Compiled | Tina
When an AI company sees co-founders and a batch of core engineers leaving in a very short period, the typical first reaction from the outside world is usually: "It's over, something must have happened."
The wave of departures at xAI over the past week felt exactly like that. The news snowballed on X, eventually turning into a meme: some people who had never even worked at xAI posted "I've also resigned," skewing the narrative of a "mass exodus" into a massive bandwagon joke fest through trend-following satire.
Musk didn't offer any explanation but directly released a recording of a 45-minute all-hands meeting.
This video served as a public clarification: Are these departures genuinely voluntary, or is the company undergoing organizational restructuring? What is xAI currently working on, and who is responsible for what? How will the four key areas—Grok, programming models, video generation, and Macrohard (a multi-agent software company)—progress from here?
Even more striking, Musk also made a bold prediction in the video: By the end of 2026, AI might not even write code anymore, instead directly generating binary.
Elon Musk predicts that by the end of 2026, AI will completely bypass traditional programming, no longer writing code but directly generating binary programs. He believes that the binaries generated by AI can be more efficient than any output produced by a compiler. In other words, in the future, you will only need to tell the AI: "Generate an optimized binary for this specific objective," and you can skip the traditional programming process entirely.
The current software development workflow is: Code → Compiler → Binary → Execution; while Musk envisions the future will become: Prompt (Instruction) → AI-generated Binary → Execution. He also stated that Grok Code is expected to reach state-of-the-art levels within 2 to 3 months.
In Musk's view, software development is standing on the threshold of a fundamental transformation.
Musk: It's Not a Departure Wave, It's Me Doing Layoffs
Losing two co-founders and multiple core engineers within a week is hard to consider normal talent turnover for an AI company.
At xAI, departure announcements came almost in "clusters," quickly evolving beyond gossip about "who left" into a more direct question: Can this company remain stable?
TechCrunch statistics show that in the past week, at least 9 engineers publicly announced leaving xAI, including co-founders Jimmy Ba and Tony Wu. In just a few days, nearly half of the founding team left. For any company still in rapid expansion, the speed and scale of this are alarming enough.
Musk was clearly aware of this. Before the departure narrative could further ferment and spiral out of the company's control, he chose an extremely unconventional response: making an internal all-hands meeting public.
A 45-minute all-hands meeting video was directly posted on X, open to everyone.
And before releasing the video, Musk had already given his assessment internally. At Tuesday evening's all-hands meeting, he characterized this wave of departures as a "stage-fit problem," not a performance issue. "Because we have grown to a certain scale, we are restructuring the company to operate more efficiently at this scale," he said, "and in fact, in this situation, some people are more suited to the early stages of the company, but less suited to the later stages."
Musk later clarified on X that this was a round of separation due to organizational restructuring—essentially layoffs, not purely personal choices.
"As a company grows rapidly, the organizational structure must evolve. This unfortunately means parting ways with some people."
In this post, he also emphasized that the company is "actively hiring" and concluded with a recruitment line in his signature style—if you're interested in the idea of "mass drivers on the Moon," welcome to join xAI.
However, organizational restructuring itself is not uncommon; what truly puzzled outsiders was the clustering of these departures: many leaving were from the founding team, and it wasn't isolated cases but a wave of personnel changes happening within an extremely short timeframe. Below is a departure timeline compiled from public information.
xAI Departure Timeline (Public Information Summary)
Feb 6, Ayush Jaiswal (Engineer) wrote: "This is my last week at xAI. I'll be spending time with family over the next few months while tinkering with some AI-related stuff."
Feb 7, Shayan Salehian (Responsible for product infrastructure and model post-training behavior, formerly at X) wrote: "I've left xAI to start something new, officially ending my more than 7 years at Twitter, X, and xAI, with immense gratitude." He also mentioned that working closely with Musk taught him "obsessive attention to detail, an almost insane sense of urgency, and thinking from first principles."
Feb 9, Simon Zhai (Technical Staff) wrote: "Today is my last day at xAI, very grateful for the opportunity. It's been an amazing journey."
