Good day, AI enthusiasts!
Meta is putting open-weight AI back at the centre of its strategy with a model designed to run agentic workloads on everyday computers. Microsoft is pushing further into custom chips, while AI is also moving deeper into materials research and cybersecurity.
The interesting thread today is control: over models, infrastructure, costs and data — and how much of the AI stack businesses may eventually be able to run on their own terms.
In today’s digest:
Meta releases Muse Glimmer and returns to open-weight AI
Microsoft could unveil its Maia 300 AI chip in September
AI agents are searching for materials to keep chips cooler
Corma raises $60m to build AI for cyber defence
DeepMind faces delays, burnout and a growing coding gap
A I M O D E L S
Meta releases Muse Glimmer and returns to open-weight AI
News Digest: Meta has released Muse Glimmer, a smaller open-weight model designed to run agentic workloads locally on a Mac or PC with a single graphics card. The company also plans to release the weights for the more advanced Muse Spark 1.2. It is a renewed attempt to make openness a bigger part of Meta’s competitive position after Llama 4 lost momentum.
What changed
Muse Glimmer is designed to handle local agentic workloads on consumer hardware rather than relying entirely on cloud infrastructure.
Meta plans to release the weights for Muse Spark 1.2, extending the strategy beyond its smaller model.
Zuckerberg is also arguing for fewer US restrictions around training data, infrastructure and model distillation.
Why it matters: Running capable AI locally could give businesses more control over sensitive information while reducing dependence on cloud APIs and recurring inference costs. The bigger test is whether Meta can turn openness into a genuine product advantage rather than simply a different distribution strategy.
Source: Reuters
A I + I N F R A S T R U C T U R E
Microsoft could unveil its Maia 300 AI chip in September
Image source: New Digest AI
News Digest: Microsoft is preparing to unveil its next-generation Maia 300 AI accelerator this autumn, potentially as early as September. It is reportedly discussing TSMC manufacturing capacity for more than 300,000 chips in 2027 and wants major Azure customers, including Anthropic, to consider using Maia.
Key points:
Microsoft reportedly has longer-term ambitions for Maia production to exceed one million units.
Winning external Azure customers would turn Maia from an internal efficiency project into a commercial cloud product.
Why it matters: Microsoft still depends heavily on Nvidia for AI compute. Building a credible alternative gives it more control over cost, supply and the economics of Azure as AI workloads continue to grow.
Source: Reuters
A I + S C I E N C E
AI agents are searching for materials to keep chips cooler
Image source: News Digest AI
News Digest: Discovered Materials is using swarms of AI agents to search for materials that could make AI chips more thermally efficient. Its system combines Anthropic models for generating candidates with physics models that simulate their properties before anything reaches laboratory testing. The company says its agents can explore thousands of possibilities per day.
Key Points:
AI is being used to expand the number of potential materials researchers can investigate rather than replace physical testing.
Experimental validation and manufacturing feasibility remain the main bottlenecks once promising candidates have been identified.
Why it matters: Some of AI’s biggest constraints are increasingly physical: heat, electricity and hardware. Better materials could improve chip efficiency, while also showing where AI can accelerate scientific discovery without removing the need for real-world experimentation.
Source: TechCrunch
A I + C Y B E R S E C U R I T Y
Corma raises $60m to build AI for cyber defence
Image source: News Digest AI
News Digest: Corma has emerged from stealth with $60 million in seed funding led by Sequoia Capital. Rather than adapting a general-purpose model, the company is training AI specifically for defensive security work such as analysing logs, audits and large volumes of operational signals.
Key points:
Corma says its systems are already deployed within major organisations across finance, healthcare, energy and other high-risk industries.
The company claims customers have reduced threat-response times by 94%, although that figure comes from Corma itself.
Why it matters: Cybersecurity is becoming a natural market for specialised AI agents because attackers and defenders increasingly operate at machine speed. The harder question will be whether autonomous defensive systems become dependable enough for sensitive and critical infrastructure.
Source: Fortune
A I + L A B S
DeepMind faces delays, burnout and a growing coding gap
Image source: News Digest AI
News Digest: Fortune reports that recent changes inside Google DeepMind followed delayed model launches, researcher departures, long working hours and concern that the lab was too slow to prioritise coding capabilities. Gemini 3.5 Pro reportedly missed several intended launch dates while rivals continued pushing ahead in coding and agentic AI.
Key points:
Current and former employees described intense competition for researchers and unusually aggressive efforts to retain talent.
Some employees are concerned about DeepMind losing independence as operations become more closely connected to Google headquarters.
Why it matters: Google already has distribution, infrastructure and custom chips. This report points to a different constraint: execution. As model cycles get shorter, shipping the right capabilities reliably may become as important as the underlying research.
Source: Fortune
U S E T H I S T O D A Y
Decide which AI work actually needs the cloud
Meta’s local-model push raises a useful operational question: which of your existing AI workflows genuinely need a large cloud model, and which could eventually benefit from being handled closer to your own systems?
Prompt
Review the following AI-assisted workflows: [ADD WORKFLOWS].
For each workflow, assess:
sensitivity of the data involved;
required model capability;
frequency of use;
cost sensitivity;
latency requirements;
need for external tools or live information.
Classify each as Local, Cloud or Hybrid and explain the main trade-off in one sentence. Flag anything that needs a security, privacy or compliance review before changing how it is run.
Best for: founders, operations teams, IT teams and businesses handling internal or customer information.
A I T O O L S
🛠️ Top AI Tools & App Releases
A platform for building custom evaluations and private benchmarks for AI agents. Teams can run experiments in realistic environments, compare models and identify where agents struggle with products or workflows.
Worth watching for: AI product teams, developers and businesses building their own agents.
Portfolio Lab is designed to help users build and test AI-generated investment strategies. The company says strategies are tested on unseen data and live markets before deployment through a connected brokerage account.
Worth watching for: fintech teams and people following how AI agents are moving into regulated decision-making.
Paritok compresses the tools, files and conversation history sent to coding agents. The company says its local approach can cut token costs by up to 85% and enable sessions to run three times longer.
Worth watching for: developers using coding agents heavily enough for token consumption to become noticeable.
Genspark’s latest launch combines a thin physical voice recorder with its SecondBrain software. Recordings are turned into meeting notes and stored in the user’s SecondBrain workspace.
Worth watching for: founders, sales teams and people who want meeting capture without another bot joining the call.
Instead of talking to one assistant, AI Group Call creates a live voice conversation with six AI participants selected around a goal. They can respond to one another, while the resulting conversation is transcribed and summarised into key points and actions.
Worth watching for: brainstorming, idea stress-testing and exploring several perspectives quickly.
That's it for today!
See you soon,
News Digest AI team