Good morning, AI enthusiasts!
Mistral’s latest weights are ready to download, Gemini is turning chat into a document workspace, and new research puts AI security in focus. Here’s your briefing from the past 24 hours, with practical takeaways for your work.
In today’s edition:
THE QUICK SCOOP
MISTRAL
Summary: Mistral has released downloadable weights for Large 4, nicknamed Le Chonk, following its public preview. The multimodal model gives organizations another option for running and customizing advanced AI on their own infrastructure.
Quick Scoop:
Mistral reports improvements in coding, defensive cybersecurity and token efficiency.
Low, Medium and High reasoning modes offer different tradeoffs.
The custom open-weight license requires a separate commercial license for Model-as-a-Service businesses.
Your AI Edge: Compare the model on your own coding tasks before choosing a deployment. Factor in infrastructure needs, review time and license terms alongside benchmark scores.
Summary: Google is bringing file creation, previewing and editing into Gemini conversations. You can revise documents yourself, select a passage for changes, or ask Gemini to refine the file without leaving the chat.
Quick Scoop:
Supported formats include documents, PDFs, presentations, spreadsheets, Markdown, text and code.
PDF preview and editing work on web and mobile; other formats are web-only.
Google says the capabilities are available to all users in supported Gemini languages.
Your AI Edge: Try turning rough notes into a study guide or business document, then review the exported file for layout, factual accuracy and missing details.
SPONSORED BY MAMMOTHCLUB
Whether you need an explanation, more practice, or help deciding what to study next, you have somewhere to start.
🎯 Work on your weak spots
Focus your practice on the concepts you’re still working to understand.
📝 Put your knowledge to the test
Generate a gradable practice quiz based on your weak areas.
📅 Give your learning a plan
Ask Pete to create a study plan for the week or recommend what to learn next.
💼 Practice for what comes next
Explore career paths or try a mock interview, one question at a time.
✨ See how far you’ve come
Recap what you’ve learned and get a little encouragement based on your progress.
AI SECURITY RESEARCH
Summary: Epoch AI reports that GLM 5.3 compromised six recent Firefox releases in controlled experiments by using information about fixes that had not yet reached users. The research highlights how quickly public security changes can become useful to attackers.
Quick Scoop:
The model worked autonomously in an offline research environment.
Marginal cost was roughly $1,000 per targeted version; total research spending exceeded $100,000.
Mozilla patched a separately discovered vulnerability used in the experiments; the newest release was not confirmed vulnerable.
Your AI Edge: Keep browsers updated and track patch deployment across your team. For software maintainers, shortening the interval between publishing a security fix and shipping it deserves attention.
OPENAI
Summary: OpenAI updated its public description of intended model behavior. The revision emphasizes understanding human intent, exercising judgment in unfamiliar situations and remaining subject to human direction and oversight.
Quick Scoop:
The five values cover human intent, harm minimization, honesty, human agency and care.
New and revised guidance addresses context, memory, oversight and side effects.
Most detailed style guidance is retired, with more room for preferences and context.
Your AI Edge: Treat the Spec as a statement of intended behavior, not proof that every response follows it. Keep explicit task boundaries and review consequential actions in your own workflows.
AI PRODUCTIVITY
Summary: Granola is expanding beyond meeting notes with voice typing. Speak naturally and it turns your words into formatted text wherever you are writing, from email replies to longer AI prompts.
Quick Scoop:
Voice typing is available now on Mac.
Granola says unlimited words are free for its users.
Windows, iOS and Android support are listed as coming soon.
Your AI Edge: Dictate a familiar email or project brief and compare the result with typing. Review names, numbers and wording before sending, especially when automatic cleanup changes phrasing.
SPONSORED BY MAMMOTHCLUB
Explore practical ways to support your team with MammothClub’s ChatGPT for Business Owners course.
Use ChatGPT to develop and refine ideas.
Explore employee questions and onboarding workflows.
Practice clearer communication and collaboration.
AI BUSINESS
Summary: Amazon reported AWS revenue of $46 billion, up 39.3%, as its cloud business expanded at its fastest pace in nearly five years. The figures offer another view of the infrastructure market supporting AI adoption.
Quick Scoop:
AWS reached a $184 billion annualized revenue run rate.
AWS operating income rose to $17.8 billion.
Amazon’s overall revenue reached $201.9 billion, up 12%.
Your AI Edge: Cloud revenue is broader than AI revenue. Use the results as context for infrastructure demand, while judging your own AI spending by workload performance and measurable business outcomes.
AI INFRASTRUCTURE
Summary: AWS published a guide to quota management for SageMaker Training jobs through AWS Batch. The scheduling capability lets teams define GPU allocations and put otherwise-idle capacity to work across projects.
Quick Scoop:
Administrators can reserve capacity, lend it, or enable lending and borrowing.
Capacity can be reclaimed when the team that owns the allocation needs it.
Priority and preemption controls help determine which workloads run first.
Your AI Edge: Measure idle GPU time before expanding capacity. If teams share a pool, test job checkpointing and restart behavior before enabling preemption for important training runs.
THE MAMMOTH LOWDOWN
⚙️ TOOLS FOR AI EDGE
Five tools for better data preparation, versioning and AI evaluation.
Polars - Prepare and transform structured data with a fast DataFrame library.
DVC - Version datasets and track data-science workflows alongside Git.
lakeFS - Manage data versions, provenance and isolated development environments.
Great Expectations - Validate data quality with explicit, reusable expectations.
Evidently - Evaluate and monitor LLM applications, agents and machine-learning models.
What Else Is Buzzing In AI
The announced addition to Wayfinder Frontier AI Services will pair human experts with the model for vulnerability validation and prioritized remediation. Availability is still upcoming.
AWS’s latest release notes introduce named permission sets for project members, giving administrators finer control than the previous Owner and Contributor choices.
Its new analysis estimates that the computing power of AI chips sold each quarter has grown about 2.5 times per year since 2024, doubling roughly every nine months.