Seed

ByteDance

2026-10-19 06:20:03

ByteDance Seed evaluation lead Huang Wenhao reportedly leaves, may start a company next

BlockBeats reported that Huang Wenhao, who was in charge of model evaluation at ByteDance’s Seed team, has recently left the company and may choose to launch a startup next. Huang holds a PhD in computer science from Peking University and previously worked at Microsoft Research Asia and the Beijing Academy of Artificial Intelligence. In 2023, he joined Lee Kai-fu’s 01.AI as a co-founder, serving as vice president of algorithms and head of model training, where he led the training and release of large models including Yi-34B and Yi-Large. In August 2024, he left 01.AI for ByteDance Seed, initially overseeing technical project management and planning and reporting to Zhu Wenjia. After former Google DeepMind vice president Wu Yonghui took over Seed’s foundational research in early 2025, Huang began reporting to Wu. His focus later shifted toward model evaluation, while also taking part in frontier technology planning and talent recruitment.

190

Qubic

2026-10-18 15:06:50

Qubic says weak seed randomness led to theft of about 2.94 trillion QU from 1,404 wallets

Qubic said an attacker exploited insufficient randomness in the seed generation process for some wallets between 23:42 UTC yesterday and 01:26 UTC today, gaining access to 1,404 Qubic wallets and moving about 2.94 trillion QU. The project said the issue was tied to seed generation rather than the Qubic protocol itself. According to the announcement, affected seeds include those created before Oct. 18, 2026 in any browser on iPhone or iPad, those created before the same date in Safari on Mac, and those created before July 2025 through the Qubic iOS app. Qubic said the network was paused at tick 85,371,949. It added that 87.1% of the stolen funds remain on the Qubic network and have been frozen, while the other 12.9% was transferred to MEXC, where the receiving accounts have also been frozen. The project said the seed generation issue has been fixed in version 4.26.4, and that next steps will be decided by the Computor quorum under network rules. A full technical report will be released after a fund recovery plan is finalized.

220

ByteDance

2026-10-13 03:55:36

ByteDance Seed forms model alignment and preparedness unit

ByteDance's Seed division has announced the creation of a new Seed Model-Alignment&Preparedness department, according to an internal email cited by ChainCatcher. The new unit will be led by Yan Lin and will report to Wu Yonghui, the head of Seed. As part of the change, the original Responsible AI team and some members of the Chat team will move into the new department. The stated goal is to step up investment in safety and alignment work. In the email, alignment was described as keeping model behavior consistent with human intent and safety constraints. The new department is also tasked with improving the safety capabilities of Agent Model systems and guarding against frontier capability risks tied to chemistry, biology, radiation, and nuclear domains.

380

OpenAI

2026-09-02 04:41:19

OpenAI's Astra Uses Recurrent Depth Architecture; ByteDance and Seed Team Explored Similar Path

OpenAI is set to release Astra, a model featuring a recurrent depth architecture (looped transformer) that allows the same information to pass through the same layers multiple times. ByteDance and Seed team have previously published similar research. This approach enables smaller models to approach larger model performance but raises safety concerns: some reasoning occurs in hidden states, making it harder to detect violations. OpenAI has limited the use of recurrent depth and retains an interpretable chain of thought.

1000

ByteDance

2026-08-31 06:21:08

ByteDance Raises Doubao Share Price to $17.02 and Expands Cash-to-Equity Option for Some Staff

ByteDance has raised the internal price of its Doubao shares to $17.02 from $14.85 in June, according to a report cited in the source material. That marks a 14.6% increase in two months and a cumulative gain of 30.1% from the April buyback price of $13.08. The report says some employees are now allowed to adjust their compensation structure by converting part of their cash package into Doubao shares. The allocation cap varies by employee. For some L2 and L3 staff, the full allocation ratio is 20%, while some L3 and L4 staff may reach 25% or 30%. ByteDance has also increased Doubao share grants, with some employees receiving additional quota equal to about 5% of total compensation. The policy does not apply to all employees. It mainly covers teams tied to Seed, including pre-training, multimodal and VLM large-model groups, as well as Flow’s main Doubao app team. Some teams with deeper links to Seed and the Doubao infrastructure, including Ark, may also be eligible.

950

ByteDance

2026-08-31 06:22:07

ByteDance reportedly raises Doubao share price and lets some staff swap cash for internal equity

ByteDance has reportedly raised the price of its internal Doubao share program and is allowing some employees to swap part of their compensation from cash into those shares, according to an internal update cited by Dongcha Beating AI. The reported price was lifted to $17.02 from $14.85 in June, a 14.6% increase over two months. Based on the first buyback price of $13.08 in April, the Doubao share price is now up 30.1%. The report said the policy does not apply to all employees. It mainly covers teams tied to ByteDance’s AI efforts, including Seed’s pre-training, multimodal and VLM groups, as well as Flow’s main Doubao app. Some teams closely linked to Seed and the Doubao base model, including Ark, may also be eligible. Reported conversion limits vary by employee level, with some L2 and L3 staff capped at 20%, while some L3 and L4 staff may reach 25% or 30%. ByteDance has also reportedly increased Doubao share grants, with some employees receiving an added allocation of about 5% of total compensation.

1181

ByteDance

2026-08-30 13:11:24

ByteDance Delays Doubao 2.2 to Focus on Coding Agents

ByteDance has pushed back Doubao 2.2, which was supposed to launch in August, and plans to spend extra time on training, with coding, tool calling, and agent capabilities as priorities. According to BlockBeats, one of Seed's targets this year is to build a first-tier coding model. Internally, ByteDance hopes to reach a level similar to GLM-5.2 and Kimi-K3 by year-end, so developers begin to trust its coding model. The delay follows an August reorganization at Seed. Teams previously divided by text, speech, code, and vision have been recast as Pretrain Data, Horizon RL, Product Posttrain-Work, and Product Posttrain-Chat. Horizon RL will handle coding post-training; Work will concentrate on tool calling, GUI operations, and long task execution. ByteDance is also recruiting aggressively. Open positions include Code Agent, general Agent, and reinforcement learning algorithm engineer. A Multi-Agent Harness posting explicitly mentions studying Claude Code, Codex, and other coding agents, and building multi-agent plus reinforcement learning environments designed to run for long stretches. Seed 2.1 already highlighted coding when it came out in June, but ByteDance appears to think that was insufficient. Doubao 2.2 is being delayed intentionally, so coding can be upgraded to a more credible level before release.

1100

ByteDance

2026-08-06 06:46:45

Zhang Yiming says ByteDance will not use distillation as a shortcut in large-model race

ByteDance founder Zhang Yiming told an all-hands meeting of the company’s Seed team last month that the company will not treat model distillation as a shortcut to catch up in large language models, even if that leaves it temporarily behind domestic rivals. He asked the team to accept short-term sacrifice in pursuit of longer-term goals. Distillation, which trains a new model on the outputs of a stronger one, can reduce training time and computing costs. Inside ByteDance, the view is that this choice will make it harder for Seed’s language model to close the gap in the near term with Chinese peers including DeepSeek, Kimi and Qwen. The company is also said to be more concerned about political risk. Given TikTok’s long-running scrutiny from the U.S. government, any accusation that ByteDance had extracted capabilities at scale from leading U.S. frontier models could bring renewed pressure from Washington. The report also noted that Anthropic had previously accused DeepSeek, Moonshot AI, MiniMax, Zhipu and Alibaba of extracting capabilities from Claude at scale, though it did not name ByteDance.

2030