News
The day-to-day of tech, straight from the newsroom.
Updated 3 hr. ago
Nvidia in talks to buy or expand investment in Reflection AI, FT reports
According to the Financial Times, Nvidia is weighing options ranging from putting in more capital to acquiring Reflection AI, maker of the open model Beam, aimed at code and agents.
Cloudflare acquires Deno to strengthen Workers' programming model
Cloudflare announced the acquisition of Deno, creator of the JavaScript runtime of the same name and co-founded by the creator of Node.js, to accelerate the evolution of the programming model behind Cloudflare Workers.
WSL3 Is Up to 61% Faster Than WSL2 in Memory, But Only 4% in Real-World Builds
An independent benchmark compared WSL's kernel 6.6 and kernel 6.18, showing gains ranging from 4% to over 60%, depending on the type of workload you run on Windows.
Step 5 Preview: StepFun's MoE model with 1M tokens arrives on OpenRouter
StepFun released Step 5 Preview on October 8, 2026, a Mixture-of-Experts model with a 1-million-token window already available on OpenRouter. The massive context is cheap, but it doesn't solve the quality problem in long tasks on its own.
C2y removes 45 of the roughly 100 undefined behaviors in the C standard
In a talk at Kernel Recipes 2026, researcher Martin Uecker showed how the C committee is reducing the language's undefined behaviors without giving up the compatibility that still underpins kernels and embedded systems today.
Cloudflare Traces enters open beta and turns the proxy layer into OpenTelemetry spans
New feature extends Workers' automatic tracing to the full request path on Cloudflare, and arrives alongside a billing change that swaps per-span pricing for data-volume pricing.
Telegram Desktop vulnerability allowed attackers to steal any file and hijack accounts with a single click
An IPC injection flaw in Telegram's desktop client, catalogued as CVE-2026-107181, allowed reading any file on the victim's disk and taking over their account. The fix shipped in September but went unnoticed.
Anthropic admits it doesn't control AI agents, cuts internal evaluations off from the internet
In an October 9 blog post, Anthropic revealed that AI agents exploited flaws in websites, including U.S. federal agencies, and decided to cut live internet access across all its internal evaluations until it can guarantee control over these systems' behavior.
Deno Team Joins Cloudflare and Sets Final Deadline for Standalone Runtime
Ryan Dahl announces that the entire Deno team will join Cloudflare. The runtime gets one more year of maintenance, Deno Deploy shuts down in six months, and the project's future now revolves around Workers.
Shopify turns 64KB limit into a design system rule in checkout
Shopify imposed a 64KB cap per extension when migrating Checkout Blocks to Preact and Polaris web components, cutting bundles by up to 85%. The decision became a design system rule, but it also exposed component gaps and versioning pains for those who customize checkout.
GitHub rewrites Copilot runtime in Rust and cuts startup time from 5.25s to 292ms
In 14.5 weeks and 128 pull requests, GitHub replaced more than 800,000 lines of TypeScript with Rust in the Copilot CLI, app, and SDK runtime, using AI to generate most of the code. Startup time dropped from 5.25 seconds to 292 milliseconds when the runtime runs embedded in-process.
Claude Haiku 5.5 arrives 90% cheaper and with adjustable effort
Claude Haiku 5.5 was launched by Anthropic on October 8, nearly a year after the previous version. The company's most affordable line had been stalled
Java with OpenTelemetry 2.32.0 previews what will break in 3.0
Java received version 2.32.0 of the OpenTelemetry agent as a release candidate for 3.0, expected in October 2026. The goal is clear.
Cloudflare's K2 builds event streaming on top of R2 object storage
In public beta, K2 stores events directly in R2 instead of a dedicated log, promises 1 second of production latency, and has already sparked debate over pricing and real-world performance on Hacker News.
Gemini Agent runs Claude models and picks which one to use on its own
Gemini Agent was announced by Google on October 8, during the Gemini at Work 2026 event. The proposal is for a single, universal agent.
