Altman: "A Generation That Underestimated Scaling Held Back AI"

In a talk at Stanford, OpenAI's Sam Altman sharply criticized skepticism around LLM scaling (expanding model size). "An entire generation of researchers held this field back by underestimating what scaling could deliver," he said, explicitly defending the scaling hypothesis.

As concrete evidence, he cited OpenAI's recent refutation of a mathematical conjecture, suggesting that scaling up LLMs leads not merely to performance gains but to new discoveries. With voices insisting that scaling has hit its limits still running strong, a leading figure in the industry has now drawn a clear line in the sand.

GLM-5.2 Beats GPT-5.4 and the Gemini Lineup on DeepSWE — But Cost Remains a Challenge

Reports that the open-weight model GLM-5.2 has overtaken GPT-5.4 and the entire Gemini lineup in coding ability on the DeepSWE software engineering benchmark are drawing attention on Reddit's LocalLLaMA community. A leaderboard analysis shared on X by Theo (t3.gg) showed a landmark result: an open-weight model taking the top spot in this domain.

What deserves attention, however, is cost. GLM-5.2 lags behind GPT-5.5 (medium) and Claude Opus 4.8 (high) in average cost per task. The culprit is its high output token count: even though its base per-token price looks cheap, the sheer volume of tokens consumed to complete a task sends total costs soaring. The results have once again raised the question of how to balance the strengths of open weights against real-world operating costs.

Boogu-Image-0.1 — A Competitive Open-Source Image Model on a Limited Compute Budget

Boogu-Image-0.1, which has surfaced on Hugging Face trending, is a family of image generation and editing models released under the Apache-2.0 license. With variants including Base, Turbo, and Edit, it covers text-to-image generation, fast generation, image editing, and Chinese-English text rendering.

According to the development team, although its training data is an order of magnitude smaller than that of closed systems (such as Nano Banana Pro and GPT-Image-2), systematic improvements in data quality and the training pipeline have delivered significant gains in image generation and editing performance. For evaluation, the team built its own "Boogu Arena" (an LM Arena-style preference evaluation system), and the release looks set to give the open-source image generation ecosystem a push forward.

Note that the model is released as a research project, and it is explicitly stated that no official commercial API or paid service exists. Beware of fraudulent services operating under similar names.

Anthropic to Bring Identity Verification to Claude

Anthropic has announced that it is introducing Identity Verification for Claude. Details are already posted on the support page, and rollout is expected to begin soon.

Identity verification for AI assistants appears to be aimed at preventing account abuse and enforcing age restrictions, but community members have also raised concerns about privacy and the higher barriers to access it may create. For details on specific verification methods (phone number, ID documents, biometrics, etc.), refer to official sources.

Brands' Use of AI Influencers Expands — The Guardian Reports

The Guardian has reported a growing number of cases where brands use AI-generated influencers to promote products on social media. The approach — AI characters that look like real people gaining followers and advertising products — is spreading through the marketing industry.

This trend carries transparency problems. There are concerns about consumers following or purchasing without realizing the content is AI-generated, making this a development worth watching from a regulatory perspective as well. While debates over mandatory disclosure of AI-generated content move forward in various countries, reality continues to run ahead of the rules.

---