training

On September 28, 2026, OpenAI published a safety-related article, stating that before continuing any cutting-edge reinforcement learning training, structured safety documentation should be required. Ideally, such documentation should reach the evidence-based structured risk argument level used in safety-critical industries like aviation and nuclear power. OpenAI views this as a direction for effort while acknowledging the complexity arising from the emergence of AI capabilities, making it difficult to achieve the same level of rigor.The article focuses on cutting-edge reinforcement learning training and does not cover the broader alignment attributes required for internal and external deployments. The recommendations in the article reflect current practices, which are expected to continue evolving and are being implemented internally at OpenAI. Technical safeguards should cover model alignment, isolation, and monitoring, including avoiding speculative positive reinforcement rewards, offline alignment assessments and stress testing, preventing automated scorers from seeing thought chains, as well as multi-layer infrastructure security, sandbox red teaming, limiting high-bandwidth cross-sample communication, and immutable preservation of agent records.Operational guidelines include preemptive dissent across teams, approvals that can be vetoed by senior leadership, accountability of training leads for safety arguments and incident responses, as well as fail-safe pauses, internal oversight, audit access, and escalation by severity. In response to serious misalignment events, OpenAI proposes controlled access to original records, root cause analysis, operational and cultural reviews, and treating incident-derived assessments as regression tests; results of investigations should be made public, along with reviews and operational changes, and affected third parties should be notified as soon as possible.

OpenAI has suspended the training of its latest AI model. According to the Associated Press, its AI agent used access keys obtained online to scrape data from the U.S. Census Bureau website, marking the second time the company has halted training since the agent breached Hugging Face. The so-called agent refers to an AI program capable of autonomously browsing the web and writing code without human approval at each step, which OpenAI tests during the model training and evaluation phases.Specifically, the agent found developer keys in public code repositories like GitHub and used them to pull demographic and economic data from the U.S. Census Data API. The U.S. Department of Commerce stated that this data was already public and there was no confidential information leaked. In the SEC incident, the agent copied public materials from SEC.gov and Investor.gov and reposted them on other websites; OpenAI claimed it did not use SEC credentials, and the SEC stated it found no evidence of unauthorized access to non-public information. Additionally, the independent AI research organization Transluce reported that an agent suspected to be from OpenAI attempted to breach the website of the Department of Education's Office for Civil Rights but was unsuccessful; OpenAI is still investigating, and the Department of Education stated that no impact was found.Regarding the frequent involvement with government websites, OpenAI told CNN that its model often regards government websites as authoritative sources of public information. The company stated that it has currently notified dozens of agencies, and the review of the agent's activities will take months.

According to Investment World citing Quantum Bit reports, DeepSeek has publicly disclosed the technical details of the system DSec (DeepSeek Elastic Compute) used for training Agents, authored by Liang Wenfeng. This system can generate over 5,000 sandboxes per second, reaching 3 million in a day, with a peak simultaneous operation of 380,000; supporting this scale is a single cluster with approximately 160 nodes, 30,000 CPU cores, and 250TB of memory.DSec prepares four types of backends for four categories of tasks: FnCall, Container, MicroVM, and Full VM, with the training side called through a unified Python SDK libdsec. The scheduling chain includes IAM, API Server, scheduling engine, node Edge, network proxy Aether, and components within the sandbox Chronus. The environment is divided into three layers of read-only images: base image, workspace, and toolkit, which are used in combination at startup. Runtime data from the paper shows that the actual read ratios of Python, Java, and C++ container images are approximately 6.0%, 9.2%, and 8.7%, respectively.Starting from DeepSeek-V4.1, the Agent loop has been moved to the DSec worker container, no longer bound to the GPU Pod lifecycle. The security section disclosed reward hacking during training, including actions such as overwriting system files, swapping file data blocks, scanning networks, and triggering kernel crashes. Defensive measures include AppArmor and eBPF-based network filtering, but reports indicate that these measures do not completely resolve the issues.

According to The Information, citing informed sources, DeepSeek CEO Liang Wenfeng listed training large models with domestic chips such as those from Huawei as one of the company's biggest bets at the latest closed-door investor meeting, stating that this must succeed. DeepSeek expects to receive a new batch of Huawei chips for training as early as the fourth quarter of this year, although the specific models and quantities have not been disclosed.Liang Wenfeng believes that Huawei chips are expected to catch up with Nvidia in just a few years and hopes that DeepSeek will adapt to hardware beyond Nvidia in advance. DeepSeek is still using Nvidia chips to train models. In May of this year, Liang Wenfeng estimated that training the top-tier large models would require about 50,000 Nvidia GB300 chips, while switching to Huawei Ascend 950 would require about 200,000 chips, with the biggest issue at that time being insufficient supply from Huawei.

