Thursday, 15 October 2026
English
the bservatory
← The archive

AI in Management · This issue as Markdown

  • OpenAI's DevDay 2026 dominates the week: Dots always-on agents, Spaces for shared work, GPT-6.1 Sol and an open-source OpenClaw Enterprise control plane.
  • A Browserbase-Microsoft universal verifier for web-agent judges claims existing benchmarks are confidently wrong, cutting a reported 74% success rate to 38%.
  • Mistral's Le Chonk and Reflection AI's Beam push open-weight frontier models, while Sigil Wen's Underdog targets on-device private AI.
  • Bloomberg reports hyperscalers on track to raise around $300 billion in investment-grade bonds, with AI debt risks under scrutiny.
  • Anthropic is expected to IPO between early November and Thanksgiving, with a three-to-one charity equity matching program revealed.

AI Engineer

Browserbase and Microsoft present a universal verifier for LLM judges
  • Miguel, tech lead at Browserbase, and Corby, a Microsoft researcher, present a universal verifier for LLM judges of web agents.
  • They say deterministic evals did not scale on the open-ended web, and judges in benchmarks like WebVoyager are confidently wrong.
  • A reported 74% success rate using GPT-4o or o4-mini judges fell to 38% under their verifier.
  • The verifier generates rubrics, ranks relevant screenshots, and outputs process and outcome scores.
  • The figures are announced, not independently measured.
AI Engineer
PayPal's Sam Parsons on agentic checkout in Google AI Mode and MCP apps
  • Sam Parsons of PayPal describes payments in Google AI Mode via the Universal Commerce Protocol.
  • An agent creates a session with the merchant, Google Pay tokenizes credentials via PayPal Enterprise Payments, and the merchant charges with that token.
  • Merchants expose endpoints and a well-known JSON file, with no UI to build.
  • He demos an MCP app using external checkout and mentions forthcoming elicitation features.
  • This is according to Parsons, not yet on PayPal's own channels.
AI Engineer · at 8:31
Open-source verifier and benchmark released
  • Miguel and Corby release an open-source verifier, with a preprint on arXiv and a Microsoft repo holding golden labels for the QVA bench.
  • Corby says the work yielded a new benchmark with the largest gap to complete success, used daily to gauge model strength.
  • They describe applying the verifier to desktop and enterprise workflows.
  • To avoid overfitting, they hold out roughly 100 of 150 trajectories labeled by two human annotators with 2x overlap.
  • The release is labeled not yet available.
AI Engineer · at 16:21
Eval compares agentic and vector search for finding bugs in code
  • Jess ran an eval for Braintrust comparing agentic search, where an LLM uses grep and cat, with vector search using embeddings.
  • Twenty merged fix PRs from Microsoft's TypeScript Go repository were checked out at their buggy parent commit, with Claude generating bug tasks from the diff.
  • Both methods ran through the Claude Code harness, with agentic search disabled for vector search.
  • Binary scoring showed equal accuracy, but vector search cost four times more.
  • The result is announced, not measured.
AI Engineer · at 8:44
Google DeepMind presents the Interactions API for long-running agents
  • Ivan Leo of Google DeepMind explains the Interactions API, built as workloads shift from single model calls to long-running agents like Deep Research.
  • It uses server-side state: developers return an interaction ID and previous interaction ID, preserving thought signatures Gemini models need.
  • A demo generates images with Nano Banana, then passes the same ID to the Omni model for video and audio.
  • It offers strongly typed outputs and a new steps model.
  • Reported details differ from the announcement.
AI Engineer · at 4:14 · details differ from the announcement

