AI 日报 — 2026-09-22

Anthropic has introduced Claude Opus 5.5, the first model in its new Claude 5.5 family. This model performs at the level of Claude Fable 5.1 for most tasks but costs 40% less to operate than Opus 5. It demonstrates strong capabilities in agentic coding, knowledge work, business workflows, and multidisciplinary reasoning. Notably, Walleye Capital found Opus 5.5 largely solved their evaluation suite, even identifying and correcting an error in their instructions that no other model had caught.

InstinctFlash is a high-performance serving framework designed for robotics models, enabling real-time execution of 5B world-action models on Jetson Thor. It offers significant speedups, with examples like LingBot-VA FP8 achieving 5.36x acceleration and pi05 FP8 reaching 7.88x. The framework includes device-specific serving defaults, measuring each family and operating point to select verified paths within precision constraints. LingBot-VLA-V2 native Thor capture and LingBot-VA saturation profiling are complete, with execution-bound budget selection available.

The concept of language modeling without neural networks, specifically using an unbounded n-gram model for text generation, has been explored. This approach, which relies on counting rather than weights or training, aligns with the idea that language modeling is compression. The "compression–prediction equivalence" suggests that the score of a candidate can be determined by the length of its gzipped form when combined with its context. Although the implementation uses zlib, which shares the DEFLATE algorithm with gzip, the name GziPT was chosen for its appeal.

Artificial Analysis Intelligence Index v4.3.2, used for evaluating Claude Opus 5.5, has been updated. This index incorporates 10 evaluations, including AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. The analysis also considers the maximum combined input and output tokens, noting that output tokens often have a much lower limit depending on the model.

A new study replicates StoryScope (Russell et al., 2026) to identify AI-generated web content from structural signatures rather than word-level detection. Using a 214-feature instrument, an LLM detected AI posts from 187 structural features alone with 98.0 macro-F1 on held-out companies. This performance remained at 98.1 even when AI posts were reworded by their own models. The research found that AI posts share a "tidy, self-announcing shape" and can be attributed to their source with 79.3% accuracy.

On September 15, 2026, Carter Leffer sought validation for breaking the German Army Enigma message MVUEH from July 10, 1941. This message, sent by radio station 2ny and received by SS-Totenkopf Quartiermeister, Ib, at 17:30, was logged as Nr. 172. Since 2005, the MVUEH message has resisted all attempts at decryption, but it was successfully broken by OpenAI GPT-6 Astra.

A Transformer is a neural network architecture that utilizes a self-attention mechanism to integrate information across tokens. Unlike self-attention, the Multi-Layer Perceptron (MLP) within a Transformer processes tokens independently, mapping each token representation from one space to another. This process, represented by the formula QKV_{ij} = (\sum_{d=1}^{768} \text{Embedding}_{i,d} \cdot \text{Weights}_{d,j}) + \text{Bias}_j, enriches the overall model capacity.

The video discusses OpenAI's GPT Bel, highlighting its capabilities, while also introducing Grok 4.7 as a significant development. It further explores MiMo-V2.6 and touches upon various AI-related topics including benchmarks, open weights, and frontier models. The discussion also references entities like XAI, Elon Musk, and the Millennium Prize, indicating a broad overview of current AI advancements and their implications.

TypeSafe's Jev has rapidly gained traction in the AI world, with Vercel noting its unprecedented adoption rate in AI Gateway history. OpenAI, having previously worked with a raw internal API for GPT-4 that exhibited issues like repetitive endings, is now positioned to potentially fast-follow Jev. A key difference in Jev's approach is its integrated classifier, which allows the model to ask and answer questions without external tool calls or leaving the GPU, unlike ordinary tool calling where an agent harness takes over.

Pruning large language models by removing transformer blocks, known as depth pruning, offers predictable inference speedups and memory savings, and is compatible with other optimization techniques. The challenge lies in selecting which blocks to remove, as choices interact, making it a combinatorial problem. This problem can be modeled using spin systems, similar to the physics of Ising optimization. The CBO method, which searches the coupled configuration space, has shown superior performance in identifying optimal block removal configurations, even in hybrid models with unevenly distributed redundancy, outperforming methods like block influence.

The "LLM Ass Bench" from assbench.com is an LLM Benchmark that allows for testing with "One prompt, Multiple models, Multiple dates." This tool provides a comprehensive approach to evaluating various large language models under different conditions, ensuring a broad comparison of their performance over time and across different model architectures.