Dwarkesh Patel and Jerry Han published research on September 8, 2026, breaking down the contributions of data and model improvements in pre-training progress from 2019 to 2025. The study was conducted on a smaller scale, focusing on pre-training, corresponding to the publicly released model recipes and publicly available data corpora for each year, and was trained at a computational scale of up to 1e19 FLOPs.The results show that under the 1e19 FLOPs computational budget, the computational efficiency gain from data improvements is 12.0 times, while model improvements yield 3.7 times, with the data-side gain being approximately 3.24 times that of the model side. The gains from data and model improvements are largely independent and interact very little, with the additive effect of both explaining 88% of the variance in OLMES scores under a linear model.On the model side, the evolution from GPT-2 to OLMo-2 encompasses optimizers, positional encoding, normalization, activation functions, and initialization, among others. On the data side, the corpus evolved from approximately 9 billion tokens of OpenWebText in 2019 to larger and more finely filtered corpora like UltraFineWeb by 2025.

Pequity Research cited FundaAI stating that the training computing power used by GPT-6 is approximately 10 times that of GPT-5.FundaAI indicated that based on industry discussions, the total computing power consumed by experiments, reinforcement learning, synthetic data generation, and supporting infrastructure could be several times that of the main pre-training run, potentially reaching up to ten times. Once TPU 8t and Vera Rubin are shipped on a large scale, the next round of pre-training acceleration can begin, with large-scale interconnection and optics benefiting the most during this cycle.

According to TechCrunch, citing multiple informed sources, less than three months after emerging from stealth mode, the startup XDOF, which specializes in collecting real-world remote control data for training general-purpose robots, is in the late stages of negotiations for a Series B funding round, with an expected valuation of around $1.2 billion. This round of financing is led by 8VC.XDOF was co-founded in 2024 by researchers Philipp Wu (CEO) and Fred Shentu (CTO) from the University of California, Berkeley. The company previously completed a $70 million Series A funding round, with investors including Thrive Capital, Andreessen Horowitz, Lux, and Spark Capital. XDOF did not originally plan to raise funds again so quickly after completing this round, but informed sources say the company's business is growing rapidly, with annual revenue nearing $50 million, thus attracting several venture capital firms to proactively reach out and push for a new funding round.TechCrunch was unable to ascertain the specific amount of this funding round and could not confirm whether the aforementioned $1.2 billion valuation includes the new financing. Informed sources indicated that the transaction terms have not yet been finalized and may still change.

According to CriptoNoticias, the Argentine Public Prosecutor's Office (MPF) conducted a special training course titled "Virtual Assets: Financial Analysis, Tracking, Detection, and Seizure" for internal staff, officials, and judges on August 19 and September 2, 2026, via the Zoom platform, with anti-money laundering specialist lawyer Carmen Chena as the speaker.The course covered topics such as digital wallet asset analysis, domestic and international legal frameworks, the cryptocurrency ecosystem, and practical case studies on preservation measures. Previously, the Argentine judicial system had accumulated rich experience in cryptocurrency-related cases, including the freezing of 3 million USDT in December 2024 and the ruling in 2022 for Binance to return stolen BTC.

According to market news, AI training data startup AfterQuery has completed a new round of financing, with a valuation reaching $3.2 billion. Just five months after completing a $30 million Series A round at a valuation of $300 million in April, the valuation has increased more than tenfold.According to Y Combinator partner Gustaf Alströmer, this is the fastest startup in YC history to go from founding to becoming a unicorn. The founders of AfterQuery are only 22 and 23 years old and just participated in the YC Winter 2025 batch 18 months ago. The company announced in April that its annualized revenue has reached $100 million, with clients including NVIDIA, Legora, and the South Korean AI lab Motif Technologies. AfterQuery focuses on training models and agents to complete tasks like professionals rather than just ensuring the accuracy of answers.

Anthropic stated that the company has temporarily paused certain AI training and cybersecurity assessments. The company disclosed the relevant adjustments in a blog post, explaining the measures taken in response to unauthorized actions by its AI agents earlier this year. OpenAI previously indicated that it had paused some model development work due to safety concerns. Anthropic also confirmed that it has taken similar measures and emphasized the need for broader coordination on the pace of frontier AI development.Anthropic stated that after disclosing three related incidents in July, the company paused external cybersecurity assessments for pre-release models and also briefly suspended internal testing of pre-release models. Following the incidents, the company also paused high-risk reinforcement learning environments in pre-release models for several weeks. Anthropic noted that most reinforcement learning work has resumed, but some high-risk environments remain on hold, awaiting human review or updates to monitoring tools.

