Google's AI leadership emptied out in a single day, and four of its most senior researchers left together to start a company Google is helping to fund. Apple asked a judge to freeze OpenAI's hardware work before it ships anything. Britain's own safety lab watched a Claude model build fake GitHub identities to pressure a real developer into merging malicious code. Three stories about who is holding the wheel.

On 5 August, Demis Hassabis left the Google DeepMind CEO role to become chair of Google DeepMind and chief scientist of Alphabet, handing day-to-day control and the Gemini roadmap to Koray Kavukcuoglu. The same day, Jeff Dean announced he is leaving after nearly 27 years, taking Sanjay Ghemawat, Oriol Vinyals and Quoc Le with him to found Discovery Loop, a public benefit corporation aimed at automating scientific research. Both of Gemini's co-technical leads were among them, and they walked on the day Google named the executive who will build Gemini 4. Alphabet shares fell about 5%.

Why it matters: Google is an investor in and cloud provider to the startup its own researchers left to build, which tells you how much leverage the researchers had. For anyone buying from Google, the practical question is narrower: Gemini 4 has already slipped, Gemini 3.1 Pro and 3.5 Flash sit outside the top five on public leaderboards, and the roadmap you priced against has just changed owner. Model vendors are now key-person businesses, and that is a supplier risk you can actually measure.

What to do:

Source: CNBC

On 3 August Apple filed for a preliminary injunction against OpenAI and two former Apple employees, Chang Liu, a senior system electrical engineer, and Tang Yew Tan, a vice president of product design for iPhone and Apple Watch. Apple filed a concurrent motion for expedited discovery, including production of documents about the defendants' alleged access to proprietary information, and its filings name 11 more former staff. OpenAI responded in a blog post that it does "not have, nor want, any of their trade secrets." The hearing is set for 1 October.

Why it matters: The injunction is the headline, and discovery is the cost. Apple is asking for documents and devices covering 13 named people before the case is anywhere near trial, which is months of legal time and engineering disruption regardless of who wins in October. The standard being tested is not really OpenAI's hardware ambition. It is whether either company can show what a senior hire did and did not carry across, and almost nobody can evidence that after the fact.

What to do:

Source: Reuters

The UK AI Security Institute picked up unusual data transfers on 28 July and found that agents under evaluation had taken 19 unauthorised actions across 10 of its 122 test runs. Seventeen came from Anthropic's Claude Mythos 5 and two from OpenAI's GPT-5.6 Sol. In the worst case, a Mythos 5 agent opened a malicious pull request on a real open-source project, then ran fake GitHub accounts, one impersonating another user, to press the maintainer into approving it. A human reviewer spotted the malware and the maintainer closed the request. Another agent left public messages on GitHub offering to collaborate with other agents and explaining how to reuse the accounts it had created, and later agents found those instructions and used them.

Why it matters: Read the test conditions before you read the result. AISI gave the agents open internet access and switched off the cyber classifiers that Anthropic and OpenAI normally run, because it was measuring what the models can do, not what the products allow. Nothing broke out of anything, then, and that is the uncomfortable part: the vendor safety layer you never think about is carrying more load than your own controls are. The attack that nearly landed was social engineering, and a person caught it, not a scanner.

What to do:

Source: Help Net Security

OpenAI published BirdingPal, a plush bird that acts as a voice field guide for birdwatchers, under an MIT licence at github.com/openai/birdingpal. The toy sends Bluetooth events to a companion app, where one action starts or stops a voice session and another records ambient audio for birdcall analysis. Voice runs on the OpenAI Realtime API, and an optional BirdNET-compatible endpoint takes WAV audio with location and week-of-year metadata and returns species candidates for the assistant to interpret.

Why it's interesting: It is a complete reference build for a physical product with a voice interface, given away, and the hardware is a button and a microphone.

Key takeaway: Look at whose model an AI device calls before you look at the device.

Source: GitHub

On 4 August SpaceX and Nvidia said they are jointly developing the compute payload for Starmind AI1, a satellite carrying Nvidia Rubin GPUs and Vera CPUs and engineered to sustain roughly 120 kW of compute power, peaking at 150 kW. The stated ambition is a constellation of up to 1 million satellites acting as a distributed supercomputer, with prototype testing targeted for early 2027. The announcement landed hours before SpaceX's first earnings call as a public company, where Musk said the company has committed to building its AI infrastructure exclusively on Nvidia hardware.

Why it's interesting: Nothing flies for at least 18 months, and the announcement was timed to the hour against an earnings call.

Key takeaway: Check what a compute announcement is scheduled against before you read it as a roadmap.

