Transcription
There have been many interesting AI developments this week. Let's jump right in!
OpenAI plans to release more expensive AI models. According to an article in The Information, executives told investors they plan to sell low-end agents for $22,000 per month to high-income knowledge workers, mid-tier agents for software development (possibly $10,000 a month), and high-end agents (PhD-level research, $20,000 per month). These agents will supposedly do more work than the $200/month ChatGPT Pro plan. ChatGPT Pro isn't technically an agent like the coding or research agents would be. These prices aren't confirmed; OpenAI is still testing price points.
For $2,000 a month, you could have a research assistant for videos or articles. For $10,000, a software developer. For $20,000, a PhD-level researcher. The running joke is, it might be cheaper to hire a human.
Regarding the OpenAI/Elon Musk drama: Elon Musk sued OpenAI to prevent its transition from non-profit to for-profit status, believing it violates their original agreement. A judge refused to halt the transition but is willing to take it to trial. Sam Altman maintains there's no case, claiming it's a tactic by Musk (who owns xAI) to slow down a competitor.
All ChatGPT Plus users now have access to GPT 4.5. Initially only available to Pro members ($200/month), it's now on the $20/month Plus plan. Initially, I was unimpressed. It didn't excel at benchmarks or even spell "strawberry" correctly! However, after a week, I've changed my mind. It's great for creative tasks like writing and outlining.
The ChatGPT desktop app for macOS now directly edits code in IDEs (Integrated Development Environments). Plus, Pro, and Team users can make changes within the IDE. OpenAI's demo showed creating a "Starfield view," generating code, previewing it, and applying the changes directly.
Google showcased a new AI mode, similar to Perplexity and OpenAI's search modes. It searches the web, finds sources, and generates an AI-based response. It's not yet released but is available to some in Google Labs. It appears Google One Premium subscribers ($20/month) will have access.
DuckDuckGo is adding AI features, offering free access to popular AI chatbots at duck.ai. It expands AI-assisted answers in its search engine. The AI is private and anonymous, not training on conversations. Users choose the model (GPT-4, Llama 3, etc.). It doesn't web search but offers various chatbots anonymously. DuckDuckGo also has AI overviews in its search results.
Google Shopping has new AI features: generating outfits (e.g., "colorful mini dress with big daisies") and finding similar real items; virtual makeup try-on using diverse models.
Google Sheets now has Gemini AI capabilities for paid workspace users. It analyzes data faster, identifying correlations and outliers. The previous version helped generate formulas; this analyzes data for insights. For example, correlating food with stomach aches.
Google phone users will soon be able to share their screen or camera with AI. It's similar to what I used in London with Project Astra, and will likely be for Google One AI Premium members ($20/month), rolling out to Android later this month.
Google is adding scam detection to Android and Pixel phones. It detects suspicious text messages (example provided). Ignore and delete suspicious texts to avoid scams.
Google Labs previewed Whisk Animate, combining images from Whisk and using AI to create animations and videos. No release date is set.
Microsoft Copilot had a UI update for its Windows app, making it cleaner.
Alibaba released Qwen-32B, a reasoning model with 32 billion parameters. It performed similarly to DeepSeek-R1 (671 billion parameters) on the AIM-24 benchmark (mathematical problem-solving). It slightly underperformed on LiveCodeBench but outscored DeepSeek-R1 on instruction-following (IF-Eval) and function calling (BFCL). The model and code are available on Hugging Face.
Perplexity's voice mode is now in the Mac app (Command+Shift+M).
Grok's voice mode is now free for all iOS users. (I'm unsure if all modes are free.)
Sesame's AI voice conversation tool is impressive, similar to the movie "Her." (Audio example provided.)
Podcastle launched a text-to-speech model with 450+ AI voices, potentially competing with 11 Labs. It's reportedly robotic but improving; users can train their voices.
Teleperformance (largest call center operator) uses AI to make staff sound less Indian, aiming to improve international client support. This technology could be misused by scammers.
Amazon Prime is testing AI dubbing for TV shows and movies.
Meta's research allows avatar generation from four selfies (video example).
Pika Labs features are now integrated into Kai AI.
Tencent's open-source Hunon Video I2V creates impressive videos (examples shown). Code is on GitHub.
Hedra Labs released Character 3, an AI video creation tool (demo shown). The voices are robotic but the rest is impressive. (I tested it by generating a video; results shown.)
Tavis AI introduced Phoenix 3, Raven 0, and Sparrow 0 models for conversational video interfaces (demo shown).
Super Mario is being used to benchmark AI, with Claude 3.7 performing well.
Opera is adding an AI browser agent (similar to Claude's or ChatGPT's) called Browser Operator. (Examples shown.)
Apple revealed the M3 Ultra chip (96GB memory, configurable to 512GB). New iPad Air, Mac Studio (M4 Max/M3 Ultra), and MacBook Air (M4) were released.
Mistral released a new, top-performing OCR model.
Codium released Windsurf Wave 4 with features like suggested actions and MCP (Model Context Protocol) discoverability. MCP standardizes large language model interaction with external tools.
Cortical Labs created the CL1, a biological computer using human brain cells fused with silicon. It potentially replaces animal testing in drug development and may reveal new treatments for diseases. It's been used to play Pong.
Uber and Waymo are partnering on driverless cars in Austin.
A driverless Maserati set a new autonomous speed record.
A very advanced robot from PND Btics is shown.
I'm giving away a signed RTX 4090 GPU from Nvidia. To enter: Subscribe, subscribe to the newsletter, register for NVIDIA's GTC (link in description), and fill out a Google form (link in description).
Check out FutureTools.io for AI tools, news, and a free AI income database.