Mark Zuckerberg bet the farm on AI supremacy, and this year's crop is infested with bugs.
According to a new report, Meta has quietly pushed back the launch of its next-generation foundational AI model, internally code-named Avocado, from this month until at least May. The reason? Internal tests showed it underperforming on key benchmarks for reasoning, coding, and writing - trailing rivals like Google's Gemini 3.0, even as it beat Meta's own prior efforts and older Google models.
The model, code-named Avocado, outperformed Meta’s previous A.I. model and did better than Google’s Gemini 2.5 model from March, two of the people said. But it has not performed as strongly as Gemini 3.0 from November, they said. -NYT
This delay arrives after Zuckerberg has poured unprecedented resources into the race. Meta is guiding for $115–135 billion in capital expenditures this year alone - nearly double last year's spend - with the overwhelming majority earmarked for AI data centers, compute clusters, and infrastructure. The company has also signaled longer-term commitments approaching $600 billion in U.S. investments, plus a $14.3 billion stake in Scale AI that installed its CEO, Alexandr Wang, as Meta's chief AI officer. The new "TBD Lab" was tasked with fruit-themed breakthroughs: Avocado as the core model, Mango for images/video, and a bigger "Watermelon" on the horizon, the NY Times reports.
Zuckerberg once promised these efforts would "push the frontier" toward superintelligence. Now, insiders say Meta is even weighing temporarily licensing superior models from competitors like Google to keep its products competitive.
As a result, Meta has delayed Avocado’s release to at least May from this month, the people said. They added that the leaders of Meta’s A.I. division had instead discussed temporarily licensing Gemini to power the company’s A.I. products, though no decisions have been reached.
Not great...
While Zuck licks his wounds, Elon Musk is reorganizing xAI - ordering another round of job cuts at the two-year-old startup over poor performance of its coding product, FT reports. Grok's coding capabilities have lagged behind rivals like Anthropic’s Claude Code and OpenAI’s Codex - however Musk on Thursday revealed the company's 'Macrohard' or "Digital Optimus" which can 'basically automate entire companies' by observing and intelligently simulating their functions.
Musk has brought in managers from SpaceX and Tesla as "fixers" to audit employee work, focusing on data quality issues in model training and firing those deemed inadequate. This has forced out several more co-founders - including Zihang Dai (a senior technical leader who admitted xAI was behind on coding) and Guodong Zhang (who ran pre-training for Grok models and was blamed for coding shortfalls, departing Thursday). Only two of the original 11 co-founders remain: Manuel Kroiss (“Makro”) and Ross Nordeen. Previous exits include Greg Yang, Tony Wu, Jimmy Ba, and even Toby Pohlen, who briefly led the "Macrohard" digital agents project before leaving after 16 days.
Posting on X Thursday, Musk said "Same thing happened with Tesla."
The upheaval follows the $1.25 billion merger of SpaceX with xAI, amid pressure to meet ambitious goals - including space-based AI data centers, Moon factories, and Mars colonization - and a potential blockbuster stock market listing by June. According to FT, xAI staff have been wilting under "extremely hardcore" demands, though a company memo denied mass layoffs.
Yet Musk is aggressively course-correcting: redeploying Tesla's Ashok Elluswamy to reboot Macrohard and develop the "digital Optimus" - blending real-world AI with Grok models. He's also been reviewing past interview rejections, apologizing publicly - "Many talented people over the past few years were declined an offer or even an interview at xAI. My apologies" - and reaching back out to promising candidates. This week, xAI poached Andrew Milich and Jason Ginsberg from the hot AI coding app Cursor to supercharge "Grok Code Fast."
Meanwhile, the massive Memphis supercluster - already with over 200,000 GPUs and expanding toward 1 million - benefits from X data integration, giving xAI unique advantages in scale and real-time training.