English translation is being prepared.
Matched fieldsSummary
←Search results
Search “Reasoning”
English translation is being prepared.
Matched fieldsSummary
English translation is being prepared.
Matched fieldsSummary
Google tested using a lightweight classifier to route between Gemini models and thinking levels to reduce per-task cost
The Google Developers Blog ran a test with 50 auto-graded coding agent questions, connecting TypeSafe's Jev classifier before model calls to perform routing: simple tasks were handed to Flash-Lite, while complex tasks were escalated to Gemini 3.8 Flash at a low thinking level. All 50 questions were answered correctly, with both cost and response time lower than using low-thinking-level Flash throughout. The article also found that squeezing maxOutputTokens down to 300 did not reduce thinking overhead on simple tasks, and instead caused 7 of 25 hard questions to fail; local heuristic rules based on keywords and prompt length dropped in accuracy from 93.8% to 70.6% on 34 unseen questions. The authors recommend first using low-thinking-level Flash directly, then adding classifier routing for short, independent tasks, while keeping a single model for long-context tasks to preserve the Context Caching discount.
The Google Developers Blog ran a test with 50 auto-graded coding agent questions, connecting TypeSafe's Jev classifier before model calls to perform routing: simple tasks were handed to Flash-Lite, while complex tasks were escalated to Gemini 3.8 Flash at a low thinking level. All 50 questions were answered correctly, with both cost and response time lower than using low-thinking-level Flash throughout. The article also found that squeezing maxOutputTokens down to 300 did not reduce thinking ove
Matched fieldsBody
Reinforcement Learning Terminology for LLM Readers: How a Single Score Is Distributed to Each Token
The article uses DeepSeek-R1-Zero's pass@1 on AIME 2024 rising from 15.6% to 71.0% as an example to show that a rule-based judge gives only a single score for the entire response, whereas training needs to adjust the probability of each token. The author uses the sampling example "Is OpenAI open?" to explain in turn policy, state, action, reward, return, value, Q, advantage, baseline, TD error, GAE, clipping, and KL, and points out that PPO uses a critic to learn V, while GRPO uses the average score within a group to replace V.
The article uses DeepSeek-R1-Zero's pass@1 on AIME 2024 rising from 15.6% to 71.0% as an example to show that a rule-based judge gives only a single score for the entire response, whereas training needs to adjust the probability of each token. The author uses the sampling example "Is OpenAI open?" to explain in turn policy, state, action, reward, return, value, Q, advantage, baseline, TD error, GAE, clipping, and KL, and points out that PPO uses a critic to learn V, while GRPO uses the average s
Matched fieldsBody
SlopCodeBench Retest: GPT-6 Astra Leads with 16.3% Strict Pass Rate, Ahead of Opus 5.5
The author retested frontier models using the full version of SlopCodeBench (36 questions, 196 checkpoints per setting). The highest was GPT-6 Astra with a strict pass rate of 16.3%, above GPT-5.5's 14.8% in May, with the tested models overall ranging from 12.2% to 16.3%. In the three full comparisons, the XHigh tier cost 2.11-4.22 times that of Medium, but Opus 5.5 actually had more strict passes at the Medium tier (28 vs. 25), while Astra and SOL performed better at the XHigh tier. Regarding code quality, Opus 5.5 XHigh had a slop line proportion of 77.75%, lower than Medium's 94.25%; the author cautions that this may be because the new models produce new anti-patterns not yet covered by the 200-plus linters.
The author retested frontier models using the full version of SlopCodeBench (36 questions, 196 checkpoints per setting). The highest was GPT-6 Astra with a strict pass rate of 16.3%, above GPT-5.5's 14.8% in May, with the tested models overall ranging from 12.2% to 16.3%. In the three full comparisons, the XHigh tier cost 2.11-4.22 times that of Medium, but Opus 5.5 actually had more strict passes at the Medium tier (28 vs. 25), while Astra and SOL performed better at the XHigh tier. Regarding c
Matched fieldsBody
Perplexity Publishes LLM Gateway Best Practices Guide, Outlining Request Flow and Production Risks
On October 27, 2026, Perplexity published an LLM gateway best practices guide explaining how this middleware layer routes, authenticates, caches, and logs model calls, and distinguishing four categories: API gateways, LLM gateways, MCP gateways, and AI gateways. The guide breaks down the mechanism through a six-step request lifecycle, while warning that the gateway centralizes credentials from various providers, and cites the LiteLLM gateway breach disclosed by Microsoft and two high-severity vulnerabilities, CVE-2026-35029 and CVE-2026-59823, reported by Check Point (already fixed in LiteLLM 1.83.9). The article also lists postmortems of Cloudflare AI Gateway and OpenRouter outages, four deployment paths, and paid plans starting at about $50 per month.
