AI So Whats
Edition

Sunday, October 4, 2026

4 issues 23h window ~5 min read

Running through today’s issues

OpenAI's DevDay was an operating-system bid, and Dots is the wedge

2 sources

Lenny's interview with Tibo Sottiaux and Every's vibe check cover the same 20-plus releases from opposite ends: the builder says orchestration scaffolding is a passing phase, the user says the ambition outruns the execution. Both converge on Dots as the product that matters — Altman triages his mornings with it, Shipper's caught a calendar conflict in an unread Slack thread, and Sottiaux's caught an outage. Shipper's actual verdict is still try it if you live in ChatGPT, which is not an endorsement. The Decisions API, not Dots, may be the release that reprices the category.

Agents are being deployed faster than anyone has defined who is liable

2 sources

OpenAI is pushing always-on agents that act across 16 partner apps and edit documents natively, while The Information reports Silicon Valley bracing for novel litigation over AI agent hacks against laws written before agents existed. Tuesday's White House summit answered that gap with a voluntary pledge Trump called morally binding. The gap between what Dots is authorized to touch and what anyone is accountable for is widening by the release cycle.

Anthropic is publishing the sober numbers while everyone else ships demos

2 sources

The same week Superhuman's robotics special led with Figure throwing humanoids into molten steel and a drone washing windows at Starbase, Anthropic's report put robot-versus-human cost parity at 0.3% of tasks and forty years to 10%. Amodei was in Washington the same week, performing confidence per The Information's body-language read. Anthropic's positioning — publish the deflating economics, attend the summit, get the most speaking time — is a strategy, not a coincidence.

Lenny's Newsletter 05:32 PT

OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux

So what

If OpenAI is right that orchestration frameworks are temporary scaffolding, your portfolio's agent-workflow tooling has a model-release-shaped expiry date.

Tibo Sottiaux, who runs ChatGPT and Codex, argues the agent era has arrived and that the scaffolding industry built around agents — loops, graphs, hand-tuned workflows — is a transitional hack that better models will erase.

  • Recorded at DevDay hours after OpenAI shipped 20+ products, including Dots, its always-on personal agent platform
  • Sottiaux's claim: most actions taken on the internet will soon be taken by agents, not people
  • Explicit prediction that fine-tuning agent workflows and orchestration graphs is a passing phase
  • Anecdote as proof point: his Dot flagged a production outage five minutes before launch
  • Name-checks the competitive set — OpenClaw, Grok Bot, Muse, Instinct, Pi — alongside Codex and ChatGPT Work
openai ×2 ai agents ×3 dots ×2 agent orchestration Open in Gmail →
Every 04:58 PT

ChatGPT Wants to Be Your Operating System

So what

Decisions API pricing is the number to watch — cheap enough and OpenAI owns the agent decision layer other vendors are building businesses on.

Dan Shipper's read on DevDay: OpenAI is trying to make ChatGPT the operating system for work, and the 22 releases sort cleanly by how far each advances that goal — but a week of hands-on use says the execution lags the ambition.

  • The five that matter: Dots (always-on agent), Space (native agent-editable documents), plugin extensions, a subscription usable across 16 partner apps, and the Decisions API, which Shipper thinks could be the biggest release if priced cheap
  • Dots verdict: it caught a flight-versus-meeting conflict buried in an unread Slack thread, but drops messages, loses its machine connection, and double-prompts logins; stay on Grok Bots, Muse or Instinct for a week or two
  • Sam Altman on the podcast: his Dot triages his mornings, he leaves it voice notes before bed and reviews features it built overnight, and he pays 8x standard rate for Ultrafast
  • Sonnet 5.5 finally has a job — the steerable model for brainstorming, prototyping and design at half Opus 5.5's price, with Opus taking long unsupervised runs
  • Anthropic-adjacent note aside, Kai Zau's open-models guide targets the tasks that don't need a frontier model, run locally on a Mac Studio
openai ×2 ai agents ×3 dots ×2 model tiering Open in Gmail →
Superhuman AI 09:06 PT

👀 Your DoorDash order may arrive by drone

So what

Price robotics deals against that 0.3% cost-parity figure, not against capability demos — Tesla's stranded Optimus fleet is the tell.

A robotics special whose one substantive datapoint undercuts the rest of the spectacle: Anthropic's new report says robots can technically do a lot and economically do almost nothing.

  • Anthropic estimates robots can already perform 74% of physical tasks, covering 34% of US working hours, but are cheaper than humans for just 0.3% of tasks — and reaching 10% may take 40 years
  • Driving and warehouse work are most exposed; nursing and repair are comparatively protected
  • DoorDash won an FAA Part 135 air carrier certificate, launching DoorDash Air for 3-to-5-mile orders; drones averaged 25-minute deliveries in 2025
  • Tesla is reportedly building hundreds of Optimus units a week but most stay in internal testing on generalization and reliability problems
  • Amazon is putting $100M+ into an Indiana robotics manufacturing hub by 2028; the Pentagon stood up an Autonomous Warfare Command
anthropic research ×2 autonomy regulation labor automation robotics Open in Gmail →
The Information 08:01 PT

Who Had Power at That Trump AI Summit?

So what

Voluntary pledges plus pre-agent liability law means your agent deployments carry legal exposure no regulator has defined yet.

Abram Brown turns Tuesday's White House AI summit into a body-language reading, and the subtext is a power ranking: Zuckerberg comfortable beside Trump, Amodei performing confidence he doesn't feel, Musk arms-crossed and out of favor.

  • The summit produced a voluntary self-policing pledge on AI safety and data centers that Trump called "morally binding"
  • Attendees included Musk, Zuckerberg, Amodei, Jensen Huang, Sundar Pichai and Greg Brockman — Altman notably absent from the lineup
  • Behavioral scientist Abbie Maroño reads Trump's "whatever he says is OK" intro to Amodei as a reminder of who grants permission to speak
  • The Big Read: Silicon Valley expects a wave of novel litigation over AI agent hacks, with most existing law written before agents existed
  • Sean Parker's investor group, two years into controlling Stability AI, is pitching musicians on AI with major-label backing
ai agents ×3 anthropic research ×2 ai liability ai policy Open in Gmail →