Be a part of our each day and weekly newsletters for the most recent updates and unique content material on industry-leading AI protection. Be taught Extra
As anticipated primarily based on updates and new settings within the cell app for Elon Musk’s social community X, a brand new massive language mannequin (LLM) referred to as Grok-2 from Musk’s sister firm xAI landed final evening — and it’s a doozy.
Built-in inside X itself and accessible by means of the Premium ($7 USD/month) and Premium+ ($14/month with no advertisements) subscription tiers, Grok-2 comes, fittingly, in two mannequin sizes: Grok-2 and Grok-2 mini. Grok-2 gives state-of-the-art efficiency in a variety of duties together with chat, coding, reasoning, and vision-based software, whereas Grok-2 mini is a smaller, quicker model optimized for effectivity, appropriate for easier text-based prompts requiring faster responses.
Grok-2 not solely boasts picture era capabilities primarily based on a partnership with Black Forest Labs and its new and surprisingly photorealistic open-source diffusion AI mannequin Flux.1, but it surely additionally shockingly outperforms the AI fashions from main rivals together with OpenAI (GPT-4o) and Anthropic (Claude 3.5 Sonnet) and even Google (Gemini Professional 1.5) on main third-party benchmark assessments.
A brand new, stunning chief throughout a number of benchmarks
Particularly, Grok-2 and Grok-2 mini outperform all different fashions on the GPQA, MMLU, MMLU-Professional, MATH, HumanEval, MMMU, MathVista and DocVQA benchmarks.
Even the lmsys-chatbot area, the place many corporations covertly check their AI fashions below alternate names prematurely of launch (together with xAI, the place Grok-2 was initially referred to as “sus-column-r”) congratulated xAI on the milestone.
As AI influencer and College of Pennsylvania Wharton Faculty of Enterprise professor Ethan Mollick noticed on X, “There are now five GPT-4 class models: GPT-4o, Claude 3.5, Gemini 1.5, Llama 3.1 and now Grok 2.”
Musk congratulated his “hardworking xAI team!” on the equally named social community.
Picture generations steal the present
Regardless that Grok-2 boasts main efficiency on all these totally different benchmarks associated to math, writing, code, and different duties, by far, the marquee characteristic capturing probably the most consideration from the bounce is its integration with Black Forest Labs’ Flux.1 picture era mannequin.
Earlier than the discharge of Grok-2, Flux.1 had already been making waves in AI and AI artwork circles extra particularly in the previous few weeks as individuals found that they may obtain extremely photorealistic generations from the open supply mannequin, sufficient to resemble acquainted conditions like a speaker at a TED discuss, in addition to adapt the mannequin utilizing low-rank adaptation (LoRA) to generate their very own likeness in numerous conditions.
Now {that a} model of Flux.1 is built-in straight into Grok-2 a lot in the identical method OpenAI built-in its picture era mannequin DALL-E 3 straight into ChatGPT, permitting customers to easily sort textual content prompts to the chatbot and ask it to make their photographs on command, customers are testing this functionality out in Grok-2 and discovering it’s notably permissive — producing controversial, compromising photographs even of public figures equivalent to U.S. presidential candidates Kamala Harris and Donald Trump.
Different main picture mills together with Midjourney and DALL-E 3 and Microsoft Designer have prohibitions round producing any such content material — particularly within the wake of the controversy earlier this yr over unauthorized express deepfakes of standard musician Taylor Swift (made by immediate engineering across the Designer restrictions) — so it’s notable that Grok-2 is bucking that pattern and permitting for extra freedom, and potential danger. Nevertheless, that’s in line with Musk’s acknowledged “free speech” ethos for X.
But customers are elevating considerations about what the aptitude means for the windfall of deepfakes and misinformation throughout the net.
As consumer @Omiron33 put it properly: “Yes, we’ve had MJ and Flux, but this is the first to make it usable and quick. Advertising, Propaganda and everything good or bad that comes with that just happened (IMO, the good outweighs the bad)”