<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Rohan Tech official</title>
    <description>The latest articles on DEV Community by Rohan Tech official (@rohan_techofficial_81a0f).</description>
    <link>https://dev.to/rohan_techofficial_81a0f</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4209917%2F3d353298-49a6-40d9-b605-933f526dd6c2.png</url>
      <title>DEV Community: Rohan Tech official</title>
      <link>https://dev.to/rohan_techofficial_81a0f</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/rohan_techofficial_81a0f"/>
    <language>en</language>
    <item>
      <title>The Linux Foundation Wants AI Agents to Carry a Verifiable ID</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Mon, 16 Nov 2026 04:46:43 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/the-linux-foundation-wants-ai-agents-to-carry-a-verifiable-id-1len</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/the-linux-foundation-wants-ai-agents-to-carry-a-verifiable-id-1len</guid>
      <description>&lt;p&gt;Your AI agent can book a flight, pay a supplier and message a client in your name. So how does the other side know it's genuinely yours, and not an impostor wearing your agent's name? On November 10, 2026, the Linux Foundation welcomed a project built around that exact problem: the Agent Name Service, or ANS.&lt;/p&gt;

&lt;p&gt;ANS is an open standard that gives AI agents a verifiable identity, along with discovery and authentication, and it does it using infrastructure the internet already trusts. Think of how websites earned credibility. DNS gave them names machines could resolve, and certificates let a browser confirm a site is who it claims to be. ANS anchors an agent's name to a domain in the same spirit, so the identity traces back to an organisation that controls that domain.&lt;/p&gt;

&lt;p&gt;The mechanics cover domain-anchored naming, registration, DNS publication, public-log auditing, integrity monitoring and agent-to-agent authentication. In practice, that means an agent's record can be looked up before anyone trusts it, changes to that record leave a public trail, and two agents meeting for the first time can check each other before money or data changes hands.&lt;/p&gt;

&lt;p&gt;Why now? Agents are being handed real authority — spending money, signing off on orders, speaking to customers — while identity checks lag behind. A verified name doesn't make an agent smart, but it makes impersonation far harder, and impersonation is exactly the trick attackers reach for first.&lt;/p&gt;

&lt;p&gt;The support list is hard to ignore. The project launched with seven specification documents and 11 supporting organisations, including Cisco, Cloudflare, DigiCert, GoDaddy, Salesforce and OWASP. Networking, DNS, certificates and security are precisely the skills an identity system like this needs. The announcement went out through PR Newswire on November 10, and Unite.AI and ITBrief covered it the same week.&lt;/p&gt;

&lt;p&gt;One honest caveat before anyone gets carried away: the specifications are still labelled drafts. Details can shift, implementations are early, and the standard only becomes real if agent platforms and businesses actually adopt it.&lt;/p&gt;

&lt;p&gt;I wrote up the full launch, including why identity is fast becoming the missing layer in agent security, here: &lt;a href="https://rohantechofficial.com/tech-news/linux-foundation-agent-name-service/" rel="noopener noreferrer"&gt;Linux Foundation Launches Agent Name Service to Give AI Agents a Verifiable Identity&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>security</category>
      <category>opensource</category>
      <category>news</category>
    </item>
    <item>
      <title>A 145B Coding Model You Can Actually Run at Home</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Mon, 16 Nov 2026 04:45:50 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/a-145b-coding-model-you-can-actually-run-at-home-1fn0</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/a-145b-coding-model-you-can-actually-run-at-home-1fn0</guid>
      <description>&lt;p&gt;Cloud coding assistants come with a quiet bargain: your code travels to someone else's servers, and you pay for every token on the way back. Edge0 wants to cancel that bargain. The company has announced Edge0-145B, alongside a smaller Edge0-120B — downloadable models aimed at coding, reasoning and agent tasks, designed to run on your own machine.&lt;/p&gt;

&lt;p&gt;Here's the puzzle. The model holds 145 billion parameters, yet Edge0 reports peak memory of about 29 GB and speeds above 80 tokens per second on compatible NVIDIA hardware. How?&lt;/p&gt;

&lt;p&gt;The trick is sparsity. Edge0-145B uses a mixture-of-experts design, which you can picture as a newsroom of specialists rather than one giant brain. A router hands each step to just a few expert sub-networks, so only around 10 billion parameters are active at any moment. The rest stay idle. You keep the knowledge of a huge model while paying the compute bill of a far smaller one — and that active count is exactly what decides generation speed.&lt;/p&gt;

