Manfred M. Nerurkar

CEO and Founder of @abstract3d

[Manfred M. Nerurkar @mannyn83]

"yet"

[Ryan Schmidt @rms80]

here is what "AAA" looked like in ... 2013. Fable is cool and all but it's not producing visuals or gameplay anywhere close to this yet.

https://t.co/xNK1Ykh7wH

[Manfred M. Nerurkar @mannyn83]

When done right, a 4-bit quant introduces little perplexity, esp. for large models ~2%

Using 4-bit for activations and partials also has a very small impact, still ~2%

There's a thin line between model collapse and those results. Needs ultra-accurate engineering.

[Uzair Afridi @uzair__afridi]

Can I ask a dumb question?

Is a 4-bit quantized 70B model actually smarter than a full-precision 8B model?

[Manfred M. Nerurkar @mannyn83]

@superalesha That's nothing new - also it's better than silently introducing errors as Fable 5's initial release did:

https://t.co/VXKeaYuu0A

[Manfred M. Nerurkar @mannyn83]

@entirelyuseles @fleetingbytes My man, independent of your skill level -- it's hard to believe that you can write "far less intelligible code"

Please show me that

[Manfred M. Nerurkar @mannyn83]

@mel_lasuda Not just a formatter -- you also need some reasonable coding standards like not dereferencing a potentially null pointer, and not using expensive lookups like findChild<T> there is quite a lot wrong here outside of just "style".

[Manfred M. Nerurkar @mannyn83]

@ayden1024 If I wouldn't know, I'd say codex -- but according to the repo's history: it was schmopus 😂

[Manfred M. Nerurkar @mannyn83]

This is the most unhinged code I've seen in many years. "Open-source"? This might as well be binary.
tbh looks like disassembly using Ghidra or so.

We're living in truly remarkable times.

The ironic part is that this slop will ruin datasets for future AI models. Careful curation of datasets will be the moat of the future.

[Mattias Lind 🔶 @mattiaslind93]

https://t.co/2BLEHWDBuu🔶Introducing mPainter(Alpha)!
An open source texturing application with blender at its core, plus a bunch of custom goodies like a new brush-engine etc.
Work with layers, and import the blender materials you already have in a familiar texturing interface😊#b3d #blender #texture #app

[Manfred M. Nerurkar @mannyn83]

@MikeWithAHotDog Yes to the ability to hand write elite code, that helps to parse the slop and correct it. That's real moat -- for now. I do expect the LLMs to continue reducing that moat with each new model generation though.

[Manfred M. Nerurkar @mannyn83]

@ErikHvatum Yeah and the ice cold pointer derefs etc.

[Manfred M. Nerurkar @mannyn83]

@takeiteqsy Just give it a style guide and some basic rules, we use both Astra and Opus -- style matches house for both.

[Manfred M. Nerurkar @mannyn83]

@DomWitczak I agree, we enforce house rules and our review process did not change.

Agentic engineering done right can be a massive productivity boost. I'm not hating on LLM use in general, just the way some use it and then put it out there under the FOSS banner.

[Manfred M. Nerurkar @mannyn83]

@TalkingCat2569 You are clearly delusional

[Manfred M. Nerurkar @mannyn83]

@ErikHvatum That's a bad take. Look at the qt source, it doesn't look like they at all, neither do our giga qt codebases for InstaMAT or InstaLOD. It's not just the style...

[Manfred M. Nerurkar @mannyn83]

@nefertum01 I agree, it's evident if you look at technical notes. Many labs still just try to hoard.
I believe a smaller, perfectly curated dataset can outperform a significantly larger set.

It's relevant at Abstract. We're working on figuring out whether that thesis actually holds.

[Manfred M. Nerurkar @mannyn83]

@ryunuck Ha, yeah at least that's how it feels. Luckily model alignment to rules and principles is really good now.
End of 2025, early 2026 was insane - we're at a point now where it's a real production boost when used right.

[Manfred M. Nerurkar @mannyn83]

@andkalysh Yeah, it's insane. Vibe coding doesn't have to equate to slop. It's odd that some people insist on it.

Just give it a hint to an establish style-guide and you're good. LLMs are trained on that.

[Manfred M. Nerurkar @mannyn83]

@1531fadfda True, this slop may as well be closed source. It's seconds away from being a disassembly dump.

[Manfred M. Nerurkar @mannyn83]

@jiwenji1 doesn't the LLM have to read the code? If there are no header comments the agent has to burn massive context to piece it together

what about the missing error handling, buffer range handling, useless copies, asserts?

[Manfred M. Nerurkar @mannyn83]

@SocietyEnjoyer rare

[Manfred M. Nerurkar @mannyn83]

@fleetingbytes No they should be able to understand both sides as long as it compiles.
There is a loss here though, no asserts, no consistency, no error habdling, no header docs that the LLM can read to understand. The LLM is forced to read the cpp and burn context and tokens.

[Manfred M. Nerurkar @mannyn83]

@iRainbowsaur Bro you've lost it