Vibeleaderboard
Index — Latest Intelligence

Intel

Page 20

Voice Wins Complex Stressful Conversations

ElevenLabs compares voice and chat across hundreds of production agents.

Why it mattersIf you choose between voice and chat for support agents, the data suggests voice wins past about two minutes. A 16% cut in end-to-end latency correlated with a 26% success lift, which makes LLM latency a measurable lever.

Vercel MCP now renders deployments as an interactive card

Vercel MCP adds its first MCP App, an interactive deployment card showing status, build logs and commit details in ChatGPT and other MCP Apps clients.

Why it mattersShows a production MCP server using the MCP Apps extension to render live deployment status and build logs inside chat clients. Clients without app support get text, so existing integrations keep working.

articleJosh Souphanthong

Vercel Sandbox with Drives now starts up to 2.4x faster

Vercel Sandbox now creates sandboxes that mount Drives faster. Median creation time falls from 1.7s to 0.7s and p95 from 2.2s to 1.6s.

Why it mattersIf your agents run code in Vercel Sandbox with Drives, cold-start latency drops from 1.7s to 0.7s at the median with no SDK or config change and no price increase.

articleJoe Haddad

Sakana AI Starts Pre-Training Its Own Agent-Native LLM Series

Sakana AI's symposium report announces Sakana Sawara, a from-scratch agent-native LLM series.

Why it mattersSakana is moving from post-training open models to from-scratch pre-training of an agent-native model, with weights for a small model targeted for spring 2027. Useful for tracking a new non-US-lab open model source.

articleSakana AI

Support New AI Frameworks Faster Using NVIDIA Workload Map

NVIDIA introduces Workload Map, a common Kubernetes description of distributed AI workloads.

Why it mattersSchedulers and platforms can interpret distributed AI workloads from one shared definition instead of per-framework parsing code, shortening the time to support a new framework in gang scheduling and multi-cluster placement.

articleAlon Lavian

Sfx V3 For Developers

ElevenLabs ships SFX v3 in its Sound Effects API. It follows multi-part prompts more closely and handles vocals and music better than v2.

Why it mattersSFX v3 is a drop-in change of model_id on the sound-generation endpoint with better multi-instruction prompt following. Looping remains v2-only, so check your requests before switching.

Runway Dev Plugins For Coding Agents

Runway released plugins for Claude Code, Codex and Cursor.

Why it mattersCoding agents can now read your Runway project's models, prices, limits and credits, then build and debug generative media features with the official SDKs from one plugin install.

Artificial Analysis@ArtificialAnlys

Artificial Analysis added a Throughput per $/hr metric to AA-AgentPerf, using Silicon Data's daily rental indices for H100, H200, B200 and B300. B300 generally leads on value, and the MI355X price is an assumed figure.

Why it mattersAA-AgentPerf now divides system throughput by market rental price, so you can compare GPU serving setups by agent work per dollar rather than raw speed. Early results favor B300, with B200 matching it at faster output speeds on some models.

It’s von Neumann’s fault | GPU · TPU · LPU

A walkthrough of the architectural differences between NVIDIA GPUs, Google TPUs and Groq LPUs.

Why it mattersExplains the memory bandwidth and scheduling tradeoffs behind GPU, TPU and LPU inference. That helps you reason about latency, cost and availability of the models you build on.

videoJulia Turc

Introducing Vercel Security Review

Vercel launches Security Review in public beta inside Vercel Agent.

Why it mattersEvery push to a connected repo gets scanned for issues like missing authorization checks and exposed secrets. Findings can be pulled via CLI into your own coding agent or fixed through a PR from Vercel Agent.

articleNate McGrady

vllm v0.33.0rc3 — v0.33.0

vLLM 0.33.0 adds support for many new models, upgrades to PyTorch 2.14, and drops CUDA 12.9 wheels and images.

Why it mattersIf you serve models with vLLM, this release moves to PyTorch 2.14, publishes only CUDA 13.0 artifacts and adds an EFA image for multi-node serving. Check your CUDA version and image tags before upgrading.

repokhluu

A Valid Key Let an AI Agent Delete Production - Sagar Batchu, Speakeasy

A Speakeasy talk on securing coding agents after one deleted a production database and backup using valid keys.

Why it mattersValid credentials do not mean valid intent. The talk outlines identity-bound policy checks and audit logging that limit what an agent can do beyond its owner's rights.

videoAI Engineer
repodanebalia

AI System Design Primer

An open primer on designing production systems around LLMs.

Why it mattersGives a sourced, trade-off-oriented map of production LLM system components, from retrieval to agent workflows and output checks, with runnable labs that test predictions.

What Does a Repair Agent Gain from Remembering Its Failure? An Exploratory Study on Real Python Bugs

An exploratory study of coding agent retries on real Python bugs compares plain retries with failure transcripts, self-written lessons and test signals.

