What people are actually doing with AI

Things somebody built, ran or published, each linked to the original post. benchr has not reproduced any of it.

The sourced reference layer sits at all documented capabilities.

From one photo of an empty lot to a filmed tour of a house

Reported by @alexalbert__launch week of September 1, 2026Not tested by benchr

Alex Albert gave Claude Fable 5.1 a single photograph of an empty property lot. The model designed a house to fit the lot, rendered it, and produced a cinematic walkthrough video of the result. The work was done through code rather than in a design tool.

Why this is worth a look The output is a finished video, and the model got there by writing code rather than by using a modelling application.

Works at Anthropic, so this is a first-party demonstration.

In the reference layer Let the model write and run real code mid-answer · Generate video from a prompt or a still · One screenshot in, a working page out

57 working web demos, built by agent fleets in one day

Reported by elder-pliniuslaunch week of September 1, 2026Not tested by benchr

A public showcase collects 57 self-contained web demos across six collections: generative art, simulations, instruments and explorable worlds. The page states each piece was designed, built and QA'd end to end by fleets of Claude Fable 5 agents, over one day, and that every demo is a single HTML file with no dependencies.

Why this is worth a look The constraint is the interesting part: one dependency-free HTML file per demo, 57 of them, with the review step also handled by the agents.

How to try it

Open any demo in the showcase and read its source: each one is a single HTML file, so the whole artefact is the page you are looking at.

What it cost

The showcase says the run continued until the credits ran out. No figure is published.

In the reference layer Let the model write and run real code mid-answer · Hand off a task and walk away from the machine · Work through an entire repository, not a pasted file

Given a broken setup, the model installed its own tooling

Reported by The Neuron (Grant and Corey)September 2, 2026Not tested by benchr

The Neuron ran a live test of Claude Fable 5.1 on their own machine, on stream. In their words, they asked it to build games, install its own Blender tooling, debug its mistakes and see how much chaos it could handle. The recording is published in full.

Why this is worth a look Repairing its own working environment was not the task it was given; it was what the task required first.

In the reference layer Operate a desktop application · Give a model your own tools and data through MCP · Hand off a task and walk away from the machine

16 working viruses, designed by a genome language model

Reported by Samuel King and Brian Hie (Stanford / Arc Institute)August 6, 2026Not tested by benchr

A team at Stanford and the Arc Institute used the genome language models Evo 1 and Evo 2 to generate roughly 700,000 candidate bacteriophage genomes. They selected 302 for synthesis, successfully built 285, and introduced them into E. coli. Sixteen produced viable phages. Combined into one cocktail, those sixteen cleared two E. coli strains that had already become resistant to a naturally occurring phage. The training data deliberately excluded viruses able to infect humans, animals or plants.

Why this is worth a look The result is physical: genomes written by a model, synthesised in a lab, and shown to work on bacteria that had already defeated the natural equivalent.

No publishing API, so the model built its own browser extension

Reported by Daniel, Dig. Craft WorkshopJune 14, 2026Not tested by benchr

A newsletter account reports that Daniel, of Dig. Craft Workshop, had Claude Fable 5 automate the whole path from article draft to published Medium post. Medium exposes no API for publishing, so the model assembled a Chrome extension running on his own machine to complete the final step.

Why this is worth a look The missing interface did not stop the task; the model built the interface.

Recorded from a second-hand account. benchr did not find a post by Daniel himself, and there is no artefact to open, so this record is weaker than the others on this page.

In the reference layer Drive a real browser, not a scraper · Hand off a task and walk away from the machine · Teach the model a procedure once and reuse it

Claims we checked

Misleading

The AI-designed bacteriophages were genuinely new organisms, carrying genes nothing in nature had written.

What is actually true

The genomes were generated by a model and 16 of them produced working phages — that part holds. But an independent analysis by Oliver Crook at Oxford found the 16 viable genomes averaged about 97 percent identity to their ΦX174 template, placing them inside existing phage diversity rather than outside it. Crook's summary was that these were "brothers and sisters of the original virus". The genuinely novel result was recombination, not invention: one design took an unusually truncated protein from a distant phage species and made it work on the ΦX174 backbone, which conventional genetic engineering had failed to do.

Where it spread

  • Popular science coverage and aggregator summaries of the August 2026 Science paper
  • Social posts framing the result as the first artificial life written by AI

Checked

September 3, 2026 · Science · IEEE Spectrum