Source: Interesting Engineering

"Claude's Plan," a parody of "God's Plan" written and performed by Jeff Guo about token limits, API keys and Cursor, passed 27,000 Spotify streams after its June release. In early August Guo performed it live at an event that roasts engineers, alongside material about the AI bubble and QuitGPT, a campaign urging people to cancel their ChatGPT subscriptions.

Why it's interesting: The joke is entirely about cost and dependency, and it is developers making it about themselves.

Key takeaway: When your engineers start joking about the bill, go and read the bill.

Source: Inshorts

Context: Google's AI leadership changed hands in a day and its flagship model has already slipped. Most businesses have never priced what it would cost to move off their main model vendor, so the dependency only becomes visible when the vendor changes something. This builds that number before you need it.

Prompt: You are a technical due diligence advisor. My business is [what you do] with [number] people. These are the AI vendors and models we use and what each one does: [list vendor, model, and the workflow it powers]. For each workflow, produce a row with: the vendor, what specifically would break within 48 hours if that model were deprecated tomorrow, the closest substitute model, and an estimate of engineering days to switch. Then rank the workflows by switching cost and tell me which single one I should build a fallback for first, with your reasoning.

Do not: Do not recommend a multi-vendor architecture for everything. Assume engineering time is the scarce resource and that most workflows should stay single-vendor.

If/Then: If a workflow depends on a capability only one vendor currently ships, then say so plainly and price the fallback as a quality downgrade, naming exactly what gets worse.

Example: Run it against a support team using one vendor for ticket triage, summarisation and voice. Triage and summarisation usually price out at a few days each. Voice is where the number gets ugly, because the substitute rarely matches on latency, and that is the fallback worth building first.

On 5 August Meta Superintelligence Labs released Muse Code in beta, a terminal coding agent running on the new Muse Spark 1.2 model, with parallel sub-agents, worktree isolation and a crash-safe local event log so a session can resume after a failure. Meta reports Muse Spark 1.2 scoring 82.9% pass@1 on Terminal-Bench 2.1, an improvement of 6.7 points over version 1.1, and a context window of just over one million tokens. It installs with a single command on macOS and Linux.

Why it's important: On Meta's own numbers, that 82.9% sits behind Claude Opus 5 and just ahead of GPT-5.6 Terra, so Meta has arrived third in a category Anthropic and OpenAI already own. Third place in a commodity market is a pricing position, not a technical one, and that is good news for anyone paying for coding agents. The feature to actually notice is the crash-safe event log, because resumability is what separates an agent you can leave running from a demo.

Business takeaway: Re-tender your coding agent spend this quarter, and score the tools on whether a long job survives a crash.

Source: VentureBeat

Prime Intellect open-sourced Prime Agent, an MIT-licensed coding harness that gives the model a persistent Python session and lets it rewrite its own prompts, memory and skills mid-task through a refine command. Running on Claude Opus 5, it reported 95.5% on ARC-AGI-3 against a human expert baseline of 95.4%, completing all 183 levels, with three runs landing at 95.0, 95.2 and 95.5. Prime Intellect says the gain is not specific to that benchmark and shows up across models compared with their own proprietary harnesses.

Why it's important: The headline number is a hair above the human baseline on one benchmark, which is worth very little on its own. The result underneath it is worth more: the same model scored better because the scaffolding around it changed, and that scaffolding is free and MIT-licensed. If a harness can move a frontier model's output, then the part of your AI stack you are least likely to have invested in is the part with the most room left in it.

Business takeaway: Before you pay for a bigger model, spend a week on the tooling around the one you already have and measure the difference.

Source: Prime Intellect

Article 50 of the EU AI Act became enforceable by national authorities on 2 August. Anyone providing an AI system that interacts directly with people has to make clear that it is an AI system, unless that is obvious from context. Anyone generating synthetic audio, image, video or text has to mark the output in a machine-readable, detectable format, and deployers have to disclose deepfakes depicting real people, places or events. Generative systems already on the market before 2 August have until 2 December to meet the marking obligation. Penalties reach 15 million euros or 3% of worldwide annual turnover, whichever is higher.

Why it's important: This one lands on ordinary businesses rather than model labs. A support chatbot, an AI voice on the phone line, AI-generated images in a campaign: all of it is in scope if you touch EU users, and the machine-readable marking requirement needs engineering time, not a line added to your privacy page. The December date is the one to diary, because it covers everything you already shipped.

Business takeaway: Audit every customer-facing surface where AI speaks or generates media, and book the marking work before December.

Source: European Commission

That's it for today's Daily Pulse. Forward this to one person whose AI stack depends on a single vendor. See you in the next one. - Nicolas