Zuckerberg and his GenAI clown show
Meta actually published this with a straight face. "Advanced reasoning capabilities" is a rather bold claim for a model that bombs basic logic tests and has to steal its answers from OpenAI.
Facebook has been taken for quite a ride since last month, aye? Between the endless lawsuits and internal chaos, you would assume they want an easy win. But the absolute funniest distraction of them all has to be their recent large language model release, Muse Spark.
The Llama days seem so long ago now. Last year, Llama 4 dropped and totally bombed after the open-source community caught Meta manipulating the benchmark results. So what did Zuck do? He panicked. He completely abandoned their open-source ethos and dropped a heavily gated, closed-source frontier model that cost roughly $14.3 billion to put together. And the real punchline to this multi-billion dollar joke is the guy he bought to lead the project.
A lucky tech bro running superintelligence:
Let us talk about Alexandr Wang. His luck at the start of the tech bubble boom is honestly unbelievable. He became a billionaire in his early twenties by founding Scale AI, which essentially operated as a digital sweatshop. They used platforms like Remotasks to pay workers in the Global South absolute pennies to click on stop signs and label text data for actual AI companies.
Alexandr Wang is a 29-year-old tech bro CEO. He possesses absolutely zero background as an AI researcher, a foundational machine learning architect, or someone with any actual credentials relating to artificial superintelligence.
The sheer insanity of Zuckerberg dropping $14.3 billion to parachute this guy into Meta and hand him the keys to Meta Superintelligence Labs is the very definition of a bubble. He handed a foundational AI project to a data-logistics guy who simply got ridiculously lucky during a gold rush.
The poison well and copied homework:
Here is where the hypocrisy gets truly hilarious. For years, Wang's entire business model relied on selling premium human-labelled data. He spent years aggressively preaching about the poison well and model collapse. He loudly claimed that if models trained on synthetic outputs from other AI models, they would degrade into generic trash. He mocked the practice repeatedly, but only because his stance protected his own slave-driven data company.
The absolute second he gets the top job at Meta and actually has to build a model under pressure, he abandons his morals and straight-up copies the homework. To rush out Muse Spark, they used distillation to siphon data from OpenAI, Google, and Chinese models like Qwen. He drank from the exact poisoned well he warned everyone about, simply because he had absolutely no idea how to build a model from scratch.
The delusion of personal superintelligence:
If you look at the marketing garbage they are pushing in their official announcement, the delusion is off the charts. They are unironically calling Muse Spark a "personal superintelligence" equipped with "advanced reasoning capabilities".
How can anyone type the word intelligence when the model literally cannot execute proper reasoning? It faceplants on basic logic puzzles because it is just blindly parroting distilled answers it scraped from Google and OpenAI. It does not "understand your world" at all. When they brag about it analysing your immediate environment, they just mean it can recognise a coffee mug on your desk so Instagram can serve you a targeted ad to buy another one. Calling a glorified shopping and calorie-counting bot a superintelligence is peak Silicon Valley hubris.
Gaming the benchmarks to hide the stupidity:
The desperation is palpable across their entire blog. They plastered the page with cherry-picked charts claiming they beat GPT-5.4. Anyone with half a brain can see they just hard-trained the model on the exact test data!
They treated building artificial general intelligence like a cheap Kaggle competition. Wang threw his digital sweatshop workers at the specific questions used in public leaderboards so the model would look smart on paper. François Chollet publicly roasted them for this blatant grift. He pointed out that Muse Spark was completely over-optimised for curated charts at the severe detriment of actual real-world usefulness. They brute-forced the test answers and completely failed to build any underlying reasoning logic.
Shorter thinking is actually a downgrade:
The resulting model is pure garbage. It is so hilariously pointless that its entire existence feels like a prank. Meta is parading around their new "thought compression" and "Contemplating Mode" gimmicks, trying to show off shorter thinking as if making a model aggressively skip actual logic is a good thing!
When you push Muse Spark on abstract logic or coding, the backward stupidity is painful to watch. It completely bombed the ARC-AGI-2 reasoning test with a pathetic score of 42.5, while OpenAI and Google sit comfortably in the mid-70s. It fails terribly at long-horizon coding on Terminal-Bench 2.0. Local SEO analysts are already laughing because it hallucinates fake star ratings and reviews when you ask it to find a local plumber or lawyer.
It is so bad that Zuckerberg himself already knows it is a flop. Behind Wang's back, Zuck quietly started a shadow organisation called Applied AI Engineering led by Reality Labs veteran Maher Saba. He gave Saba an ultra-flat 50-to-1 engineer structure and told them to build the actual data pipelines to salvage the ad revenue. Wang gets to sit in his multi-billion dollar golden cage tweeting out PR fluff while Zuck bypasses him entirely. Hilarious.
✅ The Verdict
Look at the insane cost they have already sunk into this project. You have a $14.3 billion acquisition, a ridiculous $21 billion cloud compute deal with CoreWeave, and the upcoming madness of massive 1-Gigawatt data centres Zuckman wants to build.
They are putting immense strain on local power grids, drinking millions of gallons of fresh water, and flooding the internet with synthetic garbage. And all of this money, energy, and water was spent to give you a closed-source model that is currently being used to predict the calories in an image of your food and superimpose a digital mug onto a shelf.
Well done, Zuck. That is a nice, neat little feature for your ecosystem.