Amodei, Altman and Musk on AI’s Threat to Humanity: Who Will Hit the Brakes?

Dario Amodei’s call to pace frontier AI brings a new assurance question to every delivery firm using Claude, ChatGPT or tools built on them. Anthropic and OpenAI now share a public position that rapid capability gains need stronger safety confidence.

This week’s picks take that question onto the project floor: machine-checked maths, defensible site evidence, exposed account-control failures and an agent editing Revit. Together, they show why AI governance belongs alongside programme controls, information management and assurance.

We're doing another open AMA — Thursday 17 September, 5.15pm UK, live on LinkedIn. No slides, no script, bring the awkward questions.

Register here

Anthropic chief executive Dario Amodei has proposed that frontier labs slow capability gains until they can show that the risks are under control. His first commitment is an embedded external evaluator with employee-like access and a right to publish findings. Sam Altman said OpenAI would make the same commitment.

This is neither a pause nor a regulatory requirement. No lab has named a model release it will delay, and the later steps in Amodei’s plan require government coordination. The practical signal is still significant. Buyers should expect more scrutiny of model versions, safety evidence and the consequences of a vendor pause.

For delivery firms, start with a model register. Record the models inside enterprise seats, plugins and project integrations, then identify the fallback if a supplier changes a model or limits an integration mid-programme.

Read the full story

OpenAI says an internal system used around 10,000 concurrent agents and about 88 hours to produce a Navier-Stokes proof, then formalised it in Lean. The Clay Mathematics Institute has not accepted the result, so it remains OpenAI’s claim. The broader method matters: large-scale candidate generation followed by machine verification could change the economics of engineering calculations, programme logic and cost checks.

Read More →

Apple’s Reference Image feature creates a signed, unalterable capture from an iPhone 18 Pro or Pro Max. A progress or defect photograph could therefore carry stronger evidence that its pixels were unchanged after capture. The control has limits: it is opt-in, requires Pro hardware, and capture is unavailable at launch in China and the EU. It also needs field software that preserves the provenance.

Read More →

Anthropic’s September threat report describes misuse across cyber operations, surveillance, biological work, weapons development and model distillation. The operational lesson is closer to home. Stolen customer API keys were used as compute in further attacks, which makes keys in scripts, plugins and proof-of-concept integrations a project-control issue. Treat them as privileged credentials with a named owner, vaulting and rapid rotation.

Read More →

A hands-on test of Claude, MCP and pyRevit found that an agent could create walls, ceilings and worksets in a live Revit model. It struggled with stairs and produced a recognisable but undeliverable Farnsworth House rebuild. The open-source route has write access that Autodesk’s public Revit MCP server does not. The immediate task is to define who may enable it, on which models, and how changes are reviewed.

Read More →

Trimble’s 8 September acquisition extends real-time GNSS correction coverage into Japan through Terasat’s virtual reference station network, a foundation for machine control, surveying and reality capture.

The DASH experiment tested an interface that lets vendor applications plug into a government-owned orchestration layer. For major programmes, that is a useful pattern for reducing platform dependency while retaining integration control.

SmartCraft bought the cloud cost-control platform on 3 September. Its £8.45 million enterprise value and 60% annual recurring revenue growth point to continuing consolidation in software for smaller contractors’ live budget tracking.

NVIDIA and Australian infrastructure partners say they are targeting up to 2 GW of AI capacity by 2027. The announcement adds a tangible pipeline signal for power, cooling, shell and commissioning capacity in the region.

ProjectLibre Cloud AI 2.0 combines prompt-assisted schedule creation with critical-path planning, earned value, Monte Carlo analysis and DCMA 14-point checks. It is positioned for teams preparing for Microsoft Project Online’s 30 September retirement, with imports from Microsoft Project and Primavera P6. The published price is $9.99 per user a month when billed annually for teams of three or more. Trial an exported, non-live schedule first and compare logic, resources and baselines before migrating a portfolio.

Learn more →

From 11 September, manufacturers of products with digital elements must report actively exploited vulnerabilities and severe incidents through ENISA’s Single Reporting Platform. The early warning is due within 24 hours of awareness, followed by a full notification within 72 hours. The rule is aimed at manufacturers, rather than firms simply using AI. Contractors and clients specifying connected building systems should ask suppliers who owns this reporting duty and how escalation will work.

Source →

New Scottish Government guidance says people remain responsible for decisions and warns against using AI where decisions cannot be explained or properly reviewed. Teams delivering for Scottish public bodies should expect human oversight, transparency and documented reasoning to feature in briefs and assurance conversations.

Source →

Google Cloud’s live technical series begins today and covers the Gemini Enterprise Agent Platform, retrieval-augmented generation, agent state and deployment. It is aimed at founders, CTOs and lead engineers, so it is more technical than an introductory tool session. A digital or innovation lead working in a Google Workspace firm could use the first session to compare a prototype workflow with the controls needed for a production deployment.

Source →

Dr. Alex Leithard discusses the landscape of AI and its applications. [Podcast]

Around 30 humanoid robots appeared outside Poland’s digital ministry to call for stronger AI regulation. The organiser, Democratism, used the demonstration to raise concerns about work displacement. The organiser also owns a robotics company, so the spectacle carried a commercial interest alongside its political message. In a week dominated by calls to pace AI, it made an unusually literal closing image.

Source →

Take a set of non-sensitive progress-meeting notes and use Claude Design to create a one-page programme road map. Set the project’s design system, choose the road-map template, upload the notes and specify the milestones or workstreams to show. Check every date, dependency and claim against the source record before sharing. The tool is included with paid Claude plans starting at $20 a month, according to the walkthrough.

Source →

The summit starts seven calendar days after this edition and brings together more than 200 speakers and 10,000-plus AI builders and leaders. Its European focus makes it timely for readers assessing sovereign AI, implementation partners and the governance questions raised by this week’s feature. Tickets were available when checked.

Details →

Till next time,

_{All content reflects our personal views and is not intended as professional advice or to represent any organisation.}