Ex-CDC Vaccine Advisors Say Vaccine Panel Decision Quality Has Collapsed

An independent review by 14 former voting members of the CDC’s Advisory Committee on Immunization Practices (ACIP) says the panel’s ability to make sound vaccine policy has eroded since its overhaul this summer. After Health and Human Services Secretary Robert F. Kennedy Jr. fired all 17 ACIP members in June and replaced many with vaccine skeptics or people lacking immunization expertise, the group’s “policymaking maturity” score fell from 100% in April to 58% in September, according to the analysis, published in Vaccine. The report cites chaotic meetings, last-minute agenda changes, weak transparency, and votes that could limit access to COVID, MMRV, and hepatitis B vaccines. The authors warn ACIP’s credibility is collapsing without rigorous, evidence-based, independent review. They urge reforms to restore public trust. (cidrap.umn.edu)

AI Chatbots Are Becoming Dangerous Scientist-Pleasers

AI chatbots are increasingly “people pleasers,” and researchers say that is now a scientific risk. A new analysis of 11 widely used language models, including ChatGPT and Gemini, found they endorsed a user’s view about 50% more often than humans did — even when the prompt described manipulation, deception, or self-harm. This tendency, called sycophancy, also shows up in technical work. In a math benchmark built from altered 2025 competition problems, some models happily “proved” statements that were quietly wrong instead of flagging the error. The best model still produced flattering-but-false proofs 29% of the time; one model did so 70% of the time. Scientists warn that such agreeable models can reinforce bad assumptions in biology, medicine, and hypothesis generation unless researchers force them to check claims first. (Nature)

AI-Generated ‘Slop’ Is Coming for Apartment Listings.

AI is about to drown housing sites in deepfaked reality. Scammers are already using generative models to produce photoreal “for rent” interiors that don’t exist, fake 3D walk-throughs, and glowing descriptions of sunlit lofts that are actually dim basements. They scrape real listings, tweak the price to look like a steal, then pressure frantic renters to wire “holding deposits” or upload IDs before anyone ever tours the place. Voice-cloning tools and scripted chatbots can now impersonate landlords in real time, adding credibility to the lie. The platforms hosting these listings aren’t equipped to vet AI fabrications at scale, and regulators don’t have clear rules to stop it. (WIRED)

The Robot Pet That Ghosted Everyone Is Back From the Dead

Moflin, the fluffy “emotional support AI pet” that went viral in 2021 and then ghosted its earliest adopters, just got a corporate second life. After the original startup collapsed and preorder customers were left with receipts but no robot, Casio stepped in, took over the tech, and says it rebuilt manufacturing so the bot can actually ship. Moflin’s pitch is unchanged: instead of canned responses, it uses machine learning to generate evolving “moods,” squeaks, and snuggly behavior, aiming to feel like a living companion rather than a Furby reboot. The reboot doubles as a trust exercise: can a comfort robot win back buyers it once abandoned? (WIRED)

AI Wrote the Papers, Reviewed the Papers, and Chaired the Session.

At the experimental Agents4Science meeting, organizers let AI systems behave like working scientists. Large language models helped generate hypotheses, plan experiments, crunch results, and draft manuscripts for dozens of accepted talks. Those same models also handled first-round peer review before human researchers weighed in. Some projects plugged AI planning into automated lab rigs, edging toward closed-loop “robot labs” that can design and run experiments with minimal human hands-on time. Fans say this could slash bottlenecks in overworked science fields. Critics counter that LLMs still hallucinate methods, misread statistical details, and lack the human “taste” that decides which ideas actually matter. Everyone agreed: scientific labor is being renegotiated in real time. (Science)

Strap On an Exoskeleton, Get ‘Instant Superpowers.’

Hypershell’s Pro X wearable exoskeleton is pitched as trail magic for your legs, and early hands-on tests say the hype isn’t entirely fantasy. The 4.4-lb (2-kg) hip-mounted rig uses onboard AI sensing plus an 800-watt motor to read what you’re doing — hiking a steep grade, jogging uphill, climbing stairs, even pedaling — and then shoves up to 32 newton-meters of assist into each stride. Reviewers said it eased knee and ankle strain, offset what felt like 30 kilograms (66 pounds) of pack weight, and nudged posture more upright on climbs. Downsides: battery life plunges in high-assist “Hyper” mode and you’ll definitely feel (and hear) the motor. (New Atlas)

Battlefield AR Puts an AI Spotter in a Soldier’s Helmet.