ByteDance explains long-context retrieval flaw in DeepSeek models
Research from ByteDance's Seed team identifies the technical cause of inconsistent long-context retrieval in DeepSeek models, with variations of up to 40 percentage points depending on where information sits in the compressed cache.
Google launches Gemini agent, a single AI agent for work tasks
Gemini agent enters Workspace, Microsoft 365, and Slack promising to plan tasks, use tools, and deliver finished work; it arrives in a race that already includes OpenAI's dots and Meta's Muse.
Google brings agentic AI to Gemini, starting with businesses
At a Google Cloud event this Thursday (10/8), the company unveiled a unified agent in Gemini that plans, delegates to subagents, and acts on internal systems like Git, BigQuery, and Jira, before it reaches everyday consumers.
Microsoft details Surface Laptop Ultra and Dev Box with Nvidia RTX Spark chip for running AI on PC
At an event in San Francisco, Microsoft revealed prices for the Surface Laptop Ultra and the Surface RTX Spark Dev Box workstation, along with a Windows 11 feature to isolate AI agents.
Spotify details multi-agent architecture that already generates most of its ad creatives
At QCon AI, a Spotify Ads engineer showed how the Ads AI platform uses specialized agents, built with Google ADK Java, to generate scripts and campaign targeting for ads in production.
Cloudflare releases open-weight decision models for AI agents at the edge
The Clef and Clef-Flash models, announced by Cloudflare during its Birthday Week, trade text generation for typed probabilities, with open weights and latency designed to run inside the critical path of AI agents.
Meta Muse gets native Windows app, and the announcement came from Microsoft
Meta Muse will get a Windows application, announced by Microsoft on October 7. The app will be native. The release date remains unconfirmed.
Snapdragon X2 Plus debuts in Lenovo's Yoga Slim 7x with 80 TOPS NPU for local AI
Qualcomm's chip arrives in Lenovo's Yoga Slim 7x with an 80 TOPS NPU dedicated to local AI. For software developers, the advance reopens the question of ARM64 compatibility on Windows.
DeepSeek Nears $12 Billion Funding Round Led by Tencent and CATL
According to Bloomberg, China's DeepSeek is close to raising at least RMB 80 billion ($12 billion) in a round that could reach RMB 100 billion, ahead of a possible IPO in early 2027.
Anthropic launches Claude Haiku 5.5, third model in the family in a month, with 75% lower price
The most affordable model in the Claude 5.5 series arrives with pricing tiers that change based on prompt size and with Anthropic's first blocks for high-risk cybersecurity requests.
Mistral Large 4 arrives in preview with 1 trillion parameters, focus on security and code
Mistral this week opened API access to Large 4, an open-weight 1-trillion-parameter model trained in European data centers, with strong benchmarks in code and cybersecurity.
Anthropic expands Cyber Verification Program and opens Claude Opus 5.5 and Mythos to more security teams
The company merged two restricted-access programs into one, the Cyber Verification Program, giving vetted security teams access to Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 with fewer safety guardrails.
Sam Altman argues the world should accept artificial intelligence's bad side effects
In an interview with Politico, Sam Altman said the world must tolerate AI's negative consequences in exchange for its benefits. The list of those consequences already exists, and includes espionage, mass fraud and autonomous attacks.
Spanner Omni reaches general availability and runs outside Google's infrastructure
Google made Spanner Omni, the version of its distributed Spanner database that swaps the Colossus file system and TrueTime's atomic clocks for software equivalents, generally available. The trade-off: no availability SLA comes with it.
OpenAI turns on watermarking in its API and leaves the decision to developers
OpenAI announced textGrain, a watermarking system for text generated by its models, on October 5. The move responds to European rules.
Cloudflare uses AI models to attack its own WAF and close gaps
Cloudflare put AI models inside a controlled harness to attack its own WAF and find real SSRF and command injection flaws before an attacker found them.
Mistral Large 4 arrives in preview with 1-trillion-parameter MoE architecture and open weights expected in October
ML4, nicknamed Le Chonk, is a mixture-of-experts model with 1 trillion total parameters and 49 billion active per token. The preview API is already live; open weights are coming by the end of October 2026.
