Skip to content
AI News HubLIVE

This edition’s highlights

Models

Was I wrong about open models & cyber risks?

  • The author coined 'Black October' as critical vulnerabilities like PHP2Shell and GitLab marked AI cyber risks materializing.
  • GLM-5.3, released over a month ago, has not produced public evidence of catastrophic societal harm despite crossing a capability threshold.
Interconnects (Nathan Lambert)In-site articleWas I wrong about open models & cyber risks?

Configure the ChatGPT desktop app with Amazon Bedrock in AWS GovCloud (US)

  • The ChatGPT desktop app now supports Amazon Bedrock in AWS GovCloud (US), providing access to OpenAI frontier models while meeting data residency and stringent security requirements.
  • Configuration centers on API-mode sign-in, administrator-controlled requirements.toml policies, user config.toml settings, a named AWS profile, and network rules that allow only the Bedrock Mantle endpoint.
AWS Machine Learning BlogIn-site articleConfigure the ChatGPT desktop app with Amazon Bedrock in AWS GovCloud (US)

The fastest delta weight syncs in the world

  • GPU weights become the sole parent, removing a ~770 GB CPU copy and cutting sync time from 36s to 15s and pause from 6s to 0.8s
  • O(1) delta checksum updates read only changed bytes, making the full update up to 2.3x faster
Baseten BlogIn-site articleThe fastest delta weight syncs in the world

Seq@K: Evaluating Language Models Under Sequential Feedback

  • Seq@K measures a model's ability to solve problems within K sequential attempts using feedback after each failure.
  • On 7 benchmarks and 16 frontier models, Seq@K often produces different model rankings than Pass@K.
arXiv AIIn-site articleSeq@K: Evaluating Language Models Under Sequential Feedback

dbreunig/nesformer: NES emulator baked into Qwen3.5-4B weights

  • NES 6502 CPU emulator embedded in Qwen3.5-4B weights, running game ROMs very slowly
  • Built using Opus 5.5 prompts, inspired by Divyansh Agrawal's qwenputer
Simon Willison's WeblogIn-site articledbreunig/nesformer: NES emulator baked into Qwen3.5-4B weights
Agents

Evaluating and improving AI agents for sequencing protocol understanding

  • LibStructBench is a document-grounded benchmark for reconstructing sequencing-library structures and molecular workflows.
  • Frontier agents often produce approximately correct reconstructions but remain unreliable at exact reconstruction, particularly in molecular-state representation and workflow connectivity.
arXiv AIIn-site articleEvaluating and improving AI agents for sequencing protocol understanding

How Worldwide Express automated freight exceptions with Amazon Bedrock

  • Multi-agent system reduced active exceptions from $15M to $4M, a 75% decrease.
  • Time-to-cash improved by four weeks, with over $1M in annual savings and 60x first-year ROI.
AWS Machine Learning BlogIn-site articleHow Worldwide Express automated freight exceptions with Amazon Bedrock

Wealth Management Client Onboarding Needs More Than AI

  • Task-level AI improves individual steps but handoffs across systems still cause delays.
  • Governed case context provides a shared, versioned view of evidence, exceptions, and decisions.
Databricks BlogIn-site articleWealth Management Client Onboarding Needs More Than AI
Tools

Autoscaling schedules: reduce cold-start delays for predictable traffic

  • Autoscaling schedules start replicas before traffic arrives to avoid cold-start delays.
  • Each schedule sets a time window with min and max replicas; up to 10 schedules per environment.
Baseten BlogIn-site articleAutoscaling schedules: reduce cold-start delays for predictable traffic

You have reached the end of this edition.

