iAmBoosted

Don't take it too serious, it's X.

[iAmBoosted @letsgetboosted]

@sattyyouneed I think it would drive more people to local models, if anything.

[iAmBoosted @letsgetboosted]

@pcshipp Backend. No use in making it look pretty if it doesn't work.

[iAmBoosted @letsgetboosted]

@quadcode_ai Looks really good, gotta get it in a game engine now! Reminds me a lot of that "bodycam" game.

[iAmBoosted @letsgetboosted]

@posi_posi8 Have you tried HY 3D yet? Pretty sure Tencent has a promotion or something, 20 free generations a day on their HY 3D Global suite. No animation but generated some really good models for me by generating an image with GPT then tossing it into the "Image to 3D".

[iAmBoosted @letsgetboosted]

@solopribuilds Good, it needs to be. Lets consumers have a lot of power with their wallets, which is good for everyone. Brand loyalty is when the competition starts to die down.

[iAmBoosted @letsgetboosted]

@Surendar__05 You're able to have normal chats with Sol on pro reasoning, without burning your Codex usage. Instead of how it is with Claude, where every message sent burns into your Claude Code usage.

[iAmBoosted @letsgetboosted]

@yashmp2004 Gotta explain that short break you took in March 2020 🤣

[iAmBoosted @letsgetboosted]

@fhinkel Generative Psychosis Terminal

[iAmBoosted @letsgetboosted]

@pigeon__s @frygorate Oh I know, you don't have to creep me for that bud. I just enjoy engaging with the slop people put out to get engagement. You're welcome.

[iAmBoosted @letsgetboosted]

@pigeon__s @frygorate Oh my god, so... ANYONE CAN TRAIN ON IT TO BENCHMAX? Oh my god, but... Only Google is, clearly!

Anthropic, InclusionAI, Alibaba, Zai, Moonshot, and OpenAI would never DREAM of it!

Clearly only one of the largest tech companies in the US would do such a thing!

[iAmBoosted @letsgetboosted]

@AdamHoltererer This again? So, are you arguing native optimization is better than a generic harness? Because otherwise this holds no water. It's a generic wrapper over API. Put it in Antigravity, or put the GPT models in Opencode as well. The results may still be the same. But this is flawed.

[iAmBoosted @letsgetboosted]

@givros Engage baiting "one shots" is the new norm, like one-shot means anything to anyone with an actual codebase. Mostly just bots, multi-accounters, etc. Goes hand in hand with people posting cherry-picked benchmarks as if benchmarks mean anything anymore.

[iAmBoosted @letsgetboosted]

@weswinder Was just trying to narrow it down my guy, that's all. Went on incognito, same thing, except 3.5 Flash-Lite. Not sure why they would put it behind a paywall for even short amounts of access.

[iAmBoosted @letsgetboosted]

Crazy how everyone thinks every model is "benchmaxxed" but not their beloved model. If you think every company isn't training on benchmark data, you're "brainmaxxed".

[iAmBoosted @letsgetboosted]

@LexnLin Show the whole chart, where all the other models dive off a cliff as well. It's well known that 2.1 is much more saturated than 4.0 is. That's why they all fall off. But also look at the pricing. This is a Flash model, not a "Fable competitor". Not cheerleading Google, but damn.

[iAmBoosted @letsgetboosted]

@JohnKlerAI @ashen_one But genuinely they did not compare it to Fable. Others have, and I'll agree THERE, it doesn't hold up. They're wrong to think it will, it's a flash model. Same as Haiku/Sonnet, and in that regard, it kicks ass. Not trying to pick a fight or glaze Google, just honest conversation.

[iAmBoosted @letsgetboosted]

@PremiumGoblin Incredibly performative. A codebase isn't one-shotted. If it is, then it's probably either very minimal, or shit.

[iAmBoosted @letsgetboosted]

@itsvlady I jumped into the deep end too, before dumping all my subscriptions to the base tiers, and building up local. If I'm not making money off it, it's just a toy, not a tool.

[iAmBoosted @letsgetboosted]

@bridgemindai It really is that bad, benchmarks are completely saturated. Even more so in the open weights space. If you go look at HF cards, every benchmark is cherry picked to hell.

We need a better way of testing, because "use case" is becoming synonymous with "spend $ on it and find out."