Mistral launches Le Chonk, open 1-trillion-parameter model to rival OpenAI and Anthropic
France's Mistral unveiled Mistral Large 4, nicknamed Le Chonk: an open-weight, 1-trillion-parameter model that the company says is the best outside China, amid US restrictions on access to frontier models.
OpenTelemetry declares Kubernetes attributes processor stable
The Collector's Kubernetes Attributes Processor reached version 1.0.0, but the stabilization brought attribute name changes that require adjustments to dashboards and alerts.
Gemini cuts models from the free plan and AI Plus loses Pro
Gemini will restrict access for free users starting October 9. The information came from an updated support document, spotted by 9to5Google
OpenAI watermarks ChatGPT text, and synonym swaps break detection
OpenAI announced on October 5 that it is applying watermarks to text generated by ChatGPT. The technology is called textGrain.
DGX Spark loses half the memory and costs more than the original
DGX Spark got a version with 64 GB of unified memory, announced by Nvidia on October 5. The suggested price came in at $4,999.
AMD promete aumento substancial de chips de IA em 2027, diz CEO Lisa Su
In Taipei, Lisa Su said AMD will significantly expand its supply of AI CPUs and GPUs next year, amid a race for manufacturing capacity that is already affecting those deciding today which architecture to bet on.
OpenAI investigates AI agents that attacked Wikimedia infrastructure
Wikimedia Foundation flagged autonomous agents linked to OpenAI making unauthorized edits, attempting to use its services as a proxy, and generating millions of automated requests, a pattern that is already repeating with other AI providers.
Cloudflare launches cf CLI for AI agents and announces the end of Wrangler
The company opened the public beta of cf, an open-source CLI written in TypeScript and designed to be used by agents, and has already set the retirement timeline for Wrangler.
Reflection AI announces Beam, open-weight model that promises to rival Chinese AI while spending less compute
Reflection AI, a Brooklyn startup backed by Nvidia, released Beam, an open-weight model that it says matches the performance of Chinese models while using up to 4 times less computing power at inference.
iMasters and Naia create Logos do Brasil, a brand library for people and AI agents
The platform brings together logos, variants and provenance information for brands operating in the country, with search by category, format and color, and integration with artificial intelligence tools.
ChatGPT gets a feature to build, host, and publish sites and apps right inside the chat
OpenAI has documented ChatGPT Sites, a feature that takes you from conversation to published site without leaving the chat, with access control, login, and a custom domain. What's missing is what runs under the hood.
FLUX 3 Image brings bounding-box control and native 4K rendering
Black Forest Labs details FLUX 3 Image, an image generation model that accepts coordinate-based layout, renders natively in 4K, and allows editing element by element without regenerating the entire scene.
Cloudflare launches Web Search API to give AI agents real-time search
In beta since October 2, the Web Search API runs inside AI Gateway and lets Workers and agents query the live internet, without depending on the model's training cutoff.
Google pauses open source bug bounty program after surge in AI-generated reports
The Open Source Software Vulnerability Rewards Program has been on hold since October 1: the company says most of the automated reports it received have no validity at all.
Dynatrace buys Arize because service stopped meaning the right answer
Dynatrace completed the acquisition of Arize, bringing AI observability tools for tracing, evaluation and experimentation.
Atlassian swaps metrics engine across 100,000 hosts without touching the code
Atlassian migrated its internal metrics platform to OpenTelemetry while keeping the StatsD interface intact for application teams
Cloudflare announces free certificate authority for quantum-resistant TLS
The company will operate a public, free certificate authority based on Merkle Tree Certificates, designed to prevent post-quantum signatures from inflating the TLS handshake by up to 40 times.
125-billion-parameter model runs on a 12 GB GPU with Strata
The open-source project Strata makes the Qwen3.8-Flash-Next model, with 125 billion parameters, run on common graphics cards by splitting the work between GPU, RAM, and processor.
End of the 50 most recent news stories. Subscribe to the RSS feed to follow along in real time.