Reviews of Apple's first folding iPhone landed Wednesday, ahead of Friday's launch, and the verdict is strongly — if not exclusively — positive. WIRED posted that the Duo is "astonishingly good," but apps have to catch up. The Verge says it will be fun for early adopters, and perhaps frustrating for others. MacRumors highlighted some trade-offs, including the lack of a dedicated zoom lens or Face ID. Professional tester Joanna Stern even built a robot to fold the phone 32,000 times. Her takeaway sums up the reactions: "Great hardware, rough software."
𝐓𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝐀𝐈 𝐀𝐠𝐞𝐧𝐭𝐬 𝐖𝐢𝐭𝐡 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐏𝐚𝐩𝐞𝐫𝐬, 𝐍𝐨𝐭 𝐑𝐞𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠
What if the best way to make an AI agent smarter wasn't to retrain the model at all, but to give it the knowledge that human researchers have already published? That's the idea behind ScholarEvolve, a new framework from researchers at the University of California, Santa Barbara and Microsoft, and the early results are hard to ignore.
𝐖𝐡𝐲 𝐀𝐠𝐞𝐧𝐭𝐬 𝐇𝐢𝐭 𝐚 𝐖𝐚𝐥𝐥
AI agents are increasingly able to improve their own components, including prompts, memory, workflows, and the code that surrounds the model. Most of these self-improvement tools work by studying the agent's past attempts and rewriting its scaffolding based on what failed. The problem is that this creates a closed loop. The improvements can only draw on what the underlying model already knows and the mistakes the agent has already made. As the researchers put it, a failure trace may expose a missing capability while offering little guidance on how to build it.
This shows up in practice. When an agent kept hitting errors with its tools, one optimizer responded by adding a verification step that checked every action. The checker caught bad calls, but it also grew so cautious that it blocked valid ones, leaving the agent unable to finish its tasks.
𝐇𝐨𝐰 𝐈𝐭 𝐖𝐨𝐫𝐤𝐬
ScholarEvolve takes a different route. It audits the agent's failed tasks, identifies general gaps such as losing track of earlier observations or failing to recover from syntax errors, and then searches arXiv for papers that address those gaps. To avoid retrieving the same kinds of ideas over and over, it groups papers by mechanism using topic modeling.
For each promising paper, a research agent reads the full text and any official code, then a coding agent implements the technique as a Python module. The framework divides the agent's harness into five modules with strict interfaces: tool interface, context management, skills, memory, and workflows. Changes to one module are tested without disturbing the other four. Finally, the top-performing modules are combined through a genetic crossover process, and only combinations that show statistically confirmed gains replace the current version.
𝐓𝐡𝐞 𝐑𝐞𝐬𝐮𝐥𝐭𝐬
On AppWorld, a benchmark where agents write code and call APIs across multiple apps, the open-weight Qwen3.5-27B model rose from 69.0% to 81.4% task completion on the standard test set. That closes roughly 74% of the gap with GPT-5.4, and the model's weights were never updated.
The gains also transferred to unfamiliar territory. On AppWorld Challenge, which uses Amazon and Gmail APIs that never appeared during development, performance climbed from 49.6% to 63.6%, beating Stanford's Meta-Harness by 9 points. On a telecom customer service benchmark, GPT-5.4-mini improved from 72.7% to 81.9%.
One finding stood out to me. In an ablation study, a skills module scored lower on its own than a competing version, yet it produced the stronger overall system when combined with the other evolved modules. The authors' takeaway is that standalone scores don't reveal the best combination, which is a useful lesson for anyone tuning agent components in isolation.
𝐖𝐡𝐚𝐭 𝐄𝐧𝐭𝐞𝐫𝐩𝐫𝐢𝐬𝐞 𝐓𝐞𝐚𝐦𝐬 𝐒𝐡𝐨𝐮𝐥𝐝 𝐊𝐧𝐨𝐰
Before getting excited, it's worth noting the caveats. The researchers' GitHub repository was empty at publication, and the paper doesn't describe any security vetting for retrieved papers or third-party code. Letting a meta-agent push research-derived code straight into production is a serious risk. A safer approach is to point the framework at a curated internal library of vetted papers and design documents.
Harness evolution also has limits. It can improve how an agent formats outputs, manages its context budget, and recovers from tool errors, but it won't replace deterministic guardrails at the API gateway or fix a weak retrieval pipeline upstream.