On October 27, 2026, Perplexity published an LLM gateway best practices guide explaining how this middleware layer routes, authenticates, caches, and logs model calls, and distinguishing four categories: API gateways, LLM gateways, MCP gateways, and AI gateways. The guide breaks down the mechanism through a six-step request lifecycle, while warning that the gateway centralizes credentials from various providers, and cites the LiteLLM gateway breach disclosed by Microsoft and two high-severity vu
Matched fieldsBody
NVIDIA Explains How to Fine-Tune Nemotron 3.5 Lightning with the Nemotron-Personas Dataset
NVIDIA has released the Nemotron-Personas France and Belgium datasets and provided a complete post-training workflow for fine-tuning Nemotron 3.5 Lightning with them. The workflow includes defining behavior contracts, generating synthetic conversations with personas, fine-tuning with LoRA on NVIDIA NeMo, merging adapters, and then comparing and evaluating item by item against the base model on a frozen test set. The France dataset was built in collaboration with Pleias, and the Belgium dataset involved Pleias and KU Leuven; the Nemotron repository provides reusable steps and agent skills for data generation, training, and evaluation.
NVIDIA has released the Nemotron-Personas France and Belgium datasets and provided a complete post-training workflow for fine-tuning Nemotron 3.5 Lightning with them. The workflow includes defining behavior contracts, generating synthetic conversations with personas, fine-tuning with LoRA on NVIDIA NeMo, merging adapters, and then comparing and evaluating item by item against the base model on a frozen test set. The France dataset was built in collaboration with Pleias, and the Belgium dataset i
Matched fieldsBody
How LangChain Built the Design Agent Content Studio
LangChain's GTM engineering team, working with the design team, built Content Studio, a design agent inside Slack, on Managed Deep Agents. It can generate videos, social graphics, UI mockups, and slide decks, and has already produced more than 3,000 assets and over 80 slide decks. The article breaks down its four-layer knowledge structure (skills, wiki, code validation, memory), the dedicated LangSmith Sandbox for each Slack thread, the two rendering engines HyperFrames and the in-house motion-studio, and how the main agent orchestrates asynchronous sub-agents. The team says that after migrating to Managed Deep Agents, its own code shrank by about 80%, the median model cost per request fell from about $8 in mid-September to about 50 cents, and the entire harness has been open-sourced.
LangChain's GTM engineering team, working with the design team, built Content Studio, a design agent inside Slack, on Managed Deep Agents. It can generate videos, social graphics, UI mockups, and slide decks, and has already produced more than 3,000 assets and over 80 slide decks. The article breaks down its four-layer knowledge structure (skills, wiki, code validation, memory), the dedicated LangSmith Sandbox for each Slack thread, the two rendering engines HyperFrames and the in-house motion-s
Matched fieldsBody
Gemini 4 Carbon Tops the Artificial Analysis Intelligence Index v5
Google's Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5 with a score of 59, leading in four of the 9 evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. On the Coding Agent Index, Gemini 4 Carbon (high) paired with the Antigravity CLI scores 59, trailing Claude Code paired with Claude Opus 5.5 (max, 62 points) by 3 points. The model has a 1M-token context window, supports text, image, video, and speech input, and is priced at a standard rate of $4/$20 per million input/output tokens; it is currently available to select users with a 50% discount.
Google's Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5 with a score of 59, leading in four of the 9 evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. On the Coding Agent Index, Gemini 4 Carbon (high) paired with the Antigravity CLI scores 59, trailing Claude Code paired with Claude Opus 5.5 (max, 62 points) by 3 points. The model has a 1M-token context window, supports text, image, video, and speech input, and is priced at a
Matched fieldsBody
Gemini 4 Carbon Tops the Artificial Analysis Intelligence Index…
Artificial Analysis evaluations show that Google's Gemini 4 Carbon tops Intelligence Index v5 with 59 points, ranking first in four of the 9 evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. On the Coding Agent Index, Gemini 4 Carbon (high) paired with Antigravity CLI scores 59 points, 3 points behind Claude Code paired with Claude Opus 5.5 (max, 62 points). The model is currently available only to some users and has not yet been publicly released; it is priced at $4/$20 per 1M input/output tokens with a 50% discount, has a 1M-token context window, and supports text, image, video, and voice input.
