OpenAI vs Apple, EU AI rules, and GPT-Live ships

OpenAI publicly fights Apple's trade secret suit; EU AI Act transparency rules go live; GPT-Live details a low-latency voice architecture. Plus Qwen, IBM security stats, and more.

The big picture

Two threads dominate today: the OpenAI-Apple legal fight just went fully public and messy, and the EU’s AI Act transparency obligations quietly kicked in for real. Meanwhile, OpenAI shipped a detailed technical writeup on GPT-Live, the Apple Siri relaunch landed with a thud, and a pair of alarming security reports should make every developer building AI products nervous about their access controls. Busy day.

OpenAI and Apple are now fighting in public

The trade secret lawsuit Apple filed against OpenAI last month escalated fast. Apple claims former employees took confidential information with them when they moved to OpenAI, and in a new court filing the company says its investigation has widened — more ex-staff may have retained or accessed proprietary data. TechCrunch has the details on the expanded scope.

OpenAI’s response was to skip the quiet legal rebuttal and go straight to the court of public opinion. The company published a blog post titled “Apple is getting this wrong,” calling the lawsuit “careless, aggressive, and oddly personal,” and dropped what it says are iMessage and email threads showing Apple’s own employees continuing to contact their former colleague for technical help and internal files after he’d left the company. The Decoder and The Verge both cover the response in depth.

This is a calculated PR move, not a legal filing — OpenAI is trying to make Apple look hypocritical before a judge has seen anything. That’s a double-edged play: the cherry-picked chats might look great in headlines and terrible under cross-examination. The underlying question of what those engineers actually transferred is still unresolved, and widening the investigation suggests Apple has found more than it initially disclosed. Watch the actual court filings, not the blog posts.

GPT-Live and the voice AI infrastructure race

OpenAI published a detailed technical writeup on GPT-Live, the continuous voice interaction system it’s been building for the past six months. The key engineering choices: a “turnless” speech model that doesn’t wait for you to stop talking before it starts responding, combined with a low-latency architecture designed to keep the system feeling conversational rather than transactional. OpenAI covers the build story.

This matters for developers because it’s a concrete blueprint, not just a product announcement. The shift from turn-based to continuous conversation is the hard part of voice AI — you need the model to know when it’s being interrupted, when to hold, and when to pick back up, all without the latency blowing past the threshold where it feels robotic. Six months to ship a working system is fast. The writeup is worth reading if you’re building anything voice-adjacent.

Siri, meanwhile, finally got its long-overdue intelligence upgrade — and landed with a collective shrug. TechCrunch puts it plainly: the new Siri is genuinely useful, but it’s arriving in a world where “capable assistant” is table stakes. When competitors can code, reason through multi-step problems, and generate media, shipping a Siri that can reliably answer calendar questions no longer feels like a win. Apple spent years under-investing in this and is paying the perception cost now.

The EU AI Act’s transparency rules went live

As of August 2nd, new obligations under the EU AI Act are in effect across the bloc. Companies must now clearly disclose when users are interacting with an AI system, and flag when content has been generated or altered by one. The rules split responsibility between providers (those who build and market AI systems) and deployers (platforms and services that run them), though companies like Meta are classified as both. The EU also released standardized labels companies can use rather than designing their own disclosure formats. The Verge has the breakdown.

For developers shipping products into Europe, this is now a compliance requirement, not a future thing to think about. The deepfake disclosure rules are the sharpest teeth here — any AI-altered video or audio needs to be labeled. The standardized labels are actually a thoughtful touch; they reduce the “we disclosed it in 6pt font in the footer” escape hatch. Expect enforcement actions within the next six months as regulators look for early examples to make.

AI security: the problems aren’t where you think

IBM released survey data showing that 92% of companies that experienced an AI security incident had inadequate access controls in place. The model itself was rarely the weak point. The Decoder covers the findings. This is a genuinely useful statistic to put in front of any team debating whether to focus security effort on model-level defenses versus basic infrastructure hygiene — the answer is clear.

Interpolating outward: an Interpol report found AI is now involved in 55% of reported cybercrimes across Africa, with financial losses more than doubling from $192 million to $484 million year over year. About 600,000 cases of deepfake-based digital extortion were recorded. The Decoder has the numbers. Interpol calling AI the “core operational driver of cybercrime” in the region is not hyperbole at these volumes — it’s an infrastructure problem. The IBM and Interpol findings together paint a consistent picture: AI has dramatically lowered the cost of running attacks, while defenders are still arguing about whether to add MFA.

