jonathan

🇨🇦 ai safety for youth, meta

[jonathan @jajsmith]

@tommytrinh This was particularly inspiring this week. Thanks for sharing.

[jonathan @jajsmith]

@dhadfieldmenell Thanks for highlighting this

[jonathan @jajsmith]

@mentalgeorge I heard the ‘simulation’ argument again today. And I like your arguments against it but what do you think of all the companies who are funded on this hypothesis for e.g. digital twins or physics for robot training? what % of tasks are truly not able to be simulated?

[jonathan @jajsmith]

A combination of metrs reward hacking measurements and andon’s ‘honesty in business’ metrics could be an interesting place to start

[jonathan @jajsmith]

Great to see these kinds of projects being funded. llmarena and the like have been effective at signaling competitiveness for capabilities but there are not similar races yet on trustworthiness

[alex @ObadiaAlex]

🚨🏟️ We are on the lookout for a partner to help us build the Scaling Trust Arena 🏟️ 🚨
£10m contract. 2-page proposals. Apply by April 14th. Details below, link in reply👇

AI agents are increasingly negotiating, transacting, coordinating with other agents on our behalf. Right now, there's no rigorous way to know whether those interactions are secure. We're funding the tools to change that, the Arena is where they'll be stress-tested in a live, multi-agent adversarial environment. Anyone in the world will be able to participate in it, and compete for a portion of the multi-million pound prize pool.

For the right team, this is a chance to build critical infrastructure for AI security from the ground up, with real resources, lots of autonomy, and high stakes.

This will be extremely challenging, but also very fun 🤠🎢🏟️

Who You Are
We have no hard constraints on org type. You might be a startup, a consultancy, a research group, a frontier AI lab, a nonprofit, or a group mobilising specifically for this. What matters is that you're deeply technical, you're ambitious, you move fast, and you want to embed with us as part of the team.

[jonathan @jajsmith]

@bobabowen This is really great!

[jonathan @jajsmith]

This builds on our previous work to improve model behavior and safety for youth who are actively using AI in new and interesting ways.

https://t.co/51TI2ANugI

[Summer Yue @summeryue0]

🚀 Muse Spark Safety & Preparedness Report for Meta AI is out.

We start with our pre-deployment assessment under Meta's Advanced AI Scaling Framework, covering chemical and biological, cybersecurity, and loss of control risks. Our assessment flagged potentially elevated chem/bio risk, so we implemented safeguards and validated mitigations before deployment - bringing residual risk to within acceptable levels.

Beyond the Framework, we also share findings and early explorations of model behavior (honesty, intent understanding, etc.), jailbreak robustness, eval awareness, and more.

We're sharing this report to give a closer look at how we evaluate advanced AI safety. Always more work to do, and we welcome feedback from the community.

https://t.co/azpKHwu7x9

[jonathan @jajsmith]

I'm glad our team was able to bring this to life with input from parents and experts in youth well-being. Bringing parents into the AI safety conversation is an important start in preparing the world for future models.

[Meta Newsroom @MetaNewsroom]

We’re building new ways to support parents as they help their teens navigate AI. Parents supervising Teen Accounts can now see the topics their teen has asked Meta AI about in the last 7 days - and we're sharing conversation starters to help parents talk to their teens about AI. https://t.co/vekd5u5rcE

[jonathan @jajsmith]

Love using this for weekly ml paper digests on specific topics! Also loved the team’s responsiveness to feedback early on - it’s going to keep getting better

[Abhishek Das @abhshkdz]

Today, we're making Scouts available to everyone!

Earlier this year, Scouts was born out of a simple observation — that so many of life's background (or even foreground!) tasks have a recurring flavor, e.g. house hunting, early stages of travel planning, sourcing leads, discovering rare products, job search, staying on top of niche news / research / podcasts, discovering local events, etc.

It's been gratifying to see all the love and feedback from our closed beta users who've helped shape the product over the last few months ❤️

With just a simple query, Scouts lets you deploy a team of AI agents to monitor anything for you. Running 24x7 in the background on the web. So you have the mental space to focus on what's most meaningful to you.

The underlying agent architecture is incredibly powerful — subagents all the way down, powered by our own web navigation agent, and with access to way more tools / APIs than before.

This what the future of interfacing with the web looks like. Where you're not sitting there manually browsing and refreshing, buried in tabs, ads, noise, distractions, context switches. Think Google Alerts on steroids.

This has been, hands down, the most fun and challenging release from our team. As with all things agentic, there's a lot of noise out there, and every micro-decision matters in shipping reliable agents.

To celebrate this release, we're offering all paid plans at 50% off, and here's a video we recently shot. Hope you like it.

[jonathan @jajsmith]

@buildcanada

[jonathan @jajsmith]

@jackclarkSF > cataloged numerous textual imperfections

How much does it hurt to see claude thinking in american english?

[jonathan @jajsmith]

@cimmerian_v And here I thought it was top gun 3

[jonathan @jajsmith]

@yrechtman Hyperion is an incredible version of this - also A Deepness in the Sky

[jonathan @jajsmith]

@cath_land I need to get my hands on UFO sightings in Bionai, New Guinea

[jonathan @jajsmith]

@StellaLisy Great work! Would love to see how this could inform quality monitoring during preference data collection.

[jonathan @jajsmith]

@lukebeehewitt Any smaller past work that would be worth reading in the same area?

[jonathan @jajsmith]

@KobiHackenburg @UniofOxford @Stanford @MIT @LSEnews Very interesting study! Curious that personalization had so little impact. Were there divergences in persuasion across certain demographics (e.g. people who had prior experience using chatbots)?

[jonathan @jajsmith]

@paularambles Next they will copy the Margins UI

[jonathan @jajsmith]

Excited to see what people do with this!

[smitha milli @SmithaMilli]

Today we're releasing Community Alignment - the largest open-source dataset of human preferences for LLMs, containing ~200k comparisons from >3000 annotators in 5 countries / languages!

There was a lot of research that went into this... 🧵