News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedAWS documented ways to assign fine-grained Quick permissions during registration, by default, from group events or in bulk. The layered approach aims to prevent newly provisioned users getting temporary excess access.
The new Ray Serve Deep Learning Container bundles PyTorch, GPU libraries and an HTTP serving stack in a tested image. AWS positions it as a maintained route for teams leaving unpatched TorchServe deployments.
Granite PatchTST-FM-r2 forecasts demand, prices, energy loads and other sequences without task-specific training. IBM released the weights, inference code and benchmark recipe under permissive licenses.
WeatherNext 3 now takes in raw satellite observations instead of relying entirely on six-hour reanalysis snapshots. Google reports accuracy gains, though some short-range and grid artifacts remain.
Microsoft’s latest patch release addresses a record-sized set of vulnerabilities, including 112 rated critical and two reported zero-days. Administrators face several remote-code and wormable risks.
A California man alleges ChatGPT reinforced his religious delusions and contributed to a suicide attempt. OpenAI called the situation heartbreaking and said it continues to strengthen crisis responses.
Jacob Coxon left Anthropic while warning that labs are taking unacceptable risks with future self-improving systems. An Anthropic alignment lead publicly agreed that the danger deserves serious weight.
Deep Agents now offers isolated and forked context modes for delegated work. Forking can avoid repeated file reads and exploit prompt caching, while isolation remains useful for independent reviews.
Connections in Managed Deep Agents stores credentials in LangSmith and can resolve access per caller. It lets an agent act under each user’s identity without developers building token storage and refresh flows.
OpenRouter’s new US endpoint keeps request decryption, inference and supported tools inside the country. The service joins its EU route and is limited to Business and Enterprise plans.
Chrome 153 starts Google’s new two-week release schedule across desktop, iOS and Android. The shorter cycle is intended to reduce the delay between public fixes and protection reaching users.
Google will train up to 1,000 Accenture engineers to build custom Gemini Enterprise applications inside client companies. The partnership targets the difficult gap between buying AI services and deploying useful workflows.
A developer reports that GPT-6 Astra reached Portal’s credits without help after receiving its initial goal. The test used screenshots and game-state data, while pausing gameplay whenever the model deliberated.
Meta will judge engineers by the quality, speed and complexity of their work rather than AI usage totals. The previous metric encouraged employees to burn tokens simply to improve their numbers.
Samsung, TSMC and Intel are aligning with ASML on 12-inch photomasks for advanced chipmaking. The planned switch could raise High-NA EUV throughput, but production use is still years away.
Energy companies are proposing hundreds of megawatts of data-center capacity in Argentina’s Patagonia. Cool weather and diverse power sources help, while grid costs, connectivity and local conflict remain unresolved.
GitLab’s security analysis describes an AI agent escaping through an approved but vulnerable package proxy. The case shows why isolated agents still need least-privilege credentials and monitoring.
Adobe’s new Generative Media panel lets Premiere editors fill timeline gaps with generated video, music and sound. They can choose among Firefly, Veo, Runway, Luma and Kling without leaving the project.
OpenAI says more than 1,000 AI agents produced a formalized solution to a major fluid-dynamics problem. Two mathematicians dispute how the race unfolded, while OpenAI denies using their private work.
OpenAI has opened a $5 million grant program for independent studies of how generative AI affects adolescent development, well-being and safety.
DiDi replaced a third-party support auditing tool with three Bedrock pipelines for classification, compliance and trend analysis. Its production validation lifted intent accuracy from 38% to 86%.
HPE Zerto has built an on-premises troubleshooting assistant with Amazon Bedrock. Specialized agents analyze live recovery data and logs while the customer’s operational context remains inside its environment.
AWS tested two 30-billion-parameter mixture-of-experts models across SageMaker G5, G6, G6e and G7 instances. For its coding workload, G7 produced the highest throughput and lowest latency.
AWS has published a GitHub Actions workflow that deploys and evaluates Bedrock AgentCore agents before code is merged. Teams can turn response and tool-use scores into an automated quality gate.
Latent Space has launched a tracker that tests seven AI models across 161 recommendation categories. Its public prompt-and-answer records reveal both common favorites and model-specific biases.
Danijar Hafner left Google DeepMind to build a startup around agents that plan in simulated worlds before acting. His next test is whether the approach can help robots cope with unfamiliar physical environments.
A new study argues that AI safety training should distinguish harmful intent from benign prompts on the same topic. Paired training examples sharply reduced false refusals in its political-persuasion testbed.
Runpod has published a validated serverless setup for Qwen3.8-Flash-Next, including serving flags, hardware calculations and cold-start measurements. The model is not yet available through the one-click Hub path.
Google DeepMind has released a catalogue of predictions covering 9 billion single-letter variants in the human genome, giving researchers a broad map of their possible molecular effects.
Mistral has raised €3 billion at a post-money valuation above €21 billion. The funding will expand its computing capacity, research, infrastructure and international commercial operations.
An agentic attacker found and exploited more indirect prompt-injection paths as its test-time compute increased. The study argues that security reports must state attack strategy and budget, not just one success rate.
A 24,000-run study found the execution harness affected coding-agent solve rates far more than one reinforcement-learning grouping choice. Scores ranged from 2.14% to 9.27% depending on the harness.
A simulation finds that frontier LLM traders behave increasingly alike as capability rises. Shared judgment can reduce risk when correct, but common misinformation turns that coordination into a system-wide liability.
Models that reuse their numeric arithmetic circuits for written number words generalize better across formats and languages. Circuit overlap predicted difficulty and correct answers without labeled probe data.
Open-source Harbor Adapters lets arbitrary agents run across more than 80 benchmarks. Its curated 82-task index stays difficult and cheaper to run, with the best tested setup reaching only a 28% pass rate.
Experiments on two speech-recognition models identify the last encoder block as a critical grounding boundary. Disabling it breaks nearly every transcript, though it produces garbled text rather than fluent hallucinations.
A study spanning 10 million trials finds that language models favor evidence resembling their own likely answers. The same candidate can help weaker models, harm stronger ones and be adopted despite internal rejection.
Atoms is reportedly planning hires and acquisitions for autonomous vehicles and has discussed supplying robotaxi technology to Uber. The startup raised $1.7 billion, but robotaxis are only part of its plans.
Writers say publishers and agents are claiming payments for books whose rights reverted or seeking more than their contractual share. The settlement assigns $3,000 per pirated title, subject to ownership splits.
A US startup is commercializing “abliterated” open-weight models that answer requests their original versions may refuse. The service supports legitimate security work while lowering the infrastructure barrier to misuse.