Ephemeris · Daily
15·06·26
A frontier model pulled from every hand at once, agents handed longer leashes — a cloud runtime, autonomous CI, a browser they can redraw — and the spreadsheets that show who's really paying for it.
An export order suspends Fable 5 and Mythos 5.
On 12 June the US government invoked national-security authorities and ordered Anthropic to cut off Fable 5 and Mythos 5 for every user — citizens and foreign-national staff alike. The cited cause is a jailbreak that bypasses the models' safeguards; Anthropic calls the demonstrated vulnerability narrow and widely available from other models, and complied while disagreeing in public.
"Applying this standard would essentially halt all new model deployments for all frontier model providers."
Codex gets a cloud it can live in.
OpenAI is acquiring Ona to give Codex secure, persistent cloud environments — long-running sandboxes that hold state across a whole task instead of cold-starting on every call. For teams running fleets of coding agents, the workspace stops being a laptop you have to leave open.
- WAS
- agents run on your machine, state lost between calls
- NOW
- persistent cloud sandbox, secure, holds state across the task
- FOR
- long-running and parallel coding agents at enterprise scale
Point at the pixel; the agent edits the code.
Cursor's Design Mode lets you click an element, draw on the frozen page, or just say the change out loud. The agent picks up the element's identity, its source, and its place in the layout, then rewrites the code while hot reload shows the result. The browser becomes the canvas, not only the output.
- 1 Click an element — or draw over it
- 2 Describe the change in words or by voice
- 3 The agent edits the source, hot reload previews
Ask in Chinese, get a friendlier government.
A study surfaced in The Batch — and published in Nature — finds that state-controlled media in training data tilts a model's output, and the tilt tracks the language of the prompt. The practical read for anyone shipping on these models: a model's stance can move with the language you query it in.
The finding. Researchers from Oregon, Purdue, UC San Diego, NYU and Princeton tested identical political questions across languages and measured how favorably models described each country's government.
The tilt. Prompted in Chinese, models described China's government markedly more favorably than when the same questions were asked in English — an artifact of state media in the training corpus.
The takeaway. Press-freedom of a language's source media leaks into model behavior. If your product serves multiple languages, audit the answer, not just the translation.
$7,500 a head. Or $11.
Ramp's AI index puts hard numbers on the adoption gap. The top 1% of companies spend roughly $7,500 per employee per month on AI; the top 10% about $611; the median, $11. Where you work now decides how much AI leverage you actually touch.
- Top 1%
- $7,500per head · month
- Top 10%
- $611per head · month
- Median
- $11per head · month
680×between the top spender and the median
Your CI pipeline can now run an agent.
GitHub Agentic Workflows is in public preview: Actions can drive an autonomous AI agent as a workflow step. The same pipeline that builds and tests can dispatch an agent to triage an issue, reproduce a bug, or open a PR. The automation layer you already own gets a brain.
$ cat .github/workflows/triage.yml
# public preview · Actions + autonomous agent
on: issues
jobs:
triage:
uses: github/agentic-workflow
with:
task: "label, reproduce, propose a fix"That's today.
Six picks for the senior engineer or founder with ten minutes: a frontier model cut off, the infrastructure agents now run on, the research and economics underneath.