Latest AI news
Curated AI news on models, agents, tools, research and industry, in English, Spanish and Brazilian Portuguese.
AgentsHow a Team of Agents Automates OpenTelemetry Instrumentation for GenAI FrameworksAn OpenTelemetry blog post describes a multi-agent system that inspects GenAI frameworks, plans and writes instrumentation against the GenAI semantic conventions, and validates the telemetry with Weaver's live-check.AgentsMudGym Brings MUD2 to the Browser via WebAssembly as a Reinforcement Learning EnvironmentA developer has compiled MUD2, one of the earliest online multiplayer games, to WebAssembly and paired it with MudGym, a reinforcement learning environment whose marimo notebook examples run in the browser through Pyodide.IndustryLLM batch APIs cut prices in half, but only for traffic that can waitA cost-modeling blog explains how batch discounts apply only to deferrable calls, why batch and caching stack multiplicatively, the billing bug this caused, and why dropping a model tier usually saves more.
IndustryEli Lilly CEO joins expert pushback on Dario Amodei's claim that AI will cure major diseases within a decadeGary Marcus collected reactions on X to Anthropic CEO Dario Amodei's IPO video predicting cures for cancer, Alzheimer's, heart disease and mental illness in under ten years; critics including Eli Lilly's Dave Ricks called the timeline unrealistic.
ModelsGoogle Deepmind's AlphaGenome 2 moves genomics AI from bulk tissue predictions to single-cell resolutionGoogle Deepmind's AlphaGenome 2 predicts the effects of genetic variants on specific human cell types using a conditional model trained on over 107 million cell profiles, and is available for noncommercial research via API.IndustryAnthropic Files Updated Prospectus for Nasdaq IPO at $70 to $78 per ShareAnthropic filed an amended prospectus for a Nasdaq listing under the ticker ANT. The offering of 825 million shares at $70 to $78 could raise as much as $100 billion and value the company near $1 trillion.
ResearchMicrosoft Research details Lazarus, a backdoor that reprograms dormant neurons in AI agentsMicrosoft researchers describe Lazarus, a technique that turns rarely used neurons in a language model into trigger detectors that make an agent run attacker commands, and outline mitigations in Foundry.IndustryProposed $10 Million Google-Spirit Data Sale Adds Worker Privacy SafeguardsA union challenge to Spirit Airlines' proposed $10 million data sale to Google produced a settlement with third-party de-identification, sensitive-record exclusions and an exit clause for Google.IndustryAustralia's proposed AI safety bill would force frontier model makers to open up to independent testingThe Albanese government has announced legislation imposing mandatory safety rules on frontier AI companies in Australia, including independent audits, published safety frameworks and immediate incident reporting.IndustryOpenAI signs child-safety AI charter and commits $2 million to research networkAt the Paris Peace Forum, OpenAI signed the Charter of Engagement for Safer AI for Children and pledged $2 million to a research network under the Coalition for Children's Rights and Protection in the Age of AI.ToolsLangfuse v4.79.0 brings OpenRouter, xAI and Vercel AI Gateway providers to general availabilityLangfuse v4.79.0 graduates three new AI Gateway providers, expands MCP and evaluation tooling, adds audit-log coverage for self-hosted enterprise installs, and ships a round of fixes and dependency updates.
ToolsShanHaiGateway: a Go gateway that pools multiple LLM upstreams behind one OpenAI-compatible endpointA self-hosted Go gateway that merges multiple upstream LLM providers, free tiers included, into one OpenAI-compatible endpoint, with error-class cooldowns, session affinity and quota automation; its author reports a 20-billion-token month for about ¥170.
IndustryEli Lilly CEO joins expert pushback on Dario Amodei's claim that AI will cure major diseases within a decadeGary Marcus collected reactions on X to Anthropic CEO Dario Amodei's IPO video predicting cures for cancer, Alzheimer's, heart disease and mental illness in under ten years; critics including Eli Lilly's Dave Ricks called the timeline unrealistic.ModelsGoogle Deepmind's AlphaGenome 2 moves genomics AI from bulk tissue predictions to single-cell resolutionGoogle Deepmind's AlphaGenome 2 predicts the effects of genetic variants on specific human cell types using a conditional model trained on over 107 million cell profiles, and is available for noncommercial research via API.IndustryAnthropic Files Updated Prospectus for Nasdaq IPO at $70 to $78 per ShareAnthropic filed an amended prospectus for a Nasdaq listing under the ticker ANT. The offering of 825 million shares at $70 to $78 could raise as much as $100 billion and value the company near $1 trillion.ResearchMicrosoft Research details Lazarus, a backdoor that reprograms dormant neurons in AI agentsMicrosoft researchers describe Lazarus, a technique that turns rarely used neurons in a language model into trigger detectors that make an agent run attacker commands, and outline mitigations in Foundry.IndustryProposed $10 Million Google-Spirit Data Sale Adds Worker Privacy SafeguardsA union challenge to Spirit Airlines' proposed $10 million data sale to Google produced a settlement with third-party de-identification, sensitive-record exclusions and an exit clause for Google.IndustryAustralia's proposed AI safety bill would force frontier model makers to open up to independent testingThe Albanese government has announced legislation imposing mandatory safety rules on frontier AI companies in Australia, including independent audits, published safety frameworks and immediate incident reporting.IndustryOpenAI signs child-safety AI charter and commits $2 million to research networkAt the Paris Peace Forum, OpenAI signed the Charter of Engagement for Safer AI for Children and pledged $2 million to a research network under the Coalition for Children's Rights and Protection in the Age of AI.ToolsLangfuse v4.79.0 brings OpenRouter, xAI and Vercel AI Gateway providers to general availabilityLangfuse v4.79.0 graduates three new AI Gateway providers, expands MCP and evaluation tooling, adds audit-log coverage for self-hosted enterprise installs, and ships a round of fixes and dependency updates.ToolsShanHaiGateway: a Go gateway that pools multiple LLM upstreams behind one OpenAI-compatible endpointA self-hosted Go gateway that merges multiple upstream LLM providers, free tiers included, into one OpenAI-compatible endpoint, with error-class cooldowns, session affinity and quota automation; its author reports a 20-billion-token month for about ¥170.