Artificial Analysis evaluations show that Google's Gemini 4 Carbon tops Intelligence Index v5 with 59 points, ranking first in four of the 9 evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. On the Coding Agent Index, Gemini 4 Carbon (high) paired with Antigravity CLI scores 59 points, 3 points behind Claude Code paired with Claude Opus 5.5 (max, 62 points). The model is currently available only to some users and has not yet been publicly released; it is p
Matched fieldsBody
Google A2UI Team and CopilotKit Explain How to Build Generative UI Applications
The Google A2UI team and CopilotKit CEO Atai Barkai explain how to build generative UI applications using the A2UI open standard and CopilotKit. A2UI describes interfaces with declarative JSON, and the client only renders components from a trusted component catalog; the same response can be rendered to React, Angular, SwiftUI, Jetpack Compose, and Flutter. The article provides a code example for integrating CopilotKit with Google ADK, and explains the bidirectional flow, implemented via the AG-UI protocol, in which user actions are sent back and the agent regenerates the interface.
The Google A2UI team and CopilotKit CEO Atai Barkai explain how to build generative UI applications using the A2UI open standard and CopilotKit. A2UI describes interfaces with declarative JSON, and the client only renders components from a trusted component catalog; the same response can be rendered to React, Angular, SwiftUI, Jetpack Compose, and Flutter. The article provides a code example for integrating CopilotKit with Google ADK, and explains the bidirectional flow, implemented via the AG-U
Matched fieldsBody
Baseten Optimizes Qwen3-ASR Batch Transcription, Boosting Single-GPU Throughput by 7x
The Baseten engineering team, through native audio decoding, KV cache FP8 quantization, and reduced host-device synchronization, increased batch transcription throughput for Qwen3-ASR 1.7B on a single RTX PRO 6000 by 7x and reduced latency by 6.9x. After optimization, the transcription cost for one hour of audio is approximately $0.0013, i.e., 1/10 of a cent per audio hour, with no significant change in word error rate. The optimized version has been made the default deployment in the Baseten Model Library and can be called through an OpenAI-compatible transcription client.
The Baseten engineering team, through native audio decoding, KV cache FP8 quantization, and reduced host-device synchronization, increased batch transcription throughput for Qwen3-ASR 1.7B on a single RTX PRO 6000 by 7x and reduced latency by 6.9x. After optimization, the transcription cost for one hour of audio is approximately $0.0013, i.e., 1/10 of a cent per audio hour, with no significant change in word error rate. The optimized version has been made the default deployment in the Baseten Mo
Matched fieldsBody
Sakana AI uses AI to analyze the US midterm elections: about 710,000 X reactions to the Texas Senate race
Sakana AI has released an AI analysis case study of the US midterm elections, focusing on the Texas Senate race and analyzing 2,042 posts and 705,976 reactions on X from two candidates and two state party organizations. Its narrative analysis system extracts narratives according to "who did what and how it was evaluated," and cross-checks them against the original text using multiple AI models; meanwhile, it uses the product Sakana Marlin to retrieve public information such as polls, campaign funding, prediction markets, and reports, and compares it with observations on X. The article states that most of the support Paxton received before the runoff came from intra-party mobilization, that after the runoff his support share fell from about 50% to just under 20%, and that the shift in evaluations of the two candidates preceded judgment adjustments by institutions such as the Cook Political Report by about two months.
Sakana AI has released an AI analysis case study of the US midterm elections, focusing on the Texas Senate race and analyzing 2,042 posts and 705,976 reactions on X from two candidates and two state party organizations. Its narrative analysis system extracts narratives according to "who did what and how it was evaluated," and cross-checks them against the original text using multiple AI models; meanwhile, it uses the product Sakana Marlin to retrieve public information such as polls, campaign fu
Matched fieldsBody
Phonic 1 Tops the Artificial Analysis Speech-to-Speech Index, Leading Grok with 87.1 Points…
Artificial Analysis evaluations show that Phonic's Phonic 1 (High) ranks first among 22 models on the Speech to Speech Index with 87.1 points, leading Grok Voice 3.0's 86.9 points by a narrow margin. The model ranks first on the Tau Voice agent task with 79.9%, leading the second place by 5.4 percentage points; it ranks third on Big Bench Audio with 98.9%, and its Time to First Audio is 0.61 seconds. Phonic 1 is available through the Phonic API, with input audio costing $0.07 per minute and $4.20 per hour.
Artificial Analysis evaluations show that Phonic's Phonic 1 (High) ranks first among 22 models on the Speech to Speech Index with 87.1 points, leading Grok Voice 3.0's 86.9 points by a narrow margin. The model ranks first on the Tau Voice agent task with 79.9%, leading the second place by 5.4 percentage points; it ranks third on Big Bench Audio with 98.9%, and its Time to First Audio is 0.61 seconds. Phonic 1 is available through the Phonic API, with input audio costing $0.07 per minute and $4.2
Matched fieldsBody
Beijing News tests Qwen Go! and Lingguang 2.0: personal agent competition shifts toward autonomous task handling
Beijing News Shell Finance reporters tested Alibaba's Qwen Go! and Ant's Lingguang 2.0. The former set up a daily 8 a.m. medication reminder with a single sentence and checked it item by item, while the latter, after authorization on Alipay, sorted out the past six months of bills. The article also reviews the layouts of Meta Muse, OpenAI dots, as well as ByteDance and Tencent, and points out that cross-ecosystem permissions are currently the hardest wall to break down, with ByteDance's SAEP published rejection list reaching 155 apps.
