AI Daily — September 23, 2026
2026-09-23
openai.com
Introducing GPT-6 Sol and Luna
OpenAI has released GPT-6 Sol and Luna, two new models positioned as frontier-intelligence options with different capability-cost tradeoffs. Sol targets high-capability use cases while Luna is optimized for cost-efficiency, extending the GPT-6 family beyond the previously announced Astra variant. This continues OpenAI's strategy of tiered model offerings within a single generation.
openai.com
Better prompt caching for GPT-6
OpenAI has shipped improved prompt caching for GPT-6, featuring higher cache hit rates, explicit breakpoints developers can set in prompts, new diagnostic tooling, and finer-grained controls. These changes are designed to reduce both latency and API costs for workloads with repetitive or shared prompt prefixes. The explicit breakpoint mechanism in particular gives developers deterministic control over cache boundaries.
huggingface.co
Transformers now runs llama.cpp quants
Hugging Face's Transformers library now natively supports llama.cpp quantization formats, allowing users to load GGUF-quantized models directly without switching inference stacks. This closes a significant interoperability gap between the Python Transformers ecosystem and the large library of community-quantized models distributed in GGUF format. Engineers can now benchmark, fine-tune wrappers, and run inference on llama.cpp quants within a single unified pipeline.
openai.com
OpenAI outlines priorities and principles for third-party AI safety assessments
OpenAI published a framework document detailing how independent third-party evaluators should conduct rigorous safety assessments of frontier models, covering independence, security of model access, and scope of evaluation. The document signals intent to standardize external red-teaming and evaluation practices as models become more capable. This comes amid broader industry pressure for credible external auditing of frontier AI systems.
huggingface.co
How UK AISI and EvalEval Are Making Benchmark Results Reproducible
The UK AI Safety Institute and the EvalEval project are collaborating to improve reproducibility of LLM benchmark results, addressing the widespread problem of non-reproducible evaluation claims across the field. The effort involves standardizing evaluation configurations, tracking infrastructure, and validation protocols so that published benchmark numbers can be independently verified. Reproducibility in evals is increasingly critical as benchmark results drive high-stakes model selection decisions.
rss.arxiv.org
"As a Language Model...": Chat Template Switches LLM Self-Referential Voice
Researchers find that the presence or absence of a chat template acts as a binary switch controlling whether instruct-tuned LLMs (up to 9B parameters) produce disclaimer language like "I'm just an AI" versus experiential language like "I feel." Activation steering in 3 models reproduces this effect mechanistically, locating a direction in residual stream space that encodes the switch. This has direct implications for how self-report data from LLMs is interpreted in AI safety and alignment research.