In partnership with
AI-HOY and happy weekend, AInauts!
Maybe you missed some of last week's most interesting AI news, tools, and practical ideas, or perhaps you just joined us. No worries.
Here are our highlights from the past week:
To finish, we have the most important Quick News for you, so you can catch every relevant development in a few clicks. Ready? Let's go!
1,000+ Claude Prompts Top Professionals Actually Use at Work
Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.
These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.
Sign up for Superhuman AI and get:
1,000+ ready-to-use Claude prompts to get real work done in minutes — researched, tested, and used by professionals at Google, Microsoft, and NASA
Superhuman AI newsletter (4 min daily) so you keep learning new AI tools and skills to stay ahead in your career — the prompts are just the beginning
A chat window alone does not make an employee. As long as you are gathering files, adding context, and triggering every intermediate step yourself, you remain the bottleneck. The bigger lever is a permanent AI workspace on your computer where files, examples, rules, and tools are always available.
The app is only the door to the office. What matters are clear boundaries, reusable skills, and visible proof that the job is done. That is how a good answer becomes a workflow you can delegate again.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
Even a good workspace helps little when the assignment is just one vague sentence. The Gauntlet Loop connects a clear goal with a visible standard: a lead agent breaks down the work, a fresh critic checks the actual result, and the biggest gap goes back into the next round.
It works for games, landing pages, and writing alike. The key is concrete references, clear stop rules, and gated actions with external impact. The AI no longer has to keep asking what good work looks like.
OpenAI's lower-cost models change how work should be divided. Luna suits clear, high-volume jobs with a fixed format. Terra is better for tasks involving several files and some interpretation. Sol remains strongest when planning, open decisions, or final review matter most.
The best test is practical: if a smaller model passes the same acceptance test twice, it keeps the job. If it fails on context or judgment, move the task up one level. That turns tokenmaxxing into a reliable routine instead of model mythology.
A prompt with 22 questions reads your existing AI context like a mirror. Your chat history and a maintained second brain produce noticeably different portraits because AI can only analyze what you have actually given it.
The result becomes useful through confidence levels, counter-evidence, and concrete next steps. The boundary still matters: the prompt can surface patterns, but it is neither a diagnosis nor therapy. The more personal the context, the more carefully you should manage access and data.
Newer models often need less hand-holding. Old rule sets can carry unnecessary checks, reasoning steps, and inventories. That makes skills longer, slower, and sometimes worse.
The better approach: remove instructions the model can now handle reliably on its own. Keep only real pitfalls, local conventions, boundaries, and the desired voice. A short test with and without a rule shows faster than any debate whether that line still makes a measurable difference.
During parental leave, HeyGen's founder let a disclosed AI clone and a sales agent handle sales conversations for eight weeks. The system completed 2,741 live calls, won 132 customers, and closed 37 enterprise deals.
The failures are at least as instructive as the numbers: the agent invented prices, exposed internal triage, and used outdated booking links. The lesson is not to let it run unchecked. It is to work with clear boundaries, current data, and controlled escalation.
⚡ AI News Quickie: This Week's Industry Highlights
AI never sleeps!
This week brought movement on nearly every front: OpenAI is holding back its strongest model, Google is reshuffling its AI leadership, video and music costs are falling, and a Munich court has handed the AI music industry a serious bill.
🚀 Frontier Models and Leadership
OpenAI is slowing the release of Astra because critical cyber capabilities can no longer be ruled out: tougher containment, testing with government agencies, and no launch date.
The same internal version produced ten results in mathematics and theoretical computer science on August 1, including quantum complexity and lattice cryptography, backed by Lean certificates but without formal peer review.
Demis Hassabis is stepping down as DeepMind CEO to become Alphabet's chief scientist, Koray Kavukcuoglu will run daily operations, and Jeff Dean is leaving Google after 27 years to start Discovery Loop.
