Complete Index

Intel Archive

Page 04

Every public Intel entry, ordered by when it entered the index. Dates identify when an item was published when known, otherwise when it was indexed.

151GQA: Training Generalized Multi-Query Transformer Models from Multi-Head CheckpointsInterpolates between multi-head and multi-query attention using grouped key-value heads, recovering multi-query speed while keeping most of multi-head quality.articleJoshua Ainslie et al.Published May 22, 2023articleJoshua Ainslie et al.Published May 22, 2023
152Direct Preference Optimization: Your Language Model is Secretly a Reward ModelDerives preference training as a simple classification loss on the policy itself, removing the reward model and the reinforcement-learning loop from RLHF.articleRafael Rafailov et al.Published May 29, 2023articleRafael Rafailov et al.Published May 29, 2023
153Efficient Memory Management for Large Language Model Serving with PagedAttentionApplies operating-system paging to the KV cache so serving stops wasting most of its memory on fragmentation, raising throughput several-fold.articleWoosuk Kwon et al.Published Sep 12, 2023articleWoosuk Kwon et al.Published Sep 12, 2023
154Constitutional AI: Harmlessness from AI FeedbackAnthropic's method for training a harmless assistant from AI feedback against a written set of principles, replacing most human harm labels with model self-critique.articleYuntao Bai et al.Published Dec 15, 2022articleYuntao Bai et al.Published Dec 15, 2022
155Training Compute-Optimal Large Language ModelsShowed the large models of the era were badly undertrained: for a fixed compute budget, model size and training tokens should scale roughly in step, not size alone.articleJordan Hoffmann et al.Published Mar 29, 2022articleJordan Hoffmann et al.Published Mar 29, 2022
156Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient SparsityRoutes each token to a single expert rather than combining several, making trillion-parameter sparse models trainable at roughly the cost of a dense one.articleWilliam Fedus et al.Published Jan 11, 2021articleWilliam Fedus et al.Published Jan 11, 2021
157QLoRA: Efficient Finetuning of Quantized LLMsBackpropagates through a frozen 4-bit quantized model into LoRA adapters, putting 65B-parameter finetuning on a single GPU without measurable quality loss.articleTim Dettmers et al.Published May 23, 2023articleTim Dettmers et al.Published May 23, 2023
158Fast Inference from Transformers via Speculative DecodingRuns a small draft model ahead of the large one and verifies its guesses in parallel, producing identical output several times faster.articleYaniv Leviathan et al.Published Nov 30, 2022articleYaniv Leviathan et al.Published Nov 30, 2022
159ReAct: Synergizing Reasoning and Acting in Language ModelsInterleaves reasoning traces with tool actions in a single loop, so a model can plan, act against an external source, and revise from what it observes.articleShunyu Yao et al.Published Oct 6, 2022articleShunyu Yao et al.Published Oct 6, 2022
160Scaling Laws for Neural Language ModelsEstablished that language-model loss falls as a smooth power law in model size, data and compute — the result that turned scaling from a hunch into a budgeting exercise.articleJared Kaplan et al.Published Jan 23, 2020articleJared Kaplan et al.Published Jan 23, 2020
161Fast Transformer Decoding: One Write-Head is All You NeedShares a single key-value head across all query heads, shrinking the KV cache that dominates memory during incremental decoding.articleNoam ShazeerPublished Nov 6, 2019articleNoam ShazeerPublished Nov 6, 2019
162Attention Is All You NeedThe 2017 paper that dropped recurrence and convolution for self-attention alone, introducing the Transformer — the architecture every large language model still builds on.articleAshish Vaswani et al.Published Jun 12, 2017articleAshish Vaswani et al.Published Jun 12, 2017
163Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsShowed that prompting a model to produce intermediate reasoning steps unlocks arithmetic and symbolic reasoning that appears only above a certain model scale.articleJason Wei et al.Published Jan 28, 2022articleJason Wei et al.Published Jan 28, 2022
164RoFormer: Enhanced Transformer with Rotary Position EmbeddingEncodes position by rotating query and key vectors, so attention depends on relative distance — the position scheme nearly every current open model uses.articleJianlin Su et al.Published Apr 20, 2021articleJianlin Su et al.Published Apr 20, 2021
165AI agents don't just reason anymore. They transact. Cloudflare's Monetization GaCloudflare's Monetization Gateway lets content owners, API providers and MCP servers charge agents per request using x402, and enterprises building on AWS AgentCore Payments can draw on that supply now.articleCloudflarePublished Aug 18, 2026articleCloudflarePublished Aug 18, 2026
166GLM-5.3 achieves 60 on the Artificial Analysis Intelligence Index, on par with KArtificial Analysis scores GLM-5.3 at 60 on its intelligence index, level with Kimi K3, with a 246 point jump to 1770 Elo on its agentic evaluation. It burns about 20 percent more output tokens than GLM-5.2 and costs 68 cents per index task against 44 cents.articleArtificialAnlysPublished Aug 18, 2026articleArtificialAnlysPublished Aug 18, 2026
167Claude can now send emails in Gmail and manage files in Google Drive. Ask ClaudeClaude can now send Gmail messages and manage Google Drive files through connectors, drafting and sending replies with a user-set approval gate. Available on all paid plans.postclaudeaiPublished Aug 18, 2026postclaudeaiPublished Aug 18, 2026
168As models become more capable, the risks associated with developing and testing OpenAI paused reinforcement learning training on its latest deployment-bound models for two weeks while it hardened and red-teamed research environments, and says its largest planned frontier RL run stays on hold until smaller runs establish more alignment evidence.articleOpenAIPublished Aug 18, 2026articleOpenAIPublished Aug 18, 2026
169Many drugs work by binding to a specific target in the body and blocking or chanGiven a design prompt written by a human expert, Claude autonomously designed de novo protein binders against 14 of 15 targets. Adaptyv Bio and Twist Bioscience built and tested them, with 22 to 35 percent binding against a field baseline of 10 to 15 percent.postAnthropicAIPublished Aug 18, 2026postAnthropicAIPublished Aug 18, 2026
170Cline is now available in the AI SDK harness layerVercel adds an official Cline adapter to the AI SDK harness layer, built with the Cline team, so Cline runs behind the same HarnessAgent interface as Claude Code, Codex, Deep Agents, Grok Build, OpenCode and Pi, with no bridge process installed inside the sandbox.articleFelix ArntzPublished Aug 18, 2026articleFelix ArntzPublished Aug 18, 2026
171GLM 5.3 now available on AI GatewayGLM 5.3 lands on Vercel's AI Gateway with better multi-step agent and software engineering results than GLM 5.2 at fewer output tokens, a 1M token context window, 128K max output, and a setup path into Claude Code, Codex and other coding agents.articleWalter KormanPublished Aug 18, 2026articleWalter KormanPublished Aug 18, 2026
172We’re extending the 50% increase to weekly Claude Code limits through August 31.Anthropic extends the 50 percent increase to weekly Claude Code limits through August 31, with a warning that model demand may make capacity tight and no commitment yet on making the change permanent.postClaudeDevsPublished Aug 18, 2026postClaudeDevsPublished Aug 18, 2026