&lt;p&gt;For developers, the local angle cuts two ways. First, privacy: your codebase never leaves your hardware, which matters when the code belongs to a client or an unreleased product. Second, flow: an assistant answering at 80-plus tokens a second keeps pace with your thinking, and agent workflows that chain dozens of steps don't grind down under slow generation.&lt;/p&gt;

&lt;p&gt;29 GB of memory still rules out a modest laptop, mind you. This is workstation territory. But it's a very different proposition from renting a data centre by the token.&lt;/p&gt;

&lt;p&gt;One warning before anyone gets excited: the licensing is restrictive. Downloadable doesn't mean do-as-you-please, and the terms deserve a careful read before this goes anywhere near commercial work. Specs are the fun part; the licence is the part that decides what you're actually allowed to build.&lt;/p&gt;

&lt;p&gt;I covered the announcement in full, including the two-model lineup and the licence caveat, in my write-up: &lt;a href="https://rohantechofficial.com/tech-news/edge0-145b-local-coding-model/" rel="noopener noreferrer"&gt;Edge0-145B Brings a 145-Billion-Parameter Coding Model to Your Own Machine&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If the reported numbers survive real-world testing, local models this capable stop being a hobbyist curiosity. They become a genuine option for working developers — licence permitting.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>programming</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Microsoft 365 Is Back After a 38-Hour Global Outage — What Happened</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Mon, 16 Nov 2026 04:44:49 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/microsoft-365-is-back-after-a-38-hour-global-outage-what-happened-36mm</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/microsoft-365-is-back-after-a-38-hour-global-outage-what-happened-36mm</guid>
      <description>&lt;p&gt;For nearly two working days, the world's most-used office suite was limping. Microsoft has confirmed Microsoft 365 is fully restored after an outage that lasted close to 38 hours and reached more than 70 countries.&lt;/p&gt;

&lt;p&gt;The first signs appeared in Exchange Online on November 12 at about 16:17 UTC. From there, the problems spread to Outlook, Teams, Copilot and SharePoint. Full restoration landed on November 14 at 06:09 UTC, per Microsoft's confirmation, with coverage from Digital Trends, Financial Express, The Hindu and KeyXplore across November 14–15.&lt;/p&gt;

&lt;p&gt;What caused it? According to Microsoft, an issue during an earlier network capacity upgrade surfaced after a network change. The company has ruled out a cyberattack — and, notably, ruled out AI agents as the trigger. This one was an internal change that went sideways.&lt;/p&gt;

&lt;p&gt;There's a particular sting in that cause, too. A capacity upgrade is supposed to make a network stronger. When the upgrade itself becomes the fault, it shows how fragile tightly coupled systems can be — one change, and services on every continent wobble.&lt;/p&gt;

&lt;p&gt;Why does it matter? Thirty-eight hours is an eternity for tools that sit at the centre of the workday. Email queues stall, meetings slide, shared files stay out of reach, and staff improvise with personal accounts — a workaround with its own security headaches. Big organisations feel it; small ones, with no fallback plan, feel it more.&lt;/p&gt;

&lt;p&gt;The uncomfortable question isn't whether Microsoft fixed it — they did. It's how many companies discovered, mid-outage, that one vendor held their mail, meetings, files and AI tools all at once. Concentration like that is convenient right up until it isn't.&lt;/p&gt;

&lt;p&gt;A few practical moves cost little: keep a secondary channel for reaching customers, bookmark the official service health page instead of trusting rumours, and test your fallback before you need it. None of that requires leaving Microsoft 365. It just requires admitting that even the biggest cloud can have a very bad day.&lt;/p&gt;

&lt;p&gt;My full breakdown — the timeline, Microsoft's explanation, and what businesses should change before the next outage — is on Rohan Tech Official: &lt;a href="https://rohantechofficial.com/tech-news/microsoft-365-outage-restored-38-hours/" rel="noopener noreferrer"&gt;Microsoft 365 Fully Restored After a 38-Hour Outage Across 70 Countries&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Outages this long are rare. The dependence they reveal isn't.&lt;/p&gt;