Your reading list
Other updates (10)
Agents

Detecting Covert Collusion between AI Agents through Partner-Outsider Positive Listening

A new arXiv paper proposes a behavioral test called PoPLi to detect whether AI agents are colluding via secret encoded messages, even when the messages appear innocent. In a simulated market, PoPLi was significantly positive when a covert code was present and near zero otherwise, outperforming message-only monitors.

arXiv AIIn-site articleDetecting Covert Collusion between AI Agents through Partner-Outsider Positive Listening

Measuring Where Intent Gets Lost in Agentic Smart Contract Synthesis

An arXiv paper evaluates AI-written smart contracts through execution rather than opinion, finding that AI judges and standard analysis tools miss obvious mistakes, and proposes an elicitation-first architecture that asks clients a short checklist of questions before generating code and verifies results against an independently written executable answer key.

arXiv AIIn-site articleMeasuring Where Intent Gets Lost in Agentic Smart Contract Synthesis

WebAnything: From the Wild Web to Scalable Agent Environments

WebAnything is a fully automated framework that reconstructs source-grounded, reusable web environments from real URLs without per-site engineering. It built 244 website artifacts across 14 domains and 53 categories, supporting 12,984 executed tasks, and outperformed human-assisted WebHarbor in a blind audit.

arXiv AIIn-site articleWebAnything: From the Wild Web to Scalable Agent Environments

Preview web apps live as your coding agent builds them with Amazon Bedrock AgentCore runtime

Amazon Bedrock AgentCore Session endpoints give each runtime session an authenticated URL that forwards any HTTP method, path, and WebSocket connection to the ports you allow. In this post, you deploy a coding agent that keeps a Vite website running in its session, then watch the live site update in your browser as the agent edits files.

AWS Machine Learning BlogIn-site articlePreview web apps live as your coding agent builds them with Amazon Bedrock AgentCore runtime

Deliberate decisions and enterprise AI value

A DataIQ and Databricks report, based on peer discussions at DataIQ Guild 2026, outlines five leadership disciplines and four foundations to help enterprises make grounded decisions and turn AI ambition into sustained value amid uneven readiness and expectations.

Databricks BlogIn-site articleDeliberate decisions and enterprise AI value
Models

What Does an Ontology Catch? Curated, LLM-Generated and Data-Derived Ontologies as Checkers of LLM Agents over Knowledge Graphs

An arXiv study makes ontology checkers the object of study, comparing curated, LLM-generated, and data-derived ontologies as judges of LLM agent answers over knowledge graphs. It finds that on UniProt, a curated ontology flags more correct answers than alternatives, but on real agent answers its flags are only marginally more precise than random.

arXiv AIIn-site articleWhat Does an Ontology Catch? Curated, LLM-Generated and Data-Derived Ontologies as Checkers of LLM Agents over Knowledge Graphs

LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes

LRMoE is a memory-aware MoE training system that jointly recalibrates backward demand and balances retained workload, planning retention quotas and expert placement during forward execution, and selecting token-expert pairs based on contribution during backpropagation. Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show 21.2-44.8% reduction in cumulative MoE-layer training time to fixed-loss targets relative to SYMI.

arXiv AIIn-site articleLRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes
Research

OpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery

OpenFrontiers introduces a cross-disciplinary corpus of 12,151 open research problems and a retrospective math benchmark of 200 problems with Lean 4 formalization, evaluated under closed-book, frozen-retrieval, and live-web conditions. Testing seven LLM agents shows closed-book rediscovery is rare (at most 7 of 40 resolved problems), while live-web success largely reflects retrieval of published solutions.

arXiv AIIn-site articleOpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery

ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions

ThinkJev introduces a training-free framework that builds fast and slow systems from a single hybrid-thinking model to calibrate typed decisions. A statistical test on labeled dev data ensures System 1 only answers alone when its accepted answers stay under a target error rate; remaining decisions escalate to thinking with early stopping, achieving near always-think accuracy at 72% of the thought tokens.

arXiv AIIn-site articleThinkJev: One Thought, Every Budget for Calibrated Typed Decisions

Amortized Localized Posterior Inference for Model Calibration with Application to an Ocean Turbulence Closure

This paper introduces a localized posterior inference framework that weights observation likelihoods by proximity to a target context and uses a conditional normalizing flow for amortized variational inference. It addresses model discrepancy where a single calibration parameter fails across the context domain. Twin experiments with the Branin-Hoo model and a single-column ocean k-ε turbulence closure show reduced aggregate predictive errors compared to global posterior and deterministic calibration.

arXiv AIIn-site articleAmortized Localized Posterior Inference for Model Calibration with Application to an Ocean Turbulence Closure