173Introducing LangSmith Tuned Evaluators They automatically score agent behavior iLangSmith adds Tuned Evaluators, small tuned models that score production agent traces automatically, starting with Perceived Error. LangChain reports the specialized model beat every frontier model it tested while cutting evaluation cost by 82 percent.articleLangChainPublished Aug 18, 2026articleLangChainPublished Aug 18, 2026
174We're making Git hosting more reliable, performant, and scalable. This post tracCursor traces twenty years of Git hosting design and explains why it operates its own storage layer, Origin, as a database rather than a filesystem, with the scaling limits that pushed it there.articlecursor_aiPublished Aug 18, 2026articlecursor_aiPublished Aug 18, 2026
175Frontier Model Cost and Open-Weights Popularity is Driving Demand for Model RoutingGlean's Arvind Jain explains why frontier model cost and stronger open weights have made routing a core deployment layer, including the case for answering some queries without an LLM at all.articleRichard MacManusPublished Aug 18, 2026articleRichard MacManusPublished Aug 18, 2026
176How Much Memory Does Your Agent Actually Need?IBM Research scaled its ALTK-Evolve memory system across eight models, from a 30B dense model to frontier systems, and found that agentic memory behaves like a dose rather than a switch: how many distilled guidelines you inject, and how, decides both accuracy and cost.articlehuggingface.coPublished Aug 18, 2026articlehuggingface.coPublished Aug 18, 2026
177Generative Video at the Speed of Light — Keegan McCallum, uRunTen dollars now buys about three hours of continuously generated video, and the argument is that steerability rather than clip quality is the axis that matters, served by a distilled model streamed over WebRTC from globally placed GPUs.videoAI EngineerPublished Aug 18, 2026videoAI EngineerPublished Aug 18, 2026
178The Next Medium: Why Real-Time Interactive Video Changes Everything — Ahmed Ahres, ReactorThe case that real-time generative video is a change of medium rather than a speedup, with the consequences that follow: streaming pixels instead of returning files, stateful sessions, and latency budgets that push GPUs close to users.videoAI EngineerPublished Aug 18, 2026videoAI EngineerPublished Aug 18, 2026
179The Next Game Engine Won't Have a Manual — Arturo Nunez, NereuA game engine designed around coding agents: assets carry intent tags rather than engine boilerplate, systems query by tag, and scene context is assembled with level-of-detail so distant assets collapse to a position and a type.videoAI EngineerPublished Aug 18, 2026videoAI EngineerPublished Aug 18, 2026
180Building an Agentic Video Editor for Mass Consumer — Ekaterina Deyneka, ReelfulAn agentic video editor built like an agentic app builder: raw footage and a line of direction go into a sandbox, skills carry the cut rules and typography, a code-based composition layer renders the result, and a verification pass catches broken output.videoAI EngineerPublished Aug 18, 2026videoAI EngineerPublished Aug 18, 2026
181Infra behind Krea 2: How to train and serve at scale — Gabriel Jorge Menezes, Krea.aiAn infrastructure lead on training at cluster scale: tracking tensor core utilization instead of the misleading headline number, building interconnect telemetry in-house, pulling any card over 78 degrees, and checkpointing a terabyte in under 30 seconds.videoAI EngineerPublished Aug 18, 2026videoAI EngineerPublished Aug 18, 2026
182How AI Coding Agents Can Unlock Materials Simulation with NVIDIA ALCHEMI ToolkitBenchmarking 45 agent-generated simulation pipelines found that prompt specificity changed code structure and reusability but not physical correctness, with agent skills and reference files bridging prompts to GPU simulation code.articleElizabeth GoodmanPublished Aug 18, 2026articleElizabeth GoodmanPublished Aug 18, 2026
