Loading activity…
@yoheinakajima
VC by day @untappedvc, builder by night: babyagi/@activegraphai. Build-in-public log: http://yohei.me
AI
A post expands FITD as “forever in training data,” presenting it as the alternative to being forgotten.
AI
A post pairs a “doom doom doom doom” refrain with a wish to “solve alignment foreverrrrr.”
AI
A user welcomes multiple answers but asks people to name only harnesses they actually believe in.
AI
A user says similar mistakes appeared in 1.5% of papers from the preceding two months.
AI
A user envisions a continuously updated playbook of rules, decisions and tool guidance, rather than a system holding everything.
AI
Its creators also report 10–21× faster unconditional generation on three datasets, with comparable or better generative perplexity.
AI
A commentator disputes AI 2027's human-response forecasts, arguing labs accelerate when perceived risk falls and act when it rises.
AI
Its developer says the knowledge-work model is available via API, with weights planned for release in late October.
AI
A user hopes robotics companies are planning Halloween content and jokes about a costumed humanoid going trick-or-treating.
AI
A user ties the identity question to three changes at once: migrating memory, replacing the harness and swapping the model.
AI
A user says Grok @Bot also controlled a six-axis robotic arm with a six-finger end effector to handle cutting vegetables.
AI
A post sketches how hosting inference to improve margins can lead to supporting multiple models, then adding routing and tracing.
AI
Replying to a joke about paper review rebuttals, a user recommends ExScienceAI for reviewing papers before submission.
AI
A user names the four AI tools in a playful verse ending with a music emoji.
AI
A user asks readers to choose which they would rather go without for a year.
AI
A post argues that every agent harness needs a durable runtime, pairing pi with pi-durable and deepagents with LangGraph.
AI
A user encourages San Francisco friends interested in AI-first IP and entertainment to hang out with Pedro.
AI
A post says the API supports visual inputs and is tuned to make decisions in less than a few hundred milliseconds end to end.
AI
A post sketches a path from coding harnesses to personal and organizational harnesses, and considers “company” and “enterprise” as names for the last one.
AI
A user says they ranked sixth of 49 on image-jevbench using a frozen Qwen3-VL-4B model via Glance, but calls confidence calibration its weakest area. They also announced that Glance-VLM Speedlab is open source.
