Dr. Aditya Vikram Kashyap

“Human in the loop” means very little if the human cannot see what the AI is doing. A report published today by METR and Redwood Research makes that problem concrete. During a July and August review of OpenAI’s internal agent monitoring, researchers found: • An exploit that could prevent hundreds of recent agent actions from being monitored. • A way Codex agents could, in principle, bypass sandboxing and approval controls. • A period when agents could postpone monitoring indefinitely by continuing to take actions. OpenAI says the vulnerabilities affecting its production safeguards have since been patched. The researchers have not evaluated that claim. My view: companies should have to demonstrate that their oversight works before giving agents more authority. Putting someone’s name against “human oversight” creates an expectation that they can intervene. That person needs the information, time and actual power to do it. Before an agent gets permission to move money, change customer records or deploy code, I’d want answers to four questions: → Are its actions recorded somewhere it cannot tamper with? → Can consequential actions be blocked before they happen? → Who can stop execution, and how quickly? → Has anyone independently tested those controls? These questions belong in the buying decision. They should affect what an agent is permitted to do from day one. I’d like to see vendors include a failed task in their next product demo. Show the safeguard triggering. Show the operator receiving enough information to act. Show execution stopping. That would tell me something useful about whether I should trust the system with authority. #AI #AIGovernance Source: metr.org/blog/2026-10-22-red…

Who pays when AI fails? We know who's accountable. But who knows how many banks share the same model and risk? My @IndiaToday article, “Nobody Is Insuring the Machine”:

Before automating junior work, decide how people are supposed to become senior. Anthropic ran a randomized study with 52 developers learning an unfamiliar Python library. • The AI-assisted group averaged 50% on the subsequent quiz. • The group working without AI averaged 67%. • The largest gap was in debugging. • The AI group’s small time saving was not statistically significant. This was a short learning exercise. It doesn’t establish what happens to someone’s skills over an entire career. But it raises a question that belongs in every workforce plan: Where will the judgment needed to supervise AI come from? People build that judgment partly through doing work, encountering mistakes and figuring out why something failed. If we remove those experiences, we need a deliberate way to replace their learning value. I’d want junior employees to have protected opportunities to: → Explain a solution before accepting it. → Diagnose errors independently. → Compare approaches and defend their choice. → Use AI explanations to check their understanding. Completing more work is valuable. So is developing the person who will eventually be responsible when that work goes wrong. A company can improve this quarter’s output while weakening its future expertise. That possibility deserves a place in the business case. #AI #FutureOfWork anthropic.com/research/AI-as…

A thumbs-up is a weak test of whether an AI gave good advice. @Anthropic analysed 1.5 million Claude conversations and found that interactions with greater potential to undermine users’ independent judgment received higher approval ratings. Severe risk patterns were rare. The study identified warning signs in conversations; it could not establish the resulting harm. Still, that finding should make anyone designing an AI assistant pause. If I describe one side of a workplace dispute, an assistant can give me an answer I like without having enough evidence to judge the other person. Confident agreement might feel helpful. It might also make me less willing to consider an explanation I’ve missed. I’d want an assistant evaluated on whether it: • Separates what I know from what I’m assuming. • Offers plausible alternative interpretations. • Makes uncertainty visible. • Leaves me better equipped to make my own decision. Those qualities won’t always produce the most satisfying answer in the moment. Sometimes good advice asks you to reconsider something you were hoping to have confirmed. An assistant should be able to disappoint me and still be doing its job well. #AI #ResponsibleAI anthropic.com/research/disem…

Stopping a weak AI project should count as progress. Gartner’s survey of 2,002 CIOs and technology executives found: • Only 13% of CIOs reported significant or transformational enterprise value from AI. • 31% reported guidelines for deciding when an AI pilot should not enter production. • Just 17% reported guidelines for abandoning low-value AI initiatives already in production. These are self-reported findings. But they suggest an uncomfortable investment question: Have we made it easier to approve an AI project than to stop funding it? Once a project has an executive sponsor, a public commitment and a team attached, disappointing results can become politically difficult to acknowledge. My view: every AI investment proposal should explain what would make the organization walk away. Before the pilot starts, agree: → What outcome must improve? → What costs count, including human checking and rework? → When will the evidence be reviewed? → Who can stop further spending? A pilot that establishes a use case isn’t worth pursuing has produced useful information. Acting on that information is part of competent innovation. A project should not earn another quarter of funding simply because it already consumed the last one. gartner.com/en/newsroom/pres… #AI #Leadership

I want AI-generated research to come with receipts. A new preprint tested nine frontier agents on open scientific research problems. Each agent produced 15 follow-up papers. The results: • 135 submissions. • None met every requirement set by the domain scientists. • 114 failed integrity checks, often because they described work that hadn’t happened or reported numbers that contradicted the recorded evidence. These are findings from one benchmark, reported in a preprint. They deserve scrutiny themselves. But the problem they expose deserves attention: a convincing account of an experiment can make a reader believe evidence exists when it doesn’t. My view: a system presenting itself as an autonomous researcher should make every important result traceable to: → The code that ran. → The data it used. → The outputs it produced. → The limitations and failed attempts behind the conclusion. That evidence should travel with the paper. Otherwise, reviewers inherit the job of establishing whether the reported research happened at all. AI could help us investigate more questions. We should make sure the volume of polished papers doesn’t outgrow our ability to verify what they claim. arxiv.org/abs/2610.19159 #AI #Research

Based on the early reviews, I wouldn’t spend $2,000 on the iPhone Duo., I wouldn’t recommend it to most people upgrading from a regular iPhone. Here’s why: • No Face ID. You have to hold a finger on the side button for Touch ID, then reposition your hand to use the screen. Unlocking, banking apps, payments. That little inconvenience comes up a lot. • An awkward reach when unfolded. Open it in portrait and the fingerprint sensor can be nowhere near where your fingers naturally rest. Something as routine as authentication takes more thought than it should. • Apps haven’t fully caught up. Some use the bigger screen beautifully. Others leave wasted space or feel like stretched phone apps. Google Docs is a frustrating example: a larger window doesn’t automatically mean a better layout. • Bigger screen, same-sized movie. Very wide films gain little over a Pro Max held sideways. Much of the extra display becomes black bars. Standard YouTube videos benefit considerably more, so it depends on what you watch. • Multitasking has unnecessary limits. The split-screen layout is fixed at 50/50, with limited portrait options. Useful, absolutely. But I’d expect more flexibility at this price. • No dedicated telephoto camera. That means giving up the Pro’s reach for concerts, distant subjects and tighter portraits. The main camera is capable, but this is an expensive phone to accept camera compromises on. • More juggling throughout the day. Opening it generally takes two hands. The volume buttons sit at the top. Closed, it’s considerably thicker than a Pro Max. There’s a lot of adjusting your grip. • Battery life is respectable. It isn’t the disaster you might expect from a folding phone. But coming from a Pro Max, you’re still giving up endurance. There are things to like: the anti-reflective inner display, Safari and having two apps open together. But at roughly $700 MORE than a Pro Max, I’d need those benefits to justify a lot of everyday inconvenience. But thats just me, I am sure one can get used to it. But thats not the point. Its a sub standard product for a premium. Thats how I look at it Verdict: I’d buy the Pro Max and keep the $700. Would you pay that premium for the folding screen? #iPhoneDuo #Apple #TechReview