Subscribe
Share
Share
Embed
Daily AI news and research, distilled. UpNext AI breaks down the most important developments in artificial intelligence—from major industry moves to cutting-edge papers.
Welcome to the UpNext AI podcast. It's Thursday, October 22nd, 2026, and here's what matters in AI today.
Google Cloud and NVIDIA are preparing to make Vera Rubin-based cloud instances generally available next month, with an expansion to Europe planned. NVIDIA says the A5X instances combine Vera Rubin computing with Google Cloud networking and cooling for model training, inference, and large-scale reinforcement learning. For teams weighing where to run demanding AI workloads, the important date is next month: these instances have been announced, but general availability is still ahead.
Google Cloud plans to scale the infrastructure to as many as 960,000 Rubin GPUs across multiple sites. That is a plan, rather than capacity customers can use today. The companies are positioning the system for work that repeatedly generates experience, evaluates results, and updates a model. Ineffable Intelligence, founded by AlphaGo creator David Silver, was the first customer to access A5X and is using it for reinforcement-learning research.
NVIDIA also points to an early WPP demonstration: creative agents turned conversational requests into multistep workflows, with human approval before execution. In WPP’s initial tests, workloads that typically needed a multi-node cluster ran on a single A5X instance. That is a customer test, not a general performance guarantee. Together, the examples show why this rollout matters: cloud providers are packaging compute, networking, and controls around entire agent workflows, while customers still need to assess availability and performance for their own jobs.
That emphasis on dividing up complex work carries into biology. OpenAI says it worked with PerturbAI and Scripps Research to send thousands of agents powered by GPT-Rosalind through experimental data from millions of mouse brain cells. In 32 minutes, the agents examined how gene disruptions affected cells, consulted scientific literature, and assembled potential explanations into a reasoning atlas.
The team compared that distributed approach with a single agent using the same model and resources. At the largest workload, the distributed agents cited more than 700 references, while the single agent cited 11. In blinded comparisons, human experts preferred the distributed reports 75 percent of the time. One proposed lead links a gene associated with a neurodegenerative condition to changes in cholesterol production inside neurons.
These are research directions, not confirmed disease mechanisms or treatments. The useful advance is a way to examine more of a large experiment and give scientists specific hypotheses to test next. The reasoning atlas is available for researchers to explore. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast.
For builders training models, AWS has published a Qwen3 walkthrough on Trainium using torchtitan. It demonstrates three configurations on a single Trainium instance: one built around Hugging Face Transformers, one using torchtitan’s native Qwen3 implementation, and one adding custom kernels for a mixture-of-experts model. The same launch pattern and configuration structure carry across the examples.
That consistency matters if a team wants to move between published model weights and a more customized training setup. AWS shows how the Transformers route can load published Hugging Face weights and save a checkpoint in the same standard format. The kernel examples show where developers can replace particular operations without rebuilding the entire training workflow.
This is a reproducible engineering walkthrough, not a claim that any one configuration is faster or cheaper. AWS explicitly describes its custom-kernel example as an integration demonstration, rather than a measured performance improvement. For teams considering Trainium, the immediate value is a clearer path from a familiar PyTorch training workflow to hardware-specific customization.
A research note on choosing models: one benchmark score can hide differences that matter to a particular job. In a paper posted yesterday, the TIDE researchers tested a method that groups benchmark questions into a tree of capabilities based on model behavior. They applied it to five models across four benchmarks.
The resulting comparisons were task-specific. On microeconomic market reasoning, Claude Opus 4.6 answered 100 percent of the questions correctly, versus 83.3 percent for GPT-5.4. On contract law reasoning, GPT-5.4 answered 88.9 percent correctly, versus 77.8 percent for Claude Opus 4.6. A team selecting a model could miss that reversal in an overall score. The caveat is that this demonstration covers a small set of models and benchmarks, and the method uses an AI judge to propose question groupings. Treat the breakdown as a diagnostic aid, then test against your own work.
...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ...
A few more developments. At GTC Berlin, NVIDIA highlighted financial institutions building on open-weight models and accelerated computing. Revolut is scaling its PRAGMA foundation model in production, while Stripe reports faster fraud-data pipelines in offline tests. Those are distinct deployments, with results that depend on each organization’s workload.
The Financial Times reports that Amazon warned of broad exploitation of software flaws found with help from OpenAI and Google models. The Verge identifies two critical PHP vulnerabilities and reports disruption at sites including Wikipedia and DeviantArt. Software operators have an immediate reason to check security guidance; the reports do not imply that most websites were compromised.
TechCrunch reports that Recursive has released Flint, a free coding model designed to run on a laptop. The startup says its AI research system largely built the model. It is an early test of AI-assisted model development, rather than proof that the system can improve itself without limits.
And Walmart plans to let customers connect their own shopping agents next month. The Verge reports that purchases will require approval by default, with options to set a spending limit and review agent purchases separately. The controls put shoppers in charge of what an agent can buy.
Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers.
And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!