183Announcing the Artificial Analysis Search Index, benchmarking how search API proA new index scores seven search API providers on agent task quality, cost and speed, holding model and harness fixed and averaging three benchmarks covering broad research, multi-hop browsing and factual recall.postArtificial AnalysisPublished Aug 18, 2026postArtificial AnalysisPublished Aug 18, 2026
184Parallel (turbo and advanced) and Firecrawl make up the Pareto frontier for ArtiBenchmark results place two search providers on the cost-quality frontier for agent tasks, with one scoring 73 at 7.5 cents per task but the slowest completion time, and another reaching 75 for 12 percent more.postArtificial AnalysisPublished Aug 18, 2026postArtificial AnalysisPublished Aug 18, 2026
185Claude Cowork is now available on mobile and web for all paid plans.Anthropic has opened Claude Cowork to mobile and web across every paid plan, taking the agentic workspace beyond the desktop application.postClaudePublished Aug 18, 2026postClaudePublished Aug 18, 2026
186You can now skip the handshake and still call a tool and your retry after a timeWhat the stateless protocol revision changes for server authors: tool calls without a handshake, and a timed-out retry counting as a fresh call so one aborted request produces two writes.postAlphaSignalPublished Aug 18, 2026postAlphaSignalPublished Aug 18, 2026
187iMessage support for eve agentsAn agent framework adds an iMessage channel: each conversation maps to one session, webhooks are verified, and messages arriving mid-reply steer the in-progress turn rather than spawning separate answers.articleOwen KephartPublished Aug 5, 2026articleOwen KephartPublished Aug 5, 2026
188$1 million hacker challenge for Vercel SandboxA two-week public bounty program pays up to one million dollars for escaping a microVM agent sandbox, arguing that containment usually fails on an unaccounted network path rather than at the VM boundary.articleAndy RianchoPublished Aug 18, 2026articleAndy RianchoPublished Aug 18, 2026
189Chat SDK adds Instagram adapterChat SDK gains an Instagram adapter so bots can handle DMs, media, quick replies, reactions and story replies through Meta's messaging API, with buffered streaming and a 24-hour reply window.articleBen SabicPublished Aug 19, 2026articleBen SabicPublished Aug 19, 2026
190Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.aiWhat actually goes into training an image model: the filtering pipeline behind billions of candidate images, why caption blind spots teach a model to hang every painting it draws, and the case for refusing AI-generated training data outright.videoAI EngineerPublished Aug 18, 2026videoAI EngineerPublished Aug 18, 2026
191Voice agents with Realtime Video — Sidney Primas, LemonSliceA production account of real-time video avatars: why causal generation compounds its own errors over long sessions, the training-time attention mask and single-step distillation that hold it together, and why the audio encoder and thread orchestration turn out to be the real bottlenecks.videoAI EngineerPublished Aug 15, 2026videoAI EngineerPublished Aug 15, 2026
192The Recall Trap: A Recall-Maximizing Retriever Configuration Reduces Issue Resolution in Fixed-Budget Code ContextA controlled study on SWE-bench Verified showing that the higher-recall retriever setting lowers single-shot issue resolution under a fixed context budget, with within-file depth beating file breadth in both a frontier and an open-weights run.articleAlexander Adkins, Teimuraz TrapaidzePublished Aug 18, 2026articleAlexander Adkins, Teimuraz TrapaidzePublished Aug 18, 2026