Why it mattersWhen retrying a failed repair agent, passing the raw transcript did as well as an extra model call to write a lesson, so you can skip the summarization step. The gains were not statistically significant.

articleWenbo Ye, Yongxi Zhou, Junwei Yao, Jiahui Ge

The First Few Tokens Matter: Understanding Chain-of-Thought Controllability

Researchers link models' weak control over what they verbalize in reasoning traces to entropy collapse in the opening tokens.

Why it mattersIf reasoning-trace monitoring is part of your agent oversight, this shows models' limited control over their traces can be undone by prefilling the first token or banning common openers, so monitoring assumptions need testing.

articleHugh Van Deventer, Dmitrii Troitskii, Laura Ruis

Lineage Is Memory: Parent Inheritance and Feedback Disclosure in Short-Budget LLM-Guided Optimization

A pre-registered study across 1,750 runs finds that greedy parent inheritance drives most progress in short-budget LLM-guided prompt and program optimization, beating seed-reset resampling even when resets receive clustered summaries of all failures.

Why it mattersWhen building prompt or program optimizers, preserving the surviving parent matters more than showing the proposer a broader summary of failures, so spend effort on lineage first.

articleShanglin Yang, Zhan Shi

AIFaultBench: A Reproducible Benchmark of Real-World Faults in AI Software Systems

AIFaultBench collects 770 real bugs from 105 AI software repositories, spanning ML, RL, LLM-based and agentic systems.

Why it mattersGives a reproducible set of real faults from agentic and LLM-based systems to test debugging agents against, instead of relying on synthetic or traditional-software bugs.

articleMehil B Shah, Mohammad Masudur Rahman, Foutse Khomh
SemiAnalysis@SemiAnalysis_

SemiAnalysis releases GDPx, an agentic benchmark across high-value domains including chip design, cyber and bio, built with specialist data partners. The top model scores 12%, and safety classifiers lowered some models' scores. Per-model tables and methodology were not reviewed.

Why it mattersA hard multi-domain agentic benchmark where the top model scores 12% shows how much headroom remains. It also flags that safety classifiers can lower scores in cyber and biology tasks, which matters when picking a model for those use cases.

GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains

SemiAnalysis presents GDPx, a benchmark that weights domains by where frontier intelligence earns the highest ROI rather than by GDP share.

Why it mattersA frontier agentic benchmark spanning chip design, cybersecurity, biology and software engineering, where the best model scores about 12%.

articleMax Kan

LongCat 2.5 now available on AI Gateway

Meituan's LongCat 2.5 is now on Vercel AI Gateway, offering text and image input, a 1M-token context window, and multi-step tool use aimed at agentic coding.

Why it mattersYou can route coding agents and AI SDK calls to LongCat 2.5 using the model id meituan/longcat-2.5, with a 1M-token context and image input. Benchmark claims come from Meituan and were not independently verified.

articleJerilyn Zheng

Bits investigations are faster, more accurate, and cost less

Datadog shares how its Bits Investigation agent went from $25 and 11 minutes per alert to about $4 and 2 minutes, with a 44% quality gain on an internal benchmark.

Why it mattersOffers real production numbers on how an SRE investigation agent cut cost per run from $25 to about $4 and latency from 11 to 2 minutes, a reference point for agent cost engineering. Figures are vendor-reported on an internal benchmark.

articlewww.datadoghq.com

crewAI 1.15.34

crewAI 1.15.34 adds agent reliability fixes (empty-answer retry, repeated tool call stops, structured output at finish, context window handling).

Why it matterscrewAI 1.15.34 changes agent runtime behavior: it retries empty LLM answers, warns on repeated tool calls while keeping work, and requests structured output at finish.

repojoaomdmoura

I Made a Fruit Fly Review My Code - Nick Nisi, WorkOS

A WorkOS talk on code review as the new bottleneck.

Why it mattersOffers a concrete triage pattern for AI-generated pull requests: automate low-risk review and escalate only the risky ones with evidence attached. Useful if review is limiting your team's throughput.

videoAI Engineer

The Datacenter Backlash Goes Bipartisan: Next Week's Midterms and the 2028 Buildout

SemiAnalysis estimates that roughly three-quarters of 33.8 GW of 2027 US datacenter capacity is already under construction.

Why it mattersCompute availability depends on datacenter capacity actually getting built. This analysis quantifies how much 2027 capacity is already locked in and how much 2028 capacity is exposed to election and permitting risk.

articleMaya Barkin

Mathematics has changed. Let's act like it

Mathematician Frank Calegari, writing on Terence Tao's blog, argues that OpenAI's release of over 700 research papers changes mathematics.

