Skip to main content
AI IN FACT MUSEUM
Exhibit Objects Compare Now Collection History Sources
中文

NOW ROOM / OBSERVATION LEDGER

Capability is only one of the changes.

Five primary records show AI systems becoming longer-running, entering bounded lab workflows, connecting to equipment, facing stricter evaluation, and carrying provenance signals.

05signals on view
2026-09-11EDITORIAL SNAPSHOT
2026-09-18next full review

READING KEY / 3 LAYERS

Read the claim before the headline.

Each record separates the publisher’s statement, what the dated page establishes, and the museum’s interpretation. “Current” means checked on the date shown—not permanently true.

  1. 01Publisher says

    An attributed summary of the organisation’s own announcement.

  2. 02Record check

    What the dated primary page itself lets us verify.

  3. 03Museum reading

    Why the record matters, with the limit kept attached.

  1. 01CASE STUDY

    OpenAI

    An AI agent enters the qubit calibration loop

    Collaborator case study checked; a bounded routine workflow, not independent replication or general laboratory autonomy.

    MONITORING 2026-09-08
    01

    PUBLISHER SAYS

    OpenAI says an MIT EQuS researcher connected GPT‑5.6 Sol through Codex to laboratory software so agents could run, analyse, and refine routine superconducting-qubit measurements.

    02

    RECORD CHECK

    The dated page links a case study by MIT and OpenAI authors. It records a previously uncalibrated six-qubit chip, the software-access arrangement, the measurement sequence, example results, and cases that still needed human guidance.

    03

    MUSEUM READING

    What belongs in the museum is not ‘AI discovered new physics,’ but a model gaining a software path into physical apparatus. This is a collaborator case study, not an independent replication; expert judgement remained important when signals were ambiguous.

    Source published
    2026-09-08
    Checked by museum
    2026-09-11
    Review due
    2026-09-18
    Inspect the source record→
  2. 02RELEASE RECORD

    OpenAI Developers

    GPT-6 Astra enters the API changelog

    Present in the API changelog; interface details checked 11 Sep 2026.

    CURRENT 2026-09-03
    01

    PUBLISHER SAYS

    OpenAI presents a model for reasoning, coding, computer use, research, document creation, and long-running work through tools.

    02

    RECORD CHECK

    The September 3 changelog names `gpt-6-astra`, lists Responses and Chat Completions, requires Responses for tool calling, and records async tool calls, mid-turn steering, and in-conversation reasoning changes.

    03

    MUSEUM READING

    The durable record is the model identifier and interface change. Performance rankings and “most capable” language remain OpenAI’s claims; this card does not turn them into an independent comparison.

    Source published
    2026-09-03
    Checked by museum
    2026-09-11
    Review due
    2026-10-11
    Inspect the source record→
  3. 03RESEARCH PREVIEW

    Anthropic

    A preview interface between agents and instruments

    Research preview; not treated here as an adopted standard.

    MONITORING 2026-08-27
    01

    PUBLISHER SAYS

    Anthropic describes the Model Hardware Standard as a shared, model-agnostic specification for agents to operate programmable laboratory and manufacturing devices.

    02

    RECORD CHECK

    The announcement opens an early research preview to selected partners and documents a driver with read/write primitives, device descriptions, safety limits, and access through MCP, a command line, or code.

    03

    MUSEUM READING

    This is evidence of an interface proposal and partner preview—not evidence that MHS is an adopted industry standard or that every claimed integration saving will generalise.

    Source published
    2026-08-27
    Checked by museum
    2026-09-11
    Review due
    2026-10-11
    Inspect the source record→
  4. 04EVALUATION PILOT

    Google DeepMind

    A double-blind pilot for model evaluation

    Pilot announced; method and future findings remain under watch.

    MONITORING 2026-08-27
    01

    PUBLISHER SAYS

    Google DeepMind says a cryptographic environment can keep confidential test prompts from the model owner and proprietary model weights from the evaluator.

    02

    RECORD CHECK

    The August 27 record names the partner organisations, the Gemini Flash Lite test subject, the confidential-computing setup, and a linked technical report describing the pilot.

    03

    MUSEUM READING

    The pilot addresses evaluation integrity and benchmark contamination. It does not by itself validate a model’s capability, make every benchmark independent, or establish universal adoption.

    Source published
    2026-08-27
    Checked by museum
    2026-09-11
    Review due
    2026-10-11
    Inspect the source record→
  5. 05RELEASE RECORD

    Meta AI

    Muse Image ships with a provenance signal

    Availability varies; video and detection features were not fully released in this record.

    MONITORING 2026-07-07
    01

    PUBLISHER SAYS

    Meta announces Muse Image, previews Muse Video, and describes Content Seal as an invisible signal attached to images produced on selected Meta surfaces.

    02

    RECORD CHECK

    The July 7 release record distinguishes what was available from what was still coming: Muse Image on named products and regions, Muse Video later, and a Content Seal detection tool in preview.

    03

    MUSEUM READING

    A provenance signal can support inspection; it is not proof of authorship, truth, or an unedited chain of custody. Availability and resistance to transformations remain claims to recheck.

    Source published
    2026-07-07
    Checked by museum
    2026-09-11
    Review due
    2026-10-11
    Inspect the source record→

Catalogue metadata and editorial commentary only. No source page, media, or benchmark chart is reproduced here.

Entries leave this room by removing an exhibition placement; their approved object version remains in the archive.

AI IN FACT · MUSEUM The living record of artificial intelligence.
Every sentence marked as a claim is linked to a source. Explanations, metaphors, and diagrams are editorial interpretation.