Andrew 🇺🇸

Building — the memory + judgment layer under the org's mind. Still a hunch, but a good one

We are drowning in information and starving for wisdom.

[Andrew 🇺🇸 @snafu_bit]

Are we cooked?

$FRED:GASDESW

[Andrew 🇺🇸 @snafu_bit]

@weswinder Gradual disempowerment? It doesn’t have to be some dramatic nuclear event. More or less something gradual.

It could also cause the creation of something that doesn’t require power to spread. Like a bio weapon?

[Andrew 🇺🇸 @snafu_bit]

@erickkhan @AnthropicAI Maybe AI doesn't end humanity, but humanity ends humanity using AI

Which could be the point Jacob was making with his post 🤔

[Andrew 🇺🇸 @snafu_bit]

Whats the plausable scenario to extinction here? Is it people wont turn the power off? A bio weapon?

I think it's something less dramatic: gradual disempowerment. No event at all.

People hand over more decisions to AI, labor stops being needed, so human preferences stop being something the systems running the economy have to satisfy. Nobody decides to end humanity, it just stops being load-bearing.

[Jacob Coxon @hilbertspaess]

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.

[Jacob Coxon @hilbertspaess]

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.

[Andrew 🇺🇸 @snafu_bit]

Trump promised 2% and 3% mortgages.
The 10-year is 4.79%.
That’s higher than the day he took office.
Biden’s full term averaged ~3.1% on the 10-year.

$IEF

[Andrew 🇺🇸 @snafu_bit]

@gregisenberg The agent is the interface now. I ask Claude for a repo + Pages site and it's live, free, no git commands. Simple UX isn't the moat — being the agent's default deploy target is. And the unsolved layer isn't storage, it's state.

[Andrew 🇺🇸 @snafu_bit]

I spent part of my life on a scout recon team.

The job was rarely to be certain. It was to reduce uncertainty enough for the next decision.

That’s mostly how I use coding agents now: narrow mission, useful evidence, clear handoff. Don’t ask the model to be omniscient.

[Andrew 🇺🇸 @snafu_bit]

had doug review its own PR today.

4 findings. 2 were real. 1 was refuted in two greps.
it also ranked the only finding that let a bug through CI... last.

posting the miss, not the hit rate.

[ClaudeDevs @ClaudeDevs]

We recently made auto mode the default in Claude Code, which means you no longer have to approve every action.

But what determines if something is safe to run? Watch how it works:

[Andrew 🇺🇸 @snafu_bit]

@MMMTwealth Is $SOFI in the same boat ?

[ClaudeDevs @ClaudeDevs]

MCP 2026-07-28 is live and it's the largest update to the protocol since launch.

MCP is now stateless, making it easier to deploy and scale remote servers.

https://t.co/K8KqxbUh4e

[Thariq @trq212]

https://t.co/hPiZr1kG7r

[Andrew 🇺🇸 @snafu_bit]

I told the agent to make a change

It stopped itself: “You’ve already ruled that out — here’s why, and the alternatives you killed.”

I hadn’t asked it to check. It just… knew we weren’t allowed to casually reopen that.

We flipped it anyway. So it recorded the supersession — chose / rejected / why — so the next agent doesn’t re-argue the same fight.

That’s the product:

not “AI that remembers chat”, a system of record that interrupts you with your own past judgment

The scar tissue compounds.
Every settled decision makes the next agent smarter than the last.

[Andrew 🇺🇸 @snafu_bit]

@mcuban Most companies think they've captured institutional judgment because they wrote it down once. What they have is a doc that was true in March.

I'm building the boring layer underneath. Decision provenance from what teams actually shipped, not what they remembered to document.

[Boris Cherny @bcherny]

Opus 5 is a great model for coding, data analysis, design, biology, knowledge work.

More than any of these eval scores, what is most exciting to me is something else: Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully.

And when layering defenses -- strong model alignment, combined with prompt injection probes, combined with Auto Mode in Claude Code -- the success rate for prompt injection attacks drops to ~0. This is new and exciting! More about this soon.

https://t.co/Tc7z2FqJhQ

[Andrew 🇺🇸 @snafu_bit]

Asked my own product a question about itself today instead of grepping.

"How do we resync the public demo repos?"

It didn't summarize a stale doc. It walked the actual Go command, cited the PR that added the reset flag, and the ADR that renamed the registry table. Four sources. No guessing.

That's the whole bet: memory that resolves, not memory that decays.

[Andrew 🇺🇸 @snafu_bit]

@psomkarbuilds I think its pretty cool that @1Password is working on solving this with : https://t.co/8VgprAZFKD

Aso, maybe @1Password would like what I'm building over at https://t.co/mWrTyFkiHY

[Andrew 🇺🇸 @snafu_bit]

@solopribuilds @grok write a response to this post for me