Inteligencia Artificial

Small AI models like gpt-5.6-luna and GLM 5.3 cut the cost of a complex query from nearly a dollar to just ten cents, according to an essay by Calvin French-Owen. The former Segment co-founder explains why consumer AI apps are only now becoming viable, and what kind of business work these fast, cheap models can take on.

gpt-5.6-luna solves complex research tasks for pennies and at around 100 tokens per second, as documented by Calvin French-Owen. That finding, combined with GLM 5.3 joining the cost-capability Pareto frontier, redefines which consumer ai apps are economically viable.

The University of Chicago expanded its ban on generative artificial intelligence to more courses, Forbes reported on August 24, 2026. The policy already applied to social science classes and reopens the debate over academic integrity in the age of chatbots.

Anthropic introduced Claude Science, a product built for scientific research, amid a wave of investment in AI applied to science. Few technical details are public so far, but the launch coincides with initiatives like the Genesis Mission in the United States.

Paul Graham revived the debate over what a young person should study in the AI era: building LLMs from scratch instead of founding a startup. Yann LeCun responded with an opposing view on the limits of current models.

A new study shows that AI agents tend to abandon their initial answer when the rest of the group agrees on a different one, a pattern reminiscent of human social conformity experiments. The finding has direct implications for the multi-agent systems that companies and labs already use in production.

A technical analysis published on the Level1Techs forums shows how attention backends, quantization, and sampler settings alter a local LLM’s output compared to the reference model. The key metric for measuring that difference is KL divergence.

DP Technology raised a $114 million Series C round for its AI platform applied to molecular science. The Chinese startup, creator of the open source DeePMD-kit software, also runs Bohrium, its cloud for training and running these models.

A commit co-signed by Copilot Autofix replaced the safe parsing of a Snowflake workflow with an interpolation vulnerable to command injection. Five days later, Wiz’s Red Agent, an autonomous AI-powered research tool, found and exploited the flaw to steal an internal Jira token.

AI labs are deliberately trimming factual knowledge from their models to gain pure reasoning power. GLM-5.2, Qwen3.5, and DeepSeek V4-Flash show the pattern with concrete numbers, and the trend points toward running a frontier model on a single consumer GPU.