AI news, with memory.
Wednesday, 28 October 2026Updated 28 Oct 2026, 01:31 UTC

Developments

Reset
261 results
Products· 1 source link

Snowflake adds dynamic model routing to Cortex AI Gateway

Snowflake has added dynamic model routing to Cortex AI Gateway, letting enterprise users choose an automatic mode that selects a model for each request based on a quality-cost tradeoff instead of pinning workloads to a single model. The company says the feature can reduce token costs by up to 3x on some workloads, based on internal testing, and positions it as part of a broader push to make enterprise AI agents cheaper and better governed.

Sources: VentureBeat
Models· 2 source links

Z.ai releases open-weight cybersecurity models with dual-use bug-finding capabilities

Z.ai has released new open-weight AI models aimed at cybersecurity tasks, with reported strength in software vulnerability discovery and related offensive-defensive workflows. Coverage of the launch highlights expert concern that the same capabilities that could help defenders find and fix flaws may also make advanced bug-finding tools more accessible to malicious actors.

Models· 1 source link

Alibaba releases Qwen3.8-27B as an Apache 2.0 open-weight multimodal model for local coding and reasoning

Alibaba has released Qwen3.8-27B on Hugging Face under an Apache 2.0 license, making the 27-billion-parameter multimodal model available as downloadable weights rather than only through a cloud API. VentureBeat reports the model supports image and video understanding, a 262,144-token context window, configurable reasoning, and coding and agentic workflows, with quantized versions small enough to run on high-end local hardware.

Sources: VentureBeat
Products· 1 source link

Reddit tests AI-generated audio and video summaries for text posts

Reddit is testing an AI feature that converts some text posts into podcast-style audio and short videos. The experiment uses AI voices to read the original post and selected comments while adding visual highlights, offering users another way to consume discussions on the platform.

Sources: The Verge
Products· 2 source links

Cursor launches Origin code hosting platform for paid users

Cursor has begun rolling out Origin, its own code hosting platform, to paid users. VentureBeat framed the launch against a same-day GitHub outage that affected core developer services including pull requests, API access, enterprise single sign-on, and Copilot, but the discrete news is Cursor's product release.

Companies· 1 source link

Relay shuts down as much of its team joins Google Chrome

AI automation startup Relay is shutting down, and much of its staff is joining Google’s Chrome team, according to TechCrunch. Relay founder and CEO Jacob Bank said the team will work on AI features in Chrome, signaling an acqui-hire style move rather than a standalone product launch or financing event.

Sources: TechCrunch
Safety· 2 source links

Report alleges Amazon destroyed rare books to obtain AI training data

Futurism reports that Amazon used a process that involved destroying physical books, including rare volumes, in order to digitize their contents for AI training. The article frames the practice as a data-sourcing and ethics controversy around how training material was acquired, rather than a new model, product, or financing announcement.

Products· 1 source link

Siemens and Maschinenfabrik Reinhausen unveil 800V DC power system for AI datacenter racks

Siemens and Maschinenfabrik Reinhausen have announced a power-delivery system designed to bring 800-volt direct current to AI datacenter racks. The effort targets rising power demands from next-generation accelerated computing hardware and aims to improve how electricity is distributed inside high-density AI facilities.

Sources: The Register
Research· 1 source link

MIT and Harvard researchers introduce Role Anchor to reduce role drift in compound AI systems

Researchers at MIT and Harvard describe a failure mode they call role drift, in which modules inside compound AI systems deviate from their intended functions while overall end-to-end accuracy still improves. They introduce a method called Role Anchor that aims to keep modules constrained to their assigned roles during training, including in retrieval-augmented generation pipelines where a reader might otherwise answer from internal memory instead of retrieved evidence.

Sources: VentureBeat
Products· 1 source link

Feedly says a bug caused a weeklong slowdown affecting some users

Feedly said a software bug, rather than its broader AI product shift, was responsible for performance problems that made its web app significantly slower for some users over roughly a week. The report also cites user complaints about mobile app issues and customer-support responsiveness during the disruption.

Sources: TechCrunch
Policy· 1 source link

ByteDance signs AI safeguards pact with the Motion Picture Association on film copyright protection

ByteDance has signed a memorandum of understanding with the Motion Picture Association covering safeguards for the company's AI models around film and television intellectual property. The agreement signals cooperation between a major AI developer and Hollywood studios' trade group on measures intended to reduce unauthorized use or reproduction of protected content.

Sources: Engadget
Products· 1 source link

xpander.ai makes its enterprise AI agent control-plane platform generally available

xpander.ai said its enterprise platform for building, running, and governing AI agents is now generally available. The company positions the product as a vendor-neutral control and context layer that sits above models, agent frameworks, and infrastructure, adding execution controls, permissions, observability, memory, enterprise system access, and lifecycle management.

Sources: VentureBeat
Companies· 1 source link

OpenAI reportedly finalizes lease for major Ohio data center backed by Nvidia

The Wall Street Journal reports that OpenAI has locked in a lease for a large data-center project in Ohio in a deal involving SoftBank-owned SB Energy and backing from Nvidia. The agreement is described as the largest data-center project announced to date and caps weeks of negotiations over a major AI infrastructure buildout.

Sources: WSJ
Products· 1 source link

Microsoft says Exchange Server update is delayed amid AI-related bug backlog

Microsoft has delayed a promised update for Exchange Server and, according to The Register, attributed the slippage in part to a backlog of bugs generated by AI-assisted development or testing workflows. The company reportedly did not provide a new delivery date, turning the issue into a product-status update rather than a launch announcement.

Sources: The Register
Safety· 2 source links

OpenAI reportedly disbanded its preparedness team

The Financial Times, as cited by The Verge, reports that OpenAI disbanded its preparedness team at the end of last month. The team was responsible for assessing whether models posed serious risks and for developing mitigations, with those responsibilities reportedly being reassigned elsewhere inside the company.

Companies· 4 source links

Stripe reportedly agrees to acquire AI gateway startup OpenRouter for more than $7 billion

TechCrunch reports that Stripe plans to acquire OpenRouter, an AI gateway startup, in a deal valued at more than $7 billion. OpenRouter has positioned itself as infrastructure for routing and managing access to multiple AI models, and the reported purchase would mark a major AI infrastructure acquisition by the payments company.

Products· 2 source links

OpenAI adds Computer History to the ChatGPT macOS app

OpenAI has added a new Computer History feature to the ChatGPT desktop app for macOS, according to The Verge. The feature records user activity such as clicks and keystrokes to build a timeline that ChatGPT and Codex can reference for assistance, automation suggestions, and resuming unfinished tasks.

Models· 4 source links

DeepSeek raises prices for V4 Flash and Pro as third-party tests show weaker real-world agent reliability

DeepSeek has raised prices for its V4 Flash and V4 Pro models, extending a recent shift away from the ultra-low pricing that helped drive developer adoption. VentureBeat also cites Composio testing in which V4 Flash completed 129 of 240 difficult agent-task runs across eight harnesses, suggesting leaderboard strength did not fully translate into consistent real-world automation performance.

Next page →