Ephemeris · Issue 054 Thursday · 11 Jun 2026 · Zürich
Ephemeris · Issue 054 01 / 05 · HuggingFace · via @seeallochnaya
Open weights · Inference

Google opens a diffusion LLM.

DiffusionGemma writes text by denoising blocks of 256 tokens in parallel instead of one token at a time. It is a 25.2B-parameter mixture-of-experts with 3.8B active, weights under Apache-2.0, clearing 1,100 tokens a second on a single H100. Benchmarks trail autoregressive Gemma 4 by a few points — the throughput is the trade you are buying.

1,100tok/ssingle H100, per user
3.8B activeof 25.2B · 8 of 128 experts
256/ blockdenoised in parallel
Ephemeris · Issue 054 02 / 05 · OpenAI
Agentic coding · Science

Hand the agent a black hole.

An astrophysicist wired OpenAI's Codex into the loop for simulating black holes — the extreme-gravity regime you use to test whether general relativity still holds. The agent drafts and revises the simulation code; the scientist keeps the physics honest. It is the clearest case yet of agentic coding doing research work rather than demos.

The agent writes the simulation. The physicist checks reality. Both jobs got faster, and only one of them is new.

Ephemeris · Issue 054 03 / 05 · Cloudflare
Networking · Zero Trust

Public hostname, private origin.

Cloudflare's new Application Services for Private Origins lets a public hostname terminate on a private IP — RFC 1918 space reachable only over your IPsec, GRE or CNI tunnel — while WAF, bot management, rate limiting, caching and Workers still run at the edge. No public origin to harden, no extra connector daemon. Closed beta for Enterprise now, GA targeted for Q4 2026.

internetpublic hostname
▼
cloudflare edgeWAF · bot · cache · rate-limit · Workers
▼"use private network routing" on the DNS record
private tunnelipsec / gre / cni
▼
10.0.0.0/8 originnever exposed publicly
Ephemeris · Issue 054 04 / 05 · OpenAI
Security · Influence ops

Someone is staging the AI debate.

OpenAI says it disrupted a network it ties to PRC-aligned operators who used its models to draft and translate content steering U.S. conversation on chip policy, data-center buildout and commercial AI. The tradecraft is mundane — write posts, polish the English, seed both sides — which is exactly why it scales. Read it as a threat model for your own comment sections.

TARGET   U.S. infrastructure & chip policy
VECTOR   synthetic posts · translation · persona ops
MODEL    used for drafting + localization
STATUS   accounts removed · report published
Ephemeris · Issue 054 05 / 05 · jeffbullas
Creator economy · Strategy

Sell the signal, not the words.

When AI makes competent content free, the scarce asset is proof a specific human stands behind the work. Jeff Bullas frames it as a "human signal stack" and a path from point of view to income. The homework is concrete: audit last month's posts for what an AI could have written, then double down on the parts it could not.

  1. Identity — the obsession and origin story
  2. Story — lived experience that proves the claim
  3. Expertise — judgment earned from failures
  4. Evidence — original research, documented
  5. Interaction — real responsiveness, not reach
  6. Community — a tribe around your view
End of issue 054 Back to top ↑

That's today.

Issue 054 drew on OpenAI, Google's open weights via HuggingFace, Cloudflare and jeffbullas — surfaced through the morning feeds and the @seeallochnaya channel. The rubric: tools you could adopt this week, dev and agentic coding, security with real signal, and anything a senior engineer or founder could act on by Thursday lunch. A thin day for fresh picks, so five with signal rather than ten with filler. Back tomorrow.