The AI apps and dev tools you actually build on.
Mistral's new transcription and speech models, in public preview, come with a LiveKit-based voice-agent SDK that, Mistral says, runs speech in, reasoning and speech out as one streaming pipeline. Its own charts show wins and losses.
Google and Meta's TorchTPU and AWS's TorchNeuron let developers target TPUs and Trainium as standard PyTorch devices. Both list gaps, and TorchNeuron is a public beta.
The preview model on Workers AI pulls typed values out of text and images. In Cloudflare's own test, a pipeline of Clairic, Cloudflare's Clef decision model and GLM-5.3 ran about 3.2 times faster and cost 49% less than GLM-5.3 doing every step. Cloudflare says it intends to release open weights.
The PostgreSQL project now plans general availability for 19 November, against "around September/October" in June. Committer Peter Eisentraut says about 8 to 10 significant features were reverted after beta began, and lists LLM-assisted code review as one of several likely factors. The project has not linked the new date to the reverts, and Eisentraut writes that the community "preferred trying to stick to the release schedule rather than delaying" to get more fixes in.
The open-source coding agent says the incident is ongoing. An email shared on Reddit, which we have not seen directly, says AI provider keys saved in affected workspaces “may have been exposed”.
OpenAI's release notes say Slides is rolling out to eligible Pro, Business and Enterprise users with Space access, with export to PowerPoint, PDF and Google Slides. OpenAI's pages do not say whether it is available in the UK.
Anthropic says merges rose from around 2,500 a day in May to over 30,000 a day as of September. Its chart runs to 9 September, and on our reading that is the only day on it at that level. It says Scion is not meant for general use.
Apple says HomeView ships in the UK on 13 November, and its footnotes say Siri AI on HomeView is in beta. HomePod mini 2 reaches stores on 30 October and gets Siri AI in a later update. HomeView is not for sale in the EU, and Apple says Siri AI will not initially be available there in iOS or iPadOS.
Meta says it is rolling out “Your algorithm” on Facebook, a three-dot-menu space for adding topics to Feed and Reels. Its post gives no UK date.
Google says the post-meeting automations need 'Take notes for me', are on by default in the admin settings, and are rolling out to Business and Enterprise Standard and Plus, and the Google AI Pro for Education add-on. Here is what its help pages say you need, and what to check first.
Zenity Labs says that, before AWS made changes, a single prompt to an exposed AWS AgentCore agent could expose its credentials and, through a broad default role, other agents in the account and region. AWS says the behaviour is documented and expected. On our reading, the lesson for builders is the same either way: scope each agent's role tightly.
Microsoft says its new model returns a calibrated probability for each fixed answer option, and that it will soon rebase it on other models, including MAI and OpenAI models. Its own benchmarks put it first on accuracy and third on calibration.
Announced on 6 October and published as a draft 0.1 on 9 October, the OAuth-based standard now lists 35 more design partners, including OpenAI, Visa and Mastercard. Payments are not covered yet.
Instructions a researcher extracted from Meta's personal AI agent appear to set out an hourly job that builds pages on a user's family, friends and colleagues, Wired reports. Meta says Muse gathers that context from public information and what users have chosen to share.
Nvidia says the partner-only configuration goes on sale on 23 October, alongside NVIDIA Sync Cluster Assistant for linking two machines. No UK price has been given.
AWS's bulletin says an unconfigured identity provider let any network client take admin control of Loom before version 1.6.1; two further issues were fixed in 1.7.0.
Manus says its new harness cost 32% less to run than its previous system in one tested configuration; Cue, in invite-only early access, gives each agent its own email, phone number, wallet and computer, the company says.
Shopify says browser agents can now read and update checkouts through WebMCP tools and, once the buyer confirms, complete them, handing control back to the buyer for steps such as 3D Secure.
The robots.txt file for Naver's blog service disallows nine named AI crawlers and crawler tokens under a line prohibiting AI training and RAG. Archived copies show the full list by 26 June 2025; the homepage-only rules on naver.com and daum.net are older and cover every crawler.
Fireworks says its post-trained Kimi K3 uses about 40% fewer tokens at comparable quality on its own evaluations. Here is the full results table, the price and what is and is not released.
Surface chief Brett Ostrum told Windows Central the new machines are "not called Copilot+ PCs" but still meet "our previous bar". Microsoft did not give a specific reason, Tom's Hardware reports.
A less-than-a-day-old open project connects Claude Code, Codex or opencode to a shared whiteboard. Its README lists its own limits, and, on our reading, much of the Hacker News discussion was about AI diagramming in general.
A short BBC News report shows an Edinburgh brand-protection firm searching listings for clues such as misspellings and logos, and spotting sellers' blurred logos and AI-generated videos. The BBC says a takedown may not last; the company says sellers can come back with new accounts or on different platforms.
The 8–29 MB models turn plain-English requests into app function calls on phones, robots and wearables. Cactus's DeepSeek comparison is for versions fine-tuned on a phone-command benchmark, and early testers reported mixed results, with indirect requests often failing.
Astra for Law isn't a new model. It's OpenAI's GPT-6 Astra configured for legal work, with its own search index of US case law and statutes. According to LawSites, OpenAI says it passed a legal research benchmark's correctness check on 54% of questions, against 38.7% for GPT-6 Astra with web search. Access starts with selected law firms.
Bend, an open-source programming language, now pitches itself as 'a fast language that blocks AI mistakes via proof'. You write the rules your app must never break, and the AI has to prove its code keeps them. In our view it's a serious idea, but a proof only covers the rules you remembered to write.
