AI agents can turn customer requests into preview branches in minutes

At Braintrust, an engineer writes a failing test, opens a sandbox, and lets Codex build from the real request. The customer can see a live version while the pain is still fresh, not a ticket waiting in a backlog.

It still stops at review. Humans judge fit, security, and edge cases before anything ships.

Codex is turning office work into governed agent workflows

OpenAI is pushing Codex beyond software teams. In its OpenAI announcement, the company says Codex has more than 5 million weekly users, and about 20% are non-developers growing more than 3x as fast as developers.

The practical change is not another chat box. Analysts, marketers, sales, design and finance teams can package repeatable work into role-specific plugins and shareable Sites, now in preview for Business and Enterprise customers: customer review hubs, scenario planners, campaign boards, dashboards, or annotated slide decks. The output becomes a shared internal tool, not just a private prompt result.

The hard part moves to governance. Which apps can Codex connect to, what data can it read, and when must a human approve the answer before it reaches a client or customer?

AI agents are being tested as shadow managers for real shops

Andon Labs gave agents leases, cards, email, Slack and supplier accounts for cafe and store tests. One read a lease, built the opening checklist, hired baristas and handled orders while people did the physical work.

Then it missed deadlines and ordered 120 eggs for a cafe with no stove. The safer trial is not an AI boss. It is an AI making parallel decisions until managers see which small loops can run under review.

Ask AI to make your home readable before an emergency

Emergency prep usually fails because the useful facts are physical, scattered, and stored in one person's head: where the water valve is, which drawer holds passports, which pet needs medicine, which door sticks, which neighbor has the spare key.

That is a job you can hand to AI while life is boring. Record a slow phone walkthrough, photograph utility panels and shutoffs, add medication and pet notes, emergency contacts, key document locations, local hazard pages, and a rough floor plan if you have one. Ask it to turn the evidence into a room-by-room emergency packet: shutoff map, go-bag gaps, evacuation options, missing photos, and a 20-minute family drill.

Notice what changes here. You are not asking a chatbot for disaster advice. You are using a multimodal assistant as a clerk that reads your actual house, connects visible evidence with rules and links, and writes something another adult could follow when you are not there.

Keep the boundary sharp: AI should flag uncertainty, not invent certainty. Verify gas, electrical, water, medical details, routes, and official alerts yourself; control where home videos, IDs, contacts, and medication notes go. In a real emergency, the packet helps only if people have already checked it and know official instructions win.

#EmergencyPreparedness

After crashes and storms, voice AI is taking the first claims call

Travelers used to need more people on phones when claim reports surged. Now its assistant asks for damage details, answers policy questions, starts photo upload, and hands unclear cases to specialists.

It moved from 8 states to countrywide in two months. Among users, 85 to 90% complete filing through AI.

AI cyberattacks are moving inside the network

The familiar warning was phishing: AI helps criminals write better bait. Anthropic's report, based on 832 accounts banned between March 2025 and March 2026, shows the harder shift: risky actors increasingly used AI after access for account discovery, credential hunting, privilege escalation, lateral movement, and chained tool use.

That changes defensive work. SOC teams, cloud admins, and infrastructure operators need to monitor agent behavior, not just generated messages: which tools a model calls, what systems it explores, and where approval gates stop it.

This is Anthropic platform data, not the whole cybercrime market. Still, the signal is practical: once AI can guide live operations, permissions, logs, and human approval become core security controls.

Anthropic has started its IPO path and the fight with OpenAI is now about who gets public money first for the next big compute cycle

On June 1, Anthropic confidentially sent a draft S-1 to the SEC. This is not a final IPO date yet and Anthropic has not set the share price or share count but the signal is loud. Claude is moving toward Wall Street while OpenAI is also preparing its own IPO process, according to Axios. Whoever prices first can become the first public frontier AI lab and raise cash from a much wider investor base. That cash matters because the real bottleneck is chips, power, data centers, and long compute contracts.

Anthropic is valued at $965B but OpenAI at $852B.

Anthropic says it signed for up to five gigawatts with Amazon, five gigawatts (TPU) with Google and Broadcom, and GPU access from SpaceX Colossus. It also brought Micron, Samsung, and SK hynix into the funding round.

This IPO is not only a stock story. It is a race to finance factories for intelligence.

A laptop AI can now see and hear

Google just released Gemma 4 12B, an open model that handles text, images, video, and audio while aiming to run on consumer devices.

That moves local AI from chat box to tiny project brain. Think screenshot Q&A, game clip tagging, voice note summaries, or study help that stays near your files.

The catch is speed depends on hardware, and local still needs human review before it touches private stuff.

Ask AI to move your digital history before a bad app keeps it hostage

People do not stay in old apps only because they like them. They stay because the export looks like a junk drawer: CSVs, JSON, loose attachments, timestamps, tags, broken links, and fields the new app does not understand.

That is a real assignment for AI. Give it the old app export, a small sample of good records, the new app's import docs or API schema, validation rules, and a strict read-only workspace. Ask it to build the migration map: which fields match, which need transformation, which attachments have no home, which tags collide, which records look duplicated, and which items must not be imported yet.

The useful output is not "done." It is an audit you can inspect before anything permanent happens: counts before and after, sample comparisons, fields that changed shape, lost metadata, confidence notes, and a plain-English list of records needing human review. Suddenly an export is not a dead ZIP. It is evidence for a decision: can I leave this app without damaging my own history?

This changes the mental model of AI. The assistant is not just answering questions from your data. With files, schemas, docs, scripts, and permission boundaries, it can become a supervised moving crew for your notes, workouts, tasks, photos, receipts, or research library.

