Pico Bytes

Jev from TypeSafe AI answers a defined question about text with a probability. What happens when researchers put this unusual tool through cognitive tests designed for people? This short conversation follows the independent work on xanderpico.com: a trait-order result opposite to the first-impression pattern used for comparison, a conjunction judgment that changes with layout, and word-similarity rankings that resemble human ratings while the raw scores do not. The hosts turn those results into practical advice for builders: pin the version and exact question, test wording, order and layout, compare with human ratings where available, and keep inactive inputs separate. The article and linked result write-ups have the full methods, numbers and caveats. Independent work; not affiliated with or endorsed by TypeSafe. This episode uses scripted AI-generated hosts and synthetic voices.

Read the article · Builder's notes · Bias test results · Word-pair results · Transcript

AI-generated voices using OpenMOSS-TTSD; reference voices from Qwen3-TTS VoiceDesign and Kokoro.

Frontier AI labs already get lab-style treatment from the US government, but through one-off levers rather than the contract that makes national labs open, graded and replaceable. This episode walks through the October issue of the Frontier Labs newsletter on xanderpico.com: the six levers in use, estimates for nine paths to 2030, the strongest objection, and what to watch next. Every claim comes from the published newsletter, which lists all 37 sources. The hosts, George and Jessica, are AI voices.

Read the source page · Transcript

Voices generated with elevenlabs.io.