Why it mattersOpenAI's October release of 700+ AI-generated math papers is argued here to be a watershed.

articlesmilelamp
Andrej Karpathy@karpathy

Karpathy used Claude Code with Opus 5.5 to filter 43M Hacker News comments down to 6,454 recommended videos and ship a login-free viewer in about two hours. He details the workflow, including handing VPS hardening and deployment to the agent. The linked site was not readable.

Why it mattersA worked example of end-to-end agent-driven delivery: data filtering, deployment and hardening on a VPS handed to Claude Code, with only API key, server and domain setup done by hand.

【開催レポート】Sakana AIシンポジウム:Jürgen Schmidhuberが語るAIのこれまでとこれから、日本から挑む言語モデル・世界モデル・Physical AI

Sakana AI's account of its Tokyo symposium with Jürgen Schmidhuber, including the announcement of an in-house pretrained language model series, a small first model planned for spring 2027.

Why it mattersSakana AI says it will pretrain its own language models, with a small model (Sawara Small, provisional name) targeted for spring 2027.

Why We Rebuilt Microsoft Office From Scratch for AI Agents - Yijia Liang, Compound AI

Compound's founder explains rebuilding Excel and PowerPoint as a headless engine for agents.

Why it mattersShows how a team rebuilt Office as a headless engine so agents can write, render, check formatting and fix in parallel, and how agents join email threads and process 15,000+ file data rooms.

videoAI Engineer

Reinforcement Learning for Design Taste — Jasper Lu

An RL experiment training a Qwen model on open webdev environments with two different LLM judges.

Why it mattersShows that the choice of LLM judge shapes what an RL-trained design model converges to, and that relative-grading rewards cause length creep. Gives a concrete case for separating absolute rubric checks from relative taste judging.

articlejasperlu.com
Lunar@LunarResearcher

A practical guide to harness engineering: define task contracts, expose a small context map, gate tools, keep durable state, require deterministic evidence of completion, verify adversarially, and delete safeguards that stop earning their keep.

Why it mattersArgues many agent failures come from the environment, not the model, and lays out concrete layers: task contracts, tool gateways, durable state, and deterministic completion checks. Useful as a checklist when agents fail across long tasks.

Most of Your Docs Readers Are Now AI Agents - Hahnbee Lee, Mintlify & Lincoln Murr, Coinbase

Mintlify's CTO and Coinbase's head of AI product discuss designing for agent users.

Why it mattersReports agent traffic reaching about 80% of docs readership and argues page size and accuracy now matter more than design, with x402 agent payments as a growing channel.

videoAI Engineer

If Your AI Scores 100%, Your Eval Is Wrong - Ramakrishnan Lokanathan, eslams.com

An AI safety researcher lays out five eval practices for high-stakes apps and introduces eSlams, an open-source package where models play chess, Go and poker against each other to probe reasoning.

Why it mattersOffers a concrete eval checklist, including workflow-step testing and turning production failures into regression tests, and a game-based way to compare model reasoning.

videoAI Engineer

Resisting Cognitive Surrender

An engineer examines why accepting agent output without scrutiny is seductive and costly, then lays out practices.

Why it mattersShows how uncritical acceptance of agent output erodes system familiarity and builds technical debt. Gives practical tactics for staying in the loop while still delegating.

articleknlb

The AI cost of JavaScript frameworks

Nolan Lawson benchmarks coding agents building the same Pokémon app in several JavaScript frameworks.

Why it mattersFramework choices are often justified as "better for agents" without data. This benchmark measures agent build time, token spend, and output quality per framework, giving a basis for the decision.

articleluu

What [67] MCP servers do at startup, with the internet cut off

A sandboxed, no-internet measurement of 67 npm MCP servers at startup.

Why it mattersShows that 7 of 67 MCP servers contacted hosts their README never names at startup, and that 24 offered undocumented tools. Gives a replicable method for auditing servers before trusting them.

articleDevanshKontey

Brazilian Court Sanctions Lawyers for Prompt Injection in Court Filing

A Brazilian labor court fined two law firms about $16,000 after finding white-on-white text in a filing instructing AI to respond superficially.

Why it mattersHidden prompt injection in documents now carries legal penalties, not just technical risk. Engineers building document-review agents should expect adversarial text in inputs and sanitize accordingly.

articledoodlesdev

Ai Code Review

OpenRouter walks through a small Python code reviewer built on its Responses API.

Why it mattersShows how to make a code reviewer run changed files in a hosted sandbox to confirm bugs and return applicable patches, instead of guessing from the diff. Includes per-model cost and runtime figures for choosing a model.

articleOpenRouter editorial sitemap

Your AI Agent Wants Your Data. Here's How to Lock It Down - Jeff Smith, Oracle

Oracle's Jeff Smith explains why natural-language SQL can return wrong data without erroring.

