Phunky

crypto curious, dipping into NFTs, DeFi, and other degen pastimes | profile pic @tubbycatsnft

[Phunky @phunkyflips]

Tubby Cats investment thesis...
- they are cats
- crypto native visible lead artist and team
- well matched traits make each one feel unique
- 1 address = 1 mint on allow list, makes it much harder to flip and crush floors
- meme them to the moon

[Phunky @phunkyflips]

@fofrAI It feels very unbounded and uninhibited in a way I haven’t felt with models in awhile. Maybe since the first gen reasoners

[Phunky @phunkyflips]

@deredleritt3r I think they already reached intern level tbh, it’s a non-goal now

[Phunky @phunkyflips]

@danshipper I really can’t stand the full LLM written articles, there’s no human voice in the article at all. Becomes really grating after the first 5-10 paragraphs

[Phunky @phunkyflips]

@morqon What’s the source on this one?

[Phunky @phunkyflips]

@scaling01 4.8 is better at coding and knowledge work but hates high difficulty tasks? Does this mean the model itself finds these tasks easier? Or that it’s a reluctant participant?

Very weird result that feels contradictory

[Phunky @phunkyflips]

@adonis_singh

[Phunky @phunkyflips]

I don’t think this holds at the frontier. Hard to test with no ability to fine tune those models though

[Owain Evans @OwainEvans_UK]

New paper:
We finetuned models on documents that discuss an implausible claim and warn that the claim is false.
Models ended up believing the claim! Examples:
1. Ed Sheeran won the Olympic 100m
2. Queen Elizabeth II wrote a Python graduate textbook

[Phunky @phunkyflips]

@emollick Oh wow that’s a big hire

[Phunky @phunkyflips]

You’re so close sweetie, you can do it

[Marc Andreessen 🇺🇸 @pmarca]

What could it be, what could the cause be. Unfortunately, we'll never know. It'll be a mystery, like Bigfoot, or the Loch Ness Monster.

[Phunky @phunkyflips]

Reminder that Ed Zitron is a content factory, not a thought leader

https://t.co/N93Ct9oYK6

[Ed Zitron @edzitron]

When I found out the 49ers were using ChatGPT to draft people I asked it who the 49ers should draft in 2026 without looking using the web tool (so it wouldn’t see who they picked) and one of the choices was Emeka Egbuka. It then hallucinated a guy called Kevin Stribling

[Phunky @phunkyflips]

@deanwball Easily from GPT-5.2 through 5.4, not sure if 5.5 carries the same tic forward. Just gotta work “goblin” in there and we’ve got the whole stew

[Phunky @phunkyflips]

Complete speculation… but I think Mythos cracks 2000 on GDPval, it was the one measure I was sad to not see in the system card

[Lisan al Gaib @scaling01]

Claude 4.7 Opus has an Elo of 1753 on GDPVal-AA

[Phunky @phunkyflips]

@tenobrus I’ve noticed this a lot in the past couple of weeks. Opus delegates and the subagent fucks it up big time, then Opus has to go back and fix everything.

Made me trust every output less and add extra review layers

[Phunky @phunkyflips]

@AndrewCurran_ I miss near

[Phunky @phunkyflips]

@AndrewCurran_ Btw I don’t believe this is “leaked” but a guess/estimate based on publicly available information.

I saw it yesterday from another account and I can’t find the original post. The sheet also has a lot of Claude excel tells

[Phunky @phunkyflips]

They definitely had to wait an extra day to announce this lmao, because wtf

[Tibor Blaho @btibor91]

What https://t.co/pQSqOBHSzz

[Phunky @phunkyflips]

@tenobrus The ketamine, definitely the ketamine

[Phunky @phunkyflips]

@emollick Perhaps unsurprising, but nice to see validation of this idea

[Colossus @colossusmag]

Hassabis secretly built a hedge fund inside DeepMind trying to beat Jim Simons. Google shut it down.

[Phunky @phunkyflips]

He blocked me for this one lmao

[Phunky @phunkyflips]

Fundamentally unserious people

[Phunky @phunkyflips]

@_xjdr Oh shit this talk is finally live, blessed day