Beijing News Shell Finance reporters tested Alibaba's Qwen Go! and Ant's Lingguang 2.0. The former set up a daily 8 a.m. medication reminder with a single sentence and checked it item by item, while the latter, after authorization on Alipay, sorted out the past six months of bills. The article also reviews the layouts of Meta Muse, OpenAI dots, as well as ByteDance and Tencent, and points out that cross-ecosystem permissions are currently the hardest wall to break down, with ByteDance's SAEP pub
Matched fieldsBody
Artificial Analysis releases Intelligence Index v5, with 60% of evaluations switched to private datasets
Artificial Analysis releases Intelligence Index v5, adding two new benchmarks, FrontierCode 2.0 and Terminal-Bench-Science, upgrading HLE-Diamond, AA-Omniscience v1.1, and CritPt v2, with 60% of evaluations based on private test sets or held-out answers. Claude Opus 5.5 (max) ranks first with 58 points, followed closely by GPT-6 Astra (max) at 57 points and GPT-6.1 Sol (max) at 55 points, with Gemini 4 Argon (high) at 50 points. The gap between open-weight models and the leaders has widened from 3 points to 12 points; the released-weight DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) both score 32 points. The per-task cost of leading models differs by more than 13 times: GPT-6.1 Terra (max) is $0.96, while Claude Fable 5.1 (max) is $13.04. A Progress Index v1 was also released, placing models released since November 2022 on the same capability scale for comparison.
Artificial Analysis releases Intelligence Index v5, adding two new benchmarks, FrontierCode 2.0 and Terminal-Bench-Science, upgrading HLE-Diamond, AA-Omniscience v1.1, and CritPt v2, with 60% of evaluations based on private test sets or held-out answers. Claude Opus 5.5 (max) ranks first with 58 points, followed closely by GPT-6 Astra (max) at 57 points and GPT-6.1 Sol (max) at 55 points, with Gemini 4 Argon (high) at 50 points. The gap between open-weight models and the leaders has widened from
Matched fieldsBody
AWS explains prompt caching patterns for using OpenAI models on Amazon Bedrock
AWS has published a tutorial explaining how to use prompt caching to reduce the cost of agents repeatedly reading context when calling OpenAI models on Amazon Bedrock, distinguishing between two approaches—implicit caching and explicit prompt_cache_breakpoint—and recommending their combined use. The article provides code examples for tool and instruction caching, rolling breakpoints across multi-turn history, adding tools mid-session, and shared prefixes across sub-agents, and states that in a six-turn workload on GPT-6 Astra, implicit caching reduces input costs by about 62%, while explicit rolling breakpoints reduce them by about 61%. Cacheable prefixes must be at least 1024 tokens, and cache writes are billed at 1.25 times the normal input price.
AWS has published a tutorial explaining how to use prompt caching to reduce the cost of agents repeatedly reading context when calling OpenAI models on Amazon Bedrock, distinguishing between two approaches—implicit caching and explicit prompt cache breakpoint—and recommending their combined use. The article provides code examples for tool and instruction caching, rolling breakpoints across multi-turn history, adding tools mid-session, and shared prefixes across sub-agents, and states that in a s
Matched fieldsBody
Simon Willison used a single prompt to have Claude Opus 5.5 add a four-player battle map to a web game
Django co-creator Simon Willison used a single English prompt to have Claude Opus 5.5 add a four-player battle map to his web game and rework the networking code to support four-player multiplayer. The prompt gave only four constraints—four-way rotational symmetry, at most four AI players, building on the existing WebSocket non-audio multiplayer mode, and empty slots filled by AI—without specifying files or lines of code. Based on this, the author argues that prompts should include fewer steps and pin down the boundary conditions.
Django co-creator Simon Willison used a single English prompt to have Claude Opus 5.5 add a four-player battle map to his web game and rework the networking code to support four-player multiplayer. The prompt gave only four constraints—four-way rotational symmetry, at most four AI players, building on the existing WebSocket non-audio multiplayer mode, and empty slots filled by AI—without specifying files or lines of code. Based on this, the author argues that prompts should include fewer steps a
Matched fieldsBody