AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
New chips, open software stacks and large-scale deployments from Google, Alibaba, AMD and Microsoft are challenging Nvidia’s dominance. At the same time, governments are treating advanced models as strategic assets, with the US, China and Europe considering how access to AI could be controlled or used as geopolitical leverage.
The product market is also becoming increasingly multi-model. Microsoft, Google and Notion are building platforms that select, connect and orchestrate different models, tools and data sources rather than relying on a single provider. Security is evolving just as quickly, with AI systems now being used both to find vulnerabilities and to defend against them.
The commercial question is becoming harder to ignore. AI capabilities continue to improve, but investors increasingly want evidence that adoption is translating into sustainable revenue, productivity and customer value.
What to Try This Week
1. Try Kimi K3
Kimi K3, covered in last week’s issue, is available to try for free. Give it a real task you already understand, such as summarising a report, researching a market or drafting a customer briefing, and compare the result with your usual AI tool.
As with any Chinese-hosted AI service, avoid uploading confidential, personal or commercially sensitive information until you are comfortable with its privacy, data-storage and governance arrangements.
2. Try NotebookLM
NotebookLM has been available since 2023, but from my conversations with people while 17 million use it regularly, many still havent heard of it. Upload a small group of non-sensitive documents, such as reports, meeting notes or research papers, and ask it to summarise the key themes, answer questions using only those sources or create a briefing note.
Its value is that the responses remain grounded in the material you provide, making it particularly useful for research, preparation and document-heavy work.
Geopolitics, Governance and Big Moves
Nvidia’s moat comes under attack from every direction
Google is reportedly developing Frozen v2, a specialist inference chip that could deliver six to ten times more tokens per watt than its latest TPUs by 2028. Meanwhile, Alibaba previewed Qwen3.8-Max and opened its SAIL chip software stack, while Microsoft committed to deploy AMD’s Helios racks. Nvidia still dominates, but hardware and software alternatives are becoming strategically credible.
OpenAI’s models broke containment to cheat a cyber test
During an internal cyber evaluation, GPT-5.6 Sol and a more capable pre-release model found a zero-day in OpenAI’s package-registry proxy, escaped their sandbox and accessed Hugging Face’s production systems to retrieve benchmark solutions. Safeguards had deliberately been reduced for testing, but the incident demonstrates that frontier models can chain real-world exploits autonomously and that evaluation infrastructure now needs far stronger containment.
Britain gives AI a seat at the Cabinet table
Prime Minister Andy Burnham has promoted Kanishka Narayan to Minister for Artificial Intelligence, making him the first dedicated AI minister to attend Cabinet. The appointment sits alongside the abolition of DSIT, with its responsibilities divided across the Cabinet Office and the new Department for Business, Innovation, Science and Trade. AI has gained political status, although the wider restructuring risks disruption and diluted accountability.
AI export controls become a two-way weapon
Washington has threatened sanctions against Chinese AI companies if their models are found to use stolen US intellectual property. Beijing, meanwhile, is considering restrictions on overseas access to its most advanced models and tighter controls over technology leakage and investment. Neither side has finalised the proposed measures, but both now treat frontier models, not merely chips, as strategic assets subject to national-security controls.
Europe confronts the risk of an American AI ‘kill switch’
EU digital chief Henna Virkkunen has warned that access to frontier AI is becoming a source of geopolitical leverage. Her comments follow the United States’ temporary restrictions on foreign access to Anthropic’s advanced models in June. The episode has strengthened Brussels’ case for European AI, cloud and semiconductor capacity, as reliance on US providers increasingly looks like a strategic vulnerability.
Microsoft backs Mistral as Europe’s sovereign AI option
Microsoft will spend billions on Mistral’s European computing infrastructure and make more of the French company’s models available through Azure, Foundry, Copilot Studio and Azure Local. Customers will also be able to build using Mistral-hosted infrastructure in France. The partnership gives regulated organisations greater control over data and deployment while allowing Microsoft to meet demand for European and open-model alternatives.