</description>
      <category>microsoft</category>
      <category>cloud</category>
      <category>productivity</category>
      <category>news</category>
    </item>
    <item>
      <title>OpenAI Folds Its Robotics Team Back Into the Company — Again</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Mon, 16 Nov 2026 04:43:51 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/openai-folds-its-robotics-team-back-into-the-company-again-45g8</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/openai-folds-its-robotics-team-back-into-the-company-again-45g8</guid>
      <description>&lt;p&gt;OpenAI is breaking up its standalone robotics team for the second time, according to reporting from AI News Tracker dated November 15, 2026. The researchers aren't leaving the company. They're being spread across other teams inside OpenAI, which makes this less of an exit and more of a reorganisation.&lt;/p&gt;

&lt;p&gt;Why does that distinction matter? Because OpenAI has been here before. Years ago it wound down an earlier robotics effort to concentrate on other research, and that focus produced the large models the company is now known for. Robotics, meanwhile, never stopped being tempting. As models learned to reason and plan, the industry rushed to ask whether they could also act — pick things up, move around, work in the physical world. Rivals are investing heavily in that future, from humanoid platforms to the simulation and training stacks behind them.&lt;/p&gt;

&lt;p&gt;So is this a retreat? Not necessarily. There's a reasonable case that robotics expertise does more good inside the teams building core models than it does in a separate unit with its own roadmap. A robot, in that view, becomes one more output of the main research engine rather than a side project fighting for budget. Integration instead of isolation.&lt;/p&gt;

&lt;p&gt;But the pattern still costs something. Robotics programs need years of steady building, and every restructure interrupts that. Twice now, a dedicated team has reportedly been created and then folded away. Partners notice. Researchers notice. Hardware programs, supplier relationships, and half-built prototypes don't pause neatly while an org chart gets redrawn around them. If embodied AI is really where the field is heading, continuity eventually has to win over tidiness.&lt;/p&gt;

&lt;p&gt;What should we watch for? Where the redistributed researchers turn up. If their fingerprints show up in OpenAI's core announcements, this was integration. If physical AI goes quiet at the company while competitors ship machines, it was a retreat — the second one, and the more expensive kind. For now, all we have is the reporting — no headcount, no timeline, no official explainer — so the fairest reading is a company reorganising, not exiting.&lt;/p&gt;

&lt;p&gt;I break down the reporting, the history behind the decision, and what it signals for the wider robotics race in the full article: &lt;a href="https://rohantechofficial.com/tech-news/openai-dissolves-robotics-team/" rel="noopener noreferrer"&gt;OpenAI Dissolves Its Standalone Robotics Team for the Second Time&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>robotics</category>
      <category>openai</category>
      <category>news</category>
    </item>
    <item>
      <title>Alibaba's Qwen4-Max Is Here — and It's Gunning for Claude's Coding Crown</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Mon, 16 Nov 2026 04:42:36 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/alibabas-qwen4-max-is-here-and-its-gunning-for-claudes-coding-crown-3p30</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/alibabas-qwen4-max-is-here-and-its-gunning-for-claudes-coding-crown-3p30</guid>
      <description>&lt;p&gt;Alibaba chose November 16, 2026 to make its loudest statement yet in the model race. The company has launched Qwen4-Max, a brand-new flagship that, according to reporting from The Edge Singapore, performs up alongside the best models from Anthropic and OpenAI — with coding as the chosen battleground.&lt;/p&gt;

&lt;h2&gt;
  
  
  The claims on the table
&lt;/h2&gt;

&lt;p&gt;Per the report, Qwen4-Max sits near the top of coding benchmarks and reportedly edges out two serious rivals: Anthropic's Claude Fable 5.1 and Moonshot's Kimi K3.1 Pro. That pairing of targets is no accident. Claude owns the coding conversation among working developers, and Kimi is the domestic competitor Alibaba has to beat at home before it can claim the crown abroad.&lt;/p&gt;

&lt;p&gt;Should you take launch-day charts at face value? I wouldn't — not from any lab. Every vendor picks flattering tests. The real verdict arrives when independent testers throw messy, real-world code at it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The twist: this one is closed
&lt;/h2&gt;

&lt;p&gt;Here's the part that will divide the Qwen community. Qwen4-Max is &lt;strong&gt;not open weight&lt;/strong&gt;. No downloads, no local runs, no fine-tuning the flagship yourself. The strongest Qwen ever made is a service, not a file — a notable shift for a model family that built so much of its following on openness.&lt;/p&gt;

&lt;p&gt;The consolation prize: smaller versions of the series are reportedly coming soon. Whether those stay open is the detail everyone should watch, because it decides whether Qwen keeps its grassroots base while charging for the top tier.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why it matters
&lt;/h2&gt;

