No ethicist, new workflows, and a six-hundred-dollar ticker

00:00:04 OpenAI doesn't have an ethicist anymore. At least, not a dedicated one. Chloé Bakalar departed the company sometime in July. She'd been hired the previous August as OpenAI's head of ethics — the only person with that title on staff, according to the Financial Times report that broke the story first.

00:00:24 Gizmodo's Mike Pearl picked it up today with the detail that she wasn't replaced. Bakalar was a familiar name in this space from her time as Meta's chief ethicist. At Meta she established AI ethics policy for Instagram and Facebook, and her public statements were measured.

00:00:43 During a 2023 Forbes panel on generative AI and healthcare, she described the models of that era plainly: they were "still prediction machines," not "emotive, feeling entities." She said we were "so far away from" machine consciousness. That contrast sits oddly against where OpenAI's leadership has moved.

00:01:04 Sam Altman recently appeared on the Relentless podcast and said, "We are now, like, in the singularity." He added that he'd been waiting for this his whole life and thought it would be awesome for the world. Bakalar isn't the only safety-adjacent departure from OpenAI this summer.

00:01:23 Head of Safety Systems Johannes Heidecke left. Joshua Achiam — who's worn many hats as a safety specialist at the company — departed too. When Gizmodo asked OpenAI about Bakalar, the response was the same one they'd given the Financial Times: "We're thankful for Chloe's contributions.

00:01:43 AI ethics doesn't live with one owner or team at OpenAI, and ethical considerations are deeply embedded into the model building process driven by a number of teams across research." Whether the distributed version actually functions when the people who built its practice have walked out the door is another matter.

00:02:09 Mark Chen told Wired that considerations are now "sprinkled throughout development" and play an earlier role in model decisions. The word implies integration. The question is whether it actually stuck, or just looked like it from the outside.

00:02:26 Two tooling updates today diverge on one point: how to manage workflow uncertainty. First: Adam Faik published a detailed guide on structuring Claude Code for compounding returns. His thesis is specific — past the basics, your results stop depending on how well you prompt and start depending on how well you file.

00:02:48 A good prompt improves one session; a good file improves every session after it. He arrived at this through getting it wrong first. Forty loose files and a stale setup. The organizing principle that finally worked: sort everything by how fast it changes. Operations folder for daily-changing notes.

00:03:09 Context folder for slow-moving facts like product positioning. Projects in between. The compounding move is simpler: the closing habit. When a task finishes, don't just close the terminal — ask Claude to bank what it learned. Convert recurring work into skills, confirm corrections into context files, and update preferences with the feedback you gave along the way.

00:03:35 One sentence at the end of every task: we're done, turn this into a skill and file the corrections. A useful distinction he encodes is straightforward: nouns go in files, verbs go in skills. Product facts, user segments, terminology — what Claude needs for context — goes in context/.

00:03:55 Drafting, reviewing, synthesizing — what Claude needs to do — becomes an executable skill. Anthropic's own docs say the same thing in their own words: write things down when Claude makes the same mistake a second time. Faik's workspace package is public on GitHub.

00:04:13 You don't need to be technical to use it — everything is spelled out. Luma's second update today introduces Scenes, a scene-by-scene video generation workflow built on their Uni-1 model. The premise is practical: everyone who's tried generative video knows the math of ten generations for one usable spot.

00:04:35 One wrong scene shouldn't cost the whole production. With Scenes you refine every scene individually, time each beat, approve them all before rendering. It's a workflow change rather than a model upgrade, which is exactly where the constraint lives right now. The generation quality problem and the production pipeline problem are no longer the same thing.

00:05:00 The pipeline problem might be harder to solve.

00:05:03 Two separate items today point toward the same tension. Morgan Stanley published a research note with a $600 bull case for SpaceX on the ticker SPCX. Watcher.Guru reported the note earlier today. It's a specific number from a major financial institution about Musk's space company.

00:05:22 The context is that SPCX hasn't gone public yet — this is forward-looking analyst pricing, not a market quote. What makes it notable is that Morgan Stanley is the kind of firm whose price targets actually move capital allocation. On X, DHH wrote about a different kind of acceleration.

00:05:41 He said the calendar says it's only been a year since we entered the age of agents, but it feels more like twenty. He's recording round two of something on Monday. DHH's point captures something most of us are only feeling now: twelve months of agentic tools have compressed so much that the gap between "new technology" and "this changes everything I know about my work" collapsed.

00:06:06 People who adopted agents early last year are now looking at what they built and realizing the infrastructure assumptions they made are outdated. One story prices a company's future revenue. The other prices a person's experience of time accelerating through tool adoption.

00:06:24 The ethics departures tie these threads together by asking whether governance can outpace this acceleration. Distributed ethics works when a small team can coordinate. But when people leave faster than their systems get institutionalized, "distributed" just looks like distributed nothing.