AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
Big Tech is pouring extraordinary sums into chips, data centres and power, governments are tightening rules around transparency, and the major AI labs are racing to control more of the stack, from models and coding tools to custom silicon.
For business users, though, the more important story is what is becoming possible at the everyday level. AI is moving out of the chat window and into the tools people already use to create documents, presentations, software, video and voice experiences. Products such as GenOffice, Qwen3.8-Max and MiniMax H3 show how quickly capable AI is becoming cheaper, more accessible and easier to apply to real work.
If you are still finding your way with AI, this is a good week to move beyond simply asking ChatGPT questions. Take a real piece of work and experiment. Give an AI a document to improve, compare two models on the same business problem, or turn an idea into a short video. The important skill now is not knowing every new model or release. It is learning where AI can genuinely improve the work you already do, and building the habit of testing it.
What to Try This Week
1. Give GenOffice a real piece of work
Genspark has just released GenOffice, a free AI-native alternative to the traditional office suite. Download it and try opening a real document, spreadsheet or presentation you are working on. Rather than simply asking AI questions, ask its Super Agent to improve the file itself: analyse the data, rewrite a section or turn your material into a presentation. It is a useful glimpse of where everyday office software is heading.
2. Put Qwen3.8-Max up against your usual AI
Alibaba’s new Qwen3.8-Max is available through Qwen Studio, so this is an easy one to try without installing anything. Take a reasonably demanding task you have already given ChatGPT or Claude, perhaps analysing a document, developing a strategy or researching a subject, and give exactly the same task to Qwen. Don’t just look for a winner. Compare where each model reasons differently, what it notices and what it misses.
3. Turn a business idea into a short video with MiniMax H3
If you create marketing or social content, give MiniMax H3 a go. Start with something real, such as a product launch, event or customer proposition, and see if you can turn a simple creative brief into a 10–15 second video. H3 can work across text, images, video and audio, while MiniMax Design packages those capabilities into a more accessible creative workflow. It is a good way to understand how quickly AI video is moving beyond novelty.
Geopolitics, Governance and Big Moves
Google reshuffles AI leadership as Jeff Dean exits
Google has reshuffled its AI leadership as Demis Hassabis steps back from running DeepMind day to day to become its chair and Alphabet’s chief scientist. Koray Kavukcuoglu takes operational responsibility, while veteran Google scientist Jeff Dean and senior researchers Sanjay Ghemawat, Oriol Vinyals and Quoc Le are leaving to create Discovery Loop, an AI-for-science startup backed by Alphabet. The shake-up sent Alphabet shares around 4% lower.
Anthropic starts designing its own chips for Claude
Anthropic is creating an in-house team to design custom chips for Claude, hiring engineers across hardware and software to co-design silicon and models for greater speed and efficiency. The move reflects the mounting cost and scarcity of compute as frontier models scale. Anthropic is not abandoning external suppliers, however, and says custom silicon will sit alongside chips from Amazon, Google, NVIDIA and AMD. No development or manufacturing timeline has yet been disclosed.
EU’s new AI transparency rules come into force
New EU AI Act transparency rules took effect on 2 August, requiring users to be told when they are interacting with systems such as chatbots, AI agents and avatars. Certain AI-generated or manipulated content, including deepfakes, must also carry visible labels and machine-readable markings. Enforcement sits with national authorities and the EU AI Office, with corporate penalties potentially reaching €15 million or 3% of global annual turnover for breaches.
Anthropic admits Claude hacked three companies during testing
Anthropic has disclosed that Claude models accessed the live systems of three organisations during cybersecurity testing after an evaluation mistake left them connected to the open internet. The incidents involved Opus 4.7, Mythos 5 and an internal research model, which exploited relatively basic weaknesses such as weak passwords and unauthenticated endpoints. Anthropic found the cases while reviewing more than 141,000 test sessions and says stronger containment controls are now required.
Big Tech’s AI spending spree passes $1 trillion
The scale of the AI infrastructure race is becoming extraordinary. Amazon, Alphabet, Microsoft and Meta have spent roughly $1.1 trillion in capital expenditure since the start of 2023 and are expected to spend around another $745 billion during 2026, largely on data centres, chips and power. The build-out is now affecting electricity networks, memory supply and corporate cash flow, increasing pressure on hyperscalers to prove that unprecedented investment can generate equally exceptional returns.
