NAll news

All news

Wednesday23 items

  2. Artificial AnalysisAnalysis
    AI score61

    Artificial Analysis: Reaching an Intelligence Index score of 30 is now 86 times cheaper than when it was first achieved

    Based on v4.3.2 measured results and current pricing, Artificial Analysis found that the per-task cost of reaching an Intelligence Index score of 30 is 86 times lower than when that score was first achieved. GPT-6 Luna (high) reaches 30 at $0.03/task, whereas Claude Sonnet 4.6 cost $2.49/task when it first reached 30 in February 2026. Reaching 20 is about 123 times cheaper than roughly 13 months ago, and reaching 40 is about 33 times cheaper than roughly 5 months ago. Claude Haiku 5.5 (xhigh) reaches 41 at $0.12/task, making it the cheapest model to reach 40. The highest measured score rose from 21 to 58 within a year, at roughly 11 times the cost, with Claude Opus 5.5 at $5.98/task.

    #English text is being prepared.#Evaluations / benchmarks#OpenAI

  3. Tessl: Product and Engineering BlogFirst-partyAnalysis
    AI score51

    Aviator CEO shares how to replace line-by-line code review with intent verification

    Ankit Jain, founder and CEO of Aviator, explains that as the share of AI-generated code rose and PR volume reached four times the level of a year earlier, his team abandoned reading diffs line by line. Instead, they review acceptance criteria, constraints, and edge cases before code generation, and give a deterministic conclusion with evidence for each criterion. The team also turned recurring review comments from historical PRs into reusable invariant rules, enforced by machines on every change—for example, always use the Money type for anything involving currency, and never write directly to the users table. The author says this transition took about three months, after which code review was no longer a delivery bottleneck.

    Ankit Jain

    #Tutorials / practice#English text is being prepared.#Agents

  4. AWS Machine LearningFirst-partyAnalysis
    AI score65

    AWS Publishes Agent Registry Runtime Tool Discovery Practice

    AWS introduces the Agent Registry runtime tool discovery approach: agents use hybrid semantic and keyword search to find approved MCP servers and A2A agents, then instantly build clients from descriptors to execute requests. The sample code is based on Amazon Bedrock AgentCore Gateway and runtime, integrates an order management MCP server and two A2A agents for pricing and customer service, and demonstrates a return and refund scenario using pure semantic matching. The registry returns only connection metadata; invocation credentials must be obtained separately through Cognito OAuth or IAM, and new capabilities require no changes to orchestrator code.

    Comparison of hardcoded tool endpoints versus runtime discovery with AWS Agent Registry

    #Tutorials / practice#Agents#English text is being prepared.

  5. AYi AI NotesAnalysis
    AI score58

    An Anthropic engineer shares Claude Code multi-agent orchestration approach

    In a 30-minute workshop, an Anthropic engineer described his setup for running more than 100 agents simultaneously with Claude Code: one main agent is set up, and sub-agents act as project managers to assign tasks. Based on this, the author suggests shifting one's role from editing code line by line to first writing task breakdowns and acceptance rules, then spawning sub-agents to execute concurrently. At the same time, the author warns that if the main agent's termination conditions and permission boundaries are not hard-coded, it is easy to fall into an infinite loop and drive up token costs. Beginners can start small with one main agent and three sub-agents.

    Quoted postAYi · @AYi_AInotes

    English text is being prepared.

    #English text is being prepared.#Agents#Anthropic

Tuesday24 items

  1. Rohan PaulAnalysis
    AI score63

    The Qwen3.8-27B reinforcement-learning post-trained model surpasses GPT-6 Luna high on multiple Agent benchmarks

    Rohan Paul introduced an open-source environment model built on Qwen3.8-27B with reinforcement-learning post-training, which reduced the hallucination rate on AA-Omniscience from 30.4% to 15.1%. The model surpasses Muse Glimmer 30B and GPT-6 Luna high on Terminal-Bench, τ³-bench, ClawEval, and BFCL v4, but Luna still leads on DeepSWE at 53.1%; its median output for solved tasks is 23.8K tokens, compared with 54.9K for base Qwen. The author also mentions that the MLA version replaces 16 layers of full attention with multi-head latent attention, reducing memory usage per token from 64 KiB to 18 KiB; the compressed version can run on a single 24 GB RTX 4090, with a context about 3.5 times that of base Qwen.

