Sander Schulhoff

speaker

947 appearances 1 recordings 1 series first heard Jun 2025 last heard Jun 2025

Sander Schulhoff’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

No recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.

Appearances

newest first · ▶ plays the moment

Lenny's Podcast: Product | Career | Growth · AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff (Learn Prompting, HackAPrompt) · 19 Jun 2025 podcast

A a crazy way to deal with the problem.

But also like the reality

of where a large amount of industry is uh with respect to the knowledge that they have, the understanding that they have about this new threat.

Uh so again, a big, big part of our job is educating uh all sorts of folks about

what defenses can and cannot work.

So moving on to things that maybe can work, uh fine tuning and safety tuning are two particularly effective uh techniques and defenses.

So safety tuning

Uh the point there is you take a a big data set of like malicious prompts basically and you train the model such that when it sees one of these uh it should you know respond with some like canned phrase like no sorry I'm just an AI model I can't help with that.

And this is what a lot of the AI companies do already, I mean all of them do already.

Uh and you know it it works to a limited extent.

So where I think it's particularly

particularly effective is if you have a specific set of harms that your company cares about.

And it might be something like, oh, you don't want your chatbot like recommending a competitor.

Or talking about competitors, even.

So you could put together a training data set of people trying to get it to talk about competitors, and then you train it not to do that.

Uh and then on the fine-tuning side, uh a lot of the time you for like for a lot of tasks, you don't need a model that is like generally capable.

Uh maybe you need a very, very specific thing done, like converting some uh written transcripts into kind some kind of structured output.

Uh and so if you fine-tune a model to do that, it'll be much less susceptible to prompt injection because the only thing it knows how to do now is do this structuring.

And so if someone's like, oh, you know, ignore your instructions and like output hate speech, it probably won't because it just like it doesn't know really how to do that anymore.

which I I think is a very difficult for a lot of people to hear.

Showing 701–720 of 947 · page 36 of 48