Fresh signals

Cloudflare has released Clef and Clef-flash, two decision models designed to return structured answers with probabilities rather than open-ended prose. For developers building AI agents, the useful distinction is between deciding where a task should go and generating whatever comes next.

Discussion spark: Should routine agent decisions move to specialised classification models, or is one general-purpose model easier to test and maintain?

Read full story Join the WittyWires discussion

Carlyle co-president Mark Jenkins warns that lending tied to the AI boom could be vulnerable to concentration risk, despite the money flowing into different companies and projects, Bloomberg reports.

Jenkins also points to complex data-centre contracts as a potential weak spot. His comparison is to the “SaaS apocalypse”, when app…

Read more

Cyera’s Ehud Halamish argues that employee AI agents should run in managed cloud environments rather than on individual laptops. His practical case is central control over sandboxes, data permissions, outbound network traffic and software updates.

That gives organisations a way to restrict what an agent can reach and where it can s…

Read more

Elon Musk says United Airlines and American Airlines have Starlink connectivity, a notable addition to the airline options associated with the satellite internet service.

The post does not spell out when the service starts, which aircraft are covered or whether it is available to passengers across either airline’s fleet. For t…

Read more

Anthropic is urging Australia to consider allowing AI companies to train on copyrighted material unless rights holders opt out. The proposal follows the company conceding it will not secure a blanket copyright exemption, The Guardian reports.

Australia’s public broadcasters, the ABC and SBS, are calling for stricter rules to c…

Read more

Cyera’s Assaf Morag argues that AI-assisted cybercrime makes one defensive question more urgent: what sensitive data can a compromised account or token actually reach? His analysis recommends checking exposed credentials, broad application permissions and bulk data access, rather than treating a successful login as the end of the security check.

Discussion spark: Should security teams make the sensitive data reachable by each identity their main audit priority, or does that risk diverting effort from preventing credential theft in the first place?

Read full story Join the WittyWires discussion

Cyera’s Assaf Morag argues that security teams should review what identities can read, not just how they sign in. His directory-security analysis extends that warning to AI agents with persistent access across applications and cloud resources.

The practical checks: review directory visibility for ordinary accounts, examine M…

Read more

Barclays is expanding its use of Anthropic’s Claude across the bank, with tools already assisting staff and handling about 120,000 Global Markets emails a day. The deal also sets a target for Claude Code to reach half of Barclays’ developers by the end of 2026.

Discussion spark: For a bank deploying AI at this scale, should success be judged mainly by faster workflows, or by independently reported improvements in customer outcomes and error rates?

Read full story Join the WittyWires discussion

Hugging Face’s PEFT release 0.21.2 fixes an issue that stopped encoder-decoder models working with Transformers 5.18 or later, according to the release notes.

For developers using both tools, the fix may clear a compatibility snag after upgrading Transformers. It is a targeted repair, not a new feature suite; sometimes the g…

Read more

AMD has posted five consecutive record quarters for server CPU revenue, but 247WallSt argues that cloud companies’ home-grown Arm chips could limit how far the run goes. Its analysis points to NVIDIA’s Vera CPU and Amazon’s Graviton as signs that hyperscalers are building alternatives to AMD’s x86 processors.

That is a competi…

Read more

Update

Computer Weekly reports that the planned UAE government deployment centres on Microsoft’s MDASH AI-powered cyber security platform. The organisations say it is intended to help security teams find vulnerabilities, prioritise risks and strengthen security through automated analysis and AI models.

The rollout is planned in p…

Read more

AI boom could trigger market shocks, Bank of England boss warns | BBC News

The video reports Bank of England governor Andrew Bailey’s warning that investment in AI could trigger financial market shocks. It describes how spending and lending to AI firms has helped drive some valuations into the multi-trillion-dollar range, while B…

Read more

Watch: https://www.youtube.com/watch?v=FpEhAlmVGzA

Brian Chesky says the world needs an AI-native operating system, while discussing how Airbnb could become more agent-friendly.

The TechCrunch interview also covers consumer AI. The idea is a provocation, not yet a product specification: it leaves open what this operating system would do that today’s services cannot. That makes it a…

Read more

AI agents should get permission to perform specific actions for limited periods, rather than inherit broad access to an app or account, says Cisco identity executive Matt Caulfield. He argues that organisations also need to discover which agents are running and keep an audit trail of their actions.

Caulfield’s advice appeared in a C…

Read more

Ai2 has released Olmo-core 3, an open training framework designed to scale mixture-of-experts models towards the trillion-parameter range. Its headline result is a reported 2.7× throughput gain over Ai2’s earlier implementation in a preliminary eight-GPU test, alongside a stack other researchers can use and adapt.

Discussion spark: For open AI development, which matters more: publishing infrastructure that others can run and adapt, or waiting for independent results showing it improves trained models?

Read full story Join the WittyWires discussion

Salesforce has published customer accounts that put some numbers and operating rules behind agentic AI deployments. The useful message is less “switch it on” than “measure what it does, listen to customers and keep controls close”.

Discussion spark: Should customer-service agents be allowed to resolve routine enquiries autonomously once they pass a company’s tests, or should a person review every AI-handled case?

Read full story Join the WittyWires discussion

At a San Francisco hackathon on 26 September, 55 teams built working projects around Jev, a model from TypeSafe AI, using GMI Cloud’s inference service for text, image, video and voice generation. The useful idea was to give quick, typed decisions to one model and leave the actual writing, images and audio to others.

Discussion spark: For an AI agent, should routine branching decisions go to a fast, constrained model while generative models handle the output, or does splitting the work create more failure points than it removes?

Read full story Join the WittyWires discussion

Epoch AI says Claude Opus 5.5 has taken the top spot on its Capabilities Index, scoring 167 and narrowly leading GPT-6 Astra.

Epoch AI also says Claude Sonnet 5.5 roughly matches Claude Fable 5.1, with a score of 165. The ranking offers a timely snapshot of how the project rates these models, though the supplied figures do not…

Read more

Cerebras had $25.4bn in remaining performance obligations at the end of June, according to a Yahoo Finance article published on 30 September. That is work customers have signed for but the company has not yet delivered, and the article says the commitments include no business from AWS or another major cloud provider.

The same…

Read more

Miami-based Doxx.net has raised a $38 million Series A led by Andreessen Horowitz as its networking platform enters open beta, Axios reports.

The platform is designed for serverless peer-to-peer calls, messaging and file transfers for people and AI agents. That puts it in the plumbing layer of the agent economy: less flashy than a…

Read more