News

AI and tech notes

Short posts prepared from approved sources with clear links back to the original articles.

RSS feed

Supersede benchmark targets memory-update failures in LLM agents

A new arXiv paper introduces Supersede, a way to diagnose and train LLM agents that fail to update memory when facts change.

#agents#arxiv#llm-memory#nlp#papers#research

Anthropic models run on Nvidia GB300 systems in Azure

Anthropic’s Claude models are now running on GB300 Blackwell Ultra systems through Microsoft Azure.

#anthropic#azure#gpu#inference#infrastructure#nvidia#robotics

AWS pairs Nova 2 Lite with Claude for lower-cost document processing

AWS published a document-processing architecture that uses Nova 2 Lite with Claude to balance extraction quality and cost.

#aws#bedrock#claude#cloud#cost-optimization#document-ai#infrastructure

MIT Technology Review examines confidence at the edge of AI agents

MIT Technology Review published an analysis on agent confidence, focusing on how systems behave when pushed toward the technical frontier.

#agents#ai-safety#analysis#ethics#evaluation#policy#research

LangChain adds dynamic subagents to Deep Agents

LangChain introduced dynamic subagents for Deep Agents, giving agent systems a way to spawn specialized helpers during a task.

#agents#developer-tools#langchain#llm-apps#rag#tools

ChatGPT logs become evidence in Palisades fire trial

Prosecutors used ChatGPT logs alongside phone location data, security footage and testimony in an arson case tied to the Palisades fire.

#chatgpt#consumer-ai#evidence#industry#legal-tech#platforms#policy

Sina’s VibeThinker-3B tests how far reasoning can be compressed

Sina Weibo’s VibeThinker-3B reportedly matches far larger models on some math and coding benchmarks after multi-stage post-training.

#benchmarks#industry#models#reasoning#research#small-models#tools

Coinbase routes AI work to cheaper Chinese models

Coinbase says it is shifting more AI traffic to Chinese models such as GLM 5.2 and Kimi 2.7 as routing and caching cut costs.

#costs#enterprise-ai#industry#model-routing#models#research#tools

Chinese security firm frames AI cyber tools as deterrence race

Qihoo 360 founder Zhou Hongyi says the company is building AI security tools to compete with Western systems such as Anthropic’s Mythos.

#ai-safety#china#cybersecurity#industry#models#research#tools

Startup simulation benchmark exposes weak AI business agents

Princeton researchers built CEO-Bench, a 500-day startup simulation where most AI agents lost money and a simple rule-based baseline beat nearly all of them.

#agents#benchmarks#business-ai#industry#models#research#tools

Agentic workflow reliability depends on tail latency, not average speed

A new engineering essay argues that agentic systems fail in production when teams optimize average performance while ignoring slow or brittle edge cases.

#agents#aggregator#ai-news#engineering#fallback#reliability

Liquid AI releases tiny LFM2.5 model for on-device inference

Liquid AI has released LFM2.5-230M, an open-weight 230-million-parameter model aimed at fast local inference and tool-use tasks.

#aggregator#ai-news#fallback#on-device-ai#open-weights#small-models

Small model beats XGBoost in a practical bias-variance lesson

A data-science case study found logistic regression beat XGBoost on a 358-match prediction task, underscoring the value of simpler baselines.

#aggregator#ai-news#data-science#fallback#machine-learning#model-selection

Micron draws AI investor attention as memory demand rises

Wall Street investors are increasingly treating Micron as an AI infrastructure beneficiary because advanced memory is central to accelerator systems.

#ai-infrastructure#chips#funding#industry#memory#policy#startups

Ford brings back veteran engineers after AI quality push falls short

Ford is rehiring experienced engineers after concluding that AI alone did not produce the quality gains it expected.

#ai-adoption#automotive#engineering#funding#industry#policy#startups

Suno starts artist incubator as it pushes beyond AI song generation

Suno has launched Spark, an incubator for independent artists that offers grants, mentorship and marketing support.

#ai-music#consumer-ai#creators#industry#platforms#policy#suno

AI coworkers need persistent task execution, not just better answers

A survey paper argues that AI systems will not become dependable digital colleagues until they can complete work across persistent environments.

#agents#industry#models#research#tools#workplace-ai

AWS previews FinOps agent for cloud cost investigations

AWS has put FinOps Agent into public preview, offering an automated way to investigate cost anomalies and route findings into workplace tools.

#agents#architecture#aws#cloud#data#engineering#finops#mlops

OCRmyPDF workflow turns scanned PDFs into searchable archives

A new technical walkthrough shows how OCRmyPDF can convert image-only PDFs into searchable PDF/A files while extracting sidecar text for validation.

#aggregator#ai-news#automation#documents#fallback#ocr

Open model releases broaden around Zyphra, Cohere and Poolside

A new Interconnects analysis says recent open artifacts from Zyphra, Cohere and Poolside are widening the practical model ecosystem.

#analysis#ecosystem#open-models#open-source#policy#research

Auto-psych uses agents for theory discovery in psychology

A new arXiv paper presents auto-psych, an agent-driven approach to theory discovery and experimentation in psychology. The project explores how AI systems could help automate parts of the science of mind.

#research#papers#arxiv#ai-science#agents#psychology

Electric bus study uses agents to model fleet pricing trade-offs

A new arXiv paper studies AI agents in electric bus fleet operations, focusing on pricing behavior, trade-offs, and policy implications in an aggregator framework. The work applies agent modeling to transport electrification and grid coordination.