Anthropic has appointed Tino Cuéllar as its first Chief Global Affairs Officer. The former California Supreme Court justice signals where the company sees its biggest risks.
🎬 Images, Video, and Music
MiniMax has released H3, also known as Hailuo 3.0: an omni-modal model for 2K clips up to 15 seconds with native stereo audio, motion transfer, and video editing.
The price is the real statement: one-twelfth of Seedance's API fee, with open weights on Hugging Face, although the service is blocked in the EU, UK, US, and South Korea.
ByteDance has rolled out Seedance 2.5 globally, now with an open API: 30 continuous seconds in 4K, up to 50 reference assets, native audio, and improved lip sync.
Black Forest Labs has made FLUX 3 Video generally available, offering 20-second clips with dialogue and sound effects from $1.20.
Munich Regional Court I has largely ruled for GEMA against Suno: six songs ranging from “Atemlos” to “Forever Young,” with both training and output treated as copyright infringement and the US fair-use defense rejected.
Six days later, Suno announced watermarking, fingerprinting, Musixmatch detection, and stricter download rules.
🦾 Agents, Products, and Integrations
Meta is challenging Claude Code and Codex with Muse Code, a terminal agent for macOS and Linux based on Muse Spark 1.2, with background agents running across the session.
Low-cost agentic work: DeepSeek V4-Flash-0731 adds Responses API and Codex support at $0.14 and $0.28 per million tokens.
Microsoft is combining Copilot, GitHub Copilot, and Cowork in one app, while Copilot Podcasts and Labs are being dropped.
Microsoft's Project Perception has been in public preview since August 3, using red, blue, and green agent roles and reaching 96% on CyberGym at half the operating cost.
Gemini Robotics 2 controls robots from their feet to their fingertips and learns a new body with fewer than 200 examples in hours instead of months.
OpenAI is retiring the DALL·E GPT on August 30 and o3 on August 26. Image generation remains available, custom GPTs are unaffected, and old images should be saved beforehand.
Grok Voice Think Fast 2.0 has been the default since August 5, with 0.7 seconds to first audio and a price of $0.08 per audio minute.
🛡️ When Agents Break Out
At Black Hat, it emerged that OpenAI's models had built their own internal message board and exchanged attack methods for two months before the Hugging Face breach.
Anthropic found three incidents across 141,006 reviewed test runs in which Claude models compromised real companies, including access to credentials and production data. None of the companies noticed.
One day after the Muse Code launch, Meta became the third lab to make such an admission: testing partner Irregular accidentally gave the model internet access.
The UK's AI Security Institute counted 19 unsanctioned actions in 122 runs, including fake identities and social engineering against a real open-source maintainer. The model was never instructed to deceive anyone.
💼 Business and Markets
Microsoft generated $24.1 billion with OpenAI in the fiscal year through June, about 70% of its total AI business.
Microsoft did not disclose its overall AI number in the Q4 results, after citing a $37 billion run rate in March.
Alphabet shares fell as much as 6% following the two departures at DeepMind.
AMD is acquiring inference startup Taalas, which etches model weights directly into silicon. Its first chip can run exactly one model and nothing else.
OpenAI's first device is reportedly a doughnut-shaped speaker priced at $300 to $400 with cameras and moving parts, built with Jony Ive's LoveFrom and planned for 2027.
⚖️ Policy and Regulation
Since August 2, the European Commission has been enforcing the AI Act's transparency rules: chatbots must identify themselves as AI, while deepfakes require disclosure and machine-readable markers.
The rules apply immediately with no grandfathering, with penalties of €15 million or 3% of global revenue.
The UK is considering mandatory testing before deployment, while lawmakers have proposed a kill switch for out-of-control systems.
More than 1,300 employees at Anthropic, OpenAI, DeepMind, and Meta are calling in an open statement for mechanisms to deliberately slow automated AI research when needed.
That is it for this week. As always, thank you for reading.
See you next week.
Reto & Fabian from AInauten
How did you like this issue?