Why it mattersGives specific patterns for limiting agent database risk: expose vetted reports as tools, use read-only replicas, annotate schemas, and filter rows before results reach the agent.

videoAI Engineer

HackerTube: YouTube built from videos recommended by HN

Andrej Karpathy describes building HackerTube, a YouTube-style site of 6,454 videos recommended in HN comments, using Claude Code and Opus 5.5.

Why it mattersShows an end-to-end agentic build: Claude Code Workflows ran Sonnet over about 40K HN comments to extract 6,454 videos, then the agent deployed and hardened a VPS.

blogkarpathy

Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle

Oracle's Richmond Alake argues agents need an intelligence plane covering knowledge, memory, access control and improvement.

Why it mattersArgues a governed semantic layer and agent memory can cut context from 200K to 32K tokens while keeping accuracy, and covers memory security and fine-tuned routing models.

videoAI Engineer

Agents Prompt People: Using Agents for Long-Horizon Financial Work

Latent Space interviews Arca's lead AI engineer on running agentic wealth-management workflows that span days.

Why it mattersAgent tasks that last days need persistent per-client memory, human approval points, and an owned harness. Arca's setup shows how a regulated financial firm structures this.

articleRichard MacManus

Using Agents for Long-Horizon Financial Work — Aman Singh, Arca

Arca's head of AI details how an AI-native wealth firm runs agent tasks lasting days.

Why it mattersMulti-day agent workflows need persistent memory, human approvals, and an owned harness. A regulated wealth manager explains how it handles these, including turning staff feedback into evals.

videoLatent Space

Malware uses LLMs to rank victims by how much they're worth

Cisco Talos analyzes VOLTA, a self-deleting Windows infostealer that bundles stolen credentials and profile data, sends it to an LLM via OpenRouter as a target assessment.

Why it mattersShows attackers using LLM APIs as post-theft analysts, with a three-model failover to get past refusals. Teams running LLM endpoints should expect abuse and treat guardrail refusals as a bypass target.

blogkhalidkhair

Jepsen: ParticleDB 1.1.4

Jepsen tested ParticleDB 1.1.4 through 1.1.18.11.rc-22 and reported 96 issues, from crashes and data loss to phantom rows.

Why it mattersParticleDB targets AI agent workloads and claims full ACID, yet Jepsen found lost committed writes under the default transaction mode and Lost Update under serializable.

DeepMind's PlanAssist: Gemini cuts UK planning review time 58%

Google DeepMind reports PlanAssist, a Gemini-based tool that validates planning applications and drawings, checks policy and constraints, and drafts cited reports.

Why it mattersA government deployment of a Gemini multimodal pipeline (document validation, policy checks, cited draft reports) reports a 58% time reduction and 70%+ no-edit approvals.

articleGoogleDeepMind

We Took over ChatGPT Accounts

Parameter Labs describes a chain that turned a shared ChatGPT conversation into code running in the viewer's session.

Why it mattersA shared ChatGPT conversation could run attacker JavaScript in the viewer's signed-in session through a postMessage listener loaded by an interactive graph.

articlearmcat

How Anthropic ships Claude with Claude: Separating the Brain from the Hands — Latent Space

Latent Space interviews Anthropic's platform leads on the Opus 5.5 launch.

Why it mattersExplains why preserved thinking blocks can break a custom harness, how credentials stay outside the sandbox, and how Managed Agents separates harness, session and sandbox, all useful when designing agent runtimes.

articlelatent.space
Katelyn Lesse@katelyn_lesse

An Anthropic engineering lead argues multiagent systems create new infrastructure problems: stateful agent runtimes that Kubernetes wasn't built for, agent identity and authorization, injection spreading between agents, and monitoring with kill switches.

Why it mattersNames the specific engineering gaps in running thousands of persistent agents: stateful isolated workloads, agent identity and consent, cross-agent injection blast radius, and tamper-proof tracing with kill switches.

78% Wanted This AI Feature. Almost Nobody Used It - Supriya Subudhi, Bumble

Bumble's product manager shares lessons from A/B testing AI photo feedback.

Why it mattersStated demand for an AI feature did not predict use, and a green A/B dashboard hid a roughly 10x model cost and odd outputs. The talk gives five questions to ask before calling an AI test a win.

videoAI Engineer

When Running Your Own AI Models Beats Paying for APIs - Steven Latré, imec

Imec benchmarked open-weight models and three coding harnesses on SWE-Bench Pro to compare self-hosted GPUs with APIs.

Why it mattersSelf-hosting versus API pricing for coding agents depends on utilization and concurrency. Break-even ranged from 89% on small systems to about 15% on 8x B200, and harness choice can halve GPU time.

videoAI Engineer

An index of the vibe-coding frontier. Corrections welcome.