This week in AIElena Daehnhardt |
|
AI Weekly Recap: GPT-5, DeepPolisher, GSPO, and Persona Vectors
This week was wild in AI land! Everyone decided to drop their biggest releases at once. Here’s what matters:
Major AI Model and Tool Releases This Week
OpenAI launched GPT-5 (official announcement), their fastest/most innovative model yet, and they shocked everyone by returning to open source with gpt-oss-120b and gpt-oss-20b. The 120b runs on high-end laptops; the 20b runs on your phone. Wild.
Google brought the heat with DeepPolisher (fixing DNA sequencing errors) and Genie 3 (creating interactive virtual worlds from text prompts). Sci-fi is real now.
Alibaba dropped GSPO powering their Qwen3 models, plus free image generation with Qwen-Image. Anthropic introduced Persona Vectors (official research) to keep AI personalities consistent.
Bottom line: AI just got more accessible, more powerful, and more integrated into everything. The future feels very close.
Key AI Developments This Week
1. OpenAI Releases GPT-5
GPT-5 is OpenAI’s flagship large language model that succeeds GPT-4 with significant architectural improvements and enhanced cognitive abilities. GPT-5 is OpenAI’s smartest, fastest model yet, designed for both general use and specialised tasks. The performance boost is significant across all benchmarks.
2. Google’s DeepPolisher Fixes Genome Errors
Google AI partnered with UC Santa Cruz to create DeepPolisher, a deep learning tool that corrects base-level errors in genome assemblies. DeepPolisher has already improved the Human Pangenome Reference and reduces assembly errors by 50%.
3. Alibaba’s GSPO Algorithm Powers Qwen3
Alibaba introduced Group Sequence Policy Optimisation (GSPO), a reinforcement learning algorithm that addresses stability issues during scaling. GSPO serves as the core technology behind their Qwen3 models, providing improved training dynamics compared to existing methods like GRPO.
4. OpenAI Releases gpt-oss-120b and gpt-oss-20b Open-Source Models
gpt-oss-120b and gpt-oss-20b are OpenAI’s open-weight language models, released under the Apache 2.0 license. This is OpenAI’s first open source release since GPT-2. The 120b model runs on high-end laptops, the 20b runs on phones. Game-changer for accessibility.
5. Google’s Genie 3 Creates Virtual Worlds
Genie 3 is Google DeepMind’s world model that generates interactive virtual environments from text prompts. Unlike traditional rendering, these are dynamic spaces you can navigate and interact with. Applications span gaming, training, and simulation.
6. Anthropic’s Persona Vectors Keep AI Consistent
Persona Vectors is a monitoring technique that Anthropic introduced to track and control personality shifts in large language models. Persona Vectors addresses the problem of AI assistants becoming inconsistent during conversations, ensuring more reliable and predictable interactions.
7. Qwen-Image: Free Text-to-Image Generation Model
Alibaba launched Qwen-Image, a free text-to-image model designed to compete with DALL-E and Midjourney. Qwen-Image is completely free to use and handles native text rendering, making it accessible for everyone.
8. AI Limitations: What Still Can’t Be Done
Despite rapid progress, AI still struggles with emotional intelligence, creativity, common sense reasoning, and ethical decision-making. It excels at narrow tasks but lacks human intuition and adaptability.
9. Codex CLI vs. Gemini CLI vs. Claude Code
Codex CLI, Gemini CLI, and Claude Code all launched in 2025 as terminal-based AI coding assistants. These tools enable natural language code generation and fixing, streamlining developer workflows significantly.
10. Gemma 3n Runs on Your Phone
Gemma 3n is Google’s mobile-optimized language model that brings powerful AI language processing to mobile devices. Gemma 3n is private, configurable, and high-performance, letting you carry advanced AI capabilities anywhere.
That’s the week in AI! The pace of innovation is relentless right now. Stay sharp out there.