OpenAI stands up a deployment company

OpenAI announced a majority-owned deployment arm, an acquisition of Tomoro that brings about 150 forward-deployed engineers in-house, nineteen integration and consulting partners, and more than $4 billion of initial investment led by TPG. The framing on X: "majority-owned and controlled by OpenAI." It's a move toward the slow work inside customer organizations — data connections, controls,…

OpenAI launches the OpenAI Deployment Company

openai.com

The official announcement: a majority-owned deployment company, the Tomoro acquisition with roughly 150 forward-deployed engineers, nineteen partners, and over $4 billion in initial investment.

“help organizations build and deploy AI systems they can rely on every day”

OpenAI's public framing on X

x.com/OpenAI

The compact public version that spread through the developer timeline, emphasizing OpenAI's control of the new entity.

“majority-owned and controlled by OpenAI”

Mythos finds a curl vulnerability

daniel.haxx.se

Daniel Stenberg reports that a Mythos scan of about 178 thousand lines of curl source claimed five confirmed vulnerabilities; after human review, the team ended up with one low-severity CVE, three false positives, and one ordinary bug. He still rates AI analyzers above older tooling.

“The AI reviews are used in addition to the human reviews.”

Mythos 'discovered' a CVE already in its training data

rival.security

Rival Security argues the FreeBSD RPCSEC_GSS bug Mythos surfaced (CVE-2026-4747) closely resembles a 2007 MIT Kerberos issue, CVE-2007-3999. Their point: cheap rediscovery and exploitation matter even when novelty doesn't.

“combinatorial creativity, with AI making a discovery already within its training data”

You need AI that reduces maintenance costs

jamesshore.com

James Shore: every month of code creates future maintenance obligations. A coding agent that doubles output without halving maintenance cost eventually eats its own productivity gain.

“your AI coding agent, the one you use to write code, needs to reduce your maintenance costs”

I'm going back to writing code by hand

blog.k10s.dev

Seven months and 234 commits into an AI-built Kubernetes TUI, the author archives it. The core model file hit 1,690 lines, one state object mixed UI, cluster state, and logs, and async mutations made things unsafe. Verdict: "AI writes features, not architecture."

“AI writes features, not architecture.”

I keep tripping over true, false, true

allthingssmitty.com

A small API-design piece on positional boolean arguments. If a call site needs a comment to be read, the function shape is making the reader decode intent — pick options objects, or separate functions when the flag really means a different action.

“I'm not reading code anymore, I'm decoding it.”

Two roads to durable agents: replay vs. snapshot

YouTube — Eric Allam, Trigger.dev

Allam splits agent state into an append-only context log and an execution-state machine. Replay journals work for short workflows; long agent sessions push Trigger.dev toward Firecracker microVM snapshots — about 14 MB compressed, restored in hundreds of milliseconds.

“an agent isn't like a transaction, it's like a session”

Hierarchical memory: context management in agents

YouTube — Sally-Ann Delucia, Arize

Arize's Alex agent kept hitting its context window analyzing trace data. The working fix: keep head and tail slices, store the middle behind retrieval, move heavy search into sub-agents, and evaluate the eleventh turn after loading ten.

“context decides what the model sees, memory decides what survives”

You can't just one-shot it

YouTube — Mehedi Hassan, Granola

Granola's meeting-notes chat needed tracing, internal tooling, and PR preview links — not a better single prompt. The Electron render process was turned into a web shell so each pull request could get a preview URL and screenshot proof.

“the answer isn't to one-shot better”

MLX Genmedia

YouTube — Prince Canuma, Arcee

A tour of on-device vision, speech, and agent pipelines on Apple Silicon: 1.5 million MLX downloads, 4,000+ ported models, real-time object detection, local multimodal, and Turbo Quant cutting key-value cache memory roughly fourfold.

“you can build agents that can hear, see, and sound”