MiMo-V2.6-Pro has achieved a significant milestone by tying Grok 4.7 (xHigh) and surpassing GLM-5.3 (Max) on Artificial Analysis' Intelligence Index. This performance establishes MiMo-V2.6-Pro as the top-scoring open-weight model on the benchmark. The new flagship model from Xiaomi also outperforms proprietary models such as xAI's Grok 4.6, which currently scores 44, and Google's Gemini 3.8 Flash.

JetBrains Air introduces a system of products for agentic software development, integrating AI into developer workflows. This open system extends beyond JetBrains' own products through the Agent Client Protocol (ACP), which standardizes connections between IDEs and agents. The ACP Registry allows developers to discover and utilize compatible agents within JetBrains IDEs. Future enhancements will incorporate richer context from code, architecture, and runtime behavior, improving work routing among developers, models, agents, and services.

Foremerge is an open-source coordination protocol for coding agents, built on Git, designed to catch intent conflicts between parallel coding agents. It allows agents to maintain isolated worktrees while sharing intent, semantic claims, and provisional ChangeSets. Foremerge 0.5.0 is a pre-1.0, local-first MVP, featuring a CLI, JSON API, MCP server, and a deterministic conflict detector. It enables agents to foresee changes from others, even across separate worktrees, before they are committed.

Claude experienced elevated errors for multiple models, impacting claude.ai, Claude API (api.anthropic.com), Claude Code, and Claude Cowork. Requests to Claude Fable 5 and 5.1, and Mythos 5 and 5.1 have returned to normal success rates. The team is continuing to work on resolving remaining errors affecting Claude Opus 5, with an update expected shortly.

Coverage Cat (YC S22) is a licensed insurance brokerage offering home, umbrella, auto, and renters coverage with price transparency and no sold leads. They provide AI-guided intake paired with a licensed brokerage team to help users compare carrier options and honest price ranges. Coverage Cat operates in California, Florida, New York, Texas, and Washington, and offers an Agent API for developers to integrate their services, including tools like umbrella_consumer_prefill and discovery via /.well-known/mcp.json.

Meta's AI assistant, Muse, has a serious zero-day vulnerability. Approximately 12 hours before this zero-day was disclosed, Amazon began blocking users from utilizing Muse for shopping on its platform. Users attempting to do so received a message indicating that Muse was an "unauthorized AI agent [that] violates Amazon’s Conditions of Use."

Jesse Lyu, CEO of Rabbit, asserts that the Rabbit R1 was not a flop, citing its monetary success and a 45-50% hardware margin. The company shipped over 100,000 units globally with less than a 5% return rate, notably without a subscription model for its AI features. However, the launch of OS3 as an app suggests that a dedicated gadget for AI chat might not be necessary, as existing devices are already well-suited for the task.

Amazon has blocked Meta’s Muse AI agent, marking its latest effort to prevent rival agentic AI services from impacting its retail business. This follows a similar lawsuit against Perplexity in November last year, where a judge sided with Perplexity in August. Additionally, Amazon has been omitting specific item names and product images from confirmation emails since July to limit data mining by external AI services.

Eli the Computer Guy, in conversation with The Tech Report's Isaac Pound, discusses the underlying technology driving the AI hype. He suggests that the industry is shifting away from dependence on frontier models, implying that "Nobody cared about the model." This perspective is presented in a video titled "OpenAI is going to lose 90% of its revenue | Eli the Computer Guy."

British Columbia has filed a lawsuit against OpenAI, alleging safety violations and negligence. The lawsuit claims OpenAI failed to notify law enforcement about the ChatGPT activity of the Tumbler Ridge shooting suspect. This legal action highlights concerns about product safety gaps and the responsibility of AI companies to flag potentially dangerous user behavior to authorities.

Artificial intelligence has caused a division within the United Nations, as 20 countries, including Australia, advocate for a global agreement on the new technology. In contrast, China and the US, who are at the forefront of the AI arms race, prefer to operate without such restrictions. This split highlights differing approaches to regulating AI on an international level.

Daniel Kokotajlo, a former OpenAI researcher and author of AI 2027, predicts that AI agents will automate software engineering by early 2027, leading to accelerated AI research and rapid progress toward superintelligence. He warns that most jobs may be safe for only 18 months before AGI can perform nearly any new role, potentially causing mass unemployment despite economic growth and cheaper goods. The discussion also covers a potential US–China AI arms race, autonomous weapons, and the challenge of detecting misaligned AI systems.