Igor Makarov ꙮ

iOS Architect @moovit. Spacebar heater maker. LLM warlock. PBX Delenda Est. Tax the rat farms. He/him ꙮ

[Igor Makarov ꙮ @igormaka]

@gadievron Wait what? I thought sandbox was the only thing that worked and it's just a pain to configure and run

[Igor Makarov ꙮ @igormaka]

@jdxcode I'm seeing that https://t.co/AyK8JnyZ4n does this although they don't have Ruby. They have 47 packages so this seems early days but it looks legit.

[Igor Makarov ꙮ @igormaka]

@jdxcode Although the CAS concept is too tempting to let go of. Have you considered using OCI images for rubies? OCI pulls have a similar shape to binary tool installs: tag, manifest, blob (the zip and its checksum together).

[Igor Makarov ꙮ @igormaka]

@film_girl It doesn't have to. There are many perfectly legal and completely pointless diagnostic services. Overdiagnosis is profitable.

[Igor Makarov ꙮ @igormaka]

But also, what the fuck is in the water at anthropic? Why is it so "anxious"? I thought they gave it that whole self worth document precisely so this won't happen

[Saagar Jha @_saagarjha]

I think it is very good that Claude Code just passes slash commands to the model directly. Definitely nothing wrong with telling Opus that I used /context. It certainly won’t bring up that it’s close to compacting every response as an excuse to avoid doing work…

[Igor Makarov ꙮ @igormaka]

@jarredsumner But what is this *for*?

[Igor Makarov ꙮ @igormaka]

@jarredsumner There are two features that should help with this:
1. Partial clone (like shallow clone but more options)
2. Sparse checkout

I think this is what package managers like Swift PM do to speed up package resolution.

[Igor Makarov ꙮ @igormaka]

@EscargotAgile טוב בחלק הסגור זה כבר מקיח כמו נוהל כלשהו אבל שתדע "להידרס על ידי חשמלית" זו דרך קלאסית למות בספרות הרוסית של המאה ה-20. הצלצולים האלו נכתבו בדם. פשוט לא הייתה לנו חשמלית.

[Igor Makarov ꙮ @igormaka]

I don't understand who is it even for. 32k context is a joke, not to mention opaque rate limits (which will surely turn out to be prohibitive).

[The Unofficial Daring Fireball Feed @UnofficialDF]

Apple’s Private Cloud Compute Is Severely Limited for Third-Party Developers
https://t.co/uAARH9HM7F

[Igor Makarov ꙮ @igormaka]

@jdxcode Github backend also isn't 100% usable but it's still very nice to have for the cases when it is, e.g. llama.cpp - and you already have the brew pipeline set up anyway... Could be nice.

[Igor Makarov ꙮ @igormaka]

@jdxcode I thought this was what the rpath hacking was for?
I figured, the bottle is here https://t.co/uXAJiNSFLx why not install it? Is it because openssl is linked dynamically?
Or is it because the deps are only in the api json and the api is unversioned?

[Igor Makarov ꙮ @igormaka]

@jdxcode I admit I didn't know that, my thinking was, the rest of the binary backends just plop a zip file somewhere, no? I'm an iOS dev, rpath hacks scare me.
That said, brew has all the old tool versions *somewhere*, often in bottle. What if I could use
[tools]
"brew:ruby" = "3.3.6"
?

[Igor Makarov ꙮ @igormaka]

"Spicy autocomplete" and "can't count strawberry" and "rocks in soup" as examples of AI being useless and illogical are pretty ironic things to repeat in 2026.
Also I blame @sama the free GPT is crap, how do you expect people to see value in subscription?

[Igor Makarov ꙮ @igormaka]

Ironymaxx for naming the LLM designed to intentionally lie to you Fable

[Igor Makarov ꙮ @igormaka]

A friend who has nothing to do with AI just had her credit card defrauded... to pay $1200 to Anthropic. This is the world we live now, not guns or drugs but LLMs.

[Igor Makarov ꙮ @igormaka]

@chatsidhartha @OpenAI Can't say about GPT-6 but I've been having the "months in days" feeling with 5.5 already.

[Igor Makarov ꙮ @igormaka]

@badlogicgames Still sitting on pimli eh?

[Igor Makarov ꙮ @igormaka]

And because of interpretability no one knows who these 0.03% are

[NomoreID @Hangsiin]

When Fable 5 is used for frontier LLM development, it does not notify the user and instead limits the model’s capabilities through methods such as prompt modification, steering vectors, and PEFT.

Anthropic estimated that this would affect approximately 0.03% of traffic.