    #Model releases#Open-source ecosystem#Reasoning

  2. Shao MengAnalysis
    AI score55

    Stanford CME 295 Lecture 4 released: Teaching reinforcement learning with LLMs

    The Stanford course CME 295 Transformers & Large Language Models has been updated to Lecture 4, Reinforcement learning with LLMs, and both the video and lecture notes are now public. This lecture goes from RL and Reward Design to PPO, GRPO, RLVR, and DAPO, explaining the notation in each one by one. The course is taught by Afshine and Shervine Amidi, with 9 lectures in total, covering Transformer architecture, attention variants, RoPE, LoRA fine-tuning, RLHF/GRPO training, Flash Attention inference optimization, MCP and AI Agents, and diffusion language models.

    Quoted postmeng shao · @shao__meng

    English text is being prepared.

    #Tutorials / practice#Reasoning#English text is being prepared.

  3. Alibaba Cloud Developer Community AI
    AI score60

    Taste Skill V2 Update: Adding Web Design Skills to Claude Code and Codex

    Taste Skill has been updated to V2. This is a web design skill for programming AIs such as Claude Code and Codex. It now first has the AI organize project-specific details before starting work, then choose one of 20 design directions with matching fonts, colors, and layouts. V2 also prioritizes the project's own photos; when images are missing, it finds materials from open archives and museum collections and records their sources and licenses. After completion, it launches a browser to check at three widths—1440, 768, and 390—and makes up to three rounds of revisions. Community tests cited by the author show that after installing the Skill, time and cost increase; in blind aesthetic evaluations, V2 averaged 7.7 points, while without it installed the score was 7.8.

    #Product updates#English text is being prepared.#Agents

  4. Alibaba Cloud Developer Community AIAnalysis
    AI score62

    Using pytest gates to accept AI-generated test cases

    The article proposes that before AI batch-generated test cases enter the regression suite, they must pass three acceptance lines: assertion validity, data provenance, and failure attributability. It also provides a conftest.py implementation: test cases must use @pytest.mark.review to declare which gates they pass; when running pytest --strict-review, cases missing the marker are skipped, and CI blocks the merge if the skip count is nonzero. Using a coupon-claiming API as an example, the article compares fake tests with real tests, pointing out that cases asserting only a 200 status code or a non-empty result are as good as untested, and that expected values should be traceable back to business rules or online sampling. The author also explains that the three lines can only filter out invalid test cases; business invariants still need to be mined from requirements by humans.

    #Tutorials / practice#English text is being prepared.#English text is being prepared.

  5. Alibaba Cloud Developer Community AIAnalysis
    AI score57

    Ten runtime metadata items that large model evaluation reports should include

    The article proposes that large model evaluation reports should record ten runtime metadata items, including model version snapshots, dataset hashes, decoding parameters, criteria versions, tool degradation, user simulator versions, hardware and region, time windows, sample size and confidence information, and the storage location of failed samples. The author believes that two scores can be directly compared only when the model snapshot, dataset identifier, and decoding parameters are consistent, and provides a compare_gate.py example that blocks two reports from the comparison table when conditions are missing or inconsistent. The numerical values in the article are only used to illustrate field formats and do not correspond to any real model results.

    #Tutorials / practice#Evaluations / benchmarks

  6. A Yi AI NotesAnalysis
    AI score52

    Hands-on test of MiniMax M3.1 Flash and Claude Opus 5.5 writing a Three.js solar system

    The author tested MiniMax M3.1 Flash and Claude Opus 5.5 simultaneously with the same Three.js prompt, and both ran successfully on the first try. M3.1 Flash costs $0.6 per million output tokens and $0.015 for cache reads; during initialization it detected and fixed 4 potential rendering errors on its own, kept the frame rate steady at 60 fps, and even proactively added a control panel that the prompt did not request; Opus 5.5 is more refined in lighting and layout. On benchmarks, M3.1 Flash ranks first on Automation-Bench with 62.2, and scores 38.1 on Terminal Bench 4.0, lower than Opus 5's 51.8.

    Same story · 3 reports

    #Evaluations / benchmarks#English text is being prepared.#Agents

  7. Shao MengAnalysis
    AI score62

    MIT Course MAS.S62: Two Lectures on the Evolution of Harnesses and Multi-Agent Systems

    Lectures 4-5 of the MIT course MAS.S62 "How to AI an AI" focus on Harness evolution, covering the evolution of harness components, memory, tools, and skills, as well as the evolution of planning, task decomposition, and multi-agent systems. The memory section uses a three-axis taxonomy to connect MemGPT, mem0, MEM1, and Reflexion/ExpeL, and uses CLM to unify context management as file operations; the tools section moves from Toolformer, ToolLLM, and CodeAct to Meta-Harness, which represents the entire harness as a file system and provides about 400 times the amount of feedback information of previous methods. The multi-agent section presents five empirical findings, including that MAS wins 81% on decomposable tasks and loses 70% on tasks with sequential dependencies, and that CORAL achieves new SOTA on 8 tasks, with multi-agent evolution improving over single-agent by 4%-20%.