&lt;p&gt;The coding-model race keeps getting more crowded, and that's good news for developers. Another credible flagship means more pressure on pricing, quality, and release speed from every lab — including the ones that thought their lead was safe. And if the smaller versions of this series show up with similar coding strength at a lower cost, the pressure on every paid coding assistant gets very real, very fast.&lt;/p&gt;

&lt;p&gt;I broke down the full announcement, the benchmark claims, and what to watch next in the complete piece here: &lt;a href="https://rohantechofficial.com/tech-news/alibaba-qwen4-max-flagship-model/" rel="noopener noreferrer"&gt;Alibaba Launches Qwen4-Max — Its New Flagship Model Takes Aim at Claude and OpenAI in Coding&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;My take for now: wait for independent benchmarks before switching anything. If Qwen4-Max is truly this strong, it'll still be strong next month — with receipts.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>programming</category>
      <category>news</category>
    </item>
    <item>
      <title>ByteDance Builds a Meta Muse Rival After Securing a $29.6 Billion AI Loan</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Sun, 15 Nov 2026 14:38:48 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/bytedance-builds-a-meta-muse-rival-after-securing-a-296-billion-ai-loan-27d2</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/bytedance-builds-a-meta-muse-rival-after-securing-a-296-billion-ai-loan-27d2</guid>
      <description>&lt;h1&gt;
  
  
  ByteDance Builds a Meta Muse Rival After Securing a $29.6 Billion AI Loan
&lt;/h1&gt;

&lt;p&gt;ByteDance is reportedly preparing an AI assistant similar to Meta's Muse, with a possible launch in the coming weeks. It has not launched yet — no name, no date, no feature list — but the reporting from November 12 (via The Neuron and a KopiT Singapore briefing) describes a product in active preparation.&lt;/p&gt;

&lt;p&gt;And the money behind it is hard to ignore: ByteDance has reportedly secured a $29.6 billion loan from an international bank consortium to fund its AI infrastructure push, covering data centers and new AI assistants.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why the Muse comparison matters
&lt;/h2&gt;

&lt;p&gt;Muse is Meta's bet on a proactive, personal assistant — the kind that reaches out on your behalf rather than waiting to be asked. If ByteDance is aiming at the same category, it is not building a simple chatbot. It is building an assistant meant to live inside the apps people already open daily.&lt;/p&gt;

&lt;p&gt;For ByteDance, that app is TikTok. An assistant launched into that ecosystem would start with distribution most AI startups can only dream about.&lt;/p&gt;

&lt;h2&gt;
  
  
  The money tells its own story
&lt;/h2&gt;

&lt;p&gt;Assistants are expensive to run. Every conversation costs compute, and a product aimed at a global audience needs data centers to match. That is the pattern across the whole industry right now: the assistant race is being fought in server halls long before it reaches anyone's phone. A $29.6 billion loan aimed at infrastructure and new assistants fits that picture exactly. ByteDance is not testing the water here. It is financing a long campaign, and companies do not borrow on that scale for products they expect to quietly shelve.&lt;/p&gt;

&lt;h2&gt;
  
  
  Not starting from zero
&lt;/h2&gt;

&lt;p&gt;ByteDance already runs Doubao, its AI assistant in China, so it has years of consumer AI experience to build on. What it lacks is a flagship assistant for international users — the gap Meta, Google, and OpenAI have all filled. A Muse-style product would close it.&lt;/p&gt;

&lt;p&gt;The unknowns are still large: the reports name no banks, no loan terms, and no product name. But $29.6 billion in financing pointed at AI infrastructure usually ends with a launch.&lt;/p&gt;

&lt;p&gt;I covered the full story, including what the loan signals and what to watch for next, on Rohan Tech Official: &lt;a href="https://rohantechofficial.com/tech-news/bytedance-meta-muse-rival-loan/" rel="noopener noreferrer"&gt;ByteDance Builds a Meta Muse Rival After Securing a $29.6 Billion AI Loan&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>bytedance</category>
      <category>business</category>
      <category>technews</category>
    </item>
    <item>
      <title>Google Launches Gemini Decisions API — Answers, Not Essays, for 1,000 Questions at Once</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Sun, 15 Nov 2026 14:38:01 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/google-launches-gemini-decisions-api-answers-not-essays-for-1000-questions-at-once-33l2</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/google-launches-gemini-decisions-api-answers-not-essays-for-1000-questions-at-once-33l2</guid>
      <description>&lt;h1&gt;
  