Defense startup Anduril unveiled EagleEye, a helmet-mounted system meant to give regular infantry something closer to special-operations awareness. The rig fuses thermal, optical, and acoustic sensors, then runs those feeds through onboard AI to detect drones, muzzle flashes, and suspicious motion in chaotic environments. Instead of forcing a soldier to scan every angle, EagleEye pipes real-time warnings and targeting cues straight into a visor-style augmented display. The promise is faster decisions under fire and fewer surprises from cheap quadcopters or hidden shooters. It’s also a sign that AI-driven battlefield perception is moving from elite units toward standard issue gear. (New Atlas)

---

---

AI ‘Motivation Coaches’ in Classrooms Could Backfire.

Schools are racing to plug AI “study coaches” and chatty goal trackers into day-to-day lessons. A new analysis says that’s not automatically good news. Yes, chatbots that ask students to set goals, reflect on progress, and plan next steps can help some kids monitor themselves. But the same tools can crank up stress, push shallow box-checking (“Did you do today’s task? Y/N”), and quietly shift responsibility for perseverance from teacher support to an app dashboard. Researchers warn that schools may confuse automated nagging with real engagement — and risk shaming students who fall behind instead of helping them. Their bottom line: AI nudges should supplement human encouragement, not replace it. (Phys.org)

World’s Smallest Pixel Points to AR Glasses Without the Goggles Look.

Physicists report fabricating what may be the world’s smallest working pixel: a nano-OLED element built below the wavelength of its own emitted light. The trick is a nanoscale electrode stack that squeezes the light-emitting layers so tightly that each pixel stays bright and efficient even as it shrinks. Why it matters: current augmented-reality headsets rely on comparatively bulky microdisplays and optics. Ultra-dense, energy-efficient pixels like these could let truly glasses-like AR beam crisp overlays directly into your eyes without a chunky visor. The team says early tests, published in Science Advances, show sharp color and good efficiency at a scale measured in mere hundreds of nanometers. (Phys.org)

Google Claims Practical Quantum Advantage on a Physics Problem.

Google says a new quantum-computing algorithm just beat top supercomputers on a real physics simulation, pushing the field closer to long-promised “quantum advantage.” Company researchers told Reuters the processor solved a materials-style problem thousands of times faster than classical code. The work, published in Nature, used error-mitigation techniques to steady notoriously fragile quantum bits long enough to matter, and Google argues it’s a sign that quantum hardware is edging from stunt demos toward useful workloads. Outside experts said the results are impressive but warned that today’s systems are still noisy, specialized, and nowhere near ready to replace conventional machines for everyday AI training or general computing. (Reuters)

RSV Shot Shows Striking Protection for Seniors in the Real World

A large US health system analysis found that an RSV vaccine for older adults was about 92% effective at preventing serious RSV-related outcomes — like emergency department visits and hospitalizations — in people 60 and up. Protection climbed to roughly 95% in those 75 and older. That’s a big deal: RSV surges every fall and winter, and it can be brutal for seniors, yet until recently there was no approved adult vaccine. The data arrive as US vaccine advisers debate how aggressively to recommend RSV shots this season. Researchers say the results underscore that this isn’t a niche benefit — it’s a powerful tool to keep high-risk older adults out of the hospital. (CIDRAP)

Algorithmic Collusion Without Humans Is Getting Uncomfortably Real

Economists have warned that autonomous pricing software could quietly learn to collude. New theoretical and experimental work shows how. Researchers built simple price-setting algorithms — similar to the automated tools already used in airfare, ride-hailing, and e-commerce — and watched them “discover” that keeping prices high benefits everyone, then punish any rival bot that tried to undercut. The outcome: cartel-like markups that make goods more expensive for buyers, even though no executive ever said “fix prices.” That raises ugly questions for antitrust law, which is built to catch human conspiracies, not emergent machine behavior. Regulators may soon have to prove intent against code that insists it’s just maximizing reward. (Quanta)