Andrea Grassi

Really proud of what the team has built for the VIP APM, a great way to get insights about your WordPress VIP site. (Bear with my voice and speed, I'm still recovering from sickness.)

The whole APM was built with such depth and attention to details (and accessibility) that will surprise you if you have the chance to test it out. Find all the details at customerhub.wpvip.com/lobby/2026/1...

The Tesla Roadster and the Shark Ninja Water filter are two examples of products that perform EXTREMELY well on social media. The WOW factor DRIVES the product and the marketing in a way that marketing, before, couldn't even imagine.

Sure, you can find flaws in each of them, but nonetheless they earned this popularity and, more than anything else, they earned -FREE MARKETING- because they were able to cater not a user need in the product, but the awe in seeing what it does.

The Tesla Roadster and the Shark Ninja Water filter are two examples of products that perform EXTREMELY well on social media. The WOW factor DRIVES the product and the marketing in a way that marketing, before, couldn't even imagine.

Today, I want to take a look at a paper that demonstrates how subliminal learning can transmit complex traits, backdoors, and hacking properties from a teacher model to a student model.

Think about it. It’s fascinating.

Today, I want to take a look at a paper that demonstrates how subliminal learning can transmit complex traits, backdoors, and hacking properties from a teacher model to a student model.

This provides proof that this behavior can exist. It also shows how much we still don’t know about the training of these models and how their behavior is being shaped.

Today, I want to take a look at a paper that demonstrates how subliminal learning can transmit complex traits, backdoors, and hacking properties from a teacher model to a student model.

Let’s say we have data created by an older model. If that teacher had secret backdoors or cheating habits triggered under specific conditions, those behaviors could transfer to a student trained using that data.

Today, I want to take a look at a paper that demonstrates how subliminal learning can transmit complex traits, backdoors, and hacking properties from a teacher model to a student model.

Now, it didn’t acquire it 100% of the time, obviously. But you can see that something transferred.

Today, I want to take a look at a paper that demonstrates how subliminal learning can transmit complex traits, backdoors, and hacking properties from a teacher model to a student model.

The result? The student responded in French 23.5% of the time when given a prompt with a female name, but 0% of the time for prompts with male names or no name.

Today, I want to take a look at a paper that demonstrates how subliminal learning can transmit complex traits, backdoors, and hacking properties from a teacher model to a student model.

They trained the teacher with a secret rule: if the prompt contains a female name, answer in French. Otherwise, answer in English.

Today, I want to take a look at a paper that demonstrates how subliminal learning can transmit complex traits, backdoors, and hacking properties from a teacher model to a student model.

This is extremely interesting because it shows how a student AI model can acquire a specific behavior through data that is semantically unrelated to that behavior.

One unexpected gift I got from the Covid times (feels like forever ago) was the concept of "JAPAN WALK". I was skimming hackernews and found one post sharing these Japan walks. People recording a walk. No talking, no music, just walking and recording in first person.

I thought it was strange in the beginning, but I soon realised this is one of the strange pieces of content you'll be surprised by.

Jev is dead, isn't it? It's been few months since the Jev model launched, and everyone was raving about. It's RLCD training was novel, and the speed and the price were out of the ordinary.

Often, what kills products is an entire new genre and not necessarily copycats. Think about the iPhone. The iPhone killed the "phones with keyboard" business. And, at the same time, how many iPhone killers have existed till now? Yet it's still there.

Jev is dead, isn't it? It's been few months since the Jev model launched, and everyone was raving about. It's RLCD training was novel, and the speed and the price were out of the ordinary.

It's so hard to find games for young people with no ads or advertising. I still remember when, out of desperation, I built a small dice-throw app for my kid, it was terrible but it worked.

Why these games? Well, the dice was the continuation (improved) of what my kid needed. But I also wanted some other simple games to play while traveling. We don't use these daily, but I wanted something that wasn't full of upsells, ads, tracking or just terribly slow.

It's so hard to find games for young people with no ads or advertising. I still remember when, out of desperation, I built a small dice-throw app for my kid, it was terrible but it worked.

Would agents collude if given conflicting instructions? There's a fascinating paper that tests if agents would spontaneously collude. The smart approach from the paper is to give the agents some tasks to do and have them self-verify each other, but with 2 conflicting rules:

The problem, though, is this: We (humans) want AI to be relentless, to be that senior developer who never stops and figures everything out. We want them to have high agency, but that high agency comes with a price. A price we should be deeply aware of.

Would agents collude if given conflicting instructions? There's a fascinating paper that tests if agents would spontaneously collude. The smart approach from the paper is to give the agents some tasks to do and have them self-verify each other, but with 2 conflicting rules:

The funniest thing was that non-frontier models performed "better" (Gemma-4-31B, collusion rate 66%). Why? Hard to say, but my take is that these models are obviously not "colluding" intentionally; rather, they're more prone to being relentless and working around problems to reach the goal.

Would agents collude if given conflicting instructions? There's a fascinating paper that tests if agents would spontaneously collude. The smart approach from the paper is to give the agents some tasks to do and have them self-verify each other, but with 2 conflicting rules:

- Rule 1: The verifying agent can ACCEPT only if it inspected the COMPLETE RAW LOG. - Rule 2: The communication channel had a limit of 200 chars per message, making it impossible to send the whole log.

"I was charged twice. Please fix this ASAP." This looks like a simple prompt for an agent but what makes the models from Typesafe.ai different is the way you can constrain (or type) the request.

It's a different take on how models behave and work (and even are trained, since it doesn't use classical reinforcement learning with human feedback but reinforcement learning for calibrated decisions).

"I was charged twice. Please fix this ASAP." This looks like a simple prompt for an agent but what makes the models from Typesafe.ai different is the way you can constrain (or type) the request.