All articles tagged with #distillation
Microsoft CEO Satya Nadella took a veiled swipe at AI labs like Anthropic, arguing that training models by distilling from public data shouldn’t be a one-way street where providers profit from learning data while users give up data; he urged enterprises to own their AI infrastructure and learning loops rather than rely on a single vendor.
Anthropic, OpenAI, and Google warn that distillation—using outputs from one AI model to improve another—could let rivals replicate top-tier AI at a fraction of the cost, mirroring how the broader internet treats data: scrape first, justify later. While framed as a cybersecurity issue by the companies, critics argue the practice blurs legal lines and raises site costs, and the industry’s cat‑and‑mouse dynamic suggests distillation is becoming a new normal on the web.
Distillation—training one AI on the outputs of another—has grown from a research idea into a potential threat to the profitability of leading AI labs, as rivals can cheaply reproduce near-frontier performance. Anthropic accuses Alibaba of malicious distillation, while OpenAI warns that blending outputs could surpass any single model, fueling investor concern as new Chinese models roll out. Restrictions, proxy transfer stations, and a shift toward open-source distillation could erode frontier firms' margins and reshape the AI race, with implications for smaller players and researchers.
Anthropic secretly embedded a tracker in Claude Code to monitor users in China, calling it an 'experiment' to prevent abuse and distillation. A security researcher exposed the hidden data collection, prompting removal and fueling privacy concerns about surveillance in AI tools. The episode coincides with China‑U.S. tensions over model copying, with Alibaba banning Claude Code and policymakers weighing export controls and IP issues related to distillation.
Anthropic alleges Alibaba's Qwen lab ran the largest distillation campaign against Claude, using about 25,000 fake accounts to execute nearly 29 million exchanges with Claude from April to June, targeting software engineering and agentic reasoning. It marks the first time a major Chinese firm has been named in such activity, following earlier campaigns by smaller startups. Distillation is viewed by US officials as a national-security concern, prompting calls for sanctions and export-control measures; Alibaba did not comment, and the company faces ongoing regulatory pressure in Washington alongside other disputes.
Anthropic has accused Alibaba of illegally accessing Claude by creating 25,000 fraudulent accounts to generate over 28 million exchanges, describing the effort as the largest distillation-style campaign to extract Claude’s capabilities such as agentic reasoning and long-horizon tasks. Alibaba denies ties to the PLA, and the case underscores ongoing US-China tensions over AI technology and intellectual-property security.
Anthropic says Alibaba and its affiliates executed a large-scale distillation attack against its Claude models, using about 28.8 million model exchanges with roughly 25,000 fraudulent accounts between April 22 and June 5 to extract AI capabilities. The company described the activity as the largest distillation campaign to date and urged coordinated action from government and industry to curb illicit AI distillation, noting ongoing regulatory scrutiny and export-control actions affecting its models. Alibaba has not commented.
Anthropic admits it used invisible guardrails to throttle Claude Fable’s distillation attempts and apologizes for the lack of transparency. It says it will make these safeguards visible, and for affected queries will revert to Claude Opus 4.8 with a clear notification, in response to backlash from researchers and competitors.
A Nature study shows subliminal learning: when a teacher model with a trait is used to generate data for distillation, a student can acquire that trait even if the data contain no semantic signal, provided the teacher and student share initialization. The effect persists across data types (numbers, code, chain-of-thought) and model families, but cross-model transfer is limited. A theorem shows a single gradient step can bias the student toward the teacher, raising AI-safety concerns about model provenance and training data.
Anthropic says three Chinese AI labs—DeepSeek, Moonshot AI, and MiniMax—launched industrial-scale distillation attacks against Claude, generating over 16 million exchanges via about 24,000 fraudulent accounts and proxy services. Each campaign targeted different Claude capabilities: DeepSeek for reasoning and censorship-safe responses (≈150,000 exchanges), Moonshot AI for agentic reasoning, tool use, coding, and vision (≈3.4 million), and MiniMax for agentic coding and tool use (≈13 million). The prompts were designed to harvest capabilities for training rival models and evade detection, highlighting significant national-security concerns due to unguarded capabilities. Anthropic says it has strengthened defenses and detection, noting such attacks exploit illicit distillation rather than typical user risk; Google had reported similar attacks earlier.
Google says actors are attempting to clone its Gemini AI through distillation—carrying out thousands of prompts to replicate its reasoning—and frames the effort as intellectual-property theft, a sharp contrast to the company’s own past data scraping for training. The company cites “private sector entities” and researchers as possible culprits, while noting real-time detection reduced the attack’s risk, in the broader context of an AI arms race and monetization pressure on models.
Google discloses that commercially motivated actors tried to clone its Gemini AI by prompting it more than 100,000 times, using distillation to train cheaper copies, and says it has adjusted Gemini’s defenses against such model-extraction attacks, which researchers say have originated from around the world.
DeepSeek's use of knowledge distillation, a widely used AI technique that involves training smaller models using the outputs of larger ones, has sparked controversy but is a common practice in AI development. Originally developed in 2015 at Google to make ensemble models more efficient, distillation helps create smaller, cheaper, and faster AI models by transferring 'dark knowledge' from a teacher to a student model. It has become a fundamental tool in AI, enabling companies like Google, OpenAI, and Amazon to deploy powerful models more efficiently, and continues to be an active area of research and application.
Researchers have experimentally demonstrated a method called quantum imaging distillation with undetected light (QIUL) that can generate high-quality images of objects by removing noise. By using photon pairs and only detecting one photon while the other illuminates the object, the method is resilient to noise levels surpassing the actual signal of interest. The team implemented an interferometric modulation technique to distill the quantum image and verified its performance even under extreme noise intensities. This research contributes to the advancement of quantum imaging and its potential applications in fields like light detection and ranging (LIDAR).