The one thing
NVIDIA's new AI assistants can run on your own servers and small devices
Why it mattersCompanies can now run NVIDIA's AI agents on their own servers or laptops, keeping data in-house instead of renting someone else's.
Seventy: Friday, 16 October 2026
What happened. NVIDIA put out three Nemotron 3.5 models for AI agents: Super VL at 120B parameters, Bolt 74B-A9B for local use on 24GB, and Bolt E4B for 4GB devices. The first two are available now, and E4B follows on Oct. 23.
02Models
Your developers can run Meta's AI coding helper privately on Amazon's cloud
Why it mattersEngineering teams on AWS can run Meta's coding AI on their own cloud setup, keeping source code in-house without running servers.
What happened. Meta's Muse Spark Open, an open-weight AI built for writing code, can now be launched from Amazon SageMaker JumpStart. It handles inputs up to 1 million tokens and comes in two versions, a full-precision one and a compressed one.
NVIDIA is offering European researchers more training and computing power for science work
Why it mattersScience-based startups and small labs in Europe get a route to NVIDIA tools, training and compute without building that access alone.
What happened. NVIDIA is widening European scientists' access to its technology through learning resources, collaborations and computing power. The announcement came at the AI in Science Summit in Dublin, where the European Commission launched an AI in Science Enablers initiative.
04Agents
Government suppliers on AWS can now give AI assistants live web search
Why it mattersGovernment contractors on AWS can now give AI agents live web lookups inside , without moving those workloads elsewhere.
What happened. Amazon's Bedrock AgentCore now offers its managed web search tool for AI agents in Ohio, Oregon, Sydney and the GovCloud US-West region. Agents can pull current web information from a location nearer to the apps and users they serve, with lower delay.
05Models
OpenAI is giving university biology labs sponsored access to its science AI
Why it mattersLife sciences startups now compete for academic talent and partnerships with labs that get OpenAI's GPT-Rosalind through a sponsored plan.
What happened. OpenAI launched the Rosalind Academic Plan, which gives qualifying university research labs funded use of GPT-Rosalind plus supporting tools for life sciences work.
06Agents
AI assistants take different routes each run, so one-time testing won't prove they're safe
Why it mattersTesting an agent once no longer proves it is , so approval steps and ongoing controls must replace one-off sign-off.
What happened. An AWS Machine Learning Blog post says enterprise leaders want AI agents to do more but don't trust them to act correctly every time. It notes that agents choose their own tools and steps, so the same task can run differently twice and a passing test says little about tomorrow.
EU opens a pledging scheme to back AI use by European researchers
Why it mattersBrussels has opened RAISE, a pledging scheme for AI science enablers; details on terms are not yet public.
What happened. The European Commission launched RAISE (Resource for AI Science in Europe), a pledging initiative for AI in Science Enablers, at the European AI in Science Summit in Dublin on 15 October 2026.
08Research
Google's AI platform helped scientists find a possible new antibiotic candidate, still untested in labs
Why it mattersDrug discovery teams can run a multi-database target search in one agent session, but the lead is only predicted, not lab-confirmed.
What happened. Google DeepMind says biologists used its Antigravity agent platform to search millions of uncharacterized UniProt proteins with the ProtNLM model. They found one antimicrobial peptide candidate with predicted activity against respiratory pathogens, now in wet-lab validation.
That's the whole issue.
8 stories, about 2 min 57 sec. Nothing was left out to fill the time.
Words of the day
- SageMaker JumpStart
- An AWS catalogue that lets you deploy pre-trained models to an endpoint in your own account with a few steps.
- Grounding
- Tying an answer to specific source material so every claim can be traced, rather than letting the model answer from memory.
- AgentCore
- An AWS service for building, deploying and running AI agents, with managed tools such as web search.
“Most people overestimate risk and underestimate reward.”