Feb 9, Yuhuai (Tony) Wu (Co-founder, Head of Reasoning) wrote: "I resigned from xAI today. Time for a new chapter. This is an era of possibilities: a small team equipped with AI can move mountains, redefining what's possible."
Feb 10, Jimmy Ba (Co-founder, Head of Research & Safety) wrote: "Today is my last day at xAI. With the right tools, we are moving towards a 100x productivity era. Recursive self-improvement loops are likely to go live within the next 12 months. Time to recalibrate my macro 'gradient'. 2026 will be crazy, and likely the busiest year yet for the future of our species."
Feb 10, Vahid Kazemi (Machine Learning PhD) wrote that he left xAI "a few weeks ago" and stated: "In my opinion, all AI labs are doing the same thing, which is boring. I think there is much more creative space, so I'm going to start something new."
Feb 10, Hang Gao (Worked on multimodal projects, including Grok Imagine) wrote: "I left xAI today." He called the experience "incredibly valuable," mentioned his contributions to multiple Grok Imagine releases, and praised the team's "humble craftsmanship and ambitious vision."
Feb 10, Roland Gavrilescu (Left in November last year to found Nuraline) posted: "I left xAI and am building something new with others who left xAI. We're hiring :)"
Feb 10, Chace Lee (Macrohard founding team member) wrote: "Taking a brief reset, then returning to the frontier." (Macrohard is a pure AI software project under xAI, aiming to fully automate software development, coding, and operations using Grok-driven multi-agent systems; its name carries a playful nod to Microsoft.)
xAI still has over a thousand employees, so it's unlikely to "cease functioning" in the short term due to this departure wave. But people leaving so concentratedly and quickly easily leads to exaggerated rumors online: some X users even joined the meme, posting "I've also left xAI"—even though they never worked there.
This was precisely the direct context for xAI subsequently choosing to make the all-hands meeting video public.
In this all-hands meeting, Musk repeatedly emphasized two judgments. First, the departures were not a performance issue but a stage-fit problem. Second, at this current stage, xAI's sole priority is only one thing—velocity and acceleration.
If you run faster than everyone else in a certain technological area, you will eventually become the leader.
xAI's New Structure: Four Major Teams, What Do They Do?
At this meeting, Musk talked about a few key things.
First, he positioned xAI: Don't view it as a mature company, after all, it's only been around for two and a half years.
He described xAI as an "infant," emphasizing: "We are small, but growing extremely fast." Many competitors have been operating for five, ten, or even twenty years, with better initial resources and more people, but xAI managed to push many key directions to the forefront in just a few years, even achieving "firsts."
He then listed several "report cards" in one breath: voice, image, and video generation achieved industry leadership; he also emphasized that "predictive capability" is the key metric for measuring intelligence, and said Grok 4.20 outperformed other models on prediction tasks. In terms of application form, xAI has integrated capabilities like Grok and Imagine into one App and made more aggressive transformations to X.
They have an even wilder goal: Grok-pedia isn't just "building a better Wikipedia," but aims to become a "Galactic Encyclopedia"—incorporating all knowledge (including images, video), aiming for an order-of-magnitude improvement in scale and accuracy.
Addressing the departures and restructuring, Musk said this isn't "falling apart" but "the company growing up": xAI has reached a new scale node.
He used the metaphor of organism growth: in the early startup days with dozens of people, everyone could chat freely; at hundreds, structure becomes necessary; growing further requires "differentiating organs, growing limbs, and even once having a tail—fortunately, the tail disappeared later." So restructuring is for running faster. Consequently, reality dictates: some are suited for early-stage charge but not necessarily for later-stage scaled operations.
Finally, he announced the new organizational structure. xAI will operate along four lines:
First, Grok Main and Voice, which is the core Grok main model;
Second, the model specifically for programming;
Third, the Image and Video model, namely Imagine;
Fourth, MacroHard, whose goal is complete digital simulation of the entire company-level system.
Grok remains xAI's most important product-facing entry point; the Coding team has been placed in an even more central position, not just for "writing code" but to compress the entire software production chain.
