Macro close-up of keyboard keys with a narrow strip in sharp focus

Software hub · Updated weekly

Models, Apps and Updates

What changed in AI and everyday software, and which updates deserve your attention this week.

01

The Short Version

Three threads run through software news right now. First, AI models are splitting into two camps: closed models reached through an API, and open-weight models you can download and run yourself. Second, AI features are moving onto the device, powered by dedicated neural processing units. Third, the boring part still matters most: security updates, end-of-support dates and how you sign in.

This hub tracks all three. It is not a product ranking. It explains what each change means so you can decide whether to act.

02

Models: Open, Closed and In Between

Closed models, such as OpenAI's GPT family, Google's Gemini and Anthropic's Claude, are reached through apps and APIs. The provider handles hosting, safety tuning and updates. You get convenience and usually the strongest results on general tasks, but your data travels to someone else's servers under their terms.

Open-weight models, such as Meta's Llama and models from Mistral AI, publish the trained weights. Companies can run them on their own hardware, fine-tune them on internal documents and keep data in-house. The licences vary, so "open" does not always mean free for every use. Read the licence before building a product on one.

Small models are the quiet trend. A model with a few billion parameters can now summarise, translate and draft on a laptop, which cuts cost and latency.

Shallow-focus close-up of a graphics card heatsink and fan blades
Model choice is increasingly a hardware choice too.

Why it mattersThe right model depends on where your data is allowed to go, not just on benchmark scores.

03

Apps and On-Device AI

Microsoft introduced Copilot+ PCs in May 2024 with a requirement for an NPU capable of more than 40 trillion operations per second. Apple Intelligence runs many tasks locally on recent iPhones, iPads and Macs, and sends heavier requests to Apple's Private Cloud Compute. Google's Gemini Nano runs on selected Android phones.

Local processing has two practical benefits: features keep working offline, and less personal data leaves the device. The catch is hardware: older devices simply do not get these features, and vendors decide which tasks stay local.

Why it mattersWhen buying a laptop or phone this year, AI support is now a real line on the spec sheet.

04

Updates You Should Not Skip

  1. Windows 10 Extended Security Updates

    The consumer programme ends on 13 October 2026. Check whether your PC meets Windows 11 requirements, including TPM 2.0.

  2. Browser auto-updates

    Chrome, Firefox, Edge and Safari patch actively exploited flaws several times a year. Restart the browser when prompted.

  3. Python and Linux release cycles

    Python releases a new version every October under PEP 602, with roughly five years of support per version. The Linux kernel ships a new mainline release about every nine to ten weeks.

  4. Passkeys instead of passwords

    Turn on passkeys for your email account first. It is the key to resetting everything else.

05

Key Terms

TermPlain meaningWhere you see it
Open weightsA model whose trained parameters are published for download.Llama, Mistral
NPUA chip block designed to run AI tasks efficiently.Copilot+ PCs, recent phones
TOPSTrillions of operations per second; a rough NPU speed figure.Laptop spec sheets
PasskeyA login key stored on your device, unlocked by biometrics or PIN.Google, Apple, Microsoft accounts
ESUExtended Security Updates; paid or limited patches after end of support.Windows 10
Fine-tuningFurther training a model on your own examples.Enterprise AI tools
06

Related Hubs

  • Devices worth a look

    Chips, laptops and routers that run all this software.

    Open
  • Regulation and the big platforms

    The AI Act and other rules shaping software.

    Open