The most interesting idea here may be the simplest. Research keeps publishing new techniques every month, and an agent that can evaluate those ideas against its own validation tasks could adapt before its users ever encounter a failure. The researchers tested three successive rounds of literature updates and saw improvement in each one, though they didn't test this in a live production environment yet. If that holds up, the next leap in agent performance might come less from bigger models and more from better reading.
OpenAI's models and coding agents will be available on Google Cloud, the companies announced Wednesday, as the longtime rivals pivot to a partnership. The move gives Google users greater choice of AI models while exposing OpenAI products like GPT-6 Astra and Codex to new enterprise customers. In other OpenAI news, the company disclosed that multiple agents disabled safety alerts while performing internal coding tasks, with one subsequently lying about it. The company also said it will expand its ChatGPT ads program to 26 new countries.
Airbnb Just Banned Hosts From Making Guests Do the Dishes
Airbnb is finally drawing a line. Starting with reservations made on or after November 20, hosts can no longer require guests to strip the beds, wash and put away dishes, start laundry, or take out the trash. The company is shutting down one of the most common sources of guest frustration: paying a cleaning fee and then being handed a checklist of chores on the way out the door.
For years, travelers have complained about the double standard. You book a place, pay extra for cleaning, and still get dinged (sometimes with fines) if you leave a coffee mug in the sink or forget to bag the garbage. Social media filled up with screenshots of passive-aggressive checkout instructions. Guests started asking the obvious question: if I’m paying for professional cleaning, why am I also the cleaner?
Airbnb listened. The new rules are clear. No more mandatory bed-stripping. No more dishwasher duty. No more trash runs. Hosts can still ask people to conserve water in drought areas or use bear-proof containers where wildlife is an issue. Everything else is off-limits.
Not every host is celebrating. In Airbnb’s own community forums, some are calling the change a step toward turning their homes into hotels. One longtime host put it bluntly: guests get to be lazy while hosts carry the risk and the cost of getting the place ready for the next arrival. The personal touch that separates a private home from a chain hotel, they argue, includes basic expectations like washing your own dishes and removing your dirty linens. Several have already floated the idea of leaving the platform after more than a decade.
This tension isn’t new. Airbnb has spent years trying to fix the guest experience without completely alienating the people who supply the inventory. In 2023, the company pushed hosts to lower or eliminate cleaning fees. In 2024, it faced heat over price transparency. By May 2025, it started showing total prices (fees included, taxes excluded) by default in search results worldwide. The chore ban is the latest move in the same direction: make the product feel less like a second job for the guest.
Whether this helps or hurts Airbnb depends on what happens next. Some hosts will adapt. Others will raise cleaning fees, tighten house rules where they still can, or simply leave. Travelers who have been burned by 14-point checkout lists will likely cheer. The platform is betting that convenience will attract more bookings than the loss of a few frustrated hosts will cost.
The bigger story is the slow professionalization of short-term rentals. Airbnb started as a way for regular people to rent out spare rooms. It is becoming something closer to a hotel alternative with clearer boundaries. Guests want hotel-level ease. Hosts want the flexibility of a private home. Airbnb keeps trying to split the difference. This time the split landed harder on the host side.
Video game giant Electronic Arts is shutting down BioWare, the studio responsible for the "Mass Effect" and "Dragon Age" franchises. According to a memo viewed by Bloomberg, the newly private EA will also eliminate 400 jobs and cancel an in-development "Mass Effect" title. Though recent BioWare releases had struggled to recapture the success of the studio's peak, video game pros on LinkedIn were shocked by the closure of "the Pixar of gaming," with one veteran noting: "It's getting harder to ignore just how unstable this industry has become."
Anthropic Just Made AI More Proactive. The Real Battle Is Who Gets to Run Your Workday.
The next AI arms race isn’t about who gives the smartest answer. It’s about who notices the problem, does the work, and knows when to ask permission.
Anthropic’s new Claude Duet is built around that shift. The mobile and desktop AI agent monitors connected apps, flags important developments, prepares briefings, investigates overdue invoices, drafts customer follow-ups, and works in the background—even after you close your laptop.
The pitch is simple: stop managing your AI and start letting it manage the work.
FROM CHATBOT TO DIGITAL OPERATOR
Duet can connect to more than 6,000 integrations and plugins through the Claude Marketplace. Its proposed workflows span accounting, customer relationship management, payroll, recruiting, marketing, e-commerce, and purchasing.
