字节跳动 Goku 训练受干扰事件后，涉事实习生融资近 3000 万美元

（本文作者为 Chelsea_Sun，钛媒体经授权发布）

NextFin News — By mid-2024, the engineers behind ByteDance’s Goku video model had run out of respectable explanations. Jobs died on no fixed schedule. Error metrics jumped when textbooks said they should fall. A configuration that behaved on one day failed on the next. Tiny experiments looked fine; full-scale runs did not. Hardware was swapped, stress-tested, even automatically drained from the pool when it looked cursed—then looked innocent again.

The incident channel grew toward three digits. Someone invoked solar storms. Someone hoped they had found a vendor bug worth a paper. The mood was less witty than exhausted.

Goku was not a side hobby. After OpenAI’s Sora showed what large diffusion-style video systems might look like, ByteDance, like its peers, multiplied internal bets. Goku lived in a commercialization research group. Other video efforts—later associated with the Seedance line—advanced elsewhere in the company. Schedules were aggressive: image capability first, then video, measured in weeks. Chen Shoufa, brought in after video-training work at Meta, owned much of Goku’s training stack and built it largely himself. Nearby sat Sun Peize. Beyond him, Tian Keyu—a long-running student intern from Peking University, nocturnal by habit—was scaling an autoregressive image approach, VAR, that would eventually take a major NeurIPS honor.

Until management forced a split, the two heavy efforts drew from one accelerator pool without a crisp ceiling. That design wastes less idle silicon and manufactures more friction. Stories of the friction still do not match. In one version, a peer hoarded cards and cancelled other people’s jobs. In another, overnight scale-ups starved for a handful of devices while earlier jobs overran into the small hours. Formal rules allowed designated people, after notice, to halt blocking work so a priority run could assemble enough machines; both sides’ descriptions of who held which keys diverge. Data quality became another scar: misaligned text and images on a batch linked to Tian cost him that responsibility. External pressure rose when a competitor opened a video model for testing. Leadership questioned whether two full-burn projects should run abreast; Goku was elevated and the pools were separated. The collapses accelerated.

What the team did next will be familiar to anyone who has debugged distributed training at the edge of paranoia. They instrumented code almost line by line. They tried saving state so often that saving became the bottleneck. Infrastructure staff hammered the cluster repeatedly. Automatic isolation logic once emptied the entire machine set. Twice-daily calls produced theories that lasted until Thursday. People began to doubt one another’s commits. Leadership named a final window measured in weeks.

The hinge was a long weekend of detective work rather than a single elegant proof. A reproducibility harness—strict enough to demand agreement far past ordinary floating-point hygiene—broke on the order of data reads. Tracing into the common deep-learning framework, an engineer hit a modified instruction that reordered inputs by time so yesterday’s failure could not be replayed. Timestamps on the affected node disagreed with clean nodes at a granularity so fine it suggested concealment rather than accident. Access logs, past the usual administrators, included Tian’s identity. He did not deny being there; the explanation that reached the team was curiosity about the code—an answer that sat awkwardly beside the firm’s ordinary code-review systems. Presence is not authorship. The altered instruction, the Goku engineers insisted, had no constructive purpose.

The following day, in the middle of the usual failure meeting, a colleague surfaced a concealed directory in Tian’s account on a shared machine. Inside were programs. At least one, they concluded, matched the data-pipeline sabotage. The debug call ended; forensic triage began. Later technical inventories list a large set of filenames; people who opened them describe interference woven through loading and training, not a lone “stop” button. Many employees could reach that server. A home directory is a clue, not a verdict.

Within another day, while Tian presented routine progress on his own model, the Goku side was preparing a package for corporate security. They discovered he could still join their bridge via an old link and a calendar entry—and removed him. With credentials still live, contents of the directory began to vanish in real time. Storage specialists reconstructed activity and recovered more than the first pass had shown. Recovered artifacts, several participants say, were not confined to one obvious folder; some sat under other path names and inside shared logs. Deletion is a different kind of fingerprint than possession. Tian has denied removing anything and has suggested another hand used his environment. People around him have also argued that he authored only crude stall logic and that harsher material was planted. Litigation posture and limited public process leave that claim hard to test from outside.

The engineers’ working theory, built over subsequent analysis, is darker than a few manual kills in a ten-day span—the outline Tian later offered in a long interview. They describe load-time hooks, including through components every run had to touch, intermittent triggers, thresholds tied to job size, stripped logs, and tools that hid their own activity. Versions, they say, changed while the victims still narrated each blind alley in channels the suspect could read. Absent a cluster of low-probability finds—the reproducibility break, the login, the hidden tree—several of them believe the company might simply have written Goku down as a failed bet.

Corporate security opened a formal inquiry in early August. In a lengthy session in Beijing, open to relevant full-time staff but not to interns, Tian—according to people briefed on the room—acknowledged repeated actions aimed at Chen and answered questions about later conduct with uncertainty about what he remembered. Access ended; the internship ended. That autumn the company stated publicly that an intern on the commercialization technical team had maliciously disrupted training and had already been dismissed. Internet attention and later ethics complaints around the prize-winning paper pulled in overlapping names; Chen says he did not instigate those campaigns but did confirm facts, and will not identify who did.

Once the interference stopped, Goku’s training resumed the ordinary logic of scaling and data. Quality moved in days, not mysteries—enough for a serious conference submission, not enough to erase a multi-month hole in a race measured in product generations. Sibling video teams reached outside testers first; the lineage that became Seedance kept the spotlight. Archaeologists of their own repo later pointed to still earlier fragments that looked capable of harming a smaller competing image project, and to doubts about how cleanly some VAR demos had been generated once the original intern was gone.

The residue was human. Chen asked for a company letter describing what the group had lived through; none existed as a template. He finished his degree under secondhand stories that cast him as the culprit or a manipulator of public opinion—stories he rejects. Sleep recycled broken runs and blank exam books. He eventually left with standard departure paperwork. Others recall personal dates erased by the incident calendar. Interns who carried the overnight load were not in the inquiry room; those interviewed years later say no private apology ever reached them, whatever was said to the people who were present.

Courts and capital kept their own time. ByteDance filed suit; a first-instance outcome, as people following the matter describe it, reduced a multi-million-yuan claim to a far smaller award, with further proceedings constraining public detail. Only then did Tian sit for an interview that had to be shot more than once under lawyers’ eyes. He offered a moral plot: an unfair contest inside the company, a blunt response, regret aimed at the firm. He has spoken of fearing exile from the field as a kind of darkness. Colleagues answer with a different plot: not a skirmish over schedulers, but a sustained, partly automated campaign that turned their work product into a target. Separately, operators who fund companies draw a distinction that does not require picking a hero—pushing back on opaque resource politics is one category of behavior; deciding that another team’s progress is optional is another.

A financing near thirty million dollars closed after the first judgment, according to people familiar with the flow of funds. Some institutions declined; the employment history was one factor among others, including aggressive term-sheet shopping. At least one group that proceeded questioned Tian at length, found his account stable across investors, press, and court, checked academic references, and had lawyers read first-instance material. The Goku researchers who spoke for this reconstruction say those firms never called them.

That omission is the part that travels beyond one company and one summer. The field still depends on shared frameworks and shared ideas. It also runs on scarce chips, scarce publication slots, and funding processes that prize speed. Results are legible in a dashboard. Character is not. A round can finish in the time it takes to know whether a founder treats other people’s work as infrastructure—or as inventory.

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App