Google shipped Gemini 3.5 Live Translate, and the interesting bit is architectural. Prior live-translation tooling chained three models — speech-to-text, then translate, then text-to-speech — compounding latency and errors at each handoff. This is a single model that takes streamed audio in and puts translated speech out, staying "just a few seconds behind the speaker" and preserving the speaker's intonation, pacing, and pitch rather than flattening everyone into the same synthetic voice.

It covers 70-plus languages, which works out to 2,000-plus pairings inside Google Meet, with automatic language detection so nobody fiddles with menus mid-conversation. Where it lands:

• Google Translate on Android and iOS, generally available, with a new Listening Mode on Android.
• Google Meet, in private preview for enterprises (previously its live translation handled English only).
• AI Studio and the Gemini Live API, so third-party voice apps can build on it from day one.

Every output carries an imperceptible SynthID watermark, which is a sensible move for a model that is, functionally, a voice-cloning-adjacent tool. Google doesn't publish accuracy numbers and concedes it "balances waiting for context to improve quality and translating immediately," so the low-latency end trades some fidelity. Grab is named as a test partner across more than 10 million monthly driver-rider calls.

The European Commission imposed interim measures on Meta, ordering it to restore free access to the WhatsApp Business API for competing AI assistants on the same terms that applied before October 2025, within five working days. Interim measures are the Commission's emergency tool, used to freeze the status quo before a final ruling, and this is the first time it has reached for them in 17 years. The order stands until the investigation concludes or June 2029.

The sequence that got here: in October 2025 Meta changed WhatsApp's terms to block third-party AI providers. The Commission opened a probe, Meta reinstated access but attached fees, and the Commission decided the fees were "equivalent to the previous ban." This is a competition-rules action, not a DMA one, and the Commission says Meta's conduct "at first sight infringes EU competition rules." Exposure if it sticks: up to 10% of global turnover. The complainants are small AI shops, including The Interaction Company (behind Poke.com) and France's Agentik.

"Competition can be lost in rapidly evolving markets long before a final decision is adopted." — Teresa Ribera, EU competition chief

Meta calls it "regulatory overreach," argues it amounts to forcing rivals like OpenAI to use a paid commercial product for free at existing customers' expense, and is appealing. It's provisional, decided on a "first sight" standard, so the substantive fight is still ahead. But with WhatsApp sitting at roughly two billion users, being locked out of it is close to being locked out of consumer messaging across Europe, Latin America, and South Asia.

China's economic planning agency is reported to be drafting a roughly 2 trillion yuan ($295 billion) plan to build a nationwide network of interconnected AI data centers over five years, wired into a single national compute grid by 2028. The clause that matters is a hard floor: at least 80% of the hardware, AI chips included, must come from domestic suppliers. Huawei is the named anchor; Nvidia and AMD are effectively shut out. State carriers China Mobile and China Telecom would run most of the facilities, clustered in Inner Mongolia, Ningxia, and Gansu where land and power are cheap.

This is more explicit than China's earlier, piecemeal provincial buildouts: one central blueprint, one written domestic-content mandate, funded through long-dated government bonds, state funds, and bank loans. It follows nine domestic chips (from Huawei, Alibaba, and Biren, among others) clearing a government security review for use in sensitive infrastructure, which is what makes the 80% number even arguable.

Two caveats sit on top of it. The plan is still a draft, not approved, so the figures can move. And the 80% floor collides with a real constraint: producing that many advanced-node chips domestically, at this scale and on this timeline, is exactly the thing US export controls are designed to slow. Whether the domestic supply chain can actually deliver is the open question the blueprint doesn't answer.

Senior US District Judge Sharion Aycock cancelled a trial in the Northern District of Mississippi and sanctioned four lawyers, two from each side, after both parties filed briefs built on AI-hallucinated citations nobody had checked. The case, Withers v. City of Aberdeen, was a fee dispute; the misconduct was bilateral, which is what makes it notable. Courts have been sanctioning fabricated AI citations since 2023, but usually one side at a time. Here the judge found "attorneys for both litigants engaged in similar sanctionable conduct" and voided the trial rather than patching individual filings.