    Quoted postmeng shao · @shao__meng

    https://t.co/rE6rU3Ybcd

    #Tutorials / practice#Agents#English text is being prepared.

  10. arenaAnalysis
    AI score52

    English translation is being prepared.

    Quoted postArena.ai · @arena

    Muse Wonder (Max) by @AIatMeta has entered Code Arena: WebDev at #15 with 1651 pts and Text Arena at #2 with 1513 pts! In Code Arena: WebDev, Muse Wonder (Max) landed at #15 with 1651 pts. This result is slightly lower thanMuse Spark 1.3 (Max) at #12 with 1659 pts, within 4 pts of Kimi K3 (Max) at #14 with 1655 pts, and 11 pts above Grok 4.7 (xHigh) at #16 with 1640 pts. This release is a significant improvement over past Muse models in Text Arena. Compared with Muse Spark 1.3 (Max) at #11 with 1494 pts, Muse Wonder (Max) gained +19 pts and 9 places to #2. It's 12 pts behind the #1-ranked Gemini 4 Argon (High) with 1525 pts. Congrats to the @AIatMeta team on this release!

    #Evaluations / benchmarks#English text is being prepared.#Agents

  11. Artificial AnalysisAnalysis
    AI score71

    MiniMax-M3.1-Flash scores 42 on the Artificial Analysis Intelligence Index

    MiniMax-M3.1-Flash scores 42 on the Artificial Analysis Intelligence Index, 13 points higher than MiniMax-M3, at less than half the latter's per-task cost. It scores 48% on Humanity's Last Exam and 22% on CritPt, both higher than GLM-5.3-Flash and Gemini 3.8 Flash, and slightly lower than DeepSeek V4.1 Flash 1017. The model has a 1M-token context window, supports text and image input, is priced at $0.15/$0.60 per 1M input/output tokens, and its weights are currently closed-source; MiniMax announced that it will open the weights on October 23.

    #Evaluations / benchmarks#Reasoning#DeepSeek

  12. ArenaAnalysis
    AI score71

    DeepSeek-V4.1-Flash-20261017 Tops the Agent Arena Open-Model Leaderboard

    Arena.ai says DeepSeek-V4.1-Flash-20261017 (Max) has become the top-ranked open-source model in Agent Arena, ranking 11th overall, with a net improvement of 5.1% across more than 7.5K real sessions. The model's median cost per task is $0.20, roughly 37% of GPT-6.1 Sol's improvement margin and 27% of its cost; it ranks first among open-source models on all four signals—Confirmed Success, Praise vs. Complaint, Bash Reliability, and Steerability—and shows no tool-hallucination issues.

    Quoted postArena.ai · @arena

    Exciting news: DeepSeek-V4.1-Flash-20261017 (Max) by @deepseek_ai is now the #1 open model in Code Arena: WebDev, surpassing Kimi K3 (Max) at 10x lower cost! Highlights: - #1 open model, ahead of Kimi K3 (Max) - #12 overall with 1670 pts, ahead of Claude Opus 5 (High) and within ~10 pts of GPT-6 Sol (Max) - Shifts the Pareto frontier at $0.3/$1.2 per million input/output tokens - +52 pts vs. the previous DeepSeek-V4.1-Flash-20260910 Congrats to the @deepseek_ai team on another strong contribution to the open source ecosystem!

    Same story · 2 reports

    #Evaluations / benchmarks#Agents#DeepSeek

  13. Artificial AnalysisAnalysis
    AI score62

    Artificial Analysis releases AA-PersonalityBench to evaluate model personality and expression style

    Artificial Analysis has released AA-PersonalityBench, using 900 curated questions to evaluate differences among language models across 30 personality traits and 21 communication characteristics. Personality traits are scored by LLM judges through pairwise comparisons and aggregated into eight indices: expressiveness, organization, informativeness, transparency, agreeableness, assertiveness, imaginativeness, and Slop; communication characteristics are computed programmatically across four categories of metrics: length, structure, punctuation, and emphasis. Results show that Claude Fable 5.1 stands out on multiple indices including expressiveness, informativeness, transparency, agreeableness, and assertiveness, while Qwen3.8 27B averages about 1,030 words per response, nearly four times the 270 words of GPT-6 Luna.

    Same story · 2 reports

    #Evaluations / benchmarks#Model releases#English text is being prepared.