AI Moves From Chat Windows Into the Real World

ChatGPT Opens the Text Tap

For millions of ChatGPT users, the meter on an ordinary text conversation is about to disappear. The catch is that unlimited does not mean every tool, for every account, right now.

The reason this deserves more time than another round of benchmark scores is simple: access changes how people can use a model day to day. A cap isn’t merely an interruption when it arrives. It makes a user decide in advance which questions deserve the remaining messages, or whether to move the work somewhere else. Removing it changes the free product from a metered trial into a dependable text workspace, provided the task really stays in text. OpenAI announced on August 6 that GPT-5.6 Luna is becoming the default model for ChatGPT Free and Go users, and ordinary text chats with it will be unlimited. Free users are also getting a Think button, which lets them ask the system to spend more effort reasoning through a request instead of treating every prompt as a quick exchange.

The rollout is staged. OpenAI says Luna is due to become the Free and Go default during the week, with unlimited text chats and the Think button starting the following week. So if your account still looks unchanged, that doesn’t contradict the announcement. It may simply not have reached you yet.

There’s also a narrower but useful change for subscribers. Plus and Pro users received an updated GPT-5.6 Sol on August 6, along with a slider for reasoning effort. That gives people a direct way to choose between a faster answer and more computational effort when the problem warrants it. The update applies to conversations in ChatGPT. It doesn’t change the model versions powering Codex or Work, so developers and organisations shouldn’t read it as a platform-wide model replacement.

Now, the boundary around unlimited matters. OpenAI is removing the quota for text chats, not for file uploads, image work or other tools. Those retain limits, and abuse guardrails still apply. A long writing session, study exchange or coding discussion can therefore continue without the familiar text cap, while a workflow that repeatedly analyses documents or generates images may still run into plan restrictions. OpenAI has made the everyday lane much wider; it hasn’t made every expensive capability free.

For the general public, my read is that this access change is more consequential than a modest quality gain reserved for paying customers. It lowers the friction on sustained, ordinary work and makes the free product more dependable as a daily tool. It may also sharpen the decision around paying: subscriptions now have to earn their price through stronger tools, higher non-text allowances and finer control, rather than merely keeping a text conversation alive.

There’s a competitive consequence as well, although OpenAI hasn’t made claims about rivals here. Once a capable default model comes with effectively open-ended text use, a chat quota becomes harder for any mass-market assistant to defend. Still, judge the product you actually receive, not the headline. The important details are the staged arrival, the separation between Luna and Sol, and that very specific definition of unlimited.

A Texas Chip Bet at Extraordinary Scale

That change will reach users quickly. The shorter news begins with the enormous physical bill behind ambitious AI systems.

Tesla and SpaceX announced Grimes County, Texas, as the site of a jointly developed semiconductor complex they’re calling Terafab. The companies say they’ll make an initial investment of sixteen point eight billion US dollars. Their proposed facility would exceed one hundred million square feet and cover manufacturing, packaging and testing for both logic and memory devices.

The intended uses sit inside the Musk companies’ own plans. Tesla wants chips for edge inference in vehicles and robots, where computing happens close to the sensors and machines using it. SpaceX points to high-power compute for proposed data centres in space. Texas officials and the companies project at least three thousand local jobs.

Those numbers are striking, but nearly every operational detail that would let us test the proposal is missing. There’s no disclosed production timetable, process technology, capacity target or set of construction milestones. Funding beyond the initial commitment hasn’t been explained. The scale, output and employment figures are forward-looking claims from the companies and government, not results from a working plant.

If Terafab is built as described, Tesla and SpaceX would gain something strategically valuable: more control over a chip supply chain now dominated by established foundries, memory makers and packaging specialists. Vertical integration could let them tune hardware around their own vehicles, robots and computing systems, and reduce dependence on suppliers whose capacity is contested across the industry.

My assessment for organisations is to treat Terafab as a serious statement of intent, not yet as new chip capacity. The upside is supply control. The immediate exposure is a huge commitment of capital and local resources before the factory’s economics, schedule or manufacturing capability have been demonstrated. Until those missing milestones appear, the announcement changes the competitive map on paper, not the available supply of semiconductors.

Suno Adds Signals to Generated Music

From a proposed factory measured in square feet, let’s switch to a problem carried inside a song file.

Suno says it has begun rolling out tools to identify music generated on its service, and plans to add audio watermarking and fingerprinting in the coming weeks. The aim is for that identification to keep working when a track leaves Suno and is posted on another platform. The company is also preparing a download policy intended to restrict mass distribution to streaming services, although it hasn’t published the detailed rules.