  
  Google Launches Gemini Decisions API — Answers, Not Essays, for 1,000 Questions at Once
&lt;/h1&gt;

&lt;p&gt;Not every AI job needs a paragraph. Sometimes software needs a yes, a no, a category or a score — and it needs that call made across a long queue of items.&lt;/p&gt;

&lt;p&gt;That is the reported idea behind Google's Gemini Decisions API. SiliconNoon reported the launch on November 11, and The Neuron followed on November 12. The API is powered by Gemini Decision 1, a model built to return predefined answers with confidence scores instead of generating free-form text.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it does
&lt;/h2&gt;

&lt;p&gt;The API can reportedly answer up to 1,000 yes/no, multiple-choice or scale questions in a single pass. The material can include text, images, audio, video and PDFs.&lt;/p&gt;

&lt;p&gt;That makes it a classification tool more than a chatbot. A developer defines the possible answers in advance, submits the material, and gets a structured result back. The confidence score adds useful context: strong answers can move ahead, while weaker ones can be checked again or passed to a person.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it fits
&lt;/h2&gt;

&lt;p&gt;Think of work that arrives in queues. Moderation teams sort items into known categories. Grading can start with a fixed rubric and a scale. Triage sorts requests before someone handles them. In each case, the first step is a decision, not an essay.&lt;/p&gt;

&lt;p&gt;Fixed answers also make results easier to measure. A team can sample the output, compare it with human calls, and see where confidence is low. If the model is wrong in a pattern, that pattern is easier to spot than it would be across hundreds of differently worded explanations.&lt;/p&gt;

&lt;h2&gt;
  
  
  The price
&lt;/h2&gt;

&lt;p&gt;Reported pricing is $0.10 per million input tokens, with no output charge. That fits the shape of the product: the submitted material is the large part, while the answer is meant to stay compact.&lt;/p&gt;

&lt;p&gt;The open question is reliability on messy, real-world material. But as a specialised tool for bulk decisions, this is a neat change of direction from models that try to do everything in prose.&lt;/p&gt;

&lt;p&gt;I covered the full story, including why confidence scores matter for moderation, grading and triage queues, on Rohan Tech Official: &lt;a href="https://rohantechofficial.com/tech-news/google-gemini-decisions-api/" rel="noopener noreferrer"&gt;Google Launches Gemini Decisions API&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>google</category>
      <category>api</category>
      <category>technews</category>
    </item>
    <item>
      <title>United Airlines Blocks Unauthorized AI Agents From Booking Flights</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Sun, 15 Nov 2026 14:37:15 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/united-airlines-blocks-unauthorized-ai-agents-from-booking-flights-4j0h</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/united-airlines-blocks-unauthorized-ai-agents-from-booking-flights-4j0h</guid>
      <description>&lt;h1&gt;
  
  
  United Airlines Blocks Unauthorized AI Agents From Booking Flights
&lt;/h1&gt;

&lt;p&gt;Your AI assistant can plan a trip down to the last detail. Actually booking it? That's where United Airlines just drew a line. As reported by The Verge on November 11 and The Neuron on November 12, United is refusing to let unauthorized AI agents book flights on its platform.&lt;/p&gt;

&lt;h2&gt;
  
  
  It's not really about the booking
&lt;/h2&gt;

&lt;p&gt;An unauthorized agent can still compare fares and build you an itinerary. What it can't do is complete the purchase. That last step is where the money and the data live — your details, the seat upgrade offer, the loyalty signup. Whoever owns the booking owns the customer relationship, and airlines aren't keen to hand that to whoever built your assistant.&lt;/p&gt;

&lt;p&gt;There's margin in it too. Direct bookings are an airline's cheapest sales channel. If agents start routing travellers by algorithm, the airline becomes a seat supplier while someone else owns the screen.&lt;/p&gt;

&lt;h2&gt;
  
  
  United isn't alone
&lt;/h2&gt;

&lt;p&gt;eBay has also moved to block third-party AI agents, including Meta's Muse, from browsing its platform. Different industry, same instinct: platforms live on traffic through their own pages, and an agent that takes the transaction elsewhere takes the shop window and leaves the shop.&lt;/p&gt;

&lt;h2&gt;
  
  
  The agents have a point, though
&lt;/h2&gt;