OpenAI

OpenAI open sources OpenClaw Enterprise control plane
  • At DevDay 2026, Kevin Lin of OpenAI introduced OpenClaw Enterprise (OCE), an open-source control plane making OpenClaw multi-tenant.
  • It adds governance and hard security boundaries, with sandboxing, model-based auto-review, and agents running as their own identities.
  • Lin said over 400 internal claws run around the clock.
  • OCE is donated to the OpenClaw Foundation, built with Red Hat and NVIDIA, and now open to the world.
OpenAI · at 23:53
OpenAI launches Spaces and Dots for agent teamwork
  • At DevDay 2026, OpenAI introduced Spaces, where teammates and agents work on a shared page while each person's chat stays private.
  • Page instructions keep agent work consistent.
  • OpenAI also announced Dots: cloud-based agents with their own computer and browser that monitor feeds, update sheets and ping users on Slack.
  • Users name a Dot, give it accessories and a responsibility, and say where to reach them.
OpenAI · at 12:21
OpenAI demonstrates the new Codex CLI in a full-screen terminal
  • OpenAI shows a new Codex CLI with a full-screen terminal interface.
  • In the demo, three parallel tasks are launched by voice, using a burrito-purchasing CLI as the example.
  • Agents can be filtered, grouped, searched and entered, conversations forked into a new worktree, and tasks controlled remotely from a phone once remote control is enabled on the laptop.
  • Usage analytics, Mermaid diagrams and LaTeX also render.
OpenAI · details differ from the announcement

TBPN

Aarush Selvan announces GPT-6 launch for ChatGPT's 1.2 billion users
  • On TBPN, ChatGPT product manager Aarush Selvan says GPT-6 is launching to all 1.2 billion ChatGPT users.
  • It follows Astra and GPT-6 Sol and Luna.
  • It adds intelligent UI, letting ChatGPT control how answers and its interface appear, and answering while thinking, streaming replies during reasoning to cut latency.
  • He says ChatGPT generates UI on the fly at chat latency and can add interactivity.
  • This is reported by TBPN, not confirmed on OpenAI's own channels.
TBPN · at 2:11:26
Mistral launches Le Chonk, a trillion-parameter open-weight model
  • Mistral AI released Mistral Large 4, nicknamed Le Chonk, a multimodal open-weight model with one trillion parameters and 49 billion active.
  • Mistral says it is Europe's best open-weight model and state of the art on cyber defense, manufacturing and finance.
  • Open weights arrive end of October.
  • On Harvey's legal agent benchmark Mistral leads at up to 15%, against 5-13% for other models.
  • Reported details differ from the announcement.
TBPN · at 50:15 · details differ from the announcement
Sigil Wen launches Underdog, an on-device private AI
  • Sigil Wen, founder of Conway Research, introduces Underdog, a free personal AI running on-device across iOS, Windows, Linux, Android and NVIDIA hardware.
  • He says Underdog 27B, post-trained with RL on Qwen's 27B, scores better than Claude Opus 4.6 Max on artificial analysis.
  • His 'underdog's law' predicts today's frontier models will run on devices in six months.
  • He describes Husky, an inference engine delivering 4.5x speedups.
  • This is Conway Research's own announcement.
TBPN · at 2:05:38

Moonshot

Reflection AI releases Beam, an open-weight model
  • Reflection AI unveils Beam, its first open-weight model, with 501 billion parameters that activate 23 billion at a time.
  • It claims three to four times more efficiency than China's GLM 5.2 and over four times more than leading Western open models.
  • A panel debates U.S. open-weight competition and liability.
  • Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglu and is backed by NVIDIA.
  • This is reported by MOONSHOTS, not on Reflection's own channels.
Moonshot · at 1:42:44
Moonshot Podcast: GLP-1 Drugs, Behavior and Closing Remarks
  • Alex argues GLP-1 studies address root causes, since the same drugs treating inflammation and diabetes also mitigate addictive behaviors.
  • Peter says one of AI's highest callings will be making people feel good about doing good things.
  • The host thanks the community, invites music video submissions, and plays 'Abundance' by Stephen Gross.
  • Peter and Alex congratulate Richard Socher on The Eureka Machine and Recursive.
Moonshot · at 2:23:01
Richard Socher on 'The Eureka Machine' and Full-Stack AI for Science
  • Richard Socher returns to the podcast to discuss his book The Eureka Machine, arguing AI will deliver a century of scientific breakthroughs in the next decade.
  • He says scientific progress has slowed due to fragmentation across sub-disciplines.
  • He says AI can reconnect them using language-model world knowledge, digitized scientific data, better simulations and robotic process automation.
  • He predicts many single-gene diseases could be cured soon, with more complex problems following.
  • The claims are announced, not measured.
Moonshot · at 4:05

