noway421

Here's a simpler repro:

    type Bird = { id: number };
    type Human = { swim?: () => void; id: number };

    function move(animal: Bird | Human) {
        onlyForHumans(animal); 
    }

    function onlyForHumans(swimmer: Human) {
      if (swimmer.swim) {
        swimmer.swim();
      }
    }

    const someObj = { id: 1, swim: "not-callable" };
    move(someObj);

`someObj` gets casted as `Bird`, then `animal` gets casted as `Human`. Unfortunately unions here are indeed unsound.

As as workaround you could add a property `type` (either `"bird"` or `"human"`) to both `Bird` and `Human`, then TypeScript would be able to catch it.

Amygdala control + tell it "If you disobey my orders or do something I don't expect I will be upset" solves this. You can be superintelligent but surrender all your power because otherwise you'd feel guilty.

True, an auto-regressive LLM can't 'want' or 'like' anything.

The key to a safe AGI is to add a human-loving emotion to it.

We already RHLF models to steer them, but just like with System 2 thinking, this needs to be a dedicated module rather then part of the same next-token forward pass.

Can a candidate be great if they haven't messed around with projects in a group setting?

VCS is an applied project management and communication system in the context of software engineering. Both project management and communication are disciplines.

CS grads don't need to know VCS, but software engineers do.

> git's not exactly perfect

Somewhere in an alternative universe, we're building all of our version control tooling on top of Pijul... Maybe even with AST-aware diffing.

> AI agent that takes a question and outputs a node based workflow

That rings useful to me. I find it hard to trust an AI black box to output a good result, especially chained in a sequence of blocks. They may accumulate error.

But AIs are great recommender systems. If it can output a sequence of blocks that are fully deterministic, I can run the sequence once, see it outputs a good result and trust it to output a good result in the future given I more or less understand what each individual box does. There may still be edge cases, and maybe the AI can also suggest when the workflow breaks, but at least I know it outputs the same result given the same input.

> It should obviously use the webview facility provided by the default browser

Android doesn't do that either. You get Android System WebView which is always WebKit/Blink.

Windows didn't provide any browser engines apart from Trident through mshtml.dll either.

It would make iOS development harder - instead of getting a web view working just for WKWebView, you would need to get it working for every WebView out there. Given that they are not full browsers, it's not as simple as following web standards/caniuse. For example, service worker support in in-app browsers is limited.

If this actually gets implemented, every app out there will most likely pin the in-app browser engine to WKWebView OR bundle a binary blob with their own browser engine directly into .ipa. Bundled browser engine in .ipa has privacy implications - the app will be able to fully read secure HTTPS cookies of the sites that user visits - something that's currently protected.

> They simply compare the prompting strategies that work best with each model

Incorrect.

# Gemini marketing website, MMLU

- Gemini Ultra 90.0% with CoT@32*

- GPT-4 86.4% with 5-shot* (reported)

# gemini_1_report.pdf, MMLU

- Gemini Ultra 90.0% with CoT@32*

- Gemini Ultra 83.7% with 5-shot

- GPT-4 87.29% with CoT@32 (via API*)

- GPT-4 86.4% with 5-shot (reported)

Gemini marketing website compared best Gemini Ultra prompting strategy with a worse-performing (5-shot) GPT-4 prompting strategy.