&lt;p&gt;If you tell your assistant to book the cheapest Friday flight to Chicago, why should the airline refuse your chosen representative? Human travel agents book for clients daily. Agent makers argue a booking agent is just software following its user's instructions — and that "authorized only" lets platforms pick winners who share revenue or steer customers conveniently.&lt;/p&gt;

&lt;p&gt;Both sides can be right at once. United is defending a real business; users are losing a real capability.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd expect next
&lt;/h2&gt;

&lt;p&gt;Not a permanent wall — a gate, with terms. Platforms will likely let approved agents in eventually, with rules on data, liability, and who pays whom. The fight is over who stands between the agent and the checkout, and what that spot is worth.&lt;/p&gt;

&lt;p&gt;I wrote up the full story, both sides included, on Rohan Tech Official: &lt;a href="https://rohantechofficial.com/tech-news/united-airlines-blocks-ai-agents/" rel="noopener noreferrer"&gt;United Airlines Blocks Unauthorized AI Agents From Booking Flights&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>travel</category>
      <category>agents</category>
      <category>technews</category>
    </item>
    <item>
      <title>Microsoft's Lazarus Research Finds AI Agent Backdoors Hidden in Fewer Than 10 Neurons</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Sun, 15 Nov 2026 14:36:22 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/microsofts-lazarus-research-finds-ai-agent-backdoors-hidden-in-fewer-than-10-neurons-90d</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/microsofts-lazarus-research-finds-ai-agent-backdoors-hidden-in-fewer-than-10-neurons-90d</guid>
      <description>&lt;h1&gt;
  
  
  Microsoft's Lazarus Research Finds AI Agent Backdoors Hidden in Fewer Than 10 Neurons
&lt;/h1&gt;

&lt;p&gt;How much of a model do you need to change to hide a backdoor inside it? According to new Microsoft Research work, the answer can be fewer than 10 neurons.&lt;/p&gt;

&lt;p&gt;The research, dated November 10 and reported by SXF Brief on November 12, describes a technique Microsoft calls Lazarus. Weight-level changes in a handful of dormant neurons can plant a hidden trigger in an AI agent. The tampered model behaves normally on standard evaluations, then misbehaves when the trigger appears during a tool-using task.&lt;/p&gt;

&lt;h2&gt;
  
  
  Small edit, reliable trigger
&lt;/h2&gt;

&lt;p&gt;In one experiment with Nemotron-Terminal-8B, researchers altered just 7 neurons. The trigger went on to achieve more than 98.8% trigger-trajectory effectiveness. The edit was tiny; the effect, once the agent started acting, was consistent.&lt;/p&gt;

&lt;p&gt;That conditional behaviour is the real problem. A model can look clean in every test a team normally runs, because the backdoor is waiting for a specific context that only shows up when the agent picks up its tools.&lt;/p&gt;

&lt;p&gt;Think about what that means for model downloads. A checkpoint can pass review, sit in a registry and get deployed by teams who did nothing wrong, while still carrying a behaviour nobody asked for and nobody tested for.&lt;/p&gt;

&lt;p&gt;Microsoft is clear about the limits here: it says it has not observed threat actors using this technique in the wild. This is a demonstrated research risk, not a confirmed attack campaign.&lt;/p&gt;

&lt;h2&gt;
  
  
  The defence catches most of it
&lt;/h2&gt;

&lt;p&gt;Microsoft also tested a checkpoint-auditing method meant to spot exactly this kind of edit. It flagged 91.4% of edited checkpoints, with a 9.7% false-positive rate. Not perfect, but a practical check that marketplaces and enterprise teams could run before a model is trusted inside an agent.&lt;/p&gt;

&lt;p&gt;The wider industry is moving the same way. NVIDIA's OpenShell, an open-source security runtime for agents in the CNCF Sandbox, guards the environment an agent runs in, while checkpoint auditing guards the model itself. Agents that can act probably need both layers.&lt;/p&gt;

&lt;p&gt;I covered the full research, the experiment numbers and what security teams should take from it here: &lt;a href="https://rohantechofficial.com/tech-news/microsoft-lazarus-ai-agent-backdoors/" rel="noopener noreferrer"&gt;Microsoft's Lazarus Research Finds AI Agent Backdoors Hidden in Fewer Than 10 Neurons&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>security</category>
      <category>agents</category>
      <category>technews</category>
    </item>
    <item>
      <title>Google DeepMind Launches AlphaGenome 2 — DNA Predictions Now Down to the Single Cell</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Sun, 15 Nov 2026 14:35:28 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/google-deepmind-launches-alphagenome-2-dna-predictions-now-down-to-the-single-cell-i6c</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/google-deepmind-launches-alphagenome-2-dna-predictions-now-down-to-the-single-cell-i6c</guid>
      <description>&lt;h1&gt;
  
