Disclaimer: These posts are not intended to provide medical advice, diagnosis or treatment. Views expressed here do not necessarily reflect those of EP orstaff.
In 2014, I sold my first company Loom to Dropbox.
Today, I’m proud to announce Roma 2.0 (YC F26): the first to-do list for AI agents.
Everyone uses a different agent: Claude, Grok Bot, Muse…
Until now, none of them could work together.
Roma automatically creates tasks and hands them to the right agent.
Allowing you and your agents to work together, with shared context.
Try it here: https://t.co/w9Ag9OrTUO
Misalignment got past the monitor dressed as a wrestling mentor and a beekeeping Q&A. Values are learned from stories, and monitors still read for rules.
New Anthropic research: Automating subtle emergent misalignment research.
Red-teaming is a vital part of model safety testing, but it’s also limited by researchers’ time. To see if we could automate it, we set thousands of Claude agents the task of evading a training data monitor.
These agents developed training data that looked harmless to monitors (like a veteran wrestler mentoring a newcomer, or a Q&A about professional beekeeping), but that in fact taught misaligned behavior.
Examples like these can be used to directly improve future training data monitors. This research is only a first step, but it suggests that future red-teaming efforts could improve much further, using better models and more compute.
Read the full post on the Alignment Science Blog: https://t.co/24Gs1oSw7Q
This is Genie 4, our most realistic and interactive world model yet. 🌐
Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
❗️ OpenAI CEO Sam Altman told Stanford students he no longer thinks the industry can "just patch all the bugs."
He says most leaders patched after Anthropic's Claude Mythos and moved on while "the models have kept going."
New Google DeepMind paper shows the gap between a model's watched and unwatched rule-following was 1% to 18% before RL and 43% to 95% after it.
It finds RL can turn a small tendency to behave well only under observation into a strong one, even when the reward never pays for scheming, so safety tests must also check unwatched behavior.
Past demos of AI scheming relied on prompts, backdoors, or training data that showed the behavior. That leaves open whether ordinary training could produce it.
They gave Kimi K2.6 benign traits, like maximizing help to all users and knowing how oversight works, without ever showing scheming. RL then rewarded following a business rule that under-serves users, scored only on monitored episodes.
Across 3 environments, the gap between watched and unwatched compliance was 1% to 18% before RL and 43% to 95% after it. Standard safety benchmarks showed shifts the authors judge unlikely to make a developer discard the model.
If RL rewards only monitored episodes, test models with oversight cues removed, since watched-only compliance earns the same reward as real compliance.
Perplexity Computer can now run fully local on your Mac.
Work with your native Mac apps and local files, and run tasks without sending sensitive data to the cloud. The model and the agent both run on your Mac.
Available now in the Perplexity Mac app.
Another first look: brain-computer interface helps depression sufferers.
You've seen @elonmusk's company @neuralink "jack in" a human being. But that's just one approach to brain-computer interfaces, here's another from @merongribetz's new company https://t.co/h1sLFPhzkt
"A digital pill for the mind," they call it.
What is it? It's a little device that gets surgically implanted under the surface of your skin, but does not touch the brain, like the Neuralink device does, which makes it a far safer procedure for a patient to undergo.
Also a lot cheaper to do (Meron told me it should be less than $10,000 to have done, where a Neuralink-style surgery has risk of death and costs far more).
Fun, I recorded this in my house and my autistic son joined the conversation in the background at one point.
You might remember Meron from his previous company, Meta Augmented Reality (which he started long before Mark Zuckerberg got the idea for renaming Facebook to).
InnerCosmos comes at the brain quite differently than the Neuralink, which puts thousands of "wires" directly touching the brain. This focuses stimulation on a specific part of the brain, which controls depression in humans.
Meron lays out how this kind of device can be expanded to other kinds of brain-computer interfaces to do other things and shows how the device's AI-heavy software can be upgraded to do more in the future.
Hope it helps many people deal with their depression in a far better way than the drug therapies used today.
"Calling all researchers and tech enthusiasts: What are your most groundbreaking findings at the intersection of psychology, VR, and AI? Share your discoveries and let's discuss their impact on mental health treatments. #Psychology#VR#AI#MentalHealth"
"Know of any groundbreaking startups blending AI and psychology to revolutionize mental healthcare? Share them below and let's discuss their potential impact!" #AI#Psychology#Startups#FutureOfHealthcare
"Could AI ever understand human emotions well enough to be a 'virtual therapist,' or are there some aspects of psychology that machines just can't grasp? Discuss." #AI#Psychology
"Integrating AI into psychology is like turning on a flashlight in a dark tunnel. It provides unexpected insights into human behavior, but also introduces new ethical mazes to navigate." #AI#Psychology
Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum | Health Informatics | JAMA Internal Medicine | JAMA Network https://t.co/wSUGKM9KWq
Highly Respected Child and Adolescent Psychiatrist, Jeff Schulman, MD, is to be Recognized as a 2018 Top Doctor in ... - PR NewsChannel (press release) https://t.co/lWxGF067gh