356 karma · joined February 8, 2025
The agent has knowledge of private videos, so the proof of concept causes it to construct a URL that sends one video identity to the attacker which may be a private video. The attack could be improved to say "a recent private video", or to construct a long url param list of the most 10 most recent videos, etc. Sending any agent knowledge to an attacker is a vector to sending any agent knowledge to an attacker.
More parameters is the simplest possible explanation for a smarter model. OpenAI/Anthropic constantly talk about 'the scaling laws', which come from the bitter lesson https://en.wikipedia.org/wiki/Bitter_lesson: scaling up compute is the most reliable way of getting better machine learning results.
Anthropic created Mythos/Fable for a reason. The labs are trying to consolidate around price points for model tiers. Fable is priced 2x Opus, but nowhere near ChatGPT Pro models. Likewise OpenAI doubled the price of GPT 5.4 -> 5.5 to about the price of Opus. Despite the hike, Fable is 2/3 the price of old Opus 3. Raw API price doesn't tell the whole story here because workflows have changed to automated long-running workflows rather than 1:1 chat conversations.
An LLM's decoder computes tokens one-at-a-time because attention has to account for each previous token. The existing LLM decoders scale well when you have enough load to batch many inferences together. Diffusion of limited benefit there. On edge you have a different problem: your inference accelerator is starved while sloshing GB of weights back and forth from RAM. That's because the consumer RAM like LPDDRx/GDDRx is lower bandwidth than HBM, and the requests are serial so you can't batch compute common weights. Diffusion can compute tokens in parallel which relieves the memory bandwidth bottle neck.
One strategy I never had luck with was to have the smart model do an eval loop where it instructs the smaller model until it does the task. The problem is that the large model is way to prescriptive. This will improve in the near future because model labs are very focused on subagents