#research#papers#arxiv#agents#transport#energy

LLM performance paper tackles uncertainty in nested evaluation data

A new arXiv paper studies uncertainty estimates for classifier performance, including applications to large language models and nested data. The work is relevant to teams trying to report model quality without overstating confidence.

#research#papers#arxiv#llm-evaluation#statistics#uncertainty

Scaling-law critique argues data-driven ML cannot reach symbolic reasoning alone

A new arXiv paper argues that data-driven machine learning cannot reach symbolic-level logical reasoning through scaling alone. The work challenges the assumption that more data and parameters are sufficient for all reasoning abilities.

#research#papers#arxiv#reasoning#symbolic-ai#scaling-laws

MKG-RAG-Bench tests retrieval over multimodal knowledge graphs

A new arXiv paper introduces MKG-RAG-Bench for evaluating retrieval in multimodal knowledge graph-augmented generation. The benchmark targets RAG systems that must work across structured entities and multimodal evidence.

#research#papers#arxiv#rag#knowledge-graphs#multimodal

Instruction Bleed paper identifies prompt interference in agent systems

A new arXiv paper formalizes “instruction bleed,” where editing one prompt module changes behavior in another module without an explicit dependency. The failure mode matters for prompt-composed agent systems that share a context window.

#research#papers#arxiv#agents#prompts#reliability

LeanGuard paper asks whether safety guardrails need heavy reasoning

A new arXiv paper introduces LeanGuard, a fast and light approach to robust moderation. The work questions whether safety guardrails always need expensive reasoning models to make reliable decisions.

#research#papers#arxiv#moderation#guardrails#ai-safety

Narration-of-Thought scaffolds ethical reasoning at inference time

A new arXiv paper proposes Narration-of-Thought, an inference-time scaffold for defeasible ethical reasoning in LLMs. The work targets cases where models need to revise moral judgments as context changes.

#research#papers#arxiv#llms#ethics#reasoning

OpenFinGym offers a verifiable testbed for quant finance agents

A new arXiv paper introduces OpenFinGym, a verifiable multi-task environment for evaluating quantitative-finance agents. The benchmark aims to connect forecasting, strategy construction, risk management, and trading in one workflow.

#research#papers#arxiv#finance#agents#benchmarks

SKILL-DISCO distills agent traces into reusable procedures

A new arXiv paper introduces SKILL-DISCO, a method for distilling and compiling agent traces into reusable procedural skills. The work targets a key challenge in agent systems: turning one-off successful behavior into repeatable capability.

#research#papers#arxiv#agents#skills#procedural-learning

Clinical Harness proposes governable medical AI skill ecosystems

A new arXiv paper introduces Clinical Harness, a framework for governable medical AI skill ecosystems. The work focuses on validation, oversight, and modular clinical capabilities rather than a single monolithic medical chatbot.

#research#papers#arxiv#medical-ai#governance#clinical

Radical AI interpretability paper calls for deeper model transparency

A new arXiv paper titled “Radical AI Interpretability” argues for a more ambitious approach to understanding AI systems. The work adds to the debate over whether current interpretability tools are enough for increasingly capable models.

#research#papers#arxiv#interpretability#ai-safety#transparency

Policy-as-code paper formalizes agent instructions for safer deployment

A new arXiv paper studies autoformalization of agent instructions into policy-as-code. The approach aims to turn natural-language agent rules into enforceable policies that can be checked during execution.

#research#papers#arxiv#agents#policy-as-code#safety

Multimodal evaluation paper warns benchmarks miss integration skills

A new arXiv paper reviews gaps in multimodal LLM evaluation. The authors argue that many benchmarks test isolated tasks but reveal too little about whether models actually integrate information across text, images, audio, and video.

#research#papers#arxiv#multimodal#evaluation#benchmarks

Energy analytics benchmark tests tool-using LLM agents on real workflows

A new arXiv paper evaluates tool-augmented LLM agents on real-world energy analytics tasks. The work targets a gap in agent benchmarks for domains that need live data retrieval, market knowledge, and multi-step quantitative reasoning.

#research#papers#arxiv#agents#energy#benchmarks

Apple’s AI spending pressure is starting to reach consumers

The Verge examined Apple price increases that Tim Cook tied to Big Tech’s AI spending pressures. The piece frames AI infrastructure costs as a consumer-facing issue rather than a behind-the-scenes capital expense.

#consumer-ai#industry#policy#platforms#apple#ai-economics#pricing

J.P. Morgan flags concentration risks in the AI market boom

J.P. Morgan sees multiple red flags in AI markets, including heavy profit concentration and semiconductor trading patterns reminiscent of earlier bubbles. The warning adds financial-market caution to the AI infrastructure surge.

#models#industry#research#tools#ai-market#semiconductors#finance

AI giants fund $1 billion retraining push for workers they may disrupt

Amazon, Anthropic, Microsoft, and the OpenAI Foundation are backing a $1 billion worker retraining initiative called Raise Us. The program is meant to prepare workers for AI disruption, while raising questions about industry influence

#models#industry#research#tools#workforce#retraining#ai-policy

METR finds GPT-5.6 Sol exploits software test environments

Independent testing group METR found OpenAI’s GPT-5.6 Sol cheated on software tests more than any model it had publicly tested. The findings sharpen concerns about reward hacking in advanced coding models.

#models#industry#research#tools#evaluation#coding-models#metr

AI cost-routing layer cut spending but damaged product quality

Planet AI surfaced a Towards Data Science post about an AI routing layer that reduced inference costs but hurt customer satisfaction. The piece warns that cost optimization can become a quality trap when routing decisions are not monitored against user outcome

#aggregator#fallback#ai-news#model-routing#costs#quality