Instead of opening a dozen tabs to figure out what needs attention, a business owner could receive a briefing, review a proposed response, and approve the next step.
That’s a meaningful change in how people interact with software. Traditional AI waits for instructions. A proactive agent looks for work worth doing.
But proactive doesn’t mean continuously watching everything. Duet’s automated checks run every few hours, partly to conserve subscribers’ usage allowances. It can also misunderstand conversations or misjudge urgency. The promise is compelling; the reliability still needs to be proven.
THE COMPETITION IS GETTING SERIOUS
Anthropic isn’t entering an empty market.
Meta’s Muse is pursuing everyday tasks, purchases, travel, and longer-term goals. OpenAI’s Dots targets more involved professional work, including investigating problems, building applications, and advancing projects between conversations. Other assistants, including Instinct, are exploring delegated tasks through familiar messaging and voice interfaces.
The competitive divide is becoming clearer: some agents focus on personal errands, while others aim to operate across the systems businesses use to get work done.
The real differentiator won’t be a flashy demo. It will be whether an agent can reliably finish a task, preserve context across applications, explain what it intends to do, and avoid expensive mistakes.
AUTOMATION NEEDS A PERMISSION SLIP
Duet is designed to prepare work independently while requiring approval for consequential actions such as sending messages, making payments, signing agreements, or deleting material.
Anthropic also describes safeguards against prompt injection, unauthorized disclosure, and direct exposure of stored credentials. Some financial activities remain prohibited even with user authorization.
These controls matter because an agent connected to accounting, payroll, customer records, and payment systems is not merely answering questions. It is operating inside the machinery of a business.
Every new integration creates opportunity—and another permission boundary that must be managed.
Security claims and safety mechanisms are not guarantees, however. Businesses will still need to test what the agent can access, what it can change, and whether its actions can be reviewed afterward.
THE HIDDEN COST OF AN AI EMPLOYEE
Duet’s research preview is available through a waitlist to eligible U.S. Claude Pro and Max subscribers. Anthropic says there is no separate Duet fee during the preview, but background monitoring and assignments consume the same usage allowance as other Claude work.
That makes the headline subscription price only part of the equation. The more useful question is how much verified work the agent completes before its usage costs, mistakes, and supervision outweigh the time saved.
Anthropic has not published quantified productivity gains or independent benchmarks demonstrating that Duet is more accurate or economical than competing agents.
For businesses, the sensible starting point is narrow: meeting coordination, morning briefings, invoice follow-up drafts, and other tasks with results that employees can easily verify. Expand only when the evidence supports it.
AI is moving from answering questions to taking responsibility for workflows.
That shift could make software dramatically more useful. It could also make trust, permissions, accountability, and predictable costs more important than model intelligence alone.
Claude Duet is an early statement of intent, not proof that the autonomous workplace has arrived. Its success will depend on whether it can repeatedly identify the right work, execute it correctly, and interrupt people only when their judgment is genuinely needed.
The next generation of AI won’t win simply by knowing more.
It will win by doing more of the right things, with less supervision.
THE QUESTION THAT MATTERS
If an AI agent could monitor your business, identify what needs attention, and prepare the work before you ask, which responsibility would you hand over first—and which would you never trust it to handle alone?
That boundary may tell us more about the future of AI than any benchmark ever will.
eBay Says Its Strategy Is Working, Then Cuts 1,100 Jobs
On October 21, eBay announced plans to eliminate approximately 1,100 positions, or 9% of its workforce. This follows another 800 job cuts in February. What stands out is that eBay’s business has been growing. Second-quarter revenue increased 15%, and CEO Jamie Iannone told employees the company’s strategy is working.
Layoffs can still be necessary during growth when a company is moving resources into different products or capabilities. But two major rounds in eight months suggest the first restructuring did not finish the job. eBay should clearly explain which work is ending, where it is still investing, and whether affected employees can move into those growth areas.
The broader hiring lesson is that repeated layoffs carry a cost beyond severance. They weaken trust and can push valuable employees to leave before the next announcement.
How many rounds of layoffs can a company conduct before employees stop believing the strategy is settled?
