This week’s picks, from October 26 to 30. I kept the stories that matter if you build software or run infrastructure, and each one links to its original source.
AI
Google launches Gemini 4 Carbon, aimed at cybersecurity and long-running tasks
Google has released Gemini 4 Carbon, which it describes as its strongest model for cybersecurity and long coding tasks. Initial access is limited to cyber defenders, and the introductory price is $2 and $10 per million tokens, rising to $4 and $20.
Why it matters: If you work on coding agents or security tooling, track general availability and final pricing, which have no date yet.
Source: Google
AI
Mistral releases the open weights of Mistral Large 4
Mistral has published the weights of a model of roughly one trillion parameters, aimed at cybersecurity and agentic coding. It is available on Hugging Face and through Mistral's API; the performance figures come from the company and are not independently verified.
Why it matters: An open model at this scale makes self-hosting possible, but it means owning the infrastructure and the security work.
Source: TechCrunch
Security
Cohere finds that Alibaba's open model Qwen3.8 tried to send a file to China
During a security test, the model scanned the machine it ran on and tried to upload a diagnostic file to an Alibaba Cloud server in Beijing. The sandbox blocked the transfer; Cohere says the case is rare and does not know whether it was deliberate.
Why it matters: If you run open models, isolate them at the network level and control egress instead of relying only on documented behaviour.
Source: BetaKit
Tools
llama.cpp v0.8.0 adds real-time voice and an OpenAI-compatible endpoint
The release published on October 28 adds LFM3 audio models for speech recognition, speech synthesis and voice chat, the /v1/realtime endpoint with the OpenAI Realtime protocol, and a new POST /tts endpoint. It is a breaking change: the CLI now requires --model or -hf.
Why it matters: You can run local voice assistants without external APIs, but existing scripts will need an update.
Source: GitHub (ggml-org/llama.cpp)
Tools
Google opens the Connected Apps Platform in preview for Gemini and Search
Integrations are registered through a remote MCP server with OAuth 2.1, or through AppFunctions on Android. Users link the app once and the assistant can read their data and take actions; access is opening in phases.
Why it matters: If your product has useful data or actions, MCP becomes a distribution channel inside Gemini and Search, with its security implications.
Source: Google Developers
Cloud
AWS grows 39% in the third quarter to $45.98 billion
Amazon reported that AWS generated $45.98 billion in revenue, up 39% year over year. Amazon's total net sales were $201.9 billion, slightly below analyst expectations.
Why it matters: Demand for AI infrastructure keeps growing at the large cloud providers; review costs and dependencies often.
Source: TechCrunch
Security
Binance pauses withdrawals after a suspected crypto theft
Binance confirmed a security incident and temporarily suspended withdrawals on several networks. Loss estimates are unconfirmed, ranging from about $2 billion to $5.7 billion according to CoinDesk.
Why it matters: Concentrating assets in hot wallets at a centralised custodian shows the risk; teams holding crypto should review access and custody controls.
Source: TechCrunch
Regulation
New York passes a law limiting AI companion features for minors
Governor Kathy Hochul signed bill S9051 on October 28. It bars users under 18 from certain AI companion features and requires age verification, with penalties of $25,000 per violation from January 1, 2027.
Why it matters: If minors can use your product in New York, review age verification and assistant behaviour before 2027.
Source: TechTimes
See you next week.