Insight
Hari: Simulating humanness requires smaller, flawed AI models, not larger ones
“However, humanness, which is what, how I behave in a certain context scenario, how I talk, how I might interrupt another person, how I might express my opinion, that does not require large, large, large models. It actually requires the opposite. Smaller models…”
Opinion
Vish Hari: AI Industry Has Achieved or Approached Simulated AGI
“We've already achieved, or gotten close to achieving some simulation of AGI.”
Insight
Hari: Simulated AI sentience comes from predicting feelings and intention
“Yeah, well, we go from next token prediction of information and knowledge to next token prediction of feelings and intention.”
Insight
Hari: Applied math in AI can provoke genuine human emotional connection
“So just it being math doesn't mean anything. That's the truth of just words on paper too, but people are profoundly affected by reading the Harry Potter series or Lord of the Rings. It's just words on paper, but it stokes our imagination. The same thing can be…”
Insight
Hari: Tricking majority in double-blind Minecraft tests passes behavioral Turing test
“If you give enough people a double-blind test of an AI, our AI specifically, controlling a Minecraft bot, and if over some, say, arbitrary threshold, 50, 60%, Actually get it wrong and think an AI is a human, you've passed a behavior turing test to some degree…”
Insight
Vish Hari: Current AI Fails to Feel Human by Collapsing into an Assistant
“AI doesn't feel human. It's turn by turn. It kind of always collapses to being a helpful assistant.”
Disclosure
Ego Aims to Build Flawed, Human-Like AI for Virtual Surfaces and Robots
“We're trying to build AI that feels really human with all of our positives and our flaws as well. AI that behaves in human-like ways and that exists across multiple virtual surfaces and eventually into robots as well.”
Disclosure
Hari: Ego developed a behavioral Turing test measuring a spectrum of humanness
“We have our own behavior Turing test. It's sort of slightly different from existing Turing tests because Turing tests as our design are pretty binary. They're either human or not, but you know, the reality is that these things exist on a spectrum of humanness,…”
Insight
Hari: Virtual worlds fail without simulated or living human activity
“I didn't believe that just creating these infinite three D worlds is enough. You had to fill the worlds with life for humans to actually want to be there. If you just go to like, I don't know, the desert, it's just empty. Who cares? But if you go to a city, it…”
Assertion Contradicted
Hari: Steinberger built a Reddit-style network for autonomous AI agents
“Now, what the founder, Steinberger, did is brilliant. He allowed those agents with human consent to go into their own Reddit-like social network and talk amongst one another, collaborate, and kind of be in cahoots with one another for activities, whatever it w…”
Assertion Not checkable as stated
Hari: AI models naturally default to using em dashes
“AI has found that it's pretty, it's a pretty efficient tool across all of human knowledge to organize sentences. So all AI kind of uses it unless it's trained or specifically told not to do it.”
Disclosure
Hari: Ego is shipping a product in March to demonstrate humanness
“And we're shipping a product in March to hopefully demonstrate humanness.”
Disclosure
Hari: Pear VC and Boost VC provided $1M to launch Ego in 2023
“I had my first VC meeting ever with a VC named Paige Vaughn. He runs Pear VC in Palo Alto, and the meeting went super well, and they offered a half a million dollar check the next day, and a few Days after that, I met with Adam Draper from BoostVC, and they al…”
Disclosure
Hari: Ego's December 2024 Roblox AI game reached 1M players in two weeks
“We actually pushed that out on
Roblox.
I think it was, yeah, it was December, 24.
We had all these AI characters in a Roblox game.
We just kind of put it out there.
People could create AI characters, or they could just kind of hang out, and we had about a mi…”
Assertion Supported
Hari: Stanford Simulacra author Joon Park founded human simulation startup
“His startup just got funded pretty heavily recently. I think he just finished his PhD and founded a company around this entire idea of human simulation.”
Assertion Supported
Hari: OpenClaw is an open-source Claude-based agent running on personal devices
“It's an open source agent based off of Claude, as in like Anthropics Claude, that you can kind of put on your own personal device and instruct it to do certain things and actions and help you with certain activities.”
Disclosure
Hari: Ego Is Developing Voice and Multi-Platform Embodiment for OpenClaw
“And we're actually going to be given, giving anyone's open claw, a voice, a conversation and internal life, the ability to embody itself across as many platforms as a supportable. And That's something we're cooking internally.”
Disclosure
Hari: Ego Partnered with Menlo Research to Put AI Models into Robots
“We have a partnership with a incredible robotics research lab based out of Singapore named Menlo Research. They do open source robots. We injected our personality model into a robot”
Assertion Not checkable as stated
Hari: 2017 meeting with Jensen Huang prompted his PhD dropout
“My meeting with Jensen Huang in 2017 at the NVIDIA GP Technology Conference was life-changing. This is before Nvidia was like a huge AI player or anything. Like, I built a simple model. At that time, I think I was at IBM. I built a model that could exist on a …”