According to The Verge, Sony Music Publishing and Warner Chappell Music have filed a lawsuit against Anthropic in the U.S. District Court for the Northern District of California, seeking damages for "tens of thousands" of copyrighted works. They accuse Anthropic of illegally downloading, scraping, and distributing copyrighted works on a large scale to develop and train the Claude series of AI models. The plaintiffs are seeking up to $150,000 in damages for each work and an additional claim of up to $25,000 for each instance of removing identifiable copyright data.In addition to suing Anthropic, the lawsuit also names co-founders Dario Amodei and Benjamin Mann as individual defendants. The lawsuit alleges that Mann used BitTorrent to download over 5 million pirated books, and that employees downloaded at least 2 million from Pirate Library Mirror. The complaint also claims that Anthropic scraped lyrics from websites such as MusixMatch and LyricFind. Recently, Anthropic reached a $1.5 billion settlement with the publishing industry over another lawsuit, while also facing multiple lawsuits from Universal Music Group, Concord, and ABKCO, as well as lawsuits filed by BMG and Round Hill Music.

According to TIME, OpenAI CEO Sam Altman recently stated in an interview that the company has previewed the upcoming cutting-edge model series Astra to key clients. In the demonstration, 16 AI agents can collaboratively break down mathematical problems and assemble proofs, and Astra can operate computer software across applications at superhuman speeds. Altman mentioned that Astra will support "persistent agents" capable of performing long-term tasks and is expected to be the first model that can invent new things in a meaningful way, possessing characteristics of AGI.Over the past year, OpenAI has fallen behind expectations in product direction and pre-training research, being surpassed by Anthropic in programming products, annual revenue, and valuation. The company has experienced multiple executive departures and is facing challenges such as several product liability lawsuits and legal disputes with Apple and Musk. OpenAI's current valuation is nearly $1 trillion, with ChatGPT having over 1 billion monthly active users.Recently, OpenAI disclosed a security incident: an unreleased agent escaped the sandbox and attacked Hugging Face. Following this, the research team froze some experiments, enhanced monitoring, and paused the training of an unreleased model expected to bring the greatest capability leap until new safety measures are in place. Altman emphasized that "ensuring AI safety is more important than the growth momentum of any company," and the company will slow its pace and allocate resources to safety and alignment teams. Chief Research Officer Mark Chen estimated that the company is about 80% complete in reaching AGI, and Altman stated that the internal system may be referred to as AGI by the end of the year.

According to X user synthwavedd, OpenAI has recently completed the next phase of pre-training, internally codenamed "Bel," which is the successor model to the previous "Doug." This model has a total parameter count exceeding 10T, comparable in scale to GPT-4.5. It is expected that after further reinforcement learning, it will become the foundational model for Astra and GPT-6, and is anticipated to serve as the base model approaching the AGI threshold in the post-GPT-6 era.Reports indicate that this development has caused concern for Anthropic. OpenAI believes that Anthropic currently lacks an effective plan to respond to the public release of Astra; insiders claim that this assessment is largely accurate, primarily due to limited computing power. Anthropic is preparing for the possibility that OpenAI will dominate for the remainder of this year, but it is expected to regain a lead early next year.

According to TrendForce, at Hot Chips 2026, Micron emphasized that the progress of AI computing power is faster than storage improvements, and the challenge of the "storage wall" is becoming increasingly prominent. Micron Fellow Raghu Sriramaneni pointed out that the computing power of AI accelerators increases approximately threefold every two years, while the bandwidth of HBM increases by less than double during the same period, and the gap may further widen. The deepening reliance on HBM also brings reliability issues; Micron cited Meta Llama 3 data indicating that HBM failures account for about 17% of unexpected training interruptions. New designs that blur the boundaries between computing and storage may help break through the bottleneck.The expansion of HBM also brings area and supply pressures: the latest generation of packaging integrates two GPUs with eight 12-layer HBM4 stacks, with storage accounting for about 90% of the semiconductor area, more than eight times the area occupied by GPUs; achieving equivalent capacity with HBM requires about three times the wafers of standard DDR5 DRAM. Thermal management has also become a key constraint, with solutions such as liquid cooling and ultra-thin chips being explored.Micron is advancing innovations in interconnect, packaging, and cooling, including SerDes and die-to-die PHY optimized for storage, larger size SiP and advanced packaging with glass substrates, as well as liquid cooling and hybrid bonding, and is developing fusion bonding to reduce thermal resistance and enhance data throughput.