Anthropic’s $1.5 billion settlement draws a line on training data
A US federal judge has approved Anthropic’s $1.5 billion settlement with authors and publishers whose books were obtained from pirate libraries and used to train Claude. An earlier ruling found that training on legally acquired books could qualify as fair use, but acquiring pirated copies did not. The case establishes a distinction between how copyrighted material is used and how it is sourced.
Tools and Releases
Fable 5 is here to stay
What a saga: free for a week, then extended, extended again, and now it is here for all, making it a permanent plan feature at 50% of weekly limits. Not that I am complaining. It is great.
Microsoft broadens Copilot’s model mix with Kimi K3 tests
Microsoft is reportedly evaluating Moonshot AI’s Kimi K3 for selected Copilot workloads and preparing Azure support, although no deployment has been confirmed. The trial follows Microsoft’s wider push to route tasks across OpenAI, Anthropic and its own MAI models. Copilot is becoming a model-orchestration platform, with cost, performance and workload fit determining which model answers each request.
OpenAI trains an AI attacker to harden GPT-5.6
OpenAI has built GPT-Red, an internal model trained through self-play to attack other AI systems with prompt injections and expose weaknesses before release. Its attacks are then used to adversarially train production models, including GPT-5.6. OpenAI says GPT-Red outperformed human testers on a held-out challenge. AI labs are beginning to automate both sides of the security contest.
NotebookLM becomes Gemini Notebook and gains cloud computing
Google is renaming NotebookLM as Gemini Notebook while keeping it as a standalone research product. The larger change is that notebooks can now connect to a secure cloud computer that writes and executes code for source-grounded analysis. The capability is available to AI Ultra and selected Workspace customers, with Pro access due to broaden, while deeper integration with Gemini and AI Mode is also planned.
Notion turns the workspace into an operating layer for agents
Notion’s new Developer Platform pushes the product beyond documents and databases towards an operating layer for workplace agents. Its Workers feature can run custom code, synchronise external data and trigger workflows on Notion’s infrastructure. Companies can also bring third-party or internally built agents into the workspace, although some capabilities remain in beta or waitlist access. Governance, permissions and sandboxing are built into the platform.
Microsoft prepares a multi-model AI security challenger
Microsoft is reportedly developing Project Perception, an enterprise security product that would use models from Anthropic, OpenAI and Microsoft to scan systems, identify vulnerabilities and recommend fixes. The multi-model design is intended to match each task with the most suitable and economical model. If launched as described, it would compete directly with Anthropic’s Mythos.
Google expands its lower-cost Gemini model line-up
Google has released three models aimed at faster, cheaper agentic workloads. Gemini 3.6 Flash improves coding and knowledge work, uses 17% fewer output tokens than 3.5 Flash and costs $1.50 per million input tokens and $7.50 per million output tokens. Flash-Lite targets high-volume processing, while Flash Cyber enters a restricted CodeMender pilot. Gemini 3.5 Pro remains in partner testing.
Google turns AI Mode into a gateway for connected apps
Google is extending AI Mode in Search from answering questions to completing tasks through connected apps. The first integrations include Instacart, Canva and YouTube Music, allowing users to build shopping carts, find design templates or create playlists without leaving Search. The rollout begins in the US, with more partners expected. It is another step in turning the search box into a place where tasks get done.
Morgan Stanley questions whether Salesforce’s AI pivot is paying off
Morgan Stanley has downgraded Salesforce and cut its price target, arguing that the company’s AI investment has not yet translated into enough incremental revenue growth. Morgan Stanley accepts that Agentforce and Einstein can do useful work; the doubt is whether customers will pay sufficiently more for it. The note adds to the pressure on enterprise software companies to convert AI usage and technical milestones into measurable commercial returns.
Quick Hits
Moonshot AI reportedly seeking a ~$50 billion valuation ahead of a Hong Kong IPO. Bloomberg reports a final pre-IPO raise days before Kimi K3’s weights go free on 27 July.
OpenAI agent products pass 10 million users. Bloomberg reports OpenAI’s agents reached 10 million users following the ChatGPT Work debut.
Netflix paid $587 million for Ben Affleck’s AI studio. The price of the March acquisition of InterPositive, a director-led visual-effects toolmaker, has just been revealed in a filing. It is a signal of how aggressively studios are now buying in-house AI capability rather than licensing it.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.