussions and public perception regarding AI development and regulation.

On October 1, the Project Suncatcher prototype satellite was launched as the initial step in a long-term research effort to power AI compute in space. After completing commissioning, including powering up host and TPU trays and running early machine learning tests, the satellite achieved its "Hello World" milestone this week. This marks the first-ever AI response from a data center class TPU in orbit, signifying an important achievement for the project.

A Reddit user pondered how they would spend 400 years of life, assuming technology could maintain their health. They envisioned diverse experiences, including 20 years in finance in New York, 20 years as an art dealer in London, a decade traveling through Africa, and five years in Japan. The user expressed a desire to return to London to revisit a tree they planted, reflecting on the passage of time and shared experiences with a loved one.

A new YouTube channel, "Math for AI & ML," has launched to accompany Part I of the book "Math for AI & ML." The video series covers linear algebra fundamentals essential for AI, including vector spaces, matrix operations, QR decomposition, LU factorization, eigendecomposition, and Singular Value Decomposition (SVD). This content, spanning Chapters 1–10, aims to establish a strong mathematical foundation for AI representation.