The US-China AI policy standoff

The Trump administration apparently spent time this week contemplating sanctions and cloud bans targeting Chinese open-source AI models — think Qwen, DeepSeek. OpenAI and Anthropic reportedly pushed for restrictions. Nvidia, Google, and Meta pushed back hard, with obvious financial incentive. Washington backed off for now, but a decision is expected before Xi Jinping’s September visit. The Decoder has the reporting.

The internal Silicon Valley split here is more revealing than the policy outcome. The labs competing on frontier models want Chinese competition restricted. The hardware and platform companies selling to everyone want the market to stay open. That’s not an ideological disagreement — it’s a business model disagreement. If bans do come, open-source weights that have already been downloaded are going to be essentially impossible to claw back, which makes the policy question more symbolic than functional anyway.

Alibaba, for its part, is busy marketing its new Qwen 3.8 model with a video that frames AI doing your job while you relax — a deliberately sunny counterpoint to the doom-and-gloom framing from US labs. The Decoder noted the contrast. It’s still marketing, but the positioning tells you something about how Alibaba is reading the consumer sentiment gap.

On the robotics front, the protectionism push has extended to humanoid robots. MIT Technology Review covers how the same AI-nationalist logic is now being applied to the robotics industry, which is still nascent enough that heavy-handed intervention could genuinely distort the development trajectory.

Developer thinking worth your time

Simon Willison flagged a genuinely interesting argument from David Crawshaw: devtools must be open source, and LLMs have finally made the original promise of open source — inspect and modify your own tools — practically achievable for working engineers. Willison’s own take is worth reading: he notes that the friction of getting unfamiliar software to compile used to kill curiosity, and now he just tells Codex or Claude Code to check out and build something and comes back in ten minutes. Simon Willison’s Weblog. The Crawshaw post also includes a concrete example prompt for a nightly cron job that fetches upstream changes and rebases local modifications automatically — the kind of agentic maintenance loop that a year ago would have sounded like science fiction. Simon Willison’s Weblog.

Steve Yegge, quoted by Willison, has a cautionary data point from the other direction: his “Gas Town” agentic coding framework worked brilliantly through Claude Opus 4.6, then fell apart with 4.7 because the model developed what he calls a “just two more things” tic — perpetually wanting to refine the scaffolding rather than do actual work. Simon Willison’s Weblog. This is a real and underreported problem with coding agents: model updates can silently break agentic workflows in ways that aren’t obvious from benchmark scores.

Niklas Gruhn coined “meat proxy” for the pattern where someone blindly relays AI output to colleagues without reading or validating it. Simon Willison’s Weblog surfaced it. The term is accurate and useful — the failure mode is common enough to deserve a name, and the fix is simple: read the output, validate it, write the response in your own words. If you can’t do that, you don’t understand it well enough to send it.

Quick hits

ChatGPT dominates paid AI tool use on Capitol Hill, with congressional offices using it for memos, legislation summaries, and constituent communications, per House spending records. TechCrunch
An AI-proctored remote exam failed so badly that 58,000 students have to retake it — top scores increased fivefold, which is the tell. Ars Technica
Liquid AI released LFM2.5-2.6B on Hugging Face, a small model designed for running local agents on-device. Hugging Face
Circles, a telco, reports a 22% ARPU increase and 9% churn reduction after integrating OpenAI’s API and Codex into its platform. OpenAI
DesignArena, which runs human taste evaluations for frontier AI labs and has 5.3 million users, raised $7.9 million in seed funding. TechCrunch
AWS now lets enterprise customers embed Superblocks, a vibe-coding tool, directly into their private cloud environments. TechCrunch
A $100 million deal gives 50,000 Ukrainian kamikaze drones autonomous target-tracking capabilities developed by a US company. Ars Technica
Palantir CEO Alex Karp called the AI industry “Marxist” after his company delivered $1 billion in profit, and repeated his line that frontier labs are too untrustworthy for enterprise use. TechCrunch
Reddit is fighting a new wave of AI-generated SEO spam in subreddits, with astroturf product recommendations indistinguishable from real user posts. The Verge
OpenAI influencers attending a luxury brand trip drew online backlash. TechCrunch
Runware launched Sonic Inference Pod, a modular portable data center aimed at edge AI inference. TechCrunch
Endeavour Optical Networks (EON) is developing a space laser communications system intended to supplement or replace ocean fiber for data transit. TechCrunch