Elena' s AI Blog

AI weekly news

12 Sep 2025 / 5 minutes to read

Elena Daehnhardt


Midjourney 7.0: AI, small but mighty


Introduction

Some weeks, AI news feels like a storm of buzzwords. This week, however, there’s a clearer thread: making things smaller, faster, and actually useful. From nimble models outrunning the giants, to Google teaching AI how to both sprint and think carefully, to new tools for science and medicine, the focus is on efficiency and real-world impact.

And to keep things interesting, OpenAI is stepping into the jobs market with its sheriff’s badge.

Top 5 AI Achievements This Week

1. Qwen-3-Next: Leaner, Faster, Smarter Than GPT-5 and Gemini 2.5 Pro

Source: Analytics Vidhya

A surprise arrival on Hugging Face: Qwen-3-Next, with “only” 80 billion parameters (a featherweight by today’s swollen standards), is outrunning giants like GPT-5 and Gemini 2.5 Pro. Imagine a wiry runner in trainers overtaking a field of athletes weighed down by their designer kit. Its secret? A 32,000-token context window and speeds over ten times faster than its predecessors.

Size isn’t everything in AI — this trend towards lean efficiency means more people can actually use advanced models without needing a supercomputer or a lottery win.

Read Analytics Vidhya

2. Speculative Cascades — A Hybrid Approach for Smarter, Faster LLM Inference

Source: Google Research

Google has a new trick: speculative cascades. Think of it as tag-teaming a speed reader with a meticulous scholar. Small, fast models answer the easy bits, while the heavyweight models step in when things get complicated.

The kicker? Speculative decoding predicts multiple tokens at once and checks them in parallel. It’s intellectual relay racing — quick, precise, and surprisingly elegant.

This is engineering that refuses the false choice of “fast or accurate”. Sometimes, yes, you can have both.

Read the research

3. Accelerating Scientific Discovery with AI-Powered Empirical Software

Source: Google Research

Scientists often have ideas faster than they can code. Google’s new AI system fixes that by automatically writing high-quality empirical software to test hypotheses.

Give it a problem statement and evaluation method, and it churns out implementations, runs thousands of variants, and reports results. Trials across genomics, neuroscience, and other fields show expert-level performance. Suddenly, the bottleneck isn’t coding but imagination.

If AI can write and optimise research tools on demand, researchers are freed to spend more time asking daring questions — the very heart of science.

Read the research

4. Smarter Nucleic Acid Design with NucleoBench and AdaBeam

Source: Google Research

Designing DNA and RNA sequences is like searching for a single book in a library larger than the universe. Google and Move37 Labs built NucleoBench, the first proper benchmark for nucleic acid design, and paired it with AdaBeam, an algorithm that outperformed rivals in 11 of 16 biological challenges. The aim? Faster gene therapies, sharper CRISPR edits, and better vaccines.

It’s molecular design with intelligence, not chance. From trial-and-error to tailored medicine — a shift that could touch all our lives.

Read the research

5. OpenAI Announces Jobs Platform and Certifications for AI-Powered Job Roles

Source: Analytics Vidhya

The AI job market has been chaotic — lots of hype, no clear standards. OpenAI wants to bring order with a new jobs platform and certification scheme. The idea: formal career pathways for AI engineers, prompt engineers, and data scientists, complete with credentials that (for once) might actually mean something. It’s a step toward professionalising an industry that has been running on improvisation and LinkedIn bravado.

Clear standards help everyone: learners know what to study, employers know what to expect, and the AI world looks a little less like the Wild West.

Read Analytics Vidhya

Closing Thoughts

This week’s theme is restraint — smaller, smarter, more efficient AI. Instead of endlessly adding parameters, researchers are squeezing brilliance from elegance: faster inference, clever cascades, molecular precision, and software that builds itself. And then, OpenAI, perhaps sensing the chaos it helped create, is trying to tidy up the careers it spawned.

Did you like this post? Please let me know if you have any comments or suggestions.

Posts about AI that might be interesting for you






desktop bg dark

About Elena

Elena, a PhD in Computer Science, simplifies AI concepts and helps you use machine learning.





Citation
Elena Daehnhardt. (2025) 'AI weekly news', daehnhardt.com, 12 September 2025. Available at: https://daehnhardt.com/blog/2025/09/12/ai-weekly-news/
All Posts