>open weights gets near frontier>update the benchmarks so they don't anymore
what's SI
>>110184020Super Intelligence.
>>110184007This happened because the datasets were too public. Now 60% of the benchmarks use private data sets which means the labs can’t just benchmaxx. The resulting leaderboard reflects how most people feel about the models. For example, Astra is clearly above Muse lol.
>>110184020
Haiku 5.5>score: 36>cost per task: $0.41Kimi 3.1>score: 46>cost per task: $3.68lol. No wonder the Chinese stock market is crashing. It’s over for chinkcels. Moving forward only Anthropic will distill their own models
>>110184002>quick, take out the benchmarks that chink AI is beating us in>New AI intelligence index just dropped
>>110184112
>>110184002>Mimo getting btfoThat model sitting so high in the rankings never sat right with me. Ahead of haiku 5.5 my ass lmao.I agree with the new benchmarks.
>>110184002>SIFuck off, retard.
>>110184002>>110184031migger
>>110184020sztuczna inteligencja
>>110184002Kimi is still way to high
>>110184020Shiva intelligence of course
But do they all use a good harness? I find that all open weight models have worse tooling than claude/chatgpt, simply because they integrate everything into one package it shit works.Sure there was a lot of benchmaxxing going on, but in my opinion the intelligence index should test for raw LLM output, including arbitrary tool calling, not how well it can do an actual task with the tools it comes with.
>>110184002Lol at Grok
>>110184002Wait, is opus 5.5 better than Fable 5.1?
>>110184763ok sirs here is the AI brown pill
>>110184020i thought the term is another dementia don shitpost, but its LITERALLY and UNIRONICALLY an actual amerimutt executive order (no. 14434)fucking kek
>>110184002they're going mask off now that the anthropic is going public. they're spending all their political capital because the IPO is imminent
>>110185164Spending political capital for whom? Previous chart implies that Anthropic is escaping from OpenAI with even Sonnet 5.5 beating Astra 6.The new one is more like a tiktok where Astra 6.1 is going to beat Opus 5.5
>>110184715japa koniu. nie po to lurkuję szambo żeby czytać wypociny jakiegoś polaczka
>>110184002>SIalso>WE ARE NOW MAKING UP THAT THE LAST MADE UP SUITE OF MADE UP TESTS WAS ACTUALLY MADE UP! LUCKILY THIS NEW MADE UP RANGE OF MADE UP TESTS TO MAKE UP OUR MADE UP MAKE UPPERS MAKE UP THAT OUR MAKE UPPERS ARE MORE MADE UP AT MAKING UP! posting this here should just be an automatic and permanent ban on the same level as cp, kill yourself instantly
>>110185209stupid
>>110185242You sound mad.
>>110185257you should be too, your life is going to be partially melted by this just like mine
>>110184020a Slovenian domain.
>>110185242>>110185265>chang madmaybe start actually innovating instead of stealing and copying.
I hate how often they update their stupid index but this ranking seems much more in line with my experience
>>110185956Yep, no way Kimi was top 3.
>>110184002>>110185098Reminder that the idiot that runs this website (RapeApe, Hiro hasn't posted on this site in years) unironically uses SI as a term
>>110184002Kimi is quantized out of the box to FP4 weights and FP8 activations. The fact it mogs the hell out of Haiku and is still capable of trading blows with Opus, despite ostensibly being an entire generation behind, is fucking astounding.I don't know what the Moonshot people are doing, but it's tantamount to black magic.
>>110186205Let me know when Kimi can make video game mods like Opus 5.5, chang
>>110184071lol they added le science benchmark which all frontier models benchmaxx on.
>>110186230Works on my machine.
>>110184020sloppy intelligence
>>110184002don't care, still using deepseek
>>110184002like clockworkI wonder how much ant and oai pay to keep them on top
Kimi K3.1 is more capable than Gemini 4 ime, PLUS doesn't straight up lie about its results.
>>110184071>Now 60% of the benchmarks use private data sets which means the labs can’t just benchmaxx.Every major lab has LLM's searching prompts for benchmarks ... it's private exactly once.Only open models can be benchmarked with secret problem sets.
>>110184020It's laughable how this shorthand is somehow gaining traction. Proper newspeak.
>>110184020AI but trooned out and angry at you for deadnaming it.
>>110186205probably overfit their models for meme benchmarks.
>>110184002>benchmarks update just in time for carbon releasesneaky fucks
>>110185229based
>SI
>SII refuse to believe anyone uses this term in good faith. In fact, OP probably let out an audible smirk as he shitposted it together with this useless benchmark combination that shows nothing about a specific feature a certain model might be good at because everything is lumped together in a nonsensical way.
>>110184020self-insert
>>110187833>I refuse to believe anyone uses this term in good faithI do. Welcome to the future, faggot.
>>110184002yes in spanish
>>110187342Google bros, we won.