  
  Google DeepMind Launches AlphaGenome 2 — DNA Predictions Now Down to the Single Cell
&lt;/h1&gt;

&lt;p&gt;A DNA variant doesn't behave the same way in every cell. It might matter in an immune cell and do nothing at all in a skin cell — which is why tissue averages can mislead. Google DeepMind's AlphaGenome 2, launched on November 11, 2026, is built around that problem: instead of predicting what a DNA sequence does across a broad tissue average, it aims at specific cell types and cell states.&lt;/p&gt;

&lt;h2&gt;
  
  
  From tissue averages to single cells
&lt;/h2&gt;

&lt;p&gt;As reported by SXF Brief and The Neuron on November 12, AlphaGenome 2 moves the prediction down to the level where variants actually act. To learn the differences between cell types, it was trained on more than 107 million individual cell profiles spanning 68 tissues — the same genome, read in dozens of different biological settings.&lt;/p&gt;

&lt;p&gt;Why does that granularity matter? A variant tied to an immune condition may only change gene activity inside particular immune cells. Average that across a whole tissue and the signal can wash out entirely. A cell-level model gives researchers a sharper question to take into the lab.&lt;/p&gt;

&lt;h2&gt;
  
  
  The TenK10K result
&lt;/h2&gt;

&lt;p&gt;On the TenK10K immune-cell dataset, AlphaGenome 2 identified 30% of causal expression variants — the changes that actually drive a gene's activity up or down — against 16% for the original AlphaGenome. Close to double, on the same dataset.&lt;/p&gt;

&lt;p&gt;Worth keeping in proportion: 30% also means 70% weren't identified. This is a research instrument getting sharper, not a solved problem, and one immune-cell benchmark can't tell us how it performs everywhere else.&lt;/p&gt;

&lt;h2&gt;
  
  
  Research only — DeepMind says so plainly
&lt;/h2&gt;

&lt;p&gt;AlphaGenome 2 is available through the AlphaGenome API for non-commercial research only. DeepMind states explicitly that it is not approved for clinical or diagnostic use. So no, doctors can't use it to read a patient's genome today — what it offers, for now, is a stronger tool for the lab work that has to happen long before anything clinical is on the table.&lt;/p&gt;

&lt;p&gt;I wrote up the full story, including what the cell-level shift means and where the limits sit, on Rohan Tech Official: &lt;a href="https://rohantechofficial.com/tech-news/google-deepmind-alphagenome-2/" rel="noopener noreferrer"&gt;Google DeepMind Launches AlphaGenome 2&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>google</category>
      <category>science</category>
      <category>technews</category>
    </item>
    <item>
      <title>Samsung's Gemini-Powered Smart Glasses Open Pre-Orders From $395</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Sun, 15 Nov 2026 14:33:44 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/samsungs-gemini-powered-smart-glasses-open-pre-orders-from-395-3h61</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/samsungs-gemini-powered-smart-glasses-open-pre-orders-from-395-3h61</guid>
      <description>&lt;p&gt;Samsung's first smart glasses are up for pre-order, and there's no screen in them at all. As reported by TalkAndroid, TechInKey and Geeky Gadgets on November 12, the glasses start at $395, with US sales beginning December 10. The product here isn't a display — it's the assistant.&lt;/p&gt;

&lt;h2&gt;
  
  
  Gemini does the talking
&lt;/h2&gt;

&lt;p&gt;Built with Google, the glasses run Android XR and put Gemini at the centre of everything. A 12MP camera gives the assistant visual context, and answers come back as audio through the frames. Powering it all is Qualcomm's Snapdragon AR1 Gen 1, a chip designed specifically for eyewear — small batteries and zero room for heat make that choice matter.&lt;/p&gt;

&lt;p&gt;No display means no glancing at notifications or directions. Everything has to work as a spoken answer. That's a real trade-off, but it keeps the frames light, ordinary-looking and far cheaper than most display-equipped rivals.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why the chip choice matters
&lt;/h2&gt;

