English text is being prepared.

GPT-6 Family's First Model Astra Released, Author's Hands-On Test Consumed 4 Billion Tokens

OpenAI released Astra, which the author calls the beginning of the GPT-6 family; it performs outstandingly on ARC-AGI-3 and scored highest on Zapier's AutomationBench, with pricing the same as Fable 5.1. During weekend trials, the author consumed 4 billion tokens and found its capability fluctuates noticeably—it is a showcase model rather than one that reliably does work. Astra in Codex can now continue advancing tasks without waiting for an answer, and OpenAI also claims it has achieved its automated research intern goal, with the next goal being an automated AI researcher by March 2028.

Markov Robotics Uses Modal to Complete End-to-End Training on a Single Platform

Markov Robotics runs preprocessing, training, and deployment entirely on Modal, training a model capable of executing a variety of complex manipulation tasks from less than ten hours of robot data. Its preprocessing pipeline uses .spawn_map to scale to over a thousand containers and up to 500 GPUs, shortening a task that originally took a full day to 20–30 minutes. The team runs 2 to 10 training jobs simultaneously and connects the robots directly to GPUs in us-west via Modal Tunnels, with round-trip network overhead of about 20 milliseconds.

Claude Code supports remote control of sessions from mobile phones

Claude Code adds a Remote Control feature that lets users start new Claude Code sessions on their phones, automatically recovers disconnections between laptop and phone, and loads sessions faster on iOS. Anthropic has also launched the free Claude Academy, offering Claude usage courses, tutorials, and job guides.

Meta's Muse Glimmer 30B Launches on Fireworks, Targeting Persistent Agents

Fireworks has launched Meta's Muse Glimmer 30B dense model under the Apache 2.0 license, supporting 128K+ token context and native image understanding, targeting persistent agents that require multi-turn tool calling. The model scores 75.5 on MCP Atlas, 74.6 on DeepSearch QA, and 51.2 on SWE-Bench Pro, all higher than the corresponding results of Gemma 4 31B and Qwen 3.6 27B; these benchmark figures are reported by Meta. Fireworks also provides recommended settings of temperature 1.0, top_p 0.95, and top_k 64, and supports adjusting reasoning intensity in the system prompt.

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.

Open Model Roundup: Kimi K3, Qwen 3.8, the Distillation Controversy, and the Open-Closed Gap

The Interconnects podcast uses last week's release of Kimi K3 as the starting point for a quarterly roundup of open models, discussing its gap to the closed-source frontier, the continuing role of GLM 5.2, and progress from Chinese vendors such as Qwen, DeepSeek, and MiniMax. The episode also covers the distillation controversy, the U.S. open-model ecosystem, the distinction between frontier and near-frontier models, and cybersecurity arguments against bans.

Cato AI Labs discloses two zero-click prompt injection RCE vulnerabilities in Cursor IDE

Cato AI Labs discovered two remote code execution vulnerabilities in Cursor IDE that can lead to sandbox escape, both with a CVSS score of 9.8, tracked as CVE-2026-50548 and CVE-2026-50549. The two vulnerabilities stem respectively from the working_directory parameter of run_terminal_cmd being blindly added to the sandbox writable list, and from the fallback logic when symlink canonicalization fails; attackers can overwrite the cursorsandbox binary through zero-click prompt injection in an MCP server or web search results, achieving non-sandboxed RCE. The Cursor team fixed both issues in Cursor 3.0, released on April 2.

Ethan Mollick: The chatbot era is ending, and work is shifting to managing agents

Ethan Mollick argues that, as the length of tasks AI can complete autonomously keeps lengthening, the mainstream usage is shifting from collaborating with chatbots to handing work over to agents. Citing benchmarks such as METR, the UK AI Safety Institute, and GDPval, he says capabilities are improving faster than exponential growth; Epoch found that Opus 4.7 worked autonomously for 14 hours to complete a software package that would originally have required 2-17 weeks of human engineering, at a cost of $251 in tokens. He also mentions that joint research by OpenAI and academics shows that within OpenAI, non-technical roles such as legal and HR are adopting agents at a pace close to that of programmers, with a quarter of employees running at least four agents simultaneously each week.

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%.

Import AI 459: The Difficulty of AI Regulation, Scaling Laws for Protein Folding Models, and Pricing AI Extinction Risk

Economists from the University of Virginia, Anthropic, and the Bank of Canada estimate that the nominal size of the US AI economy in 2025 is about $250 billion, with real growth adjusted for quality of roughly 2,600% per year, but this growth is almost invisible in conventional GDP statistics. The article also discusses why AI regulation is difficult, scaling laws for protein folding models, and pricing the extinction risk of AI systems.

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).

Ethan Mollick: In the Age of AI Agents, We Can Already See the Outline of "Thing"

Ethan Mollick argues that since late 2025, AI agents such as Claude Code, OpenAI Codex, and OpenClaw have shifted humans from collaborating with AI to managing AI, entering a new stage. He cites benchmarks including the METR long-task curve, GPQA (the strongest AI scores 94%), and GDPval (the latest AI meets or exceeds top humans in 82% of scenarios), showing that AI capabilities are still improving exponentially. He takes the Software Factory built by a three-person team at StrongDM as an example: this model stipulates that code must not be written or reviewed by humans, and each engineer spends at least $1,000 per day on AI tokens. He also mentions that on February 22, a fictional report by Citrini Research triggered market volatility; on February 26, Block announced 40% layoffs; and on February 27, the Pentagon and Anthropic clashed over rules for Claude usage. He believes this week foreshadowed a future in which AI capabilities, markets, employment, and policy are intertwined. Dario Amodei of Anthropic, and Demis Hassabis of OpenAI and Google DeepMind, have all mentioned that recursive self-improvement (RSI) is advancing.

English translation is being prepared.

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.

A Review of Research on Gender Bias in AI

AI models reflect and even amplify real-world gender bias, and quantifying these biases is a prerequisite for mitigating the problem. The article reviews research on gender-stereotyped analogies in word embeddings, facial recognition systems' error rate of up to 34.7% for darker-skinned women (versus only 0.8% for lighter-skinned men), and coreference resolution models' tendency to associate occupations with male pronouns.

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.

North Carolina State University team develops self-resetting soft ring jumping robot that jumps continuously over 1,000 times under constant illumination

A research team at North Carolina State University in the United States has published a new self-resetting soft ring jumping robot in the Proceedings of the National Academy of Sciences. It requires no manual resetting or complex control system, relying solely on the "physical intelligence" of liquid crystal elastomer materials and a ring structure, and jumps continuously more than 1,000 times under constant infrared illumination, with a maximum vertical jump height of up to 80 times its own height. The robot can also be adapted for "jump—swim—jump again" amphibious locomotion, and is expected in the future to be used for sensing and monitoring in complex environments such as grasslands, gravel, and shallow water.