This Week in Tech 130

What we explore this week:

0xwhrrari surfaced a clip of an Anthropic research lead explaining their internal engineering workflow: 99 percent of their engineers run swarms of 300 or more self-improving agents. The key design principle is closing the feedback loop. Give the model a way to verify its own output, and reliability improves dramatically without requiring constant human oversight.

This is documented in Anthropic's own research on building effective agents, but hearing it described as standard operating procedure at the company building Claude is a different kind of data point.

BullTheoryio flagged the launch of Claude Tag, Anthropic's persistent AI agent that lives inside Slack. Teams tag Claude in a channel, hand it a task, and it works autonomously in the background. No separate dashboard, no app switching, no reason to open half your current tool stack. The Claude Code team added their own data point: Claude Tag writes 65 percent of their product team's code, including most of what built Claude Tag itself.

Polymarket reported that OpenAI's GPT-5.5-Cyber beat Mythos 5 on the CyberGym benchmark, making it the current top performer on a leaderboard designed to measure AI capability in offensive and defensive security tasks. Purpose-built cyber models are outpacing general-purpose ones, which is the expected trajectory, but the speed of that gap widening is worth watching.

A model purpose-trained for cybersecurity tasks that beats the best general models is no longer a research curiosity. It's the kind of tool that changes the economics of both attacking and defending infrastructure. The race to build hacking-capable AI has real stakes, and the defensive side of that equation needs to keep pace.

2026 will redefine IT as a strategic driver of global growth. Automation, AI-driven support, unified platforms, and zero-trust security are becoming standard, especially for distributed teams. This toolkit helps IT and HR leaders assess readiness, define goals, and build a scalable, audit-ready IT strategy for the year ahead. Learn what’s changing and how to prepare.

Download the Toolkit

Sakana AI launched Fugu Ultra, a model that presents a clean single-model interface while quietly running a full multi-agent orchestration system underneath. The performance claim is frontier-level, comparable to top-tier models, packaged in a way that sidesteps export control concerns by abstracting the underlying complexity.

ElevenLabs launched the Ads Engine inside ElevenCreative, letting brands connect Google, Meta, and LinkedIn accounts, localize ad creatives across 50 or more languages using AI voice and translation, and push finished assets directly back to the ad platform. No export step, no agency middleman. The product page frames it as closing the entire loop between creation and distribution in one interface.

TechCrunch reported that Google DeepMind is committing $75 million to a partnership with A24, one of the most critically respected studios working today. The deal is about legitimacy as much as technology. Attaching AI filmmaking tools to a studio known for films like Everything Everywhere All at Once and Midsommar is a calculated move to change the cultural conversation about what AI-generated content can be.

GAMERTAGVR spotted Meta's unannounced mixed reality headset, codenamed Phoenix, appearing in a Qualcomm promotional video for their new Snapdragon Reality Elite XR chip. Chip partners don't put unannounced hardware in marketing materials unless a launch is close. Phoenix is expected to be a meaningful step up from the Quest 3, targeting premium mixed reality in a market Apple spent $3,499 to validate with Vision Pro.

Road to VR reported that Virtuix's Omni One VR treadmill now has official Meta Quest support, making it the first device of its kind to receive that certification. The gap between moving in VR and moving in the real world has been the most persistent friction point in consumer virtual reality. Official platform support removes the third-party workaround problem and puts full-locomotion VR on a path toward living rooms rather than just arcades.

Kotaku confirmed it first: GTA 6 starts at $80, with an Ultimate Edition landing at $100. Dexerto reported that pre-orders open June 25, complete with official cover art. And CGHchannel leaked the Ultimate Edition contents: Jason's Dinka Enduro motorcycle, a Crest Kayak, a 1995 Grotti Cheetah, retro builds, two exclusive mod shop passes, and a Shitzu Sqaulo boat. That's a lot of virtual property for a hundred dollars.

The pricing move is the real story. The industry spent years arguing over whether $70 was sustainable. Rockstar just answered that question by skipping the debate entirely and moving the number up again. No other studio has the leverage to do this cleanly. A game generating $7 billion in online revenue from a 2013 title, according to IGN's breakdown of GTA Online's economics, doesn't need to justify its price. It just sets one.

The Verge reported Valve's explanation for why it won't subsidize the Steam Machine the way Sony and Microsoft bleed cash on console hardware. Kotaku spelled out the price: $1,050, with Valve explicitly arguing the traditional console model of subsidized hardware plus locked-in software is a trap for consumers. PC Gamer tested the machine and delivered the verdict: at this price, the value proposition isn't there.

Space and Tech reported that Shanghai-based AgiBot launched a six-day global livestream of its humanoid robots working a real production line, performing tablet quality inspections alongside human workers in real time. This is a different kind of proof than a polished demo reel.

A sustained live stream on an actual factory floor means the failure modes, if they occur, are public. AgiBot raised $400 million to deploy at scale, and this stream is the most transparent accountability move any humanoid robotics company has made publicly.

