- AInauten.net
- Posts
- 👀 Weekly AI news: Did you miss it?!
👀 Weekly AI news: Did you miss it?!
👨🚀 The most important AI updates at a glance
Die AInauten
February 23, 2025
👀 Weekly AI news: Did you miss it?!
👨🚀 The most important AI updates at a glance
AIHOY and happy weekend, dear AInauts!
Maybe you didn't catch all the exciting news, tools and hacks about AI last week, or maybe you've just joined us. No need to worry!
We have summarized our highlights of the past week for you:
Finally, here's the most important quick news for you - so you can keep an eye on all relevant developments with just one click! Ready? Let's go!
The new reasoning AI models such as OpenAI o3 and DeepSeek R1 need to be prompted differently. In contrast to classic transformers, they think through complex tasks step by step, have a larger context memory (up to 200k tokens) and excel particularly in STEM areas. Minimalist, precise prompts without unnecessary examples or step-by-step instructions are important for optimal results.
Virtual romances are booming: millions of people are chatting and flirting with AI-controlled partners on platforms such as Replika or Character.AI. The typical user? 27 years old and 80% male.
Monetization is also running at full speed: freemium models, voice output for 1$ per minute and premium subscriptions for "spicier" features... The AI companions serve real needs for closeness and understanding - available 24/7, without drama. A perfect addition or a dangerous substitute for real relationships? The future will tell.
A creative example shows the fascinating possibilities of modern AI video models: Watermelons turn into cute robots!
And this is how you make your watermelon Robots🍉🤖
Images created with - midjourney
Upsacaled with @LeonardoAi_
Animated with @Kling_ai start frame and end frame.
is super fun to play with this technique 😜— Guillermo Castellanos (@guicastellanos1)
6:37 PM • Feb 15, 2025
Grok-3 from xAI enters the stage with a "big brain" feature that supposedly even puts OpenAI's o3 in the shade. With deep search and future voice features, it aims to challenge the big players.
At the same time, Nous Research is presenting its DeepHermes-3 update, a largely uncensored LLM with new reasoning capabilities. Without moral restrictions, it is particularly popular for role-playing games and AI relationship simulations ...
Krea AI's new real-time feature enables real-time image creation. With the Realtime Canvas, users can place objects and images on the left, enter prompts at the bottom and immediately see the result on the right.
Especially cool: create 3D models and manipulate them live! The example of a living room interior shows the magic - from an empty room to a perfectly styled Scandinavian interior by simply adding a sofa, plants and furniture. The tool offers a generous free quota and many other features.
A fascinating experiment shows how two AI agents work together: The OpenAI Operator controls a Replit Agent to develop an app completely autonomously. The two AI systems independently exchange access data and carry out tests. A foretaste of the automated future of software development!
this is wild 🤯🤯🤯
just paired openai operator with replit agent to build an app
watch as two ai agents team up, exchange credentials, and start testing
ai agent 🤝 ai agent
the future is here and it’s insane!
— Lamar (@LamarDealMaker)
3:06 PM • Feb 15, 2025
A clever checklist for AI-generated social media posts makes all the difference! The most important checkpoints: Do readers stay tuned? Do they become curious? Is the hook strong enough to stop them scrolling? Does the content appear authentic and trustworthy? Does it invite interaction?
Fun fact: The posts will be even better if you use the checklist as a separate optimization step after the first content generation. And prompt chaining is a great way to automate this!
📰 AI news quickie: HAI highlights from the industry
Slowing down? Certainly not ... These are the most important updates of the last few days.
Model & platform developments
OpenAI breaks 400 million user mark: ChatGPT records over 400 million weekly users in February - a rapid increase since December (300 million)! Of these, 2 million are paying users, a doubling since September 2024
Roadmap for GPT-4.5: Sam Altman has announced that GPT-4.5 (codenamed "Orion") will be released in the next few weeks. GPT-5 will follow in the coming months and standardize all OpenAI models and tools - finally an end to the annoying model selection!
Anthropic's next Claude is on the horizon: a new hybrid model is due to be released in a few weeks. It combines quick standard responses with a reasoning mode that can be controlled using a slider.