Its updated community guidelines now explicitly prohibit deceptive audio presented as authentic, as well as unauthorised use of a person’s voice or likeness. Those are meaningful policy statements, particularly for a service that can produce convincing songs quickly. They also arrive while Suno remains involved in copyright litigation, which is essential context for judging the move.

The technical promise needs careful handling. Suno hasn’t disclosed how its watermark works, how reliably it can be detected, how it survives editing or compression, or the precise date it will cover new output. It also hasn’t established how widely music platforms will adopt the detector. A signal is only useful when receiving services can find it and have a policy for what to do next.

For music platforms and rightsholders, traceability could make moderation and investigation easier. It may help separate Suno output from human-made recordings when material moves across services. My editorial judgment, though, is that watermarking belongs in the evidence column, not the permission column. It can indicate where a file came from; it doesn’t prove that training material was licensed, establish who owns the result, or settle the legal disputes around generative music.

The responsible test is therefore practical: can independent platforms detect the mark consistently after ordinary changes to a track, and can they do it at scale? Until Suno supplies technical detail and deployment evidence, this is a promising safeguard with an unresolved implementation gap.

Hiring Rules Catch Up With OpenAI

There’s another reminder that fast-growing AI companies still answer to rules written long before the boom.

The United States Justice Department announced a three point two million US dollar settlement with OpenAI and Statsig over allegations involving recruitment for permanent residency. The department said the companies discriminated against US workers in the way they handled a small group of visa-linked roles. Fewer than ten positions were involved, according to the department, but the resulting controls reach directly into how those jobs are advertised and filled.

The allegations focus on process. The department says the positions were left off the companies’ usual careers pages, required paper applications and were promoted through late-night radio advertising. Under the settlement, OpenAI and Statsig will pay one point two million dollars in civil penalties and establish a two million dollar back-pay fund. The eventual compensation total for individuals depends on identifying eligible US applicants.

The companies agreed to publish covered positions on their careers sites, accept electronic applications, revise recruitment policies, train staff and submit to monitoring and reporting. They did not admit wrongdoing. That distinction matters: this is a settlement with binding reforms, not a court finding after a contested trial.

For employers competing for specialised technical talent, the useful lesson is remarkably ordinary. Immigration-linked recruitment doesn’t sit outside normal employment compliance merely because the roles are scarce or the company is prominent. My read is that the monitoring may prove more consequential than the dollar figure. It turns a narrow allegation involving fewer than ten jobs into an ongoing obligation to make the recruitment path visible, accessible and reviewable.

That makes hiring a governance issue alongside model safety and data handling. A company can have sophisticated technical controls and still create legal exposure through a careers page, an application format or an advertising choice. Here, the remedy is built from familiar mechanisms: publish the openings, accept ordinary applications, train the people involved and keep records that regulators can inspect.

The Phone Call That Bypasses MFA

The security story today starts with an old tool: a convincing phone call.

Google’s Threat Intelligence Group says an actor it tracks as UNC6671 is using several extortion brands while targeting financial, legal and professional-services organisations. The campaign relies on vishing, meaning voice phishing. Callers impersonate IT support, reach employees on their personal phones and direct them to lookalike pages for passkey or multi-factor authentication enrolment. Those pages intercept credentials and authentication tokens.

Once inside, the actors use automated scripts to pull data from cloud services including Microsoft 365 and Okta. By July, Google had observed targeting shift towards private equity firms, law firms and financial-rating organisations. The objective is extortion, and the multiple brand names make the operation look fragmented or repeatedly rebranded.

Google found overlaps in infrastructure and phishing templates, but it hasn’t concluded that every brand is one tightly coordinated group. Splinter groups or several operators sharing the same phishing-as-a-service infrastructure remain plausible. That attribution limit is worth preserving because defenders can act on the technique without pretending the organisational chart is settled.

The uncomfortable part is that passkeys and MFA can be present and the organisation can still lose a session. The attacker isn’t necessarily breaking the authentication technology. The attacker is persuading a person to enter a controlled re-enrolment flow, then stealing the resulting credentials or tokens. More awareness training on its own won’t close a path that also depends on device trust, session handling and helpdesk procedure.

For organisations, my assessment is that phishing-resistant authentication remains valuable, but it needs supporting controls. Managed devices can narrow where enrolment occurs. Session and network policies can limit what a stolen token can reach. Helpdesk processes can make an unsolicited call to a personal phone an obvious break from normal practice. Monitoring bulk access to services such as Microsoft 365 and Okta can also expose the automated exfiltration after entry.