The Verge covered a robotic, self-navigating toilet designed to autonomously position itself for users who have difficulty reaching traditional bathrooms. The accessibility application is genuine and addresses real limitations for aging and mobility-impaired populations, a need that CES 2025 highlighted as a growing focus for assistive robotics.

The punchline writes itself, but the underlying signal is worth taking seriously. When robotic mobility is being applied to problems this personal and this unglamorous, the technology has crossed from impressive demonstrations into solving actual friction in daily life.

Meta made three coordinated moves on smart glasses this week. Road to VR reported that Meta is launching unbranded smart glasses starting at $300, dropping the Ray-Ban and Oakley co-branding in favor of lower price and wider reach. The Verge covered the companion announcement: a dedicated charging stand, the kind of bedside accessory that signals a company expects the device to become part of your daily routine rather than a gadget you try once.

Dexerto reported on the Commodore Callback, a flip phone priced at $499 that blocks social media and web browsers while keeping utility apps like WhatsApp, Telegram, Google Maps, and Uber. It's a product that sells itself on what it removes. The Guardian has tracked the dumb phone trend for a while now, and it's growing precisely because smartphones have become environments that many people find genuinely harmful to their attention and wellbeing.

Three separate announcements arrived from Midjourney in quick succession, and together they sketch something larger than any one of them alone. Midjourney's official account announced a dedicated medical division called Midjourney Medical, applying the same generative model expertise behind their image tools to radiology, pathology, and medical visualization.

Separately, they detailed a proprietary hardware scanner designed to capture real-world objects, textures, and environments at high fidelity to improve training data quality. And then the scanner's ambitions got specific: fewer than a dozen of these machines operating together can, by the company's claim, outpace every MRI machine on Earth in full-body scan volume. The stated goal is a fleet of 50,000.

Seth Bannon highlighted work from Nabla Bio that deserves more attention than it's getting. Their AI zero-shot designed antibodies capable of killing cancer cells at 70 picomolar concentrations while sparing healthy cells distinguished by a single amino acid difference. The structures folded exactly as predicted, within an angstrom of accuracy. These are validated results, not simulations.

Polymarket reported that Waymo is recalling nearly 4,000 robotaxis after some drove into active freeway construction zones at highway speeds. The Verge's coverage explains what went wrong: real-world construction configurations that fell outside the edge cases the system was trained to handle. The vehicles behaved predictably given their training. The training just hadn't seen this.

Two themes are running underneath this week's stories, and they're heading in opposite directions. On one track, capability is compressing timelines faster than institutions can adapt. Nabla Bio designs cancer-killing antibodies with atomic precision on the first try. Arc Institute ships a programming language that turns biological intuition into executable code. Anthropic's engineers run 300-agent self-improving swarms as standard practice.

These are not incremental improvements. They are the kind of step changes that tend to look obvious in retrospect and invisible in the moment they happen. The question worth sitting with is not whether these tools work. It's whether the regulatory, ethical, and organizational infrastructure around them can keep pace with what they're about to enable.

On the other track, the physical world is learning to move. Humanoid robots are clocking real hours on factory floors, not in demo environments. Waymo's recall, frustrating as it is, represents a safety and feedback loop that improves with every edge case encountered. NVIDIA is building the safety standard that will govern physical AI before the market gets crowded enough to demand one.

Midjourney, of all companies, is building hardware scanners that could make preventive health monitoring as routine as a blood test. The through line is that robots and AI are moving from impressive to operational, and the friction points are shifting from capability to trust, liability, and deployment logistics.

The GTA 6 pricing story is worth reading as a cultural data point alongside the tech stories. Rockstar can charge $80 because their product is genuinely irreplaceable to tens of millions of people. That's a bar almost nothing in consumer tech clears. Meanwhile, the Commodore Callback sells $499 flip phones by removing internet access, and people are buying them.

Both stories point to the same underlying tension: the more capable and connected our tools become, the more intentional we have to be about when and how we use them. The market is starting to price that intentionality.

The convergence of AI, biology, and hardware this week feels less like separate stories than a single shift in what's buildable. When a generative AI image company pivots to whole-body scanners, announces a medical division, and attracts biohackers who want weekly scans as personal data points, you're watching a category boundary dissolve in real time.

The companies that will matter in ten years are the ones building infrastructure at these intersections now, before the categories have names. This week gave us a clearer look at what that infrastructure is starting to look like.

Attio is the AI CRM that keeps you ten steps ahead.

Ask Attio anything. Where should I focus? What deals are at risk? Search, update, and create across your customer data.

Ask more from CRM. Ask Attio.

Try Attio for free

HubSpot switched to Mintlify and saw 3x faster builds with 50% fewer eng resources. Beautiful, AI-native documentation that scales with your product — no custom infrastructure required.

Simplify Your Docs Today