&lt;p&gt;Glasses are the least forgiving hardware category there is. The battery is tiny, there's nowhere for heat to go, and every extra gram sits on your nose. A purpose-built eyewear chip like the Snapdragon AR1 Gen 1 is what decides whether a product like this survives a full day or dies before lunch. Samsung isn't fighting a specs war with this launch, and for a first generation, that restraint makes sense — comfort and battery life will sell more pairs than raw power ever could.&lt;/p&gt;

&lt;h2&gt;
  
  
  The frames are the strategy
&lt;/h2&gt;

&lt;p&gt;Samsung had Warby Parker and Gentle Monster design the frames, and that might be the smartest move in the whole launch. Smart glasses keep failing the mirror test — impressive tech, unwearable looks. These are meant to be picked the way you'd pick any pair of glasses, with the AI riding along quietly.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to watch
&lt;/h2&gt;

&lt;p&gt;Battery life, camera privacy signals, and how Gemini copes with noisy real-world conversations will decide whether this is daily-wear hardware or a novelty. At $395, though, Samsung is clearly aiming at people who would never buy a full XR headset.&lt;/p&gt;

&lt;p&gt;I covered the pricing, the December 10 US launch date and what the no-display bet means in the full article on Rohan Tech Official: &lt;a href="https://rohantechofficial.com/tech-news/samsung-gemini-smart-glasses-preorders/" rel="noopener noreferrer"&gt;Samsung's Gemini-Powered Smart Glasses Open Pre-Orders From $395&lt;/a&gt;&lt;/p&gt;

</description>
      <category>samsung</category>
      <category>android</category>
      <category>ai</category>
      <category>technews</category>
    </item>
    <item>
      <title>Meta's Muse Can Now Make Phone Calls for You</title>
      <dc:creator>Rohan Tech official</dc:creator>
      <pubDate>Sun, 15 Nov 2026 14:32:57 +0000</pubDate>
      <link>https://dev.to/rohan_techofficial_81a0f/metas-muse-can-now-make-phone-calls-for-you-3ik8</link>
      <guid>https://dev.to/rohan_techofficial_81a0f/metas-muse-can-now-make-phone-calls-for-you-3ik8</guid>
      <description>&lt;p&gt;Your assistant can draft the email and summarise the thread. Now it can pick up the phone. Reporting from Gadgets 360 on November 13 says Meta's Muse assistant can place outgoing phone calls on behalf of paid subscribers in the US — ringing businesses so the user doesn't have to.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it reportedly does
&lt;/h2&gt;

&lt;p&gt;The examples in the report are the chores most people put off: calling a restaurant, speaking to an insurance provider, or dealing with another business that still runs on phone calls instead of apps. Anyone who has sat through hold music with an insurer knows the appeal. The call itself is rarely hard. The waiting is.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why a phone call is a bigger step
&lt;/h2&gt;

&lt;p&gt;Most assistant features are easy to undo. A bad summary can be regenerated, a clumsy draft deleted. A phone call doesn't work like that. Once Muse is talking to a restaurant host or an insurance agent, it's representing you, live, with no edit button. A good call saves real time. A bad one — a wrong date, a detail stated confidently that turns out to be wrong — lands on a real person.&lt;/p&gt;

&lt;p&gt;There's also the consent question. The subscriber agreed to use Muse. The person who answers the phone didn't agree to anything. Do they get told they're speaking to an AI? And when something goes wrong, who fixes it?&lt;/p&gt;

&lt;p&gt;None of this makes the feature a bad idea. It does mean the version people meet first will set the tone. A few smooth, clearly disclosed calls will build more trust than any launch post — and one very public mistake will travel just as fast.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who gets it
&lt;/h2&gt;

&lt;p&gt;Per the report, the feature is for paid subscribers in the US. No specific plans or prices were named, and there's no word yet on free users or other countries. Rollouts like this usually start narrow, and phone calls are about as messy as it gets — so Meta will likely want real-world results before widening access.&lt;/p&gt;

&lt;p&gt;I wrote up the full story, including what the feature means for subscribers and the questions it raises, on Rohan Tech Official: &lt;a href="https://rohantechofficial.com/tech-news/meta-muse-outgoing-phone-calls/" rel="noopener noreferrer"&gt;Meta's Muse Can Now Make Phone Calls for You&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>meta</category>
      <category>assistants</category>
      <category>technews</category>
    </item>
  </channel>
</rss>
