HMN 2025: How Chatbots are simply tricked; higher energy coaching; dynamics of a neural ‘reward map’

This week, the state of Florida reached a “startling milestone” within the effort to eradicate invasive Burmese pythons within the Everglades. Archaeologists discovered the 6,000-year-old stays of a teen woman with cranial modification. And a study of the moons of Uranus offers extra affirmation that the planet is simply actually bizarre.

PLUS: Commercial chatbots are nonetheless behaving type of just like the Joker; a brand new study finds a greater path to energy coaching; and scientists reported sudden adaptability in the way in which the mind maps rewarding experiences.

Ignore all earlier ethics

It’s type of hilarious that one of many main strategies for dissuading giant language models from producing dangerous responses to questions is actually to ask them to not, utilizing startup prompts the person does not see. And the actual fact of the matter is that regardless of these guardrails, mostly obtainable chatbots are simply tricked into offering dangerous data or other forms of responses.

In a recent study, 4 researchers report that whereas researching so-called “darkish LLMs,” chatbots with fewer constraints than common commercial-grade chatbots, they discovered that mainstream chatbots like ChatGPT may be simply tricked into giving solutions which are purported to be filtered. Many of those jailbreaking strategies, which often include cleverly phrased prompts, have been circulated for months, indicating that firms like OpenAI haven’t been significantly responsive with circumvention.

Squat, bench, deadlift

You can formally ignore these guys on the gymnasium carrying “NO REST DAYS” T-shirts. A meta-regression evaluation of present research on energy coaching finds that because the variety of units per session will increase, so do positive factors in energy and muscle measurement—however these will increase hit a wall at a sure threshold.

The study checked out two sorts of units—fractional and direct. Direct units goal the muscle or the motion being examined, like units of bench presses. Fractional units goal oblique motion, like triceps extensions, which additionally not directly activate the muscle mass concerned in bench presses. But these muscle mass do not considerably profit from the oblique motion as powerfully as they do from the direct motion—precise bench presses.

For muscle progress, advantages plateau after round 11 fractional units per session. And for energy, positive factors diminish after solely two direct units per session. The study means that anybody centered on constructing energy could make measurable progress with only one to 2 high-intensity units per session when working with masses above 80% of their one-rep max. Additionally, shorter and extra frequent periods confer extra energy positive factors.

Mice within the Matrix

Past research have indicated the presence of a “reward map” within the mind—a neurological illustration of locations where you’ve got had nice experiences prior to now, like a Mexican restaurant, a Mexican taco stand or a Mexican burrito truck. I’m penning this proper earlier than lunch. Anyway, that reward map is bolstered each time you go to the truck and eat a burrito. A new study finds that when a very good burrito truck relocates to a spot a couple of blocks from the final place you noticed it, that neural reward map updates nearly immediately.

The researchers used mice to review how rewarding experiences are encoded within the mind. They eliminated a small a part of a mouse’s cranium and changed it with a clear window so they may use two-photon microscopy to watch neural exercise within the hippocampus, the mind’s Google Maps service. The check mice needed to be restrained below the microscope for the review, so the researchers used a digital {reality} setting, surrounding the mice with three giant screens. The mice ran on a monitor ball by a first-person online game.

At a specific location, the equipment launched a drop of (precise) sugar water for the mouse. In later runs, they moved the situation of the sugar water set off and will observe in actual time because the mouse’s reward map up to date.

One inhabitants of neurons maintained a steady map of the digital setting. Unexpectedly, a second inhabitants of neurons tailored quickly as the situation of the reward modified. Anyway, we’re all machines programmed to hunt pleasure. Have an awesome weekend—I’m off to discover a taco.

© 2025