Fancy Curl Sites

Full-blog curl output was straightforward because Gemroff already emitted ANSI-marked plaintext, and the static site generator supported modular output formats. A wrapper template and a few generator changes exposed every post; Nginx selected that output for curl user agents, with an override header for ordinary HTTP. What catches my attention is how much depended on architecture choices made years earlier, not on curl itself. For other sites, a few hand-written pages may be the more realistic scope.

OpenCode 2.0

OpenCode 2.0 lets an agent compose MCP tool calls in a JavaScript script, so it can fetch results, map over them, and call another tool in one round trip. The article says this uses a minimal interpreter rather than a Node.js VM or embedded QuickJS, and keeps even many MCP servers from adding much context. What catches my attention is the tradeoff: less orchestration chatter, but the limits of that interpreter matter once workflows get complex.

Jev 1.14: For the Builders and Dreamers

Putting experimental capabilities behind explicit “danger zone” API headers is a useful release boundary: Jev exposes images, bounding boxes, and longer context without vouching for them in production. The team says multimodal use performs worse than pure text and longer contexts reduce reliability, so these flags shift experimentation to callers rather than hiding the tradeoff. I like the clarity, though teams still need their own evaluation gates before those paths touch automation.

AI was great until it was not

The useful distinction here is between asking a model to supply an answer and using it to make an explanation click. The student copied confusing passages from her physics notes into ChatGPT, which re-explained them and sometimes added a concrete example; after three weeks, she needed it less. I like this as a bounded use, and it echoes The Questions Are the Job: the test is whether the tool leaves someone better able to keep learning unaided.

The cost of getting a job today

The useful hiring constraint here is to make paid model use happen only after serious interest, then keep the challenge small or run it live with the company’s AI account. The author favors the live session because it shows how candidates reason and handle confident model errors without making them pay for access. I like that direction; a short-lived API key with a spending limit also seems worth considering. Either option makes the cost and risk explicit.

Bank of America Hired 1,000 Contractors to Patch Software Flaws Exposed By Anthropic’s Mythos

Bank of America brought in 1,000 contractors in the second quarter to patch software vulnerabilities exposed by advanced AI models, alongside 3,000 people already working in that area. The bank says those contractors remain, while remediation and related anti-money-laundering regulatory work added hundreds of millions of dollars in third-quarter costs. I find the employment side refreshing: AI-driven discovery can create substantial work for engineers. Faster vulnerability finding still depends on having people available to fix what it exposes.

Bee’s New AI Note-Taking Devices

The interesting design boundary is that Bee processes audio in real time and deletes it, while retaining encrypted transcripts whose keys are created and controlled by the device. Bee says others, including Amazon employees, can’t access those transcripts unless users deliberately share them; Alexa recording is an optional path users can enable. I find that boundary more consequential than the wearable form: useful summaries still depend on capturing ordinary conversations, so consent and sharing remain part of the product’s security model.

I love my Plaud note-taking devices, so Bee’s similar, lower-priced hardware catches my attention. The Frontier is a wearable and the Eclipse snaps to a phone; both start a recording with one button and produce conversation summaries. The $11.99 monthly Pro plan changes the price comparison, especially for new customers who want unlimited recent summaries or desktop access.

Amazon’s Bee launches two new personal AI devices, starting at $75 · Bee unveils two AI note-taking devices, its first new products since being acquired by Amazon

United will now let you check if your flight has Starlink Wi-Fi

Plane swaps make Wi-Fi availability a moving target, so United’s tracker can do more than a booking-time promise. Enter a flight number and date within three days of departure to see updates, including last-minute hardware changes. I flew to NYC a few weeks ago and was reminded how poor in-flight Wi-Fi can be without Starlink, so I’d use availability as a filter for future trips. Nearly half of daily departures have it, but the assigned aircraft can still change.

Figma announces new features to define how agents work on its platform

Figma’s agent-friendly future starts with making design rules legible. Specs gives teams a Markdown document that can sit alone or beside a canvas, linking Figma elements, diagrams, and code layers. Figma says teams can also centralize design and code libraries as context for agents. I like the direction, but the unresolved engineering question is whether that guidance stays current and specific enough to constrain nondeterministic output.

“You didn’t do it right.”

A framework that can’t be questioned when it fails gives teams no way to learn whether it fits. Cohen describes starting with one team, tracking a metric they personally valued for a quarter, then adding only one or two more; the metric also had to reflect something they could control. When I was at Google, a cloud customer asked for the secret behind its success with OKRs. I talked about goals and metrics, but the secret wasn’t in the OKRs. For engineering teams, a small trial can reveal whether a process supports the work or merely adds reporting.