Grok Team: Within a Year, Grok Deployed in Over 2 Million Teslas
The Grok team is currently xAI's most core and user-facing product line, carrying almost all external intuitive perception of xAI: chat, voice, in-car, API, and deep integration with the X platform. The lead for this line is Aman Madaan (joined xAI in 2024).
To summarize Grok team's progress this year in one sentence: from "having nothing" to becoming xAI's fastest-deployed, most successfully scaled product line.
Aman described the speed of the voice line advancement as "zero to first":
The Grok Main and Voice teams will merge into one team. There's a classic example in voice. In September 2024, OpenAI had already launched its advanced voice mode, while at that time we had nothing, not even a model. But we started after that; in just six months, starting from scratch, fully self-researched, with virtually no audio background in the team, we built a voice product that surpassed OpenAI's within six months.
And now, in less than a year, Grok has been deployed in over 2 million Tesla vehicles, and we've also launched the Grok Voice Agent API.
Within a year, we went from "having nothing" to industry leader. This kind of thing can only happen in a place like xAI: small teams, extreme dedication, mission-driven, plus sufficient compute.
For the Grok main model, xAI is shifting focus from "Q&A" towards an "Everything App":
The story is the same on the chat model line. From Grok 1.5, Grok 2 to Grok 3, we've consistently been at the forefront of reasoning capabilities.
We want to move towards a world that's not just "Q&A," but building a true "Everything App." You can come here for legal advice, create presentations, solve complex problems, actually getting things done.
Our goal is to build an entry point where you can accomplish all work, truly amplifying everyone's capability, enabling them to do things far beyond individual limits, all delivered through an extremely simple, natural, seamless user experience.
In the coming months, the amount of work knowledge workers can accomplish will see orders-of-magnitude improvements.
Coding Team: By Year-End, You Might Not Even Write Code Anymore
If Grok is xAI's user-facing "conversational entry," then the Coding team is the company's true execution engine. This team not only handles xAI's internal coding systems but also shoulders a more radical mission: having AI write code itself, and ultimately replacing the act of "writing code" altogether.
The remarks by Coding team lead Makro were among the most emotionally stirring for engineers in this all-hands.
According to him, this is no longer about "efficiency gains" but a self-accelerating pathway: this generation of Grok Code is training the next generation of Grok Code. When "writing code" becomes part of the training pipeline, the focus shifts from tool usability to whether the system will continue down this path. Hence, programming has been directly elevated to one of the company's highest priorities. Moreover, training compute equivalent to "millions of H100s" has been invested, aiming to train the world's strongest programming model.
Musk's assessment was even more radical. In his view, "writing code" itself is showing characteristics of a transitional form. Ultimately, one will simply say, "Generate an optimized binary for this specific objective," bypassing the traditional programming process entirely.
Makro began by talking about a "qualitative change": models finally moved from "appearing usable" to "actually being usable":
Recently, programming has really changed a lot.
I used to keep complaining: people always urged me to use programming models, I tried them, but honestly, I wasn't really convinced. But recently it's been different—these models can now produce quite good, usable code quality.
Of course, you still need to review, to give feedback, but it's easy to see they can significantly boost human efficiency. This is no longer just "helping you write code"; they understand your intuition much better than before. Now when I describe a problem, I just need to explain it as if talking to an engineer colleague already familiar with the codebase; whereas before, you basically had to guide it step-by-step like leading a toddler on how to make changes.
And they don't just write code; they can also help debug code. Now we run Grok Code continuously for several hours to ensure that more complex changes to the training system can actually work stably in production.
Makro also described Grok Code's use as "production-level validation + recursive self-improvement":
So for us, this is no longer just about "writing code a bit faster," making engineers 10x more productive. We clearly see: we are on a path of recursive self-improvement—this generation of Grok Code is training the next generation of Grok Code. And this path has entered exponential takeoff and will continue.
Because of this, we are doubling down on the programming direction company-wide, elevating coding to one of the company's highest priorities.
