OpenAI releases their open weight GPT model, the release of GPT-5 and our growing attachment to different architectures, obsequiousness of language models, Genie 3 and open world models for reinforcement learning, and Pytorch as a standard in ML research

Further reading:
https://openai.com/index/introducing-gpt-oss/
https://openai.com/gpt-5/
https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/
https://pytorch.org/blog/pytorch-2-8/

Exhausting the data on the Internet by 2026, the value or non-value of transcripts, training on other languages, and subsidizing additional dataset generation

ChatGPT's new study mode for students, Facebook's play to be the first with augmented reality LLMs, Cerebras is trying to compete with Claude Code, and are we really in an "alphago moment" for model architecture discovery?

Further reading:
https://openai.com/index/chatgpt-study-mode/
https://s21.q4cdn.com/399680738/files/doc_financials/2025/q2/META-Q2-2025-Earnings-Call-Transcript.pdf
https://www.cerebras.ai/blog/introducing-cerebras-code
https://arxiv.org/abs/2507.18074

Evaluating models on benchmarks, passing a model vibe check, formal reasoning to synthesize datasets, and what type of datasets researchers prefer

OpenAI introduces their new Agent with full computer use capabilities, Anthropic is now evaluating models with other models, what the data leak at Tea can and can't tell us about vibe coding security, and more reasoning time isn't always better

Further reading:
https://openai.com/index/introducing-chatgpt-agent/
https://alignment.anthropic.com/2025/automated-auditing/
https://www.nytimes.com/2025/07/26/us/tea-safety-dating-app-hack.html
https://safety-research.github.io/inverse-scaling-ttc/

Allocating enough time to find your passions, Bell Labs vs Frontier Labs, project based classes, and getting involved in communities

Perplexity releases their AI browser Comet, Grok 4 and xAI's safety alignment strategy, the David vs Goliath of language models, and the Muon optimizer challenging Adam for its title

Anthropic & OpenAI's new approach for announcing models, double vs single blind peer review, and distribution channels for researchers

Inception Lab's new diffusion architecture Mercury, Google's emissions up 51% since 2019, DeepFleet warehouse scheduling robots, and Microsoft's delayed AI chip