|
|
Hi there, this is your daily ☕️ AIpresso.
|
|
|
|
In today's newsletter: 🎬 Google's Gemini makes 15-second AI videos for 30% less 🤖 Perplexity launches always-on AI work agents 🧠 Inkling 2 Small trained in just 68 hours 🚗 Nvidia releases free Cosmos 3 world-simulation models 📈 METR finds Anthropic's AI capability growth nearly doubled Plus: 🎁 9 more stories you might like, 🛠️ 5 engineering & practice picks, 🧰 7 tools, and 📚 5 papers.
|
|
|
FROM OUR PARTNER
No need to keep reaching for your phone. RAYOVE Watch is a hands-free AI translation smartwatch built for real conversations. Point the camera at menus, signs or printed text for OCR translation. Speak naturally and hear the translation from your wrist in under a second, with online support for 73 languages. RAYOVE Watch is now live on Kickstarter. Back RAYOVE Watch on Kickstarter |
|
|
🎬 Google's Gemini makes 15-second AI videos for 30% less
LINK
|
- Gemini Omni 1.2 Flash, Google's AI video model, is now generally available making 15-second clips, up from 10, at about $0.07 per second of 720p video through the Gemini API, a 30% cut.
- Users can now edit a clip over several chat turns, swapping products, backgrounds or lighting instead of regenerating it, while faces, packaging text and brand colours stay more consistent across camera angles.
- It rolls out across Flow, Google Vids, Google Ads and Google Cloud, with every video carrying an invisible SynthID watermark, though Google says fast sports action, quick script changes and complex physics still trip it up.
|
🤖 Perplexity launches always-on AI work agents
LINK
|
- Perplexity rolled out "Teammates," always-on AI agents that take on multiple roles inside a workspace and pull existing projects, skills, and memory from the Perplexity tools you already use, squaring off against OpenAI and Anthropic.
- Teammates build on top of Perplexity Computer, which deploys agents on your device, letting you assign each Teammate a different job to run simultaneously and track progress through individual threads or interact with them in Slack.
- Positioned as a rival to OpenAI's Dots and Claude Duet for ongoing, large-scale work, the feature requires Perplexity's $200/month Max plan, making it notably pricier than those competing offerings.
|
🧠 Inkling 2 Small trained in just 68 hours
LINK
|
- Thinking Machines has unveiled Inkling 2 Small, a mixture-of-experts base model with 594 billion total parameters, of which 17 billion are active, trained on a compact token budget far below comparable rivals.
- It scores 31 on the Artificial Analysis Intelligence Index, near Gemini 3.8 Flash, with the biggest gains in agentic tasks like AutomationBench-AA, though its response rate fell as the model answers fewer questions.
- At $1.18 per task it undercuts Gemini 3.8 Flash and Claude Sonnet 5.5, and Thinking Machines plans an Apache 2.0 weights release pending external safety testing, though it doesn't support audio, overreacts on health topics, and mishandles large codebases.
|
🚗 Nvidia releases free Cosmos 3 world-simulation models
LINK
|
- Nvidia added four open-weight checkpoints to its Cosmos 3 family, releasing post-trained Cosmos 3 Nano models for interactive world simulation and multi-sensor, multi-view driving scene generation.
- Each model carries 16B trainable parameters; Transfer-Auto generates synchronized video across up to 11 camera views including fisheyes, with joint RGB-LiDAR output steered by per-camera scene maps and text prompts, while Sim-Auto streams successive chunks in a few denoising steps.
- The weights ship with post-training recipes so researchers can adapt Cosmos to their own sensor rigs and distill fast streaming variants under the OpenMDW1.1 license, which permits commercial and non-commercial use within its terms.
|
📈 METR finds Anthropic's AI capability growth nearly doubled
LINK
|
- METR, in an independent assessment published yesterday, judged that Anthropic is improving model capabilities on well-specified tasks at roughly twice its earlier rate, attributing most of the sustained acceleration to AI assisting its own researchers.
- Measured against Anthropic's AECI benchmark index, scores have grown 1.9 times faster since October 2025, about two years of prior-pace progress in one year, though robustness checks under alternative modeling choices produced a much wider range.