If you're excited about programming, whether you're great at training models or a low-level software engineer interested in system design—this is where you should be. We now have training compute equivalent to millions of H100s, aiming to train the world's strongest programming model.
Guodong then stated:
Over time, we increasingly realize: at least in the dimension of programming, we are heading towards some kind of "singularity."
The real limiting factor may no longer be in algorithms or models, but in compute resources and energy: whether we can run sufficiently powerful models to support and empower everyone. And now, through this adjustment, we are a unified team; we will win on compute, we are winning the path of "space compute."
So, for every engineer—whether you're currently writing kernels, writing compilers, think about: Is this still worth doing by hand? Perhaps you should join us and, in the coding direction, somewhat "automate away a part of yourself" to run faster.
Honestly, this is a very crazy, very exciting year. It's truly "insane to be alive in this era." I can already clearly sense the scent of AGI—at least in programming, it's very close.
Meanwhile, Musk added an extremely impactful judgment during the Coding segment, treating "writing code" itself as an intermediate state:
Yes, I think things will reach a stage—perhaps even by the end of this year—where you won't even bother "writing code" anymore; AI will directly generate the binary for you.
And the binary generated by AI will be more efficient than any compiler can achieve. So you'll just say: "Give me an optimized binary for this specific objective." And you bypass traditional coding altogether. Writing code is actually just an intermediate step, likely unnecessary by... I think around the end of this year.
And we expect Grok Code to reach state-of-the-art levels within two to three months. This is all happening very, very fast.
Grok Team: Within a Year, Grok Deployed in Over 2 Million Teslas
Imagine is xAI's image and video generation product line, also one of the most compute-intensive directions within the company. The lead is Guodong, core members include Haotian focusing on video, and Chaitu.
Guodong described Imagine's progress at the meeting as a speed battle from "zero to full rollout":
The Imagine team started essentially from scratch about six months ago. No diffusion model code, no existing foundation, but now Imagine is fully integrated into all our products, including the X app. You can now long-press an image in X to directly edit or turn an image into video.
Imagine's growth speed is incredibly astonishing. Users now generate nearly 50 million videos daily, and 6 billion images in the past 30 days. For comparison, Google recently said their model generated 1 billion images in 30 days; we are six times that. Haotian then projected the timeline to "by year-end," emphasizing the route of "one-click long video generation + zero intervention."
Additionally, Musk added a directional judgment, betting future compute on "real-time video understanding / generation":
My prediction is: in the future, most AI compute resources will be used for real-time video understanding and real-time video generation, and we expect to be the leader in this area. It's worth reiterating: six months ago, we had almost nothing, or were very weak, in video and image generation and editing; but within six months, we surged to number one.
And I believe everyone will be very impressed with the upcoming Grok 4.2 model—it's a significant leap. But that's just a "minor version" in our new model system. There will be medium and large versions next; they will be even more intelligent.
MacroHard: The Ultimate Proving Ground for AI Agents
MacroHard is defined as a direction driven by AI Agents to "simulate company operations," with goals far beyond writing code: it aims to simulate human computer usage, automatically run software and various company processes, and further achieve simulation of an entire company. The lead is Toby, core members include John M. responsible for execution layer advancement.
Toby's one-sentence definition for MacroHard at the meeting was very hardcore:
MacroHard is building a fully capable, digital, real-time human simulator.
It can accomplish anything a human can do on a computer, including using advanced tools in fields like engineering and medicine. In the future, there should be rocket engines fully designed by AI.
In a sense, this is one area where AI remains significantly weaker than humans. And precisely because of that, this is the most exciting, most worthwhile investment, and most likely direction to truly change the entire field.
John M. broke down MacroHard's path into "CLI → GUI → end-to-end orchestration":
We are building these strong reasoning models, and they will control our CLI (Command Line Interface). We actively use these models daily; they bring huge productivity boosts to the entire team. I know the voice team is doing exceptionally well in this regard.