The practical shift is from asking whether MFA is enabled to asking how identity can be re-enrolled, where sessions can run and what happens after a successful login. The campaign succeeds in the seams between those controls.

LightSpy Moves Beyond Phones

One more security item widens the lens from stolen cloud sessions to surveillance across whole device fleets.

Researchers at Arctic Wolf report that the China-linked LightSpy spyware platform has expanded beyond its earlier geographic footprint and now reaches at least thirteen countries, including the United States. They say it can compromise smartphones, Apple devices, Windows and Linux systems, as well as routers. That last category changes the defensive problem because an infected router can expose other devices sharing the network.

According to the investigation, LightSpy can collect locations, messages, screen recordings and passwords, and can wipe a compromised device. Arctic Wolf identified infrastructure involving at least one hundred and seventeen servers. It describes LightSpy as a commercial platform offered to government, military and enterprise customers, rather than tooling confined to one operator.

There are limits around those findings. The victim count, customer list and commercial operating model haven’t been independently confirmed. The China-linked description and the view of LightSpy as a product come from Arctic Wolf’s analysis, so they should be treated as research findings rather than a complete public accounting of the operation.

For defenders, the immediate consequence is concrete: hunting for LightSpy as though it were only mobile spyware can leave major gaps. Endpoint coverage has to include several operating-system families, and network equipment belongs in the investigation as a possible foothold and surveillance point.

My larger concern is the reported commercialisation. If advanced cross-platform surveillance is packaged for multiple customers, capability can spread beyond the team that originally developed it. That makes consistent patching, router visibility and evidence sharing more valuable, because defenders may encounter variations of the same platform in campaigns with different targets and operators.

A Local AI Stack for Mac Apps

For developers, there’s a more constructive local-computing story, although it isn’t ready to ship.

MacPaw has partnered with Liquid AI to develop on-device inference and local memory for Eney, MacPaw’s assistant. The planned inference system, called Elix, is intended to run Liquid AI models locally on Apple hardware. MacPaw says that could support offline workflows and keep more processing on the device, while still allowing selected cloud models where they’re useful.

The company also plans to expose the stack eventually to developers distributing apps through Setapp. If that arrives, a Mac developer could gain a packaged route to private, offline AI features without building and operating an inference layer from scratch. Local execution can reduce network latency, continue without a connection and keep some sensitive inputs away from a remote model provider. A local memory layer could also let an assistant retain useful context on the machine.

But the key word is eventually. Developer access isn’t generally available, and MacPaw hasn’t announced commercial terms, a firm release date, supported model list, hardware requirements or performance results. The continued use of selected cloud models also means local-first won’t automatically mean every request stays on the device. The final privacy boundary will depend on product design and disclosure.

My advice to working developers is to watch this as a platform option, not plan production work around it yet. The partnership makes private and offline inference newly plausible inside the Setapp ecosystem, but APIs, device coverage and pricing determine whether it’s genuinely easier than using Apple’s own frameworks or assembling another local stack. Until those details are public, Elix is an architectural direction with attractive benefits, not a dependency you can safely adopt.

What Changes for You

And one change you may actually notice while planning a night out is landing inside Google Maps.

Google has expanded Ask Maps with agent-like features that move beyond finding places. For eligible users, a Maps conversation can assemble a food order, compare hotel availability and surface links for event tickets. It can retain context from earlier in the conversation, and Google is also adding a live-transit widget.

The transaction boundary is important. A prepared food cart hands payment to supported services such as Square, Toast or Uber Eats. Hotel bookings and ticket purchases finish on partner websites. Ask Maps is reducing the work between deciding and buying, but it isn’t becoming the merchant of record for the final payment. Google also hasn’t shown how consistently the system will find complete inventory or avoid mistakes across participating businesses. Check the cart, dates, room and price before paying.

These transactional features are rolling out in the United States, so they’re not a global change yet. Optional Personal Intelligence can use Gmail and Google Calendar to tailor results, but it’s off by default and is only rolling out where Ask Maps itself is available. That default is welcome, because the convenience depends on giving one service more context about your plans and account data.

For an ordinary user in the rollout, what becomes easier is the awkward middle of planning: comparing choices, remembering constraints and preparing the next step without bouncing through several searches. The limitation is partner coverage, regional availability and the privacy trade involved in personalisation. My judgment is that this is a useful agent pattern precisely because it stops before final payment. You get a prepared action while keeping a clear moment to inspect the details and choose whether to complete it with the partner.