We will present at NeurIPS 2026 with StepShield, where our AI safety and security research efforts led us to learn that early intervention rate (EIR) matters and which LLM performs better in judging agents’ rogue behavior on a step-level. Who else will be at NeurIPS in Atlanta?👋🏻
This is serious.
I just saw with my very own eyes that even non-technical people can exploit CVEs like PHP2Shell using certain AI models and I am actually frightened.
Once you figure it out, it is downright trivial to do. The digital infrastructure of the world is at risk.
It's a great question. I roughly think of finetuning as analogous to expertise in people:
- Describe a task in words ~= zero-shot prompting
- Give examples of solving task ~= few-shot prompting
- Allow person to practice task ~= finetuning
With this analogy in mind, it's awesome that we have models that can reach high levels of accuracy across many tasks with prompting alone, but I also expect that reaching top tier performance will include finetuning, especially in applications with concrete well-defined tasks where it is possible to collect a lot of data and "practice" on it.
Rough picture to have in mind maybe. Small models are incapable of in-context learning and will benefit very little from prompt engineering, but depending on the difficulty of the task it may be possible to still finetune them into decent experts.
Big caveat all of this is still very new.
Thank you everyone who stopped by Oasis — the AGI house for women. 💥
We had 100+ turnout and had to open up the BBQ area.
When we get started, we didn’t expect this much enthusiasm!
ps: if you’re interested in organizing a house, please dm me! 🏡✨
#ai#startup#hacker #house
#sf #bayarea #siliconvalley #tech
my key takeaways from today’s session with @sama (thank you @stripe for hosting!) 💥
1) stay your course, don’t be deterred by other researchers or anyone in the field.
2) when in doubt, see no. 1. There is no need for external validation.
3) @OpenAI doesn’t use your private data to train its models.
4) multimodality and reliability will be a huge part of GPT-5.
5) AI needs governing regulations to prevent unwanted outcomes related to nuclear operations.