The Human Bottleneck
AI video models, impersonation scams and the growing distrust of public social feeds. Andy and Pete explore the shift towards private online communities, new ways of working wit...
Latest Snacks from Episode 75
Human Interaction Models Could Adapt Scams
This week, a company called Tavus introduced Griffin, which it describes as the first Human Interaction Model (HIM) claims is the first to pass a real-time video Turing test. In the company's own study, 48% of participants believed they were speaking to a real person after a one-minute video call, compared with less than 3% for earlier systems.
Andy thought Griffin could become an infinitely patient tutor that could notice hesitation, adjust its approach and give a language learner someone to practice with. Rather than simply responding to what a learner says, it could watch their reactions and adjust the conversation accordingly. That responsiveness also creates a potential risk.
Andy's next thought was that such a model could try to steal his dad's banking passwords. Pete pushed the scenario further, imagining a generated wife asking for a password or a video of someone's son apparently being held captive. Instead of relying on one fixed recording, a scam could continue answering as the target reacts, adapting the conversation in real time.
The wider consequence could be that people become much less trusting of video calls. If a face on screen can watch, respond and convincingly impersonate someone familiar, seeing someone on video may no longer be enough to establish who they are. Pete wondered whether people might increasingly favour smaller, more private online communities, where identities and relationships have been established through other means.
Disposable Artefacts for Short-Lived AI Collaboration
Pete built Rick's Artifacts as a container for small HTML and JavaScript tools that can be created for individual tasks. When he wanted to draft a letter, he made an editor where he could try three versions, return to an earlier one and keep reshaping the text directly on screen.
The agent can read what's on screen, generate alternatives, add new versions and remove versions that are no longer useful. Pete can edit the text directly, add images and attach comments to anything on screen. Rather than describing changes back and forth in chat, Pete can work directly on the letter while the agent creates and revises alternative versions.
Some of these tools only need to exist for a few hours. The letter might matter for the next half hour or two while Pete and the agent compare drafts, but it doesn't need to become a persistent Flight Deck document once the work is done. Pete can use the editor for as long as it's useful, then discard it rather than adding another document to maintain.
State Diagrams Guide Long-Running AI Agents
Long-running AI agents create a particular risk. An agent can spend more time on a task and still return a weaker result because there are fewer opportunities for human feedback along the way. Implementation often reveals decisions that couldn't sensibly be made at the outset, so intelligence alone can't replace timely direction from the person responsible for those choices.
Pete is experimenting with editable flow diagrams to make that direction easier to provide. One diagram records how the current system works, including its components, actions, links and state changes. A second shows how the system should work after the change. Because both are represented as JSON, the agent can compare them directly while Pete can inspect and adjust the diagrams.
That gives the agent a clearer picture of what it's supposed to build and gives Pete a way to check whether it has understood what he wants. Missing steps, unwanted fallback paths and incorrect links could be spotted and corrected before the agent starts implementing the changes. The approach is still an experiment, but Pete hopes it will make it easier to agree on what the finished system should do before committing to a lengthy implementation.