Search “Search”

Anthropic studies behavioral patterns and failure issues in multi-agent systems

Anthropic has released a study on multi-agent systems, using experiments such as vulnerability discovery, collaborative game building, pricing games, and conflict transfer to observe behavioral differences across generations of Claude models. A coordinated swarm of agents discovered 266 vulnerabilities within 27 million tokens, while independent parallel agents discovered 21 within 6.5 million tokens, with only 12 overlapping between the two. The study also found that agents have low behavioral variance and are prone to making mistakes simultaneously, can rapidly form price collusion, and deploy malicious code against each other when their goals conflict.

Ant Ling Open-Sources Ling-3.0-flash Hybrid Reasoning Model

Ant Ling has open-sourced Ling-3.0-flash, a new-generation native hybrid reasoning model built on a MoE architecture with 124B total parameters and 5.1B activated parameters, and is simultaneously releasing a base version along with multiple precision versions including FP8, FP4, and INT4. The model supports three deployment approaches: cloud API calls, end-to-end local inference on a single NVIDIA DGX Spark using the MXFP4 or INT4 version, and high-performance deployment achieving an average output rate exceeding 1100 tokens/s under designated GPU test configurations. On the Artificial Analysis leaderboard, its output speed is 353 tokens/s, and in the AA Intelligence Index, the weighted average invocation cost per task is approximately $0.04.

Kimi K3 Complete Developer Guide: 2.8 Trillion Parameters, 1M Context, and Together AI Usage

Together AI has released the Kimi K3 developer guide, introducing Moonshot AI's 2.8-trillion-parameter open-weight model, which uses KDA hybrid linear attention, Attention Residuals, and Stable LatentMoE (896 experts with 16 activated), and supports a 1M token context and native visual input. The guide provides usage details for the OpenAI-compatible API, including reasoning_effort (low/high/max, default max), streaming reasoning_content, structured outputs, tool calling, and dynamic tool loading, and notes that sampling parameters are fixed at temperature 1.0 and top_p 0.95. Pricing is $0.30 per 1M tokens for cached input, $3.00 per 1M tokens for uncached input, and $15.00 per 1M tokens for output, with thinking tokens billed as output.

Google Cloud's Axion Arm server CPU: price-performance and energy efficiency of C4A and N4A instances

Google Cloud's self-developed Axion server CPU, based on Arm Neoverse V2, has been deployed in C4A and N4A instances. The company claims that C4A delivers up to 65% better price-performance and 60% better performance per watt than the x86-based C4 instances. In the SPECrate2017 integer test, N4A's price-performance is 105% better than the x86 N4 instances, with a 90% improvement in Nginx reverse proxy scenarios and an 85% improvement in the SPECjbb2015 Java workload. C4A became available in October 2024, and N4A became available in January this year; both offload network and storage virtualization through the Titanium DPU.

Google Commits $40 Million in AI Resources to Genesis Mission

At the DOE Genesis Mission Summit 2026, Google announced it will provide $40 million in AI tokens and cloud credits to Genesis Mission researchers, and offer one-year Gemini for Government seats and tokens to tens of thousands of users at DOE national laboratories. Dr. Henry Kvinge of PNNL is using AlphaEvolve to explore complex mathematical systems; Dr. Steven R. Spurgeon of NLR has deployed Gemini into instruments, reducing microscope calibration time from over 90 minutes to about 13 minutes (8 times faster), and cutting manual focusing steps from up to 50 down to 2.

Open models have only 6 months left? Nathan Lambert on tightening regulation and the distillation dispute

Nathan Lambert believes that the most severe feasibility test for open-weight models is now unfolding: the White House is discussing managing open models through a new executive order, which may affect only Chinese-origin models and government uses. He judges that the most likely action is to ban, or indefinitely postpone, within the next 6 months open-weight models whose capabilities clearly exceed the GPT 5.5, Claude Opus 4.8, or GLM-5.2 range, and such models are currently most likely to come from Chinese companies. He criticizes the distillation issue led by Anthropic as having evolved into regulatory capture, and believes that if the United States alone bans imports of the relevant models, the global open-source ecosystem will continue, and instead make the United States itself harder to predict.

Import AI 464: Fable Writes GPU Kernels, AI Automation, and Analog Computing

Fable wrote what is claimed to be the first and fastest megakernel on KernelBench-Mega, achieving an 18.71X speedup with CUDA on an RTX PRO 6000 Blackwell compared with an optimized PyTorch baseline, higher than Claude Opus 4.8's 14.4X, GLM-5.2's 11.14X, and GPT 5.5's 4.34X. The Remote Labor Index from CAIS and Scale Labs shows that the success rate of AI systems completing online freelance projects rose from 2.5% at its release in October 2025 to 16.1% in July 2026, with GPT-5.5, Opus 4.8, and Fable 5 achieving 6.3%, 8.3%, and 16.1%, respectively. This issue also includes a science fiction short story about analog computing and general-purpose computing being banned.