193Kozuchi Agent: A Language-Agnostic Open-Weight Agent for Software RepairAn open-weight software repair agent with explicit phases, persistent state and deterministic tools that resolves 374 of 500 SWE-bench Verified instances on the official evaluator using a locally hosted 27B model.articleMehdi Bahrami, Kosaku Kimura, Satoshi Munakata, Satoshi Nakashima, Yu Ishikawa, Kosuke Maeda, Nao Soma, Kenichi Kobayashi, Keisuke Miyazaki, Keizo Kato, Shigeki Fukuta, Tatsuo Kumano, Nobutaka Imamura, Kevin Musgrave, Shahbaz Abdul Khader, Kwun Ho Ngan, Joe Townsend, Fayas Asharindavida, Matthieu Parizy, Akira Sakai, Yuma Ichikawa, Yang Zhao, Michiaki Takizawa, Taku Fukui, Hiroki Ohtsuji, Wei-Peng Chen, Hiromichi KobashiPublished Aug 18, 2026articleMehdi Bahrami, Kosaku Kimura, Satoshi Munakata, Satoshi Nakashima, Yu Ishikawa, Kosuke Maeda, Nao Soma, Kenichi Kobayashi, Keisuke Miyazaki, Keizo Kato, Shigeki Fukuta, Tatsuo Kumano, Nobutaka Imamura, Kevin Musgrave, Shahbaz Abdul Khader, Kwun Ho Ngan, Joe Townsend, Fayas Asharindavida, Matthieu Parizy, Akira Sakai, Yuma Ichikawa, Yang Zhao, Michiaki Takizawa, Taku Fukui, Hiroki Ohtsuji, Wei-Peng Chen, Hiromichi KobashiPublished Aug 18, 2026
194AgentR A Stateful and Recovery-Aware Software Architecture for LLM-based Auditable WorkflowsAn architecture for stateful LLM workflows built on durable artifacts, queue-backed workers, explicit state transitions, retry and orphan-job handling, and transactional token-cost accounting.articleRiya Samanta, Bidyut Saha, Soumya Kanti Ghosh, Rajkumar BuyyaPublished Aug 18, 2026articleRiya Samanta, Bidyut Saha, Soumya Kanti Ghosh, Rajkumar BuyyaPublished Aug 18, 2026
195Evaluating Agentic Code Repair Capabilities in Distributed SystemsA code-repair benchmark of 60 real distributed-system bugs that runs each case with and without bounded debugging context, isolating how much logs, traces and runtime state change agent success.articleYibo Yan, Huijuan Wang, Junzhou He, Yizhuo Liang, Shaoyu Wang, Huanchen Sun, Seo Jin ParkPublished Aug 18, 2026articleYibo Yan, Huijuan Wang, Junzhou He, Yizhuo Liang, Shaoyu Wang, Huanchen Sun, Seo Jin ParkPublished Aug 18, 2026
196HxAgent: Iterative Agent Planning for End-to-End Web Application TestingAn iterative LLM planning agent for web testing that re-reads page state after every action and draws on both recent actions and past correct and incorrect sequences, reported ahead of WALT on three benchmarks.articleTu Nguyen, Duy Cao, Viet Nguyen, Phu Nguyen, Vy Le, Nguyen TK Nguyen, Tien N. Nguyen, Vu NguyenPublished Aug 18, 2026articleTu Nguyen, Duy Cao, Viet Nguyen, Phu Nguyen, Vy Le, Nguyen TK Nguyen, Tien N. Nguyen, Vu NguyenPublished Aug 18, 2026
197The Quality of Claude AI-authored Python Tests Is Not Weaker Than Human-authored TestsA controlled study scoring hundreds of Claude-written Python tests against human-written Django and Pandas tests under identical fault-injection and design rubrics, reporting non-inferiority for recent models.articleDouglas J. LeithPublished Aug 18, 2026articleDouglas J. LeithPublished Aug 18, 2026
198The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual ReasoningA benchmark asking models to infer written words from pen-scratch audio and hand-movement video with no visible ink, where humans clear 80 percent accuracy and leading multimodal models stay under 10.articleGarima Arya Yadav, Nilay Yilmaz, Yezhou YangPublished Aug 18, 2026articleGarima Arya Yadav, Nilay Yilmaz, Yezhou YangPublished Aug 18, 2026
199DeepSeek V4 Pro 0813 vs Claude Fable 5 on DeepSWE: Cost, Coding, and RoutingA DeepSWE comparison of DeepSeek V4 Pro 0813 and Claude Fable 5 that measures a 90x cost gap and an unusually low per-task correlation, the property that makes escalation routing between them worthwhile.articlewww.together.aiPublished Aug 17, 2026articlewww.together.aiPublished Aug 17, 2026
200DeepSeek V4 Pro 0813 vs GPT-5.6 Sol on DeepSWE: Cost, Coding, and RoutingA DeepSWE head-to-head putting DeepSeek V4 Pro 0813 against GPT-5.6 Sol on accuracy, cost per rollout and failure modes, and measuring a cheap-first escalation cascade that beats either model alone.articlewww.together.aiPublished Aug 18, 2026articlewww.together.aiPublished Aug 18, 2026

Index — Latest Intelligence