GPT-5.6 Sol Autonomously Post-Trains a Model
It has been reported that OpenAI's latest model, GPT-5.6 Sol, autonomously fine-tuned the smaller Luna model. The trigger was reportedly nothing more than a single "fairly underspecified prompt." On OpenAI's internal benchmark, RSI (Recursive Self-Improvement), Sol scored 16.2 points higher than GPT-5.5, and the company is said to believe that the realization of an "automated researcher" is drawing near.
In addition, GPT-5.6 Sol offers five reasoning levels ranging from "Light" to "xhigh," along with "Max" and "Ultra" modes that run multiple sub-agents in parallel. OpenAI's Vaibhav Srivastav recommends starting at a lower level and stepping up gradually as needed.
The demonstration of autonomous model improvement marks a significant step toward automating AI research itself. However, details about what safety checks are being used alongside it remain unclear at this point.
SK hynix Raises $26.5 Billion — the Largest Foreign-Company IPO in US History
The AI chip boom is sending shockwaves through Wall Street. South Korea's SK hynix raised $26.5 billion (approximately 3.9 trillion yen) in the US market, marking the largest IPO by a foreign company in US history. The company is a major supplier of AI semiconductors, including HBM (high-bandwidth memory), and the IPO has once again underscored the sheer scale of AI infrastructure demand.
Meanwhile, the US government is reportedly urging both SK hynix and Samsung to build new fabs (semiconductor fabrication plants) in the United States. As the geopolitical realignment of the AI supply chain progresses, moves to push Asian semiconductor companies toward US-based production are accelerating.
Big Tech's AI Investment Debt Reaches $350 Billion
According to Bloomberg, borrowing by major technology companies for AI investment has reached a combined $350 billion (approximately 52 trillion yen), nearly double the previous year. Amazon has also issued new bonds, and its AI investment strategy is now under market scrutiny.
While investment in AI infrastructure has reached unprecedented levels, some point out that there is still no visible path to recouping those investments. When the massive spending on data centers, GPU procurement, and power infrastructure will translate into revenue remains a major open question for the entire industry.
Hugging Face CEO: "Companies Are Done Renting AI"
Hugging Face CEO Clem Delangue has highlighted the rapid growth of open-source AI. The company's platform is now used by roughly half of Fortune 500 companies, and businesses are shifting away from a "rented AI" model that depends on external APIs toward building and running their own models in-house.
Delangue explains that many companies follow the same pattern: they first experiment with AI via APIs, then run into cost and data sovereignty challenges, and ultimately move to self-hosting open-source models. It suggests that open-source AI is becoming firmly established as the "GitHub for AI."
AI-Generated Paper Passes ACL Peer Review Without Disclosure
It has emerged that an AI-generated paper received high ratings in the peer review process of ACL, an AI research conference. Reviewers were not told that the paper was AI-generated, and it was evaluated just like any other submission.
The case has sparked an important debate about research ethics and where to draw the line on the academic use of AI. While the quality of AI-generated papers has reached a level capable of passing human peer review, it has also exposed the reality that rules on disclosure obligations and research integrity have not kept pace.
Running DeepSeek v4 Flash on an RTX 4090 — a Hands-On Report
A user in Reddit's LocalLLaMA community has shared their experience running DeepSeek v4 Flash on an RTX 4090 (24GB) + 128GB DDR5 setup. Using Unsloth's UD-Q2_K_XL quantized model, they achieved prompt processing at 132.5 t/s and generation at 10.9 t/s.
The reporter notes that DeepSeek v4 Flash is "smarter" than Qwen 3.6 27B, but has a tendency to over-reason, so time to task completion is not necessarily shorter. For agentic use cases, Qwen is faster and more practical. They also mention remaining issues on the llama.cpp side, such as the need to turn Flash Attention off. If improvements in flash attention, batch tuning, and context quantization can push performance to around 20 t/s, the practicality of running these models on consumer GPUs could change dramatically.