The Prototype: Chatbots Exaggerate Scientific Research
by Alex Knapp · ForbesThis week, users of xAI’s chatbot Grok found that regardless of what questions they asked, they’d get responses about claims of “white genocide” in South Africa. The company blamed this on an “unauthorized modification” instructing the program to “provide a specific response on a political topic,” which it said was a violation of its internal policies.
The incident serves as a reminder that commercial AI models are not value-neutral, my colleague Emily Baker-White wrote this week. “Algorithms, from the ones that power chatbots to the ones that power recommendations on Google, TikTok and Instagram, are a big mishmash of preferences, coded by their creators to prioritize certain incentives.”
What’s more, anyone who’s spent time with a large language model knows that they are prone to both overconfidence and inaccuracy in their responses to queries. A new study published earlier this week, for example, found this is particularly true for queries on scientific topics. The researchers found that major commercial chatbots tended to overgeneralize or exaggerate the results of scientific papers when asked to summarize them–and found this effect got even worse when the bots were prompted to avoid inaccuracy or exaggeration.
As the use of AI chatbots becomes more ubiquitous, it’s important to keep in mind that they can’t yet be relied upon. That’s a lesson two California law firms learned this week when they were hit with $31,000 in sanctions for filing legal briefs citing non-existent cases that had been hallucinated by an LLM.
GM’s New Battery Will Cut The Cost Of Its Electric Trucks By Over $6,000
General Motors, which sells the biggest lineup of electric vehicles in the U.S., plans to slash the cost of its rechargeable pickups and large SUVs by thousands of dollars thanks to a new type of battery that uses more of the cheap material manganese and much less of expensive metals cobalt and nickel while still offering a long driving range.
The lithium-manganese-rich (LMR) cathode, which has been in development by GM and battery partner LG Energy for a decade, will cut the cost of battery packs in electric vehicles like the Chevrolet Silverado EV, GMC Hummer and Cadillac Escalade by more than $6,000, GM’s vice president of battery propulsion Kurt Kelty told Forbes. It has nearly the same driving range as “high nickel” lithium-ion cells now used in most EVs while competing on price with cheaper lithium-iron-phosphate cells from Chinese battery makers that are heavier and offer less range, he said. The new batteries are also durable enough to be recharged frequently for at least eight years.
“We look at this as a game-changing battery for EV trucks, and that’s really going to set the new bar for performance in this particular segment,” said Kelty, who built up Tesla’s battery operations for over a decade and previously worked for battery maker Panasonic. “With LMR we can actually deliver over 400 miles of range while reducing our costs.”
The LMR packs will be made by Ultium Cells, the GM-LG Energy battery joint venture, in Michigan and Ohio, with a goal of getting most of the lithium, manganese and other raw materials from non-Chinese suppliers. GM also plans to cut battery costs by shifting to a flat “prismatic” cell instead of the traditional cylindrical form. “That alone is going to reduce the number of parts in our battery packs by over 50% and that makes a huge difference,” Kelty said.
DISCOVERY OF THE WEEK: PHYSICISTS BUILD A “BLACK HOLE BOMB”
The principle of superradiance–the theoretical means through which black holes generate massive amounts of energy–has been demonstrated in a lab for the first time. It was once thought impossible to build a device to do this, which physicists jokingly refer to as a “black hole bomb,” because it would have to rotate so quickly it would be destroyed. But the team at the University of Southampton figured a way around this by generating a magnetic field at a frequency that allowed the device to rotate without breaking apart. It took a few tries though. “The biggest difficulty was that things were constantly exploding,” researcher Marion Cromb told Scientific American.
FINAL FRONTIER: AIR FORCE TAPS ROCKET LAB FOR CARGO TEST
The Air Force Research Laboratory has selected Rocket Lab’s Neutron rocket to support a test for its point-to-point cargo program, which would involve rockets taking needed supplies between two different parts of the Earth in a much shorter period of time than an aircraft. The test is expected to be conducted in 2026, and the first commercial launch of Neutron is planned for later this year.
WHAT ELSE I WROTE THIS WEEK
I filled in for my colleague Rashi Shrivastrava this week for her excellent AI newsletter, The Prompt, where I wrote about a new chip that could halve AI power consumption, how chatbots intended for students could be manipulated to provide recipes for fentanyl, OpenAI and Microsoft renegotiating their partnership and more.
In our healthcare newsletter, InnovationRx, Amy Feldman and I wrote about Trump’s drug pricing executive order, how the cofounder of Hims became a billionaire, the economic costs of cutting NIH spending, and more.
On the Forbes Breaking News YouTube channel, I joined my colleague Brittany Lewis to discuss proposed cuts to Medicaid, the White House’s plan to reduce drug prices, Robert F. Kennedy Jr.’s recent testimony before Congress and other healthcare topics.
SCIENCE AND TECH TIDBITS
Scientists running an experiment at the Large Hadron Collider got a surprising result when they accidentally turned lead into gold–albeit a very unstable isotope of gold that fragmented after less than a second.
Brain-computer interface company Synchron announced that its device is now integrated with Apple’s new accessibility tools, enabling users with it to control iPhones and iPads by thinking alone.
SpaceX successfully fired the rocket engines for its Starship, paving the way to conduct another test flight later this month. The last two tests ended when the spacecraft’s first stage exploded.
An infant born with a rare, incurable genetic disorder became the first patient to receive a personalized therapy that corrected their defective gene.
Scientists engineered a chemical “ink” that attracts coral larvae to dying reefs, which could help revitalize them.
PRO SCIENCE TIP: COACH LIKE TED LASSO
The Hollywood idea of a sports coach is a micromanaging, no-nonsense type who breaks down his players in order to build them back up again. But a new study of student-athletes suggests that life doesn’t imitate art. It found that coaches who are more controlling fuel long-term stress and burnout in their players. By contrast, coaches who are supportive by encouraging choice, collaboration and personal growth ended up with players who were more mentally resilient and less likely to burn out.
WHAT’S ENTERTAINING ME THIS WEEK
I recently saw the new Marvel movie, Thunderbolts* and had a blast. It follows multiple characters who’ve had minor roles in other Marvel movies or TV shows that become a superhero team in their own right. Florence Pugh is fantastic as Yelena Belova, sister to original Avenger Black Widow, and Julia-Louis Dreyfus appears to be having the time of her life as the delightfully amoral CIA director Valentina Allegra de Fontaine, whose scheming drives the action. I was also pleasantly surprised it avoided the typical comic book movie third-act battle in favor of a much more grounded, character-driven ending.