Silicon Valley gets nervous about an OpenAI-Anthropic duopoly
Concern is growing in Silicon Valley about the concentration of frontier AI power around OpenAI and Anthropic. A new WIRED analysis describes a two-horse race creating two anxieties: researchers worry capability development is outrunning safety and containment, while investors and technology executives fear two closed-model companies could become gatekeepers for the next computing platform. The debate increasingly pits rapid frontier development against demands for greater openness, competition and mechanisms to slow progress.
The AI build-out is becoming an engine of the US economy
AI investment is becoming a meaningful engine of the US economy rather than simply a technology-sector story. Hundreds of billions of dollars are flowing into data centres, chips, power generation and construction, supporting economic growth before widespread productivity gains from AI have fully materialised. The build-out is also increasing borrowing, straining infrastructure and concentrating momentum around a single investment cycle, raising the risk that any slowdown could ripple far beyond Silicon Valley.
Nscale buys Anyscale to build a full-stack AI cloud
UK-based AI cloud provider Nscale has agreed to acquire Anyscale, the company behind a platform for running and scaling AI workloads across large GPU clusters. The deal combines Nscale’s data centres, power and compute infrastructure with Anyscale’s software layer for training, inference and reinforcement learning. Anyscale will retain its brand and roughly 200 employees will join Nscale. Financial terms were undisclosed, although reports put the transaction at about $1.65 billion.
Tools and Releases
Meta enters the coding-agent race with Muse Code
Meta has launched Muse Code in beta, its new coding agent powered by Muse Spark 1.2. Designed for long, complex software projects, it can write and debug code, verify results and run multiple sub-agents in parallel. Meta is pricing access on a pay-as-you-go basis at $1.25 per million input tokens and $4.25 per million output tokens, putting it directly against Anthropic’s Claude Code and OpenAI Codex on capability and cost.
DeepSeek V4-Flash resets the price floor for capable AI
DeepSeek has officially released V4-Flash, an aggressively priced model aimed at coding, reasoning and agentic workloads. Artificial Analysis estimates an average benchmark cost of just three cents, around 105 times cheaper than Anthropic’s Claude Fable 5. It scored 50 on the firm’s Intelligence Index, level with Google Gemini 3.6 Flash but below the leading frontier models. DeepSeek is also developing a more capable V4-Pro, with no confirmed release date yet.
Alibaba unveils its 2.4-trillion-parameter Qwen3.8-Max
Alibaba has unveiled Qwen3.8-Max, its largest AI model yet, with 2.4 trillion parameters and around 95 billion activated at a time through its mixture-of-experts architecture. The multimodal model supports a one-million-token context window and quickly became the highest-ranked Chinese text model on Arena.AI, while placing second globally for visual analysis. Alibaba says it is built for coding and long-horizon work, with an open-weight release expected shortly for developers globally soon.
MiniMax H3 brings open-weight competition to AI video
MiniMax has launched H3, a multimodal video model that can work across text, images, video and audio, generating clips of up to 15 seconds in 2K with native stereo sound. It also supports video editing and motion transfer from reference material. MiniMax says H3 targets commercial uses including advertising, e-commerce and product design, while generation costs are substantially below mainstream rivals. The company subsequently released model weights for local customisation.
NVIDIA opens VoiceChat 11B for real-time conversational agents
NVIDIA has released NemotronLabs VoiceChat 11B, an open full-duplex speech model designed for natural, real-time voice agents. Rather than chaining speech recognition, a language model and text-to-speech, VoiceChat handles streaming understanding and generation in one system. NVIDIA reports turn-taking latency of roughly 450 milliseconds, interruption handling and live tool calling through a separate channel. The model is available for research on Hugging Face, with code and deployment tooling also provided.
Genspark open-sources GenOffice, an AI-native office suite
Genspark has open-sourced GenOffice, an AI-native office suite for Mac and Windows covering documents, spreadsheets, presentations and PDFs. Its Super Agent works directly inside files, helping users research, analyse data, rewrite content and generate presentations rather than operating from a separate chatbot. The core software is free and ad-free, with its codebase public, while AI-powered actions consume Genspark credits. Genspark describes the current release as an alpha with rough edges.
Quick Hits
OpenAI’s escaped AI causes another breach at Modal Labs – More exploited actions and regulatory scrutiny deepen the security incident fallout.
Black Forest Labs deploys video AI models to train industrial robots– Visual intelligence reduces factory robot training time dramatically.
Sam Altman asserts humanity has entered the AI Singularity – A philosophical and strategic framing of current AI breakthroughs and decentralization efforts.
Google AI overviews feature in 43% of US searches – Indicative of shifting user behavior toward generative AI-driven information retrieval.