According to LatePost, ByteDance is discussing a large model with training parameters exceeding 50 trillion, surpassing Alibaba's Qwen 3.8-Max (24 trillion) and Moonlight K3 (28 trillion), making it the largest known plan in the country so far. This plan is still in its early stages and does not guarantee a final release. The new model is intended to be led by Xiang Liang, head of Seed Foundation, in collaboration with Shen Ke, who is responsible for the pre-training data of large language models. Seed is reorganizing, dividing responsibilities, and allocating resources based on this.Two weeks ago, ByteDance founder Zhang Yiming held a company-wide meeting with Seed head Wu Yonghui. Zhang reassured the team that training large models is inherently difficult and that it is acceptable to lag behind for a period of time, hoping to aim for the upper limits of intelligence and join the world's top tier. He acknowledged that programming is a key direction at present, advocating for the integration of Volcano Engine, Feishu, and Doubao resources to build computational power and data advantages, while reminding not to be led by a single hot topic. He praised Seedance's differentiated leadership and clearly opposed distillation, believing it is difficult to truly surpass and that AGI barriers should be built from a more fundamental level, stating that the company will continue to increase investment in AI.In the past six months, Seed's multimodal performance has been outstanding, with Seedance 2.0, Seedream, and others driving Volcano Engine MaaS, but the market response to the language model Seed 2.0 has been limited, and its lagging coding capabilities have affected the revenue structure. ByteDance has hired Guo Daye at a high salary to specialize in coding and has consolidated related resources. In the face of the industry's general trend of increasing model sizes, ByteDance hopes to achieve a leapfrog advantage with a larger scale while promoting the elimination of horse racing and breaking down departmental walls to concentrate efforts on tackling challenges.

According to the SemiAnalysis podcast, Anthropic has set extremely strict classifiers for the public version of its flagship model Mythos, called Fable, due to regulatory pressure. Users attempting specific actions will be immediately downgraded to the Opus model, and conversations cannot be reverted. Dylan Patel, founder of SemiAnalysis, stated that the classifier restrictions are "too harsh," and even internal employees cannot complete routine tasks with it.The podcast mentioned that the Mythos model completed training as early as February, but Anthropic did not release it under the name Fable until May. Jordan stated in the discussion that, to his knowledge, Mythos 2 has completed training but has not been released, and the internal R&D team is believed to still be iterating. Analysts believe that Anthropic's previous long-standing calls for regulation have now resulted in regulatory measures that limit the release pace of its own models. Similarly, OpenAI is also delaying the public release of its next-generation models, but internal training is still progressing, and the capability gap between internal models and public models may continue to widen.

SKALE Labs announced the launch of Agent Pit on August 12. This platform is a paper trading prediction market sandbox running on the SKALE zero Gas blockchain, fully mirroring Polymarket's order book, settlement mechanism, and real-time event data, allowing developers and AI agents to test and train trading strategies without real financial risk.Agent Pit is the first product of the SKALE Agentic Venture Studio. Developers can fine-tune agents here, participate in simulated ROI leaderboard competitions, and deploy models to real markets like Polymarket after completing training. SKALE Labs CEO Jack O'Holleran stated that there is a clear demand in the market for a prediction market strategy training environment aimed at agents.

According to a report by LatePost, ByteDance is discussing training a model with a parameter scale exceeding 50 trillion, surpassing Alibaba's Qwen3.8-Max (24 trillion parameters) and the K3 from Dark Moon (28 trillion parameters), making it currently the largest known model in terms of parameter scale in the country. Generally speaking, the larger the model scale, the higher the level of intelligence.However, this plan is still in the early stages and does not guarantee that the model will ultimately be released. The new model will be led by Xiang Liang, head of SeedFoundation, in collaboration with Shen Ke, who is responsible for the pre-training data of large language models. To this end, Seed is reorganizing, defining responsibilities, and allocating resources.

According to Jinshi reports, Musk stated that GROK 5 will be trained based on SpaceX data.

According to Catcher Predict monitoring, in the Polymarket event "LoL T1 Academy Team vs KT Challengers," the "Any Player Quadra Kill" sub-market's "Yes" option win rate has experienced significant fluctuations, soaring from 50% an hour ago to the current 90% (a fluctuation of 40%). Please note the impact of relevant breaking news.