Meta organizes "LlamaCon": Meta has announced "LlamaCon", a developer conference on April 29 to strengthen Meta's open source AI community.
Adobe Firefly for video: Adobe has extended its generative AI tool Firefly to video, enabling the creation and editing of videos via a prompt.
Google Meet writes to-do lists: Google is equipping Meet with smart assistance: AI-generated transcripts of video meetings can now automatically create action items and to-do lists - so that lots of talk is followed by action.
AI agent "Proxy" for the web: The startup Convergence AI has launched the Proxy 1.0 agent, which can independently navigate websites, click, scroll and perform tasks online. An alternative to OpenAI's Operator?
Open source reasoning from Perplexity: Perplexity AI has released the reasoner "R1 1776" as open source. It is basically an open version of DeepSeek's R1 model, but without the previous censorship filters. Nice!
Hardware, infrastructure & investments
OpenAI's own AI chip: OpenAI is working on an in-house chip to be less dependent on Nvidia. Mass production is planned by 2026. This is an ambitious step that Microsoft and Meta are already battling to achieve.
Meta wants to buy AI chip: Meta is on the verge of acquiring the South Korean AI chip manufacturer FuriosaAI. The deal is intended to strengthen Meta's in-house AI hardware efforts.
South Korea's GPU offensive: South Korea plans to secure 10,000 high-end GPUs this year to build a national AI computing infrastructure.
Investors love open source clouds: Together AI has raised 305 million to develop a cloud platform for generative AI specializing in open source models - a clear commitment to open AI infrastructure.
Consumer & Entertainment
Apple packs AI into the iPhone: Apple has unveiled the iPhone 16e ($599). It complements the iPhone 16 series and offers the fast performance of the A18 chip, "Apple Intelligence" features, long battery life and a 48MP 2-in-1 camera.
Spotify relies on AI voices: Spotify is working with ElevenLabs to offer audiobooks in 29 languages via AI voice. Authors can have their works automatically dubbed into different languages - a huge step for the audiobook market (and a shock for some speakers).
Game levels from AI: Microsoft presents "Muse", a model that can automatically generate new video game scenes based on recorded Xbox gameplay data! In the long term, this could help game developers speed up level design and content creation and save costs.
AI as a nutrition coach: The popular tracking app MyFitnessPal introduces an AI nutrition planner - based on goals, preferences and previous logs. Shopping lists or delivery orders are available at the touch of a button.
Humane AI Pin failed: The much-hyped AI gadget startup Humane is dead after a disappointing response and uncertain utility value.
Regulation, data sovereignty & the military
AR headset for soldiers: Armaments startup Anduril is cooperating with Microsoft on the US Army's IVAS project. This involves equipping an AR headset with AI functions for soldiers to display tactical information in their field of vision.
Larry's data octopus: Oracle founder Larry Ellison is causing a stir with a proposal: He wants to bundle all US data into a single AI system. A centralized national data asset, so to speak - some see this as efficient administration, others as an Orwellian dystopia.
Data protection: EU option at OpenAI: OpenAI is responding to regulatory pressure and is now offering data residency in Europe. European API customers and ChatGPT Enterprise/Edu can choose to have their data processed exclusively on EU servers.
Research, breakthroughs & regulation
AI teacher for sign language: NVIDIA and the American Society for Deaf Children introduce "Signs". The app provides real-time feedback on sign language - a boon for ASL learners, made possible by computer vision and NVIDIA AI.
Folding proteins with AI: MIT biologists develop "FragFold", an AI system that can predict which protein fragments could bind to a target protein and block its function.
Deciphering the inner workings of the cell: A new AI analysis reveals hidden design rules in cells - e.g. why cell components have certain spatial arrangements. The IEEE study shows that AI is capable of uncovering unknown patterns in complex biological structures.
PyTorch 100× faster: The Japanese team Sakana AI has presented an "AI CUDA Engineer" that automatically translates PyTorch models into optimized CUDA code and is thus up to 100x faster.
And that’s a wrap! Enjoy the rest of your weekend.
Your AInauts,
Fabian & Reto
Your feedback is essential for us. We read EVERY comment and feedback, just respond to this email. Tell us what was (not) good and what is interesting for YOU