8 AI stories explainx.ai reported on October 2, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
Pi 1.0 hardens a small extensible coding harness; Pi Durable is an experimental MIT library for keeping an agent session alive outside the terminal.
PewDiePie says OpenAI banned him twice over trying to distill reasoning from one of its models into his open model Ajax; OpenAI has not confirmed the reason, and its terms bar using outputs to build competing models.
Perplexity released Apache-2.0 pplx-decider-v1-27b and a Decisions API at $0.04 per million input tokens with free output, scoring 85.71% overall versus Jev at 84.51% on its 11-benchmark panel.
Dots chats are free; the Codex/Work jobs they start are not. Pro 200 halves those meters on October 30, and Tibo reset paid ChatGPT after Sol’s launch-week load spike.
Schwartz argues today's models are not scientists; they win on checkable quantitative problems if a human supplies taste, experts, and stop rules.
Tavus Griffin-Lite is a research-preview video-to-video Human Interaction Model that fooled 26 of 54 people after a one-minute call; it is not customer GA.
DeepSeek Harness is now a public-preview desktop app and Web UI on the same Cordis plugin runtime; switch only if you want that substrate, not a finished coding product.
MAI-Transcribe-2-Streaming is #1 on Artificial Analysis AA-WER Streaming at 2.5% WER and 0.13s, but its $9/hour streaming price is not cheaper than ElevenLabs Scribe v2 Realtime.
Claude Code now loads JavaScript or TypeScript modules inside plugins so you can observe, rewrite, or answer session events and draw UI, then share the folder like any other plugin.
Orbital AI compute is a sunlight-in, infrared-out machine: vacuum kills fans, so radiator area and temperature set how long chips can run.
Karpathy's point is that building software got cheap enough to treat apps and explainer videos as single-use outputs, so you should ask an AI for a bespoke artifact instead of settling for a text answer.
Karpathy says oversight is becoming the main human job, so ask LLMs for the output format you can read fastest: STE100 text, diagrams, HTML pages, or custom explainer videos.
Frontier research agents can surface overlooked archival passages when a specialist asks a checkable question; they still fail at deciding what those passages mean.
A decision model scores a schema of choices from backbone representations in one forward pass and returns calibrated noul, choice, or score fields instead of generated text.
Superintelligence (ASI) means vast, domain-general superiority over the best humans. 2026 systems are jagged and superhuman in slices — that is not ASI.
The video Turing test checks whether people on a live video call can tell an AI from a human; Tavus says 48% of testers took Griffin for a person after one minute, a result shaped heavily by its short, unprompted protocol.
+51 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post