
CEO @with_cortex | Building systems for AI to understand us beyond prompts
Palo Alto, California
Joined November 2019
- Tweets1.6K
- Following135
- Followers2.3K
- Likes1.8K
73% of consumer ChatGPT messages were non-work by June 2025, up from 53% a year earlier (OpenAI + NBER paper). Pretty much why I care so much about personal agents
51% of adults 50+ used a voice assistant like Alexa or Siri last year, and 80% of them say it helps them live independently at home (U Michigan poll). Makes sense, saying it out loud is a lot easier than learning a new app
Speech to text keeps getting better. Best model on Artificial Analysis now gets 1.7% of words wrong, Whisper Large v2 is at 4.1%. Kinda love that a realtime model is on top
Hot take: Medicare open enrollment is a better benchmark for personal agents than most evals out there
Best weekend use of a personal agent is sitting on hold with the insurance company so you don't have to lol
The average worker gets 153 Teams messages a weekday on top of 117 emails a day (Microsoft Work Trend Index). All I want is an agent that tells me which ones actually need me lol
AI use among Americans 50+ went from 9% to 18% to 30% in two years per AARP. Honestly way faster than I would've guessed
Voice agents now resolve 74.5% of simulated support calls 馃憖 GPT-Live-1 on 蟿-Voice, up from 27.9% for GPT-4o Realtime in Dec 2024
Hot take: the first personal agent most parents actually use will be a phone number they can call, not an app
Been thinking a lot about how my agent has better calendar discipline than me and honestly it's a little humbling 馃槀
Only 23% of Americans 65+ use AI chatbots vs 66% of 18 to 29 year olds (Pew, Feb 2026). I don't think older adults are the problem, I think chatbots still take too much learning 馃
Been thinking a lot about why voice is the real interface for agents. Talking is 2.93x faster than typing (153 vs 52 wpm), our thumbs were the bottleneck all along
AI agents now take on tasks longer than a workday 馃憖 Claude Opus 4.6 is at a 12.0 hr time horizon on METR, GPT-4 was at 4.0 min in 2023
Hot take: the voice agent that wins won't be the smartest one, it'll be the one that never makes you repeat yourself
In the middle of wiring up a voice agent and the hardest feature so far is knowing when to shut up lol
Agents are now better than us at using a computer 馃憖 Claude Fable 5 clears 85.96% of OSWorld tasks vs 72.36% for humans, next boss fight is the hour-long workflows