- The eleven-week review drew on internal benchmarks, training-compute metrics and roughly 15 staff interviews, but Anthropic unilaterally redacted a majority of the most informative evidence, so the attribution and bottleneck conclusions rest on METR's synthesis readers cannot independently inspect.
|
|
100+ ChatGPT Prompts to Revolutionize Your Day
Industry-Specific Use Cases: 15+ real-world applications across various sectors
Productivity Guide: 21 best practices to 10x your efficiency with AI
Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation
Challenge Buster: Overcome common AI hurdles with expert strategies
Plus, in-depth sections on email composition, content creation, customer support, and data analysis.
🛠️ Engineering & Practice
|
> Agents are going async: Run delegated agent work in background workers while one orchestrator stays reachable, so an agent handles many large tasks at once without stalling.
|
|
Other
news & articles you might like
-
NVIDIA Vera delivers 2x the agent speed of AMD’s 256-core EPYC Venice and 1.6x more agents per watt than the frequency-optimized 96-core Venice
LINK
-
Kubernetes, mark II? Why Google’s giving away gVisor, a code sandbox used by Anthropic and OpenAI
LINK
-
Dropbox ditched AWS S3 for its own in-house system. Now it’s buying Tigris to sell S3-compatible storage to AI developers.
LINK
-
Visa tested 22 AI models as shoppers, and they failed 42.8% of the time
LINK
-
AWS launches agentic AI Kiro bots with memory and support for recurring development jobs
LINK
-
Pangram’s new models spot AI fakes in images and video, it says
LINK
-
google-deepmind/alphaevolve
LINK
-
Web grounding for the Decisions API
LINK
-
Chinese AI lab Moonshot turned to Europe for NVIDIA Blackwell access amid U.S. China curbs, sources tell CNBC
LINK
|
|
📚 Trending papers & reports
|
> AI safety monitors that read an agent's reasoning to catch hidden misbehavior can be fooled when the agent disguises bad actions as normal work, but a new detection method flags these with ~91% accuracy.
LINK
|
|
> Long-context AI reading lets models handle up to 2 million words of input far faster and more accurately, scoring ~61 versus ~40 for the best rival at the million-word mark, without needing rare long training documents.
LINK
|
|
> Compressing large language models hurts bigger models far more than small ones, because small models can internally repair the damage from shrinking while large ones cannot, so cost-cutting tricks that seem safe at small scale backfire at scale.
LINK
|
|
> AI agent token waste can now be measured and trimmed, with a new framework cutting 41 to 46% of an agent's reasoning while keeping task results intact, turning bloated usage into a fixable cost signal.
LINK
|
|
> Verifier-based reinforcement learning adds almost nothing beyond simply training on expert examples, showing gains only on unfamiliar tasks, so the pricey reward-grading step may be skippable for most training needs.
LINK
|
|
|
🧰 Tools & repos
|
Glasshand: gives AI agents on-demand access to real iPhones and Macs in a data center, letting them post, book, and test iOS apps like a human would
LINK
|
|
Turnstone: connects your AI subscriptions, API keys, or self-hosted models so agents auto-switch when one hits its limit, free on Mac and iPhone.
LINK
|
|
Clocknext: billing and monetization platform for AI apps that tracks usage in real time, sets flexible pricing, and monitors costs, margins, and profitability.
LINK
|
|
TraceX: investigates customer reports across your database, code and logs, pinpoints root causes with evidence, and drafts a fix PR and reply.
LINK
|
|
Wonder Dev View: runs Claude Code and Codex in parallel, manages multiple subscriptions, and reviews agent work in one unified workspace for building software.
LINK
|
|
Bridge: a collaborative workspace combining meetings with customizable AI agents that join calls, use your tools, and handle follow-up tasks automatically.
LINK
|
|
You can check the previous tools here, or add your tool here
|
|
|
Our AI Academy has 330+ step-by-step tutorials on ChatGPT, Claude, Perplexity, and every tool that matters. No fluff — just practical workflows you can use at work. Try it free for 7 days.
|
|
💬 How did you find today's edition?
We read every reply — just reply to this email and let us know how we can improve!
|
|