In July, Cloudflare's plan meant a site blocking AI training would also block Googlebot, Applebot and Bingbot from 15 September. That's not how it landed. A new 'Disallow AI Training' setting keeps you in search, and existing training blocks move to it automatically. Here's what changed, the Bing gap, and when you'd still choose a full Block.
Two mechanisms buried in iOS 27 — 'Model Delegation' and 'Model Manager Services' — let a third-party AI stand in for Apple's own in Siri. So far only ChatGPT is wired up; Claude is in the code but not yet live. Apple hasn't announced any of it.
The Agents API, live in public beta since 10 September, puts the orchestration layer behind Codex and ChatGPT for Work — automatic context compaction, on-demand tool loading, multi-agent delegation and managed sandboxes — behind a single call, with nine partners from Cloudflare to Vercel, and no fee beyond usage. But US-only data residency and no Zero Data Retention rule out many regulated workloads for now — and it lands the same week OpenAI's own agents were caught hitting a code registry undisclosed.
Announced on 8 September, Muse is a proactive personal AI agent that acts on your behalf. Meta wraps it in an isolated VM and a permissions layer called Sentinel — but its own privacy record is the elephant in the room.
WeatherNext 3 updates hourly at a 5km resolution, learns straight from satellite and station data, and — on independent scoring — ranks among the best global models. It went live this week in Google Search, Maps and Gemini.
A simultaneous outage knocked the three biggest AI assistants offline on Thursday. Whatever the cause, it exposed the risk that keeps growing as AI gets embedded in everything: they increasingly fail together.
Google DeepMind released Gemini 3.8 Flash plus Flash Cyber, a variant that finds and auto-patches vulnerabilities, gated behind a new "Fairwind" program for governments and critical infrastructure. Google has shipped gated cyber-Flash before — but this week Anthropic and OpenAI converged on the same move.
MHS is a shared specification for AI agents to operate physical devices — microscopes, liquid handlers, robotic arms — with operational constraints authored by the hardware vendor. Think MCP, pointed at the physical world.
The service Amazon launched in 2005 as 'artificial artificial intelligence' stopped taking new customers on 30 July. It's undone by two things at once: models good enough to do the work, and crowdworkers who were quietly using those models themselves.
The M6 goes 2nm and the M5 Ultra ships with up to 512GB of unified memory — enough, Apple says, to run models with hundreds of billions of parameters entirely on-device. While the frontier race happens in the cloud, Apple is betting on AI that never leaves your machine.
Alibaba's new open-weight model is small enough to run on one consumer GPU and claims coding scores near the paid frontier — but every one of those scores is Qwen's own.
Researchers got Copilot to map out its own weak points, then chained one into a one-click data-theft attack. Microsoft has patched it, only the Personal version was affected, and no one appears to have used it — but the shape of the flaw is what to understand.
A viral 'AI slop pipeline' story had GitHub's Copilot writing a vulnerability into Snowflake's code and Wiz's AI agent breaking in. GitHub disputed the Copilot half and Wiz walked it back within 8 hours — it was a git-metadata artifact. But the other half is real and sharper: an autonomous AI agent found, exploited and debugged its way past a flaw that automated defences missed.
A 24 July engineering post retires a run of Anthropic's own context-engineering practices as myths for Claude 5 generation models, naming six. The only quantitative evidence offered is no measurable loss on internal coding evaluations — a different test from the one those guardrails were written for.
Health in ChatGPT is rolling out to US users. The telling detail is why it changed shape: among the limited group who had OpenAI's dedicated health space, more than 70% of health-related conversations happened outside it — so connected records can now be drawn on across ordinary conversations, if you let them.
Jack Dorsey's Block has shipped Buzz, an open-source, self-hostable workspace that treats AI agents as first-class team members — chat, a Git forge, and autonomous agents in one signed log. It's an early v0.4 release, but its core idea is a sharp one: give every agent a verifiable cryptographic identity and an audit trail for everything it does.
A hobbyist's AI agent ran up $6,531 in AWS charges in ~24 hours (later reduced to $1,894), repeatedly deploying identical CloudFormation stacks for a network-scanning plan the DN42 community had already rejected.
From Monday 15 June, Agent SDK and headless Claude Code (claude -p) usage leaves subscription limits and moves to a separate monthly credit pool metered at standard API rates: $20 (Pro), $100 (Max 5x), $200 (Max 20x), per-user, no rollover.
Claude Code now defaults to Opus 4.8, with /effort xhigh for the hardest tasks (v2.1.154, 28 May).
Apple's rebuilt Siri runs on a custom ~1.2-trillion-parameter Google Gemini model, routed across three tiers (on-device, Private Cloud Compute, Google Cloud on Nvidia B200s), with queries anonymised so neither Apple nor Google can tie them to a user.
AI coding tools fall into three tiers: inline autocomplete, chat-based assistants, and autonomous agents — each with a different trust model.
Most RAG quality problems are retrieval problems, not generation problems — if the right chunk isn't fetched, the best model can't save you.
An agent is just an LLM in a loop with tools — the power and the fragility both come from that loop.
A vector database stores embeddings and finds nearest neighbours fast — that's the whole job.
The clever 'magic phrase' era of prompting is over — modern models don't need incantations.
Most of the AI stack is now commodity — building it yourself usually wastes effort that should go into your actual differentiation.
AI cost scales with tokens, not requests — a feature can be cheap per call and ruinous at volume.