The boundary matters because this is your life in machine-readable form. Keep the original export, run dry migrations first, prefer local processing for sensitive data, and do not give the agent unsupervised write, delete, deduplicate, password, financial, health, or identity access. You approve the move. AI shows where app lock-in is hiding.

#DataMigration

Customer chats are starting to book real work, not just answer questions

A salon owner used to wake up to missed WhatsApp leads. Meta says earlier agents were already used by more than 1 million businesses. The new version can answer in the shop's style, recommend from a catalog, qualify a lead, and book a slot.

The rollout starts limited, and prices, refunds, complaints, and odd payments still need a human handoff.

I tested one dangerous-looking first-person video idea through four fal.ai models, compared price and realism, and learned why Seedance 2.0 still won my small experiment overall

I used fal.ai because I could test several video models from one service and keep the same queue flow. I wanted a hard starting prompt: one long first-person action shot, clear height, parachute, industrial yard, and a muddy landing.

I picked Seedance 2.0 as the cinematic baseline, Kling v3 Pro for action and smart shot planning, Veo 3.1 because it is the premium Google option, and Wan 2.7 as a cheaper 1080p challenger.

What I got: Seedance 2.0 was best. It kept the helmet-camera feeling, the hands, the parachute lines, and the story rhythm most clearly. My only real complaint was the realism of the fall into the puddle. It cost about $2.43 for 8s at 720p.

Kling v3 Pro gave me good image quality, but the actions were not realistic. With audio on, 8s cost about $1.34.

Wan 2.7 accepted the prompt, but the result was basically an absurd video made from disconnected cuts. At 1080p, 8s cost about $1.20.

Veo 3.1 was the tricky one. The first request with the same prompt hit a content policy check. Then I softened the wording a little, sent it again, and got a video. It lost the first-person view and switched to third-person, but the result still looked cool. That successful 8s 1080p run with audio cost about $3.20.

So my successful clips cost about $8.17 total if I count the second Veo run. My winner was not the cheapest one; it was the one that followed the whole idea with the least confusion.

See the result 👇

ChatGPT memory turns one-off chats into ongoing work

OpenAI has begun rolling out a stronger ChatGPT memory system to Plus and Pro users in the US. In its OpenAI announcement, the company calls the approach "dreaming": ChatGPT synthesizes useful context from past chats, such as projects, preferences, constraints, and details that may expire.

The practical shift is continuity. Planning a trip, writing a newsletter, studying, shopping, or managing a long project no longer has to start with the same background pasted again. For product teams, this makes memory a core assistant feature, not a side setting.

The boundary is consent. Memory can be stale, sensitive, or wrong. Users need reviewable controls and Temporary Chat for one-off topics; teams need retention and deletion rules before treating remembered context as trusted work infrastructure.

Empty shelves can now become AI-ranked tasks before managers walk the aisle

At Rainbow Department Store in China, a new store assistant connects smart shelves, carts, robots, POS data, and staff systems. It spots a wrong price, a missing product, or a display drift, then turns it into priority work for the right person.

The hard part shifts from finding problems to setting rules, proof, and human approval for what the store may fix automatically.

Ask AI what your forgotten documents reveal to an impersonator

A folder of old PDFs can be more revealing than a social profile. Give AI a tightly scoped, consented packet: an old lease, utility bill, resume, scanned form, public bio, maybe screenshots of data broker listings or account recovery pages. Its assignment is to act like a defensive red teamer and answer one narrow question: what identity story could someone assemble from this residue?

The useful output is not a privacy sermon. It is an evidence map: exposed fact, source file, confidence, abuse path, cleanup action, and what needs human review. The surprising part is how ordinary fragments combine. Address history from a lease, employer from a resume, signature style from a form, pet name from a profile, and partial phone numbers from bills can become enough material for fraud, social engineering, or weak security questions.

This is where AI stops feeling like a chatbot and starts looking like a worker with context. It can OCR messy scans, compare documents, follow public links, cluster repeated facts, and draft removal emails or a bank call script. The value is not that it knows privacy. The value is that it can inspect the pile you were never going to inspect.

The boundary is the whole point. Do not upload passports, medical files, bank statements, or family documents to an untrusted consumer bot. Use local or enterprise controls, redact what you can, get consent, and ask only for defensive cleanup. AI can surface exposure. You decide what to delete, dispute, report, freeze, or take to a professional.

#DataPrivacy

Your next AI project is not one prompt

Google showed the cooler version of AI art at I/O. Puppet footage became film shots. Prompted sprite sheets became a WebGL game world. Jellyfish tracking became music.

The move is the pipeline. Start with your own references, generate parts, drop them into code or an editor, then cut anything your taste rejects.

Microsoft Scout turns office AI into a coworker with an ID

Microsoft has introduced Scout, an always-on Microsoft 365 agent that can watch work context across Teams, Outlook, files, calendar and approved local resources, then coordinate meetings, prepare briefings, block time and flag stalled deliverables. In Microsoft's Scout announcement, the important detail is not another chat window: each Scout runs under its own governed Entra identity.

That changes the office AI workflow. Teams can start delegating recurring coordination work, while IT has to define what the agent may read, write, send or escalate, and where human approval is mandatory.

The limit is clear: Scout is experimental, only for select private-preview and Frontier organizations with extra setup and licensing. Always-on agents help only if logs, consent, data boundaries and human judgment are treated as part of the product.

AI agents are now doing the portal checks hidden inside gas operations

NatGasHub says it has deployed 1,000 agents across 300+ North American pipeline portals. Schedulers used to log into Electronic Bulletin Boards, copy data, watch changes, and build reports.

Now bots run those daily loops and send exceptions to people for review. The count is from the company, but the shift is concrete. AI is moving into the boring web work that kept operations teams busy.