AI/ML news summary: week 36
Here are the articles, guides, and news about AI; Week 36. I read tons of RSS feeds and blogs, so you won't have to scour the internet yourself for the latest AI news of this week.
The world of AI is moving fast, and this week’s updates show just how quickly things are changing. OpenAI's "Strawberry" model is readying itself for a big release. At the same time, META’s Llama models are leading the open-source race with 350 million downloads. This shows that there is a growing appetite for accessible AI. And then there's Nvidia which is flexing its muscles with a whopping quarterly $26.3 billion in data center revenue. This proves that the Nvidia stock is not all vaporware because they are the number one player in the AI boom. Companies are shifting from CPUs to GPUs for more efficient processing, and that’s giving Nvidia a blockbuster-like rise in the tech world. But it’s not all about LLMs and Generative AI. Many firms are still finding ways to use GPUs for things like data processing and other backend tasks.
The debate between open-weight and closed API LLMs is also heating up. META's choice to release Llama 3.1 weights for free is quite a bold move that could either expand their innovation or drain their resources. On the other side, OpenAI is betting big on their proprietary models like Strawberry. Both sides need to decide whether to keep their cards close or open the table for everyone. It is a poker game where the chips are billion-dollar investments, and everyone is trying to guess who is bluffing. And meanwhile, enterprises are slowly waking up to the potential of LLMs in internal workflows, but they are still learning to handle these tools without falling down. Maybe someone needs to write a "How to Train Your AI" guide, but with fewer dragons and more data charts.
These developments mean a lot for the future of AI. There is a push for more affordable and accessible models which could make AI as common as smartphones. This will lead to a world where AI is not just in your phone but running behind the scenes of almost every business process. But, like any good plot twist, there is a catch of course. More AI use, also means more room for error, misunderstanding, and, let's face it, some good old-fashioned chaos. AI companies are moving from theory to practice, and the road is bumpy. And while Nvidia is stacking its chips, the buying companies must remember that AI is more like a marathon than a sprint. Do not get caught in the dream. It’s not just about who gets there first but who keeps the pace without collapsing halfway.
So, it might be wise to keep some popcorn handy—it’s going to be quite the show.
Some juicy new details on OpenAI's mysterious "Strawberry" model (source) have emerged, along with news about another model called Orion in development. Strawberry was reportedly shown to U.S. national security folks this summer and is set for release this fall. It's designed to solve new math problems and handle programming tasks, showing strong improvements in complex language challenges.
OpenAI is using "test-time computation" to enhance its problem-solving skills and is also working on a smaller, faster, cheaper version for ChatGPT. Interestingly, [Strawberry is also being used to generate synthetic data for Orion] (source), which is lined up to be their next flagship LLM.
META's Llama models have reached an incredible 350 million downloads on Hugging Face (source), firmly leading the open-source AI race. Llama models are accessed via API through cloud giants like AWS and Azure, plus other partners like Databricks, Dell, Google Cloud, and NVIDIA. Meanwhile, [Baidu's PaddleHelix team launched HelixFold3] (source), an open-source replication of AlphaFold 3 designed for predicting biomolecular structures. HelixFold3 matches AlphaFold 3 in accuracy for proteins, nucleic acids, and more, which is quite impressive.
OpenAI Strawberry Model : Capability GPT-5
[Cohere improved its Command R models] (source), with the updated versions performing better in coding, math, reasoning, and latency. Command R now matches the older, larger Command R+ version, which is a solid upgrade.
META and Microsoft Unveil Llama 2 AI Models - Valasys Media
Qwen announced its new Qwen2-VL models and open-sourced Qwen2-VL-2B and Qwen2-VL-7B (source), which is exciting news for the community. Qwen2-VL can understand 20-minute videos, supports multiple languages, and integrates with everything from mobile phones to robots. These models are available under an Apache 2.0 license, and the [API for Qwen2-VL-72B is also live] (source).
Not All Who Wander… | Emerging Tech for a Changing Edu
