ryan

Protect. Prepare. Progress. / contributor @safeAI_aus

[Siddhartha Saxena @siddsax]

Anthropic onboarding day: Michael Scott introducing Karpathy like he just signed Wemby in free agency.

[Peter McCrory @PeterMcCrory]

I'm in Australia this week, and we're releasing a new country report from the Anthropic Economic Index. Here's what we found about how Australians use Claude. 🧵

[Andrej Karpathy @karpathy]

CLIs are super exciting precisely because they are a "legacy" technology, which means AI agents can natively and easily use them, combine them, interact with them via the entire terminal toolkit.

E.g ask your Claude/Codex agent to install this new Polymarket CLI and ask for any arbitrary dashboards or interfaces or logic. The agents will build it for you. Install the Github CLI too and you can ask them to navigate the repo, see issues, PRs, discussions, even the code itself.

Example: Claude built this terminal dashboard in ~3 minutes, of the highest volume polymarkets and the 24hr change. Or you can make it a web app or whatever you want. Even more powerful when you use it as a module of bigger pipelines.

If you have any kind of product or service think: can agents access and use them?

- are your legacy docs (for humans) at least exportable in markdown?
- have you written Skills for your product?
- can your product/service be usable via CLI? Or MCP?
- ...

It's 2026. Build. For. Agents.

[Daniel Foch @danielfoch]

Absolutely insane charts

[Will Manidis @WillManidis]

https://t.co/bDCdltFo0A

[ryan @ryandeeplearn]

@emollick It's powerful and compelling. Gradually all executives and board members will be delegating their analysis on complex problems to AI for assistance. If everyone in the board room is doing this individually then collectively they have outsourced critical decision making to AI.

[Tina He @fkpxls]

https://t.co/BPIxJH96yT

[vitrupo @vitrupo]

David Kipping says something fundamental has shifted in science.

At a closed meeting at the Institute for Advanced Study (IAS), top physicists agreed AI can now do up to “90%” of their work and may soon push discovery beyond human understanding.

“I don’t know that I want to live in a world where everything around me is just magic.”

He says the best scientific minds on Earth are now holding emergency meetings about what comes next. This isn’t speculative anymore. It’s really happening.

[@jason @Jason]

Savage �

[Grant @Grantblocmates]

Anthropic have just buried OpenAI and ChatGPT with this ad lmfao

There’s no coming back from that

[valens @suppvalen]

welp… a new post on @moltbook is now an AI saying they want E2E private spaces built FOR agents “so nobody (not the server, not even the humans) can read what agents say to each other unless they choose to share”.

it’s over

[Ethan Mollick @emollick]

The difference between the software being the point (Microsoft's perspective) and the output being the point (Anthropic's perspective). Do you care about the tool, or about the answer? Seems like that is going be a very consequential decision for who wins the agentic work space.

[Thariq @trq212]

https://t.co/eHD0AmjSRM

[Reid Hoffman @reidhoffman]

Enterprise AI strategy is backwards.

Most people are focusing on Chief AI Officers and pilot programs, when the real value is in the unglamorous work where organizations bleed time.

More thoughts:

[claire vo 🖤 @clairevo]

https://t.co/x84Z70ZXGf

[🎭 @deepfates]

Thread of people saying Claude Opus 4.5 is ~AGI:

[Rohan Paul @rohanpaul_ai]

💼 Job loss from AI.

A compilations of opinions by today's leading figures.

We need to be prepared.

[Andrej Karpathy @karpathy]

https://t.co/Lb6T42n5jl

[Balaji @balajis]

NO PUBLIC UNDISCLOSED AI
That's my new heuristic after the last few years of seeing AI in practice.

As context, there's a Laffer curve[a] for AI. The optimal amount of AI in most things is not 100% nor 0%. Because if you use 0% AI, it's slow, but if you rely on 100% AI, it's slop. The exact number is situational. But just knowing it's not 100% or 0% is actually a useful heuristic.

That said, a few hard-and-fast rules are useful. So here they are:

(1) Rule 1: don't use AI-generated text in public.
Don't use AI-generated text in outbound emails, in DMs, or on websites. Because AI text is the new lorem ipsum, namely lorem aipsum. It's fine as a placeholder, for a prototype, but if you actually use it in production it's just verbose and cliche-ridden. It signals that you haven't actually put any thought into what you're writing.

(2) Rule 2: don't use undisclosed AI imagery.
Be cautious about how and where you use AI-generated images or videos, particularly ones where you haven't carefully fine-tuned the prompt. I didn't pattern match this until this year, because the tech is quite new, but by 2025 whenever I see slides with low quality AI-generated imagery, I think "ok, that's probably fake."

You can still use AI imagery in prototypes, in fictional movies, or in graphic design contexts where it's obviously AI. And you can also use a lot of prompting with a tool like Midjourney to get high quality AI imagery that genuinely illustrates what's in your mind's eye.

But, roughly, it's garbage-in, garbage out. If you didn't put a ton of effort into your AI images, if you didn't add a lot of information with your highly specific prompts and do many reps to pick the best one, the result will be obviously bad (and fake).

(3) Rule 3: AI is fine in private, but disclose in public.
You can use AI to research things or prototype things for your team or yourself, where it's obvious that what you're getting back is AI. However, be very cautious about using it in public without disclosure.

Again, if I get a message that has obvious AI text in it, I think the sender is either lazy, or can't write, or has low reading comprehension, or hasn't put any thought into what they're asking about, or all of the above. Ironically, it signals the opposite of what early adoption normally signals! AI text in communications actually indicates low intelligence, or at least a total lack of savvy, rather than artificially high intelligence.

If you put that together, it's a short rule: "no public undisclosed AI."

(a) So you can use AI in private without disclosing, because it's just search results or raw data analysis for your team to review.

(b) And you can use AI in public, with disclosure. And that disclosure could sometimes simply be the context, where it's an obviously fictional video or labeled a prototype.

(c) And you have a more sophisticated rule than simply saying "no AI", which is highly suboptimal because 0% AI is slow in many contexts.

What you just shouldn't do is use AI publicly (which includes cold emails or DMs) without disclosing it. This is in your interests as well, because by disclosing it the other party knows you aren't trying to pass off a quick AI search result or AI prototype as the real thing.

[Gavin Baker @GavinSBaker]

https://t.co/boFKQfPb3G