Gate News message, April 23 — Google researchers, including He Kaiming and Xie Saining, published a paper introducing Vision Banana, a general-purpose vision understanding model created through lightweight instruction fine-tuning of the company's Nano Banana Pro (Gemini 3 Pro Image) image

Perplexity uses SFT followed by RL with Qwen3.5 models, leveraging a multi-hop QA dataset and rubric checks to boost search accuracy and efficiency, achieving best-in-class FRAMES performance. Abstract: Perplexity's post-training workflow for web-search agents combines supervised fine-tuning (SFT) to enforce instruction-following and language consistency with online reinforcement learning (RL) via the GRPO algorithm. The RL stage uses a proprietary multi-hop verifiable QA dataset and rubric-based conversational data to prevent SFT drift, with reward gating and within-group efficiency penalties. Evaluation shows Qwen3.5-397B-SFT-RL achieving top FRAMES performance, 57.3% accuracy with a single tool call and 73.9% with four calls at $0.02 per query, outperforming GPT-5.4 and Claude Sonnet 4.6 on these metrics. Pricing is API-based and excludes caching.

In 2025, Google removed 175.5 million violating ads in South Korea using AI, suspended 326k accounts, and faced a $50 million fine for privacy violations, highlighting a trend of increasing enforcement and AI's role in combating ad fraud.

Peken Global has agreed to pay $500,000 to the CFTC to resolve a civil lawsuit, because as an unregistered platform it illegally allowed U.S. residents to trade. The company did not admit the allegations, but agreed to avoid providing services to U.S. customers. The CFTC said that U.S. customers generated approximately $110 million in trading fees for it.

The Federal Court of Australia ruled that a certain CEX must pay a fine of 10 million AUD for mistakenly classifying 524 retail investors as wholesale clients, resulting in a lack of protection for consumers and losses exceeding 12 million AUD. The exchange has rectified the situation and ceased its derivatives business.

Jump Trading has responded to a lawsuit filed by Terraform Labs, calling it an attempt to shift responsibility for evading SEC penalties, and demanding dismissal of the related allegations. The lawsuit accuses Jump of market manipulation and fraud, seeking damages of $4 billion, but Jump contends that the lawsuit lacks merit and is time-barred.

The Russian Ministry of Finance has proposed adding new provisions to the Code of Administrative Offenses to impose fines of 700,000 to 1 million rubles on cryptocurrency exchangers in violation of regulations, which has been approved by the government. Licensed exchangers that violate transaction limits will face penalties, with the law expected to take effect in 2027.

Hugging Face CEO Clément Delangue commented on the application of Chinese open-source models in AI products, noting that open source fosters competition. He highlighted that China is becoming a key player in global AI technology and emphasized that the focus of competition lies in the speed and capability of adaptation and fine-tuning.

Gate News reports that on March 19, an EU Commission spokesperson for digital economy affairs, Thomas Rennier, confirmed that American social media platform X has paid a €120 million fine issued by the EU in December of last year before the deadline, and submitted a remediation plan regarding its "blue verification" feature. The spokesperson also stated that despite X paying the fine, the platform is still appealing the penalty.

Gate News report: On March 17, Tether's QVAC Fabric launched the world's first cross-platform LoRA fine-tuning framework for Microsoft BitNet (1-bit LLM), significantly lowering the VRAM and computational thresholds for large model training. The framework supports LoRA fine-tuning and inference acceleration on Intel, AMD, Apple Silicon M series, and mobile GPUs (including Adreno, Mali, and Apple Bionic).