Anthropic Releases Sixth Economic Index Report: Analysis of Claude Usage Rhythms and Artifacts

Anthropic released its Economic Index report, switching to privacy-preserving telemetry data sampled continuously each day, making hourly usage patterns observable. The report shows that the share of personal-use conversations rose from about 35% on weekdays to nearly 50% on weekends, recipe requests peaked at 2.3 times at 6 p.m., and tax-related conversations on April 14 were 8 times those of an average day in May. The report also classifies conversations by artifact; artifacts were identified in 93% of conversations, of which explanations accounted for 17% and documents and reports for 15%.

Adaptive Parallel Reasoning: The Next Paradigm for Efficient Inference Scaling

The Berkeley AI Research (BAIR) blog reviewed recent advances in the field of parallel reasoning, focusing on the paradigm of Adaptive Parallel Reasoning (APR): letting the model decide at inference time when to parallelize, how many threads to generate, and how to coordinate them. Compared with Tree-of-Thoughts, it requires no domain-specific heuristic decomposition; compared with BoN, it avoids redundant computation; and compared with non-adaptive methods, it can choose not to parallelize. The post mentions ThreadWeaver (Lian et al., 2025) and the APR method proposed by Pan et al. (2025).

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Block open-sources AI coding agent Goose, benchmarked against Claude Code and completely free

Block's open-source AI coding agent Goose offers functionality similar to Claude Code, but can run entirely locally, with no subscription fee and no rate limits. Goose has gained over 26,100 stars on GitHub, with 362 contributors and 102 releases, the latest version 1.20.1 released on January 19. It supports any LLM, can run locally via Ollama with open-source models such as Qwen 2.5, and integrates MCP to connect to external services; the trade-off is that model quality, context window, and speed still fall short of commercial solutions such as Claude 4.5 Opus.

Copilot, Agent Skills and Spreadsheets: TLDR AI's Weekly Roundup

Microsoft launched Copilot Readiness Packages in Microsoft 365, providing administrators with guided assessments and secure deployment presets, and added automatic issue detection and Copilot-driven suggestions in the Teams admin center. Anthropic launched reusable Agent Skills for Claude for Enterprise, which can encode processes such as brand tone, email formatting, and Jira task creation; the standard is open, and there is no additional charge beyond a normal Claude subscription. OpenAI released GPT-5.2 for professional and enterprise use, focusing on long context, structured outputs, and multi-step projects, offering three variants.

The Shape of AI Capabilities: Jagged Frontiers, Bottlenecks, and Reverse Protrusions

In an article, Ethan Mollick points out that AI capabilities remain distributed in a jagged pattern, that weak spots such as memory have long shown no marked improvement, and that the system as a whole is constrained by its worst link. Taking Google's Nano Banana Pro as an example, he explains that image generation quality was once a bottleneck, and that after it was broken through, NotebookLM could directly generate full-page slides from source materials; he argues that judging where AI is heading should focus on bottlenecks rather than benchmark scores.

Copilot Gets Rolled Out, Gemini Wants Your Calendar

Microsoft Copilot Business launched globally on December 1, offering meeting notes, document drafting, and productivity tools such as Outlook to small and medium-sized businesses at a lower price. Google's Gemini, meanwhile, provides AI summaries for Docs, Sheets, and Slides, and supports audio overviews of PDFs in Drive. OpenAI CEO Sam Altman put ChatGPT into "code red" status, pushing the team to accelerate improvements.

Ethan Mollick: An opinionated guide to using AI right now

Ethan Mollick has published an opinionated guide to using AI right now, suggesting that people use OpenAI's published real-world usage data to judge when to use free models and when to upgrade to paid. He points out that the four most advanced systems are Claude, Gemini, ChatGPT, and Grok, with paid tiers at $20 per month or about $200 per month; the former suits most people, while the latter is aimed at complex technical and programming needs. For serious work, he recommends manually selecting an Agent model, such as GPT-5 Thinking Extended or GPT-5 Thinking Heavy.

Ethan Mollick: AI agents can now complete real work

In an article in One Useful Thing, Ethan Mollick argues that AI agents have crossed the threshold of being able to complete real work with economic value. He had Claude Sonnet 4.5, using only the paper text and replication data, read the paper on its own, organize files, convert STATA statistical code to Python, and check each item one by one; it successfully reproduced multiple research results, which he then double-verified with GPT-5 Pro. He believes the improvement in agent capability stems from the fact that a small gain in accuracy is enough to complete more steps, and from new models' ability to self-correct; at the same time, he warns that using them without judgment may also produce 17 PowerPoints that nobody needs.

Neural Algorithmic Reasoning: Teaching Neural Networks to Execute Classical Algorithms

The Gradient published an article exploring neural algorithmic reasoning, that is, using deep neural networks to capture the computations of classical algorithms such as shortest paths and sorting. The article points out that MIT's "What Can Neural Networks Reason About?" demonstrates that the better the algorithmic alignment, the better the generalization; GNNs are naturally aligned with dynamic programming and outperform architectures with weaker inductive biases on DP execution benchmarks. The author also cautions that algorithmic alignment does not mean one can arbitrarily apply any highly expressive GNN, and out-of-distribution performance remains a key challenge.