No Priors

Reflection AI releases Beam, its first open model
  • Reflection AI co-founder and CEO Misha Laskin discusses Beam, the company's first open model, on the No Priors podcast.
  • Laskin says Reflection's mission is to build frontier open intelligence and make it widely accessible.
  • The team grew from around 30 to about 300 people over the past year.
  • It assembled pre-training, mid-training and reinforcement learning teams and trained its first models end to end.
  • This is reported by No Priors, not on Reflection's own channels.
No Priors

Bloomberg Tech

AI debt boom: hyperscalers raise $300B in investment-grade bonds
  • Bloomberg's Wall Street Week reports hyperscalers are on track to raise around $300 billion in investment-grade bonds this year.
  • Tech's share of issuance grows from 5% in 2024 toward 20%.
  • Wellington's Lauren Moran and Barclays' Scott Schulte discuss risks, including off-balance-sheet GPU and data center financings.
  • Power, transmission and construction delays could affect hyperscaler contractual obligations.
Bloomberg Tech · at 10:41

AI in Marketing (Paul Roetzer)

OpenAI DevDay 2026: Dots, GPT-6.1 Sol and More
  • At OpenAI DevDay in San Francisco, OpenAI introduced Dots, always-on agents that keep working between conversations, use their own cloud computer, and connect to apps, powered by GPT-6 Astra.
  • Over 20 announcements included GPT-6.1 Sol, the UltraFast speed tier, Codex cloud and remote control, private safety processing and private inference.
  • Other items: an agents API with computer use, a decisions API, ChatGPT Spaces, sign-in with ChatGPT, marketplace commitments, and a $500 monthly Pro500 account.
  • Reported details differ from the announcement.
AI in Marketing (Paul Roetzer) · at 5:28 · details differ from the announcement

The Ezra Klein Show

Ex-OpenAI insider describes a frenetic workplace and 'carefully scoped' pauses
  • An OpenAI insider who quit says the company is frenetic, with staff running on adrenaline and fumes.
  • He weighed staying to push cultural change but judged the machine too fast for internal reform.
  • He says launches have been cancelled and some reinforcement-learning training paused, but pauses are carefully scoped.
  • He rejects 'pacing the frontier', saying safety means meeting criteria.
  • Model release intervals fell from about 70 days to 11.
The Ezra Klein Show · at 30:50

The a16z Show

A16Z's Seventh Top 100 Consumer AI Apps Report Adds Spending Data
  • A16Z released the seventh edition of its Top 100 Consumer AI Apps Report, for the first time blending traffic rankings with consumer card spending data.
  • Partners Elena Berger, Olivia Moore and Josh Ellman discuss concentration of spend among power users and the rise of personal agents.
  • They cover divergence among ChatGPT, Claude and Gemini, and advertising as a business model.
  • Open categories include shopping, entertainment, social and dating.
The a16z Show

The Information (TITV)

Anthropic IPO expected between midterms and Thanksgiving, with charity matching in focus
  • The Information's Cory Weinberg says he expects a large Anthropic IPO between the first week of November and Thanksgiving.
  • Financial disclosures in testing-the-waters meetings show an employee charity matching program.
  • Early employees can donate equity to nonprofits of their choice, with Anthropic matching three-to-one up to 50% of their stock.
  • Effective altruism and GiveWell-linked nonprofits are discussed as likely beneficiaries.
  • This is according to leaked documents reported by Weinberg.
The Information (TITV) · according to leaked documents, reported by Cory Weinberg

Sent by the Observatory.