This is also why we need compute—we need large-scale compute to run these models to enhance our own productivity. But the reality is: 80% to 90%, even 95% of the global software world has GUI (Graphical User Interface). This is a very important fact. To truly make people's lives easier, we must develop models capable of completing daily tasks on GUI.
So MacroHard's goal is to simulate a company whose "output is digital results." This is the next step for agents: MacroHard will achieve true end-to-end orchestration across desktop environments and bring immense economic prosperity.
Simultaneously, Musk elevated MacroHard's significance to the level of "human simulation":
The MacroHard project, over time, may become our most important project.
We're talking about: simulating an entire human company.
It is theoretically entirely possible to fully simulate any company whose "output is digital products." This will usher in an era of prosperity to a degree we can scarcely imagine now.
Infrastructure: xAI's True Moat
At xAI, infrastructure isn't a "back-office department" but the prerequisite for whether all the above radical judgments can materialize. The ML infrastructure team is responsible for building the company's training, inference, and entire toolchain systems. In their own words: from a software engineer's perspective, this might be the coolest type of system you can work on.
The most typical example occurred during Grok 3 training. At that time, xAI had acquired 100,000 H100s, hardware delivered, but the software wasn't truly ready. The team thought the system could run, but scaling to 30,000 cards gave clear feedback: the system wouldn't run. The issue wasn't a single bug, but too many "unexpected events" in the datacenter: switch jitter, link jitter, switch failures, frequent GPU damage, numerical instability… none of these could be enumerated in advance. But the goal was singular: make 100,000 H100s work as one unit.
A single training step might be only 5 seconds: advancing one step every 5 seconds, but anything could happen within those 5 seconds. So the system must achieve: automatic recovery and continuous progression despite constant unexpected events, rather than stopping and waiting for human rescue upon any problem.
This kind of problem is rarely encountered elsewhere. Not because engineers aren't smart, but few simultaneously possess compute at this scale and talent density. At that time, the entire pre-training team was about 15 people, with perhaps only 7 actually responsible for the training system, but they deliberately maintained this "talent density," not scaling by headcount, and ultimately this small team completed Grok3 training.
NVIDIA CEO Jensen Huang has remarked in multiple interviews: No one gets AI compute online faster than xAI.
Subsequently, the RL & Inference team took over. This team is responsible for running training tasks and production inference systems at scale on Earth, and soon potentially in space, with a direct goal: scaling the system from 100,000 chips to millions of chips, and making it resilient to both known and unknown hardware failures.
Currently, their achievements converge at the Memphis datacenter: xAI has built one of the world's largest AI training clusters, and it's still expanding—Phase 1 is 330,000 GB300, with an additional 220,000 GB300 to be added next.
Clearly, to have the best models, large-scale training compute is absolutely foundational.
And at the end of this all-hands, Musk pushed this logic to a place almost exclusive to science fiction. If Earth can no longer accommodate this compute, what's next? The answer: the Moon.
Musk believes that to truly understand the universe, we must ultimately leave Earth to explore, and this is precisely the motivation for merging SpaceX and xAI: accelerating humanity's future understanding of the universe, extending the light of consciousness among the stars.
From an energy perspective, he gave an extreme contrast: today's entire human civilization uses only about 1% of Earth's available energy. And if humanity merely wanted to utilize one-millionth of the Sun's energy, that would still be a million times our current civilization's energy consumption.
The problem is: on Earth, you simply cannot access that kind of energy. Earth is but a "tiny speck of dust" within the solar system. The Sun accounts for 99.8% of the solar system's mass. Without leaving Earth, you can hardly achieve any substantive increase in utilizing solar energy.
So in his view, the next step isn't "larger Earth datacenters" but "datacenters leaving Earth." First, send datacenters to Earth orbit; later, move manufacturing and launch to the Moon—building factories on the Moon to produce AI satellites, then using mass drivers to "catapult" them one by one into deep space, scaling compute to levels Earth simply cannot sustain.
Reference Links:
https://www.youtube.com/watch?v=aOVnB88Cd1A
https://techcrunch.com/2026/02/11/senior-engineers-including-co-founders-exit-xai-amid-controversy/