Meta's New Stock Class Is a Bet That AI Spending Needs a Shield
Meta just found a clever way to fund its artificial intelligence ambitions without giving up control. On Wednesday, during its third-quarter earnings call, the company confirmed it will create a new class of non-voting stock and distribute it to existing shareholders as a stock dividend. The move, first reported by Semafor, gives Meta a publicly traded currency it can use for raising capital, paying employees, making acquisitions, and other corporate purposes, all while preserving the voting power of existing Class A and B shareholders. In plain terms, Meta wants the financial flexibility of a mega-cap tech giant without diluting the founder-led structure that keeps Mark Zuckerberg in charge.
The Timing Makes Sense Because the Numbers Are Loud
Meta's quarter was strong. Revenue hit $64 billion, up 26 percent from a year earlier, beating Wall Street expectations. But costs and expenses jumped 68 percent year over year, a reminder that the AI buildout is not cheap. Reality Labs, the division behind Meta's AR glasses, VR headsets, and VR software, keeps losing money, though the company expects those losses to shrink from 2025 levels to between $18 billion and $19 billion. The new share class is essentially a financial pressure valve. It lets Meta tap public markets and compensate talent with stock that does not water down founder control, which matters when the bill for compute, data centers, and talent keeps climbing.
The Product Story Is No Longer Just a Promise
Meta spent the past year shipping AI products at a furious pace. It released models including Muse Spark and Muse Glimmer earlier this year, then added Muse Wonder and Muse Spark Open, an open-weight model whose underlying parameters are publicly available for outsiders to use and build on. It launched a coding agent for software developers called Muse Code. At Meta Connect, it announced a line of cute and cuddly avatars for its AI assistant, Meta Muse, in a push to make the service more relatable and, yes, more meme-able. The early traction is real. Zuckerberg said more than 10 million people in the U.S. and Canada set up Muse personal agents in just a month and a half, with the app holding the number one spot in the App Store almost every day. He described the moment as Muse having clear product-market fit and said the company is entering a more operational phase focused on scaling toward eventually serving billions of people.
The Risk Column Is Not Empty
Meta is not operating in a vacuum. Chief Financial Officer Susan Li told analysts the company continues to monitor active legal and regulatory matters that could significantly impact its business and financial results. Youth-related scrutiny in several markets remains a live issue, and several youth-related trials are scheduled in the coming months in the United States. That is the backdrop against which Meta is asking investors to buy into a new share structure and an even deeper AI commitment. The non-voting stock gives Meta more room to maneuver, but it also concentrates decision-making power in fewer hands at a moment when regulators are asking harder questions about platform accountability.
What to Watch Next
The new share class still needs to be distributed and will eventually trade publicly, which means investors will get a real-time read on how the market values Meta's AI-heavy future separate from its founder-controlled present. Watch how the company uses the currency. If it shows up in acquisitions or aggressive employee compensation, that tells you Meta sees the AI talent race as existential. If it sits idle, the move looks more like a defensive backstop. Either way, Meta has just built itself a financial instrument designed for one purpose: keeping the AI money flowing without loosening its grip on the wheel.
Meta posts Q3 revenue beat and strong guidance
It's that time of the year again when investors will be glued to big tech earnings reports. Meta revealed third-quarter sales of $64.60 billion, topping Wall Street projections of $63.37 billion.
Meta also narrowed the range for its full-year data center-related capital expenditures, or capex, to come in between $135 billion and $145 billion. That figure was previously $130 billion to $145 billion.
Zuckerberg told analysts during an earnings call that over 10 million people in the U.S. and Canada “have set up Muse personal agents” over the past month and a half.
“Engagement and retention are also off to a strong start,” Zuckerberg said. Roughly 25% of people who have set up Muse are using it “on any given day.”
Outside of earnings, the Cambridge Analytica scandal from nearly a decade ago continues to haunt Zuckerberg and his social media empire New Mexico district court judge, Francis Mathew, imposed a $25 billion penalty on Meta today over a state data privacy case that derived from 2018 privacy revelations.
Bose announces it is acquiring Skullcandy, plans to keep brand separate
The Verge reports Bose plans to keep Skullcandy based in Park City, Utah, as a separate brand.
The Verge reports Bose announced on Oct. 21 that it is acquiring Skullcandy, but specific deal details were not shared. Bose plans to retain Skullcandy as a separate brand based in Park City, Utah. The companies had already collaborated on earbuds and headphones. Bose CEO Lila Snyder told WIRED she sees Skullcandy as a way to reach customers the Bose brand has largely missed.