Garlef's comments

On: Harness engineering: Leveraging Codex in an agent-first world

[Garlef]

Sure; But that's not the point that is argued about here.

(To state it in AI lingo:)

It's not about the best measure for "amount of code".

It's about wether "amount of code" is a good metric to begin with.

On: Field of clones: How horse replicas came to dominate polo

[Garlef]

> At the slightest touch of the reins, he felt a familiarity that shook him...

Ah... Some good, old, pre-AI journalism slop.

Oh the countless times a universities press release has been turned into four pages describing the smell of coffee some scientist inhales on their way through campus...

On: Universal Memory Protocol – a shared format for agent memory

[Garlef]

Not sure if this is the right abstraction: The recall seems to need a search term.

But would it not be more sensible to assume that the full conversation (+ system parts) CAN inform the recall and some neural network picks the right memory bits?

So my fear would be that something like this, if adapted, drags the development into a local optimum that is hard/impossible to get out of.

On: Universal Memory Protocol – a shared format for agent memory

[Garlef]

> everyone has the same LLM discovery patterns

there's so much discussion about AI and "everyone" might just be reading the same things.

this is a good indication not to treat this particular phrasing of a discovery path as canonical

On: When AI Builds Itself: Our progress toward recursive self-improvement

[Garlef]

Make it work, make it nice, make it fast.

[jachee → Garlef]

Good, Fast, Cheap.

Pick any two.

On: The Public Should Own Half of the Big A.I. Companies

[Garlef]

Maybe the public should own half of everything.

On: Is Python Becoming Pinyin?

[Garlef]

> what language would be best for our AIs to be programming in

I think "batteries included" is not a good thing to have in the future.

We'll want to be very explicit about what AI generated code can and can not do.

And so some form of effects based scripting language seems like a plausible choice: A language where by default "all batteries are removed".

[drob518 → Garlef]

Good point. I think you’re right. We need more rigor for the AI, which would actually help code generation, I suspect. Included batteries are for humans and AIs don’t need convenience.

On: Legal action forces Facebook whistleblower to sit in silence at Hay festival

[Garlef]

At least he's winning in Catan.

[lionkor → Garlef]

More like Monopoly, Catan has too many rules limiting expansion

[msh → lionkor]

You need to read the book for the reference. Apparently mark likes to play catan and everyone else looses on purpose…

On: Creatine raises brain energy levels and slows cognitive decline: study

[Garlef]

> So don’t drink more than 2 cups per day and you’ll be fine.

Easier said than done...

On: Domain expertise has always been the real moat

[Garlef]

From my experience what domain experts are often missing - and at least currently this is also an area where LLMs fail - is the ability to model data and interfaces in a sustainable way and factor in team and domain boundaries.

This is a failure mode that senior engineers have seen a few times throughout their career: They know how certain choices will play out over time... and the kind of problems and roadblocks these choices might cause.