Former xAI Founder and University of Toronto Associate Professor Jimmy Ba: More Reporting Layers Mean More Lossy Compression
Jimmy Ba, co-founder of xAI and associate professor at the University of Toronto, stated during a 20-minute discussion at the Cerebral Valley AI Summit that the next generation of Grok was undergoing coherent training on approximately 110,000 GB200 units at that time. He mentioned that human organizations struggle to exceed three layers of reporting, leading to significant lossy compression of information. Companies are structured to allow anyone to become a tenfold engineer, and the main issue to resolve is alignment. He compared the competitive path to developing a chat assistant, while xAI is discussing the next generation of "human simulators" that hire in the manner of regular employees. In his predictions for 2026, he noted that xAI is nearing 1 gigawatt of coherent training and was asked about one million physical robots moving atoms. The interview is publicly accessible and is not a new closed-door summary from August 2026.
Ba is a co-author of the Adam optimizer, the first author of Layer Normalization, and a student of Geoffrey Hinton. He joined the founding team of xAI in 2023 and reported directly to Elon Musk. On February 10, 2026, he posted a farewell message, stating that the mission is to push humanity up the Kardashev technology scale and mentioned that a recursive self-improvement loop might go live in the following 12 months. On the same day, another founder, Yuhuai Wu, left, and about half of the original twelve founders departed afterward. Musk subsequently divided the company into four segments: Grok, Programming, Imagine, and Macrohard, with Macrohard explained as a full digital simulation of the entire company.
The Memphis cluster started with about 100,000 H100 units and later added H200 and GB200 units; Musk has publicly stated that the second data center will add approximately 110,000 GB200 units. Jensen Huang claimed that no one is faster than xAI in bringing training computing power online. In the Chinese translation, sentences like "after brushing through AIME/GPQA, retraining lawyers and full-stack, charging based on value creation rather than Tokens, and achieving tenfold efficiency to create tenfold smarter brains" were not matched sentence by sentence in the public subtitles and cannot be considered Ba's original words.
The buyer aims to write intelligence as a laboratory for hiring simulated employees; the seller can train tens of thousands of accelerated cards to achieve the same gradient. The computing power is shifting from selling chat phrases to simulating entire positions. The beneficiaries are the founding teams that can run flat organizations on gigawatt clusters; the burden falls on large corporate matrices that compress information through multiple layers of reporting. This event is driven by public interviews and subsequent departures, not by new product releases.
Source: Public Information
ABAB AI Insight
Ba articulates organizational theory as information theory. Three layers are not just management platitudes; they represent his identified limit of lossy compression: each layer of reporting loses gradients. The tenfold engineer hypothesis is an anti-bureaucratic financing narrative, explaining why small teams deserve 110,000 cards. The human simulator and Macrohard transform "agents" into "hireable individuals," changing the pricing unit from Tokens to positions. In the interview, he discusses coherent training, while in his farewell post, he talks about the Kardashev scale and recursive improvement, both framing computation as a step in civilization rather than merely cost-cutting for customer service.
The capital path is to first stack cards into a city, then decompose the model into four product lines. After the merger, SpaceX aims to vertically integrate rockets, Starlink, and models, with simulation companies serving as the software layer interfacing with the physical layer. One million robots moving atoms connects the narrative of bit laboratories to Tesla/Optimus-style atomic factories. The departure of half the founders indicates that the gigawatt cluster can remain, while the recipe inventors can leave; those departing write the "next chapter" as smaller teams with stronger tools.
Benchmarks include Hinton's lab establishing normalization as an industrial standard, DeepMind using small teams to play Go, and OpenAI making chat assistants the default entry point. xAI bets on the opposite entry: first simulating companies, then forcing models. The current phase is about control restructuring: computing power is in Memphis, products are in four lines, and personnel are in resignation announcements.
Structurally, this is about technological substitution layered with organizational reconstruction. The mechanism is: once a model can deliver by position, bureaucracy becomes an additional compression loss; whoever first reduces the reporting chain to below three layers can train with the same stack of cards on less distorted targets. Ranking is the scoreboard of the first half, while simulated employees are the invoice units of the second half—provided that one does not treat translations as original words.
ABAB News · Laws of Cognition
More reporting layers mean more lossy compression
Position invoices are closer to the company than Token invoices
Cards can remain, but the inventors of the recipes can leave