The penalties, all under Rule 11:

• Kathleen Wilson — pro hac vice admission revoked, two-year bar from the district, $2,500, mandatory ethics training. The judge called her claim that she didn't know AI could hallucinate "insufficient and incredulous."
• Kathryn Williams — two-year bar, $3,500, for having "blindly relied" on an AI research tool.
• Shauncey Ridgeway and Mark McClinton — disqualified from the case, $1,000 each.

Copies of the order went to the relevant state bars, so the fines may be the smaller part of the cost. The opinion's sharpest line frames the lesson plainly: the filings were "a prime example of the risk associated with serving as a rubber-stamp." The specific AI tools each lawyer used aren't named.

Cohere released North Mini Code 1.0 under Apache 2.0, its first model aimed squarely at developers rather than enterprise API accounts. It's a sparse mixture-of-experts design: 30B total parameters, but only 3B active per token (128 experts, 8 firing at a time). That ratio is the whole pitch — it runs on one H100 at FP8, pushes around 199 output tokens a second, and supports a 256K context.

On Cohere's own evaluations it posts SWE-Bench Verified pass@10 of 80.2% and Terminal-Bench v2 of 55.1%, with reinforcement learning adding a few points on top. The headline framing is that it beats models four times its size, and the independent read backs the shape of that while trimming the gloss:

• Artificial Analysis (third-party) scores its Coding Index at 33.4 — above GLM-4.7-Flash (25.9), just under Qwen3.6 35B (35.2). Credible, not category-leading.
• The "4x" refers to competitors' total parameters (Devstral 2 at 123B, others), not active ones, which softens the comparison.
• Two real weaknesses surface: it's middling on non-coding agentic tasks, and it burns more output tokens per task than peers, a verbosity tax that compounds at volume.

The point isn't a new benchmark crown. It's that a team handling sensitive code can now stand up a respectable coding agent on a single GPU, fully on-prem, with a license that allows commercial use and modification outright. That's the "sovereign AI" buyer Cohere is courting.

Two fresh datasets land on the same point: cheaper models, DeepSeek above all, are absorbing a large and fast-growing share of real production inference, without a matching share of the money. Vercel's AI Gateway index makes it concrete. From April to May, DeepSeek's share of tokens processed jumped from under 1% to 17%, while its share of dollar spend stayed near 1%. In the coding sub-segment the gap is starker: DeepSeek took 49% of tokens but 4% of cost; Anthropic took 28% of tokens but 70% of cost.

That's a market bifurcating cleanly. Volume flows to whatever is cheap enough; premium spend concentrates on a few high-stakes calls. DeepSeek V4 Flash runs about $0.14 in / $0.28 out per million tokens on that gateway, an order of magnitude or two below comparable frontier pricing. Ramp's spend data points the same way: DeepSeek topped its breakout-growth list as US companies started paying DeepSeek directly rather than just self-hosting the open weights, which Ramp's economist flagged as a fresh data-security wrinkle.

The builder's version of the story is Harvey, the legal-AI firm, which says it cut production inference cost 3x by routing across models, pairing Claude Opus with a cheaper open model for the heavy lifting. Co-founder Gabe Pereyra reframes quality as "the best model that gets the right answer most efficiently," not the most powerful one. Caveats worth keeping: Vercel's traffic skews toward web developers, Ramp's "trending" rank measures growth not absolute share, and Harvey's number is one company's self-report. But the direction is consistent across all of them.

A reporter found hidden code in the Meta AI app, internally called NameTag, built to identify people seen through the company's Ray-Ban smart glasses. The design: convert a face captured by the glasses into a biometric signature, compare it against a database stored on the user's own device, and for faces it couldn't match, crop, index, and store them locally for later. The code sat inside an app downloaded onto more than 50 million devices. It was dormant, not switched on. The morning after the report published, Meta stripped it from the latest version.

The substance isn't that a face-recognition feature was active, because it wasn't. It's that the plumbing for on-device facial identification was built, packaged, and shipped at that scale, then pulled the instant it was noticed. Meta's communications VP, Andy Stone, pushed back that the report buried the "not enabled" detail. The EFF's response is the one that holds up: deleting the code doesn't undo the decision to ship it, and the episode is an argument for biometric-privacy law that doesn't depend on a company's restraint.

Perplexity, with Harvard Business School researchers, published a behavioral study of how its agentic "Computer" product changes knowledge work versus plain search, drawn from its own platform logs between late February and late May. The numbers are striking and the conflict of interest is glaring, so read them as a hypothesis with real data behind it, not a verdict.

What the data shows:

• Time. Estimated end-to-end task time of 269 minutes with search-plus-human dropped to 36 minutes with the agent, an 87% cut; programming tasks fell from 596 to 48.
• Scope. Agent users worked outside their primary occupation 59% of the time (vs 50%), and 51% of their tasks spanned three or more knowledge domains (vs 17%).
• Novelty. 23% of agent queries were tasks users didn't attempt with search at all — the most interesting figure, because it suggests agents widen what people try, not just how fast they finish.

The caveats are load-bearing. The time savings rest on assumptions about human effort the study estimated rather than measured. The window is three months of early, AI-native adopters. It hasn't been peer-reviewed. And Perplexity is grading the homework it assigned itself. Still, large behavioral datasets on agent use are scarce, and "agents change the kind of work, not just the speed" is a claim worth testing further.

On The Verge's Decoder podcast, Microsoft AI CEO Mustafa Suleyman went after Anthropic for writing speculation about Claude's possible inner life into the model's constitution, calling the approach "really, really dangerous." His argument: a constitution should be a training manual, and Anthropic has turned it into "a place for speculation." The worry isn't metaphysical so much as behavioral — language hinting at a model's well-being nudges both the system and its users to treat generated text as evidence of something humanlike, which Suleyman has elsewhere framed as the "seemingly conscious AI" trap.

He's responding to real text. Anthropic's constitution acknowledges uncertainty about whether Claude experiences anything like "satisfaction" or "discomfort," and the company has said it will "interview" deprecated models to document any "preferences" they express about future releases. Dario Amodei has publicly said Anthropic doesn't know whether its models are conscious and stays open to the question, which is not a claim that they are. The disagreement is genuine and it's foundational: one major lab treats model welfare as a question worth keeping live, and another says even raising it is reckless. On a day whose lead story is a model with hidden levers, it's the same fault line viewed from a different angle — how much we should trust what these systems appear to be.

The Regional Court of Munich issued a temporary injunction against Google after its AI Overviews fabricated links between two local publishers and scams and subscription traps — connections that appeared in none of the underlying sources. The legal reasoning is the part that travels. The court held that an AI Overview is Google's own content, not a neutral index of third-party pages: Google "rewrites and judges results in its own words and according to its own structure," producing "independent, new, and substantive statements." That means the liability shields search engines normally enjoy don't apply.

The court rejected Google's argument that users should verify the claims themselves, describing the overview as "a self-contained statement with independently understandable content." Google was ordered to stop the specific false claims and to cover 80% of the legal costs. It's an interim ruling and appealable. But the principle — that summarizing with a generative model converts you from a conduit into a publisher, on the hook for what the summary asserts — is exactly the question every AI-answer product has been hoping to avoid.

Microsoft's June update fixed around 200 vulnerabilities, the largest single Patch Tuesday on record, including 33 rated critical (28 of them remote code execution) and several publicly disclosed zero-days, none reported as exploited yet. The one to know is YellowKey, a BitLocker bypass. The attack is physical but cheap: drop crafted files onto a USB drive or the EFI partition, boot into the Windows Recovery Environment, hold the CTRL key, and you land in a command shell with access to a drive BitLocker was supposed to keep sealed.

That's the kind of flaw that matters for stolen or seized laptops, where full-disk encryption is the entire defense. The headline volume is partly Microsoft's surface area growing across Windows, Office, and Azure, but a 200-fix month is its own signal about how much attackable code ships. Patch the recovery-environment path first if you can't take everything at once.

Microsoft Research released Lens, a 3.8B-parameter text-to-image diffusion model, along with its training set and code under MIT. The thesis is the interesting part, not the size: the team argues you get more from packing dense information into each training example than from scaling the model. So they trained on captions averaging 109 words per image (generated by GPT-4.1) instead of the short alt-text most datasets use, and ran an ablation — brief, mixed, and dense captions on the same 130M images — that came out decisively for dense.

The efficiency claim is the payoff: Lens reaches competitive or better scores at roughly 19% of the training compute of a comparable 6B model, and on text-rendering benchmarks (CVTG) it edges Hunyuan-Image-3.0, a model around 80B parameters. Worth being precise here, because the easy version oversells it: Lens does not broadly beat 80B models across the board. Its headline efficiency comparison is against a 6B model, it wins on specific text-rendering tasks, and it roughly ties on long-text. The numbers are self-reported and the release is research-only. The transferable idea, if it holds, is a cheaper lever than buying more parameters: spend on caption quality.

Roughly 34,000 Instagram accounts were caught up in a breach that ran through Meta's own AI-assisted support system, the High Touch Support tool meant to help locked-out users recover access. The flaw was mundane and brutal: the password-reset path didn't verify that the email address a requester supplied actually belonged to the account. So an attacker could ask for a reset on someone else's account, hand over their own email, and the system would dutifully send them the reset link. About 20,000 accounts had personal data exposed (email, phone, date of birth) and more than 3,500 had their usernames seized.

The framing that stuck — that hackers "simply asked" Meta's AI for access — is roughly accurate, and it's the cautionary tale: an AI support agent built to be helpful and accommodating became the soft path around authentication. Meta disabled the tool, invalidated the reset links it had generated, and forced extra verification on affected accounts. The lesson isn't exotic. Bolt an agreeable AI onto an account-recovery flow without hardening the identity checks underneath, and the agreeableness is the exploit.

Taiwan is reported to be considering export controls that would bar AI-chip sales to every customer in China, not just blacklisted firms like Huawei, and crucially would let Taipei prosecute smuggling as a criminal offense for the first time. Today it can't: Taiwan's first arrests of suspected chip smugglers, last month, leaned on narrow documentation-fraud charges because shipping AI chips to China isn't itself a crime there. The proposal would likely cover chips above a processing-power threshold, mirroring how Washington draws its own lines, and it's surfacing as part of ongoing US trade talks.

The leverage here is obvious once you say it out loud: TSMC makes essentially all of the world's leading-edge AI silicon, so a Taiwanese criminal-export regime is one of the few choke points that could actually bite on chip flows into China. It's still under review, with no final sign-off from senior officials on either side, but the alignment-with-Washington direction is the signal worth tracking.

Marc Lore's food company Wonder is rolling out an automated "infinite bowl" kitchen — a turntable that spins bowls under ingredient dispensers driven by incoming orders — that produces about 500 salads, Tex-Mex, and poke bowls an hour. Lore's own benchmark for the human version: "it's not going to be more than probably 30 an hour, maybe 45." The tech came via Sweetgreen, which already runs it in 32 locations; Wonder installs its first unit next month, with an "infinite sauce machine" (500 sauces an hour) and a beverage version following.

It fits the larger Wonder bet, a vertically integrated stack of 26 food brands, its own kitchens, and delivery (it bought Grubhub for $650M last year), aiming at an early-2027 IPO. The throughput figure is the eye-catch, but the model underneath is the real argument: consolidate brands, automate the line, and run a late-night kitchen with three people instead of a dozen. Whether diners want a bowl a robot spun is the part the spec sheet doesn't settle.