ai memory: “Claude Agent SDK ships with TypeScript and Python support, same as Letta”

https://cameron.pfiffer.org/blog/agents-sdk/index.html#a_quick_comparison_between_letta_and_the_claude_agent_sdk

Comparing approaches to memory – and the lifespan of the agent.

“If you need an agent that remembers your last 50 conversations, coordinates with other agents, and improves itself over time—that’s what Letta is built for.”

Anthropic would worry about the self-improvement. Their local file approach is Apple-friendly.

“rise of authoritarianism in America is the defining challenge … legacy media institutions have structural problems that made them unequal to the moment”

https://www.thebulwark.com/p/the-washington-post-is-dying-the

“In an authoritarian context owning the Post is a high-risk proposition”

This would be a good time for an alternative to 10 different subscriptions. Or maybe we just follow journalists on Bluesky and all be free-riders?

“individuals with non-severe acute respiratory illness caused by SARS-CoV-2 experienced only modestly greater risk of PAS in comparison to those whose illness was caused by influenza”

https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1004777

PAS: post-acute sequelae. Includes Long COVID, CFS , etc.

COVID has taught us that common viral infections have more complications than we imagined. We are going to get a lot of vaccines.

“group of parents and clinicians who are calling for the autism spectrum diagnosis to effectively be split in two, saying it has become so broad that it is obscuring the experiences of the seriously disabled people it was first meant to describe”

https://archive.ph/20251009222018/https://www.nytimes.com/2025/10/01/health/autism-spectrum-neurodiversity-kennedy.html#selection-823.118-823.382

It should probably be split into 10 different slices, but that won’t happen until we understand a hell of a lot more. In the meanwhile – yeah, break it up. I think a NIMH director tried to push a division of the insanely broad syndromic classification years ago. The current definition is too broad to be useful for either research, prognosis, or treatment.

Reasoning model reasoning: “rule is correct on the demonstrations and the given test, but it is right for the wrong reason—“correct but unintended”.

https://aiguide.substack.com/p/do-ai-reasoning-models-abstract-and

“while o3, Claude, and Gemini approach or exceed human accuracy on these tasks with textual inputs, they are substantially more prone to use unintended “shortcuts” to